* ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator
@ 2011-11-09 20:22 j.glisse
2011-11-09 20:22 ` [PATCH 01/13] drm/ttm: remove userspace backed ttm object support j.glisse
` (13 more replies)
0 siblings, 14 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom
So i did an overhaul of ttm_memory, i believe the simplification i did
make sense. See patch 5 for a longer explanation.
Thomas with the ttm_memory change the allocation of pages won't happen
if the accounting report that we are going over the limit and bo shrinker
failed to free any memory to make room.
The handling of dma32 zone is done as post pass of ttm memory accounting.
Regarding the pagefault comment i removed, it doesn't make sense anymore
because now we populate the whole page table in one shot. So there is
no more prefaulting few pages but a full prefaulting. Thought i can
add a comment stating that if you like.
For the ttm_tt_dma struct to hold page allocator specific informations
i think it can be done as an followup patch but if you prefer to have
that in this patchset let me know i will respin with such changes.
I am in the process of retesting this whole serie and especialy the
while memory accounting.
Cheers,
Jerome
^ permalink raw reply [flat|nested] 28+ messages in thread
* [PATCH 01/13] drm/ttm: remove userspace backed ttm object support
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 02/13] drm/ttm: remove split btw highmen and lowmem page j.glisse
` (12 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
This was never use in none of the driver, properly using userspace
page for bo would need more code (vma interaction mostly). Removing
this dead code in preparation of ttm_tt & backend merge.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
---
drivers/gpu/drm/ttm/ttm_bo.c | 22 --------
drivers/gpu/drm/ttm/ttm_tt.c | 105 +--------------------------------------
include/drm/ttm/ttm_bo_api.h | 5 --
include/drm/ttm/ttm_bo_driver.h | 24 ---------
4 files changed, 1 insertions(+), 155 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 617b646..4bde335 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -342,22 +342,6 @@ static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
if (unlikely(bo->ttm == NULL))
ret = -ENOMEM;
break;
- case ttm_bo_type_user:
- bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
- page_flags | TTM_PAGE_FLAG_USER,
- glob->dummy_read_page);
- if (unlikely(bo->ttm == NULL)) {
- ret = -ENOMEM;
- break;
- }
-
- ret = ttm_tt_set_user(bo->ttm, current,
- bo->buffer_start, bo->num_pages);
- if (unlikely(ret != 0)) {
- ttm_tt_destroy(bo->ttm);
- bo->ttm = NULL;
- }
- break;
default:
printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
ret = -EINVAL;
@@ -907,16 +891,12 @@ static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
}
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
- bool disallow_fixed,
uint32_t mem_type,
uint32_t proposed_placement,
uint32_t *masked_placement)
{
uint32_t cur_flags = ttm_bo_type_flags(mem_type);
- if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
- return false;
-
if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
return false;
@@ -961,7 +941,6 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
man = &bdev->man[mem_type];
type_ok = ttm_bo_mt_compatible(man,
- bo->type == ttm_bo_type_user,
mem_type,
placement->placement[i],
&cur_flags);
@@ -1009,7 +988,6 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
if (!man->has_type)
continue;
if (!ttm_bo_mt_compatible(man,
- bo->type == ttm_bo_type_user,
mem_type,
placement->busy_placement[i],
&cur_flags))
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 58c271e..82a1161 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -62,43 +62,6 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm)
ttm->dma_address = NULL;
}
-static void ttm_tt_free_user_pages(struct ttm_tt *ttm)
-{
- int write;
- int dirty;
- struct page *page;
- int i;
- struct ttm_backend *be = ttm->be;
-
- BUG_ON(!(ttm->page_flags & TTM_PAGE_FLAG_USER));
- write = ((ttm->page_flags & TTM_PAGE_FLAG_WRITE) != 0);
- dirty = ((ttm->page_flags & TTM_PAGE_FLAG_USER_DIRTY) != 0);
-
- if (be)
- be->func->clear(be);
-
- for (i = 0; i < ttm->num_pages; ++i) {
- page = ttm->pages[i];
- if (page == NULL)
- continue;
-
- if (page == ttm->dummy_read_page) {
- BUG_ON(write);
- continue;
- }
-
- if (write && dirty && !PageReserved(page))
- set_page_dirty_lock(page);
-
- ttm->pages[i] = NULL;
- ttm_mem_global_free(ttm->glob->mem_glob, PAGE_SIZE);
- put_page(page);
- }
- ttm->state = tt_unpopulated;
- ttm->first_himem_page = ttm->num_pages;
- ttm->last_lomem_page = -1;
-}
-
static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
{
struct page *p;
@@ -325,10 +288,7 @@ void ttm_tt_destroy(struct ttm_tt *ttm)
}
if (likely(ttm->pages != NULL)) {
- if (ttm->page_flags & TTM_PAGE_FLAG_USER)
- ttm_tt_free_user_pages(ttm);
- else
- ttm_tt_free_alloced_pages(ttm);
+ ttm_tt_free_alloced_pages(ttm);
ttm_tt_free_page_directory(ttm);
}
@@ -340,45 +300,6 @@ void ttm_tt_destroy(struct ttm_tt *ttm)
kfree(ttm);
}
-int ttm_tt_set_user(struct ttm_tt *ttm,
- struct task_struct *tsk,
- unsigned long start, unsigned long num_pages)
-{
- struct mm_struct *mm = tsk->mm;
- int ret;
- int write = (ttm->page_flags & TTM_PAGE_FLAG_WRITE) != 0;
- struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
-
- BUG_ON(num_pages != ttm->num_pages);
- BUG_ON((ttm->page_flags & TTM_PAGE_FLAG_USER) == 0);
-
- /**
- * Account user pages as lowmem pages for now.
- */
-
- ret = ttm_mem_global_alloc(mem_glob, num_pages * PAGE_SIZE,
- false, false);
- if (unlikely(ret != 0))
- return ret;
-
- down_read(&mm->mmap_sem);
- ret = get_user_pages(tsk, mm, start, num_pages,
- write, 0, ttm->pages, NULL);
- up_read(&mm->mmap_sem);
-
- if (ret != num_pages && write) {
- ttm_tt_free_user_pages(ttm);
- ttm_mem_global_free(mem_glob, num_pages * PAGE_SIZE);
- return -ENOMEM;
- }
-
- ttm->tsk = tsk;
- ttm->start = start;
- ttm->state = tt_unbound;
-
- return 0;
-}
-
struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
uint32_t page_flags, struct page *dummy_read_page)
{
@@ -452,8 +373,6 @@ int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
ttm->state = tt_bound;
- if (ttm->page_flags & TTM_PAGE_FLAG_USER)
- ttm->page_flags |= TTM_PAGE_FLAG_USER_DIRTY;
return 0;
}
EXPORT_SYMBOL(ttm_tt_bind);
@@ -469,16 +388,6 @@ static int ttm_tt_swapin(struct ttm_tt *ttm)
int i;
int ret = -ENOMEM;
- if (ttm->page_flags & TTM_PAGE_FLAG_USER) {
- ret = ttm_tt_set_user(ttm, ttm->tsk, ttm->start,
- ttm->num_pages);
- if (unlikely(ret != 0))
- return ret;
-
- ttm->page_flags &= ~TTM_PAGE_FLAG_SWAPPED;
- return 0;
- }
-
swap_storage = ttm->swap_storage;
BUG_ON(swap_storage == NULL);
@@ -529,18 +438,6 @@ int ttm_tt_swapout(struct ttm_tt *ttm, struct file *persistent_swap_storage)
BUG_ON(ttm->state != tt_unbound && ttm->state != tt_unpopulated);
BUG_ON(ttm->caching_state != tt_cached);
- /*
- * For user buffers, just unpin the pages, as there should be
- * vma references.
- */
-
- if (ttm->page_flags & TTM_PAGE_FLAG_USER) {
- ttm_tt_free_user_pages(ttm);
- ttm->page_flags |= TTM_PAGE_FLAG_SWAPPED;
- ttm->swap_storage = NULL;
- return 0;
- }
-
if (!persistent_swap_storage) {
swap_storage = shmem_file_setup("ttm swap",
ttm->num_pages << PAGE_SHIFT,
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 42e3469..8d95a42 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -122,17 +122,12 @@ struct ttm_mem_reg {
* be mmapped by user space. Each of these bos occupy a slot in the
* device address space, that can be used for normal vm operations.
*
- * @ttm_bo_type_user: These are user-space memory areas that are made
- * available to the GPU by mapping the buffer pages into the GPU aperture
- * space. These buffers cannot be mmaped from the device address space.
- *
* @ttm_bo_type_kernel: These buffers are like ttm_bo_type_device buffers,
* but they cannot be accessed from user-space. For kernel-only use.
*/
enum ttm_bo_type {
ttm_bo_type_device,
- ttm_bo_type_user,
ttm_bo_type_kernel
};
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 94eb143..37527d6 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -118,8 +118,6 @@ struct ttm_backend {
struct ttm_backend_func *func;
};
-#define TTM_PAGE_FLAG_USER (1 << 1)
-#define TTM_PAGE_FLAG_USER_DIRTY (1 << 2)
#define TTM_PAGE_FLAG_WRITE (1 << 3)
#define TTM_PAGE_FLAG_SWAPPED (1 << 4)
#define TTM_PAGE_FLAG_PERSISTENT_SWAP (1 << 5)
@@ -146,8 +144,6 @@ enum ttm_caching_state {
* @num_pages: Number of pages in the page array.
* @bdev: Pointer to the current struct ttm_bo_device.
* @be: Pointer to the ttm backend.
- * @tsk: The task for user ttm.
- * @start: virtual address for user ttm.
* @swap_storage: Pointer to shmem struct file for swap storage.
* @caching_state: The current caching state of the pages.
* @state: The current binding state of the pages.
@@ -167,8 +163,6 @@ struct ttm_tt {
unsigned long num_pages;
struct ttm_bo_global *glob;
struct ttm_backend *be;
- struct task_struct *tsk;
- unsigned long start;
struct file *swap_storage;
enum ttm_caching_state caching_state;
enum {
@@ -618,24 +612,6 @@ extern struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev,
struct page *dummy_read_page);
/**
- * ttm_tt_set_user:
- *
- * @ttm: The struct ttm_tt to populate.
- * @tsk: A struct task_struct for which @start is a valid user-space address.
- * @start: A valid user-space address.
- * @num_pages: Size in pages of the user memory area.
- *
- * Populate a struct ttm_tt with a user-space memory area after first pinning
- * the pages backing it.
- * Returns:
- * !0: Error.
- */
-
-extern int ttm_tt_set_user(struct ttm_tt *ttm,
- struct task_struct *tsk,
- unsigned long start, unsigned long num_pages);
-
-/**
* ttm_ttm_bind:
*
* @ttm: The struct ttm_tt containing backing pages.
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 02/13] drm/ttm: remove split btw highmen and lowmem page
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
2011-11-09 20:22 ` [PATCH 01/13] drm/ttm: remove userspace backed ttm object support j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 03/13] drm/ttm: remove unused backend flags field j.glisse
` (11 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
Split btw highmem and lowmem page was rendered useless by the
pool code. Remove it. Note further cleanup would change the
ttm page allocation helper to actualy take an array instead
of relying on list this could drasticly reduce the number of
function call in the common case of allocation whole buffer.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
---
drivers/gpu/drm/ttm/ttm_tt.c | 11 ++---------
include/drm/ttm/ttm_bo_driver.h | 7 -------
2 files changed, 2 insertions(+), 16 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 82a1161..8b7a6d0 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -69,7 +69,7 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
int ret;
- while (NULL == (p = ttm->pages[index])) {
+ if (NULL == (p = ttm->pages[index])) {
INIT_LIST_HEAD(&h);
@@ -85,10 +85,7 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
if (unlikely(ret != 0))
goto out_err;
- if (PageHighMem(p))
- ttm->pages[--ttm->first_himem_page] = p;
- else
- ttm->pages[++ttm->last_lomem_page] = p;
+ ttm->pages[index] = p;
}
return p;
out_err:
@@ -270,8 +267,6 @@ static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
ttm_put_pages(&h, count, ttm->page_flags, ttm->caching_state,
ttm->dma_address);
ttm->state = tt_unpopulated;
- ttm->first_himem_page = ttm->num_pages;
- ttm->last_lomem_page = -1;
}
void ttm_tt_destroy(struct ttm_tt *ttm)
@@ -315,8 +310,6 @@ struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
ttm->glob = bdev->glob;
ttm->num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
- ttm->first_himem_page = ttm->num_pages;
- ttm->last_lomem_page = -1;
ttm->caching_state = tt_cached;
ttm->page_flags = page_flags;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 37527d6..9da182b 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -136,11 +136,6 @@ enum ttm_caching_state {
* @dummy_read_page: Page to map where the ttm_tt page array contains a NULL
* pointer.
* @pages: Array of pages backing the data.
- * @first_himem_page: Himem pages are put last in the page array, which
- * enables us to run caching attribute changes on only the first part
- * of the page array containing lomem pages. This is the index of the
- * first himem page.
- * @last_lomem_page: Index of the last lomem page in the page array.
* @num_pages: Number of pages in the page array.
* @bdev: Pointer to the current struct ttm_bo_device.
* @be: Pointer to the ttm backend.
@@ -157,8 +152,6 @@ enum ttm_caching_state {
struct ttm_tt {
struct page *dummy_read_page;
struct page **pages;
- long first_himem_page;
- long last_lomem_page;
uint32_t page_flags;
unsigned long num_pages;
struct ttm_bo_global *glob;
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 03/13] drm/ttm: remove unused backend flags field
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
2011-11-09 20:22 ` [PATCH 01/13] drm/ttm: remove userspace backed ttm object support j.glisse
2011-11-09 20:22 ` [PATCH 02/13] drm/ttm: remove split btw highmen and lowmem page j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 04/13] drm/ttm: use ttm put pages function to properly restore cache attribute j.glisse
` (10 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
This field is not use by any of the driver just drop it.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
---
drivers/gpu/drm/radeon/radeon_ttm.c | 1 -
include/drm/ttm/ttm_bo_driver.h | 2 --
2 files changed, 0 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 0b5468b..97c76ae 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -787,7 +787,6 @@ struct ttm_backend *radeon_ttm_backend_create(struct radeon_device *rdev)
return NULL;
}
gtt->backend.bdev = &rdev->mman.bdev;
- gtt->backend.flags = 0;
gtt->backend.func = &radeon_backend_func;
gtt->rdev = rdev;
gtt->pages = NULL;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 9da182b..6d17140 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -106,7 +106,6 @@ struct ttm_backend_func {
* struct ttm_backend
*
* @bdev: Pointer to a struct ttm_bo_device.
- * @flags: For driver use.
* @func: Pointer to a struct ttm_backend_func that describes
* the backend methods.
*
@@ -114,7 +113,6 @@ struct ttm_backend_func {
struct ttm_backend {
struct ttm_bo_device *bdev;
- uint32_t flags;
struct ttm_backend_func *func;
};
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 04/13] drm/ttm: use ttm put pages function to properly restore cache attribute
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (2 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 03/13] drm/ttm: remove unused backend flags field j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 05/13] drm/ttm: overhaul memory accounting j.glisse
` (9 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
On failure we need to make sure the page we free has wb cache
attribute. Do this pas call the proper ttm page helper function.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
---
drivers/gpu/drm/ttm/ttm_tt.c | 5 ++++-
1 files changed, 4 insertions(+), 1 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 8b7a6d0..3fb4c6d 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -89,7 +89,10 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
}
return p;
out_err:
- put_page(p);
+ INIT_LIST_HEAD(&h);
+ list_add(&p->lru, &h);
+ ttm_put_pages(&h, 1, ttm->page_flags,
+ ttm->caching_state, &ttm->dma_address[index]);
return NULL;
}
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (3 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 04/13] drm/ttm: use ttm put pages function to properly restore cache attribute j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-10 10:27 ` Thomas Hellstrom
2011-11-09 20:22 ` [PATCH 06/13] drm/ttm: convert page allocation to use page ptr array instead of list V4 j.glisse
` (8 subsequent siblings)
13 siblings, 1 reply; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
This is an overhaul of the ttm memory accounting. This tries to keep
the same global behavior while removing the whole zone concept. It
keeps a distrinction for dma32 so that we make sure that ttm don't
starve the dma32 zone.
There is 3 threshold for memory allocation :
- max_mem is the maximum memory the whole ttm infrastructure is
going to allow allocation for (exception of system process see
below)
- emer_mem is the maximum memory allowed for system process, this
limit is > to max_mem
- swap_limit is the threshold at which point ttm will start to
try to swap object because ttm is getting close the max_mem
limit
- swap_dma32_limit is the threshold at which point ttm will start
swap object to try to reduce the pressure on the dma32 zone. Note
that we don't specificly target object to swap to it might very
well free more memory from highmem rather than from dma32
Accounting is done through used_mem & used_dma32_mem, which sum give
the total amount of memory actually accounted by ttm.
Idea is that allocation will fail if (used_mem + used_dma32_mem) >
max_mem and if swapping fail to make enough room.
The used_dma32_mem can be updated as a later stage, allowing to
perform accounting test before allocating a whole batch of pages.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
---
drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
drivers/gpu/drm/ttm/ttm_memory.c | 517 +++++++++---------------------
drivers/gpu/drm/ttm/ttm_object.c | 3 +-
drivers/gpu/drm/ttm/ttm_tt.c | 2 +-
drivers/gpu/drm/vmwgfx/vmwgfx_fence.c | 8 +-
drivers/gpu/drm/vmwgfx/vmwgfx_resource.c | 8 +-
include/drm/ttm/ttm_memory.h | 23 +-
7 files changed, 168 insertions(+), 395 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 4bde335..92712798 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -1252,7 +1252,7 @@ int ttm_bo_create(struct ttm_bo_device *bdev,
size_t acc_size =
ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
- ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
+ ret = ttm_mem_global_alloc(mem_glob, acc_size, false);
if (unlikely(ret != 0))
return ret;
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index e70ddd8..b550baf 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -35,21 +35,10 @@
#include <linux/module.h>
#include <linux/slab.h>
-#define TTM_MEMORY_ALLOC_RETRIES 4
-
-struct ttm_mem_zone {
- struct kobject kobj;
- struct ttm_mem_global *glob;
- const char *name;
- uint64_t zone_mem;
- uint64_t emer_mem;
- uint64_t max_mem;
- uint64_t swap_limit;
- uint64_t used_mem;
-};
+#define TTM_MEMORY_RETRIES 4
static struct attribute ttm_mem_sys = {
- .name = "zone_memory",
+ .name = "memory",
.mode = S_IRUGO
};
static struct attribute ttm_mem_emer = {
@@ -64,140 +53,141 @@ static struct attribute ttm_mem_swap = {
.name = "swap_limit",
.mode = S_IRUGO | S_IWUSR
};
+static struct attribute ttm_mem_dma32_swap = {
+ .name = "swap_dma32_limit",
+ .mode = S_IRUGO | S_IWUSR
+};
static struct attribute ttm_mem_used = {
.name = "used_memory",
.mode = S_IRUGO
};
+static struct attribute ttm_mem_dma32_used = {
+ .name = "used_dma32_memory",
+ .mode = S_IRUGO
+};
-static void ttm_mem_zone_kobj_release(struct kobject *kobj)
-{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
-
- printk(KERN_INFO TTM_PFX
- "Zone %7s: Used memory at exit: %llu kiB.\n",
- zone->name, (unsigned long long) zone->used_mem >> 10);
- kfree(zone);
-}
-
-static ssize_t ttm_mem_zone_show(struct kobject *kobj,
- struct attribute *attr,
- char *buffer)
+static ssize_t ttm_mem_global_show(struct kobject *kobj,
+ struct attribute *attr,
+ char *buffer)
{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
- uint64_t val = 0;
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
+ unsigned long val = 0;
- spin_lock(&zone->glob->lock);
+ spin_lock(&glob->lock);
if (attr == &ttm_mem_sys)
- val = zone->zone_mem;
+ val = glob->mem;
else if (attr == &ttm_mem_emer)
- val = zone->emer_mem;
+ val = glob->emer_mem;
else if (attr == &ttm_mem_max)
- val = zone->max_mem;
+ val = glob->max_mem;
else if (attr == &ttm_mem_swap)
- val = zone->swap_limit;
+ val = glob->swap_limit;
else if (attr == &ttm_mem_used)
- val = zone->used_mem;
- spin_unlock(&zone->glob->lock);
+ val = glob->used_mem;
+ else if (attr == &ttm_mem_dma32_used)
+ val = glob->used_dma32_mem;
+ else if (attr == &ttm_mem_dma32_swap)
+ val = glob->swap_dma32_limit;
+ spin_unlock(&glob->lock);
- return snprintf(buffer, PAGE_SIZE, "%llu\n",
- (unsigned long long) val >> 10);
+ return snprintf(buffer, PAGE_SIZE, "%lu\n", val >> 10);
}
static void ttm_check_swapping(struct ttm_mem_global *glob);
-static ssize_t ttm_mem_zone_store(struct kobject *kobj,
- struct attribute *attr,
- const char *buffer,
- size_t size)
+static ssize_t ttm_mem_global_store(struct kobject *kobj,
+ struct attribute *attr,
+ const char *buffer,
+ size_t size)
{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
- int chars;
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
unsigned long val;
- uint64_t val64;
+ int chars;
chars = sscanf(buffer, "%lu", &val);
if (chars == 0)
return size;
- val64 = val;
- val64 <<= 10;
+ val <<= 10;
+
+ spin_lock(&glob->lock);
+ /* limit to maximum memory */
+ if (val > glob->mem)
+ val = glob->mem;
- spin_lock(&zone->glob->lock);
- if (val64 > zone->zone_mem)
- val64 = zone->zone_mem;
if (attr == &ttm_mem_emer) {
- zone->emer_mem = val64;
- if (zone->max_mem > val64)
- zone->max_mem = val64;
+ glob->emer_mem = val;
+ if (glob->max_mem > val)
+ glob->max_mem = val;
} else if (attr == &ttm_mem_max) {
- zone->max_mem = val64;
- if (zone->emer_mem < val64)
- zone->emer_mem = val64;
- } else if (attr == &ttm_mem_swap)
- zone->swap_limit = val64;
- spin_unlock(&zone->glob->lock);
+ glob->max_mem = val;
+ if (glob->emer_mem < val)
+ glob->emer_mem = val;
+ } else if (attr == &ttm_mem_swap) {
+ glob->swap_limit = val;
+ } else if (attr == &ttm_mem_dma32_swap) {
+ glob->swap_dma32_limit = val;
+ }
+ spin_unlock(&glob->lock);
+ ttm_check_swapping(glob);
+ return size;
+}
- ttm_check_swapping(zone->glob);
+static void ttm_mem_global_kobj_release(struct kobject *kobj)
+{
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
- return size;
+ kfree(glob);
}
-static struct attribute *ttm_mem_zone_attrs[] = {
+static struct attribute *ttm_mem_global_attrs[] = {
&ttm_mem_sys,
&ttm_mem_emer,
&ttm_mem_max,
&ttm_mem_swap,
+ &ttm_mem_dma32_swap,
&ttm_mem_used,
+ &ttm_mem_dma32_used,
NULL
};
-static const struct sysfs_ops ttm_mem_zone_ops = {
- .show = &ttm_mem_zone_show,
- .store = &ttm_mem_zone_store
+static const struct sysfs_ops ttm_mem_global_ops = {
+ .show = &ttm_mem_global_show,
+ .store = &ttm_mem_global_store
};
-static struct kobj_type ttm_mem_zone_kobj_type = {
- .release = &ttm_mem_zone_kobj_release,
- .sysfs_ops = &ttm_mem_zone_ops,
- .default_attrs = ttm_mem_zone_attrs,
-};
-
-static void ttm_mem_global_kobj_release(struct kobject *kobj)
-{
- struct ttm_mem_global *glob =
- container_of(kobj, struct ttm_mem_global, kobj);
-
- kfree(glob);
-}
-
static struct kobj_type ttm_mem_glob_kobj_type = {
.release = &ttm_mem_global_kobj_release,
+ .sysfs_ops = &ttm_mem_global_ops,
+ .default_attrs = ttm_mem_global_attrs,
};
-static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
- bool from_wq, uint64_t extra)
+static bool ttm_above_swap_target(struct ttm_mem_global *glob,
+ bool from_wq, uint64_t extra)
{
- unsigned int i;
- struct ttm_mem_zone *zone;
- uint64_t target;
+ unsigned long target;
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
-
- if (from_wq)
- target = zone->swap_limit;
- else if (capable(CAP_SYS_ADMIN))
- target = zone->emer_mem;
+ if (from_wq) {
+ if (glob->used_mem > glob->swap_limit) {
+ return true;
+ }
+ if (glob->used_dma32_mem > glob->swap_dma32_limit) {
+ return true;
+ }
+ } else {
+ if (capable(CAP_SYS_ADMIN))
+ target = glob->emer_mem;
else
- target = zone->max_mem;
-
- target = (extra > target) ? 0ULL : target;
-
- if (zone->used_mem > target)
+ target = glob->max_mem;
+ if (extra > target) {
return true;
+ }
+ if ((glob->used_mem + glob->used_dma32_mem) > target) {
+ return true;
+ }
}
return false;
}
@@ -208,31 +198,31 @@ static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
* Note that this function is reentrant:
* many threads may try to swap out at any given time.
*/
-
-static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
+static void ttm_shrink(struct ttm_mem_global *glob,
+ bool from_wq,
uint64_t extra)
{
- int ret;
struct ttm_mem_shrink *shrink;
+ int ret, nretries = TTM_MEMORY_RETRIES;
spin_lock(&glob->lock);
if (glob->shrink == NULL)
goto out;
- while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
+ while (ttm_above_swap_target(glob, from_wq, extra)) {
shrink = glob->shrink;
spin_unlock(&glob->lock);
ret = shrink->do_shrink(shrink);
spin_lock(&glob->lock);
if (unlikely(ret != 0))
goto out;
+ if (--nretries < 0)
+ goto out;
}
out:
spin_unlock(&glob->lock);
}
-
-
static void ttm_shrink_work(struct work_struct *work)
{
struct ttm_mem_global *glob =
@@ -241,127 +231,10 @@ static void ttm_shrink_work(struct work_struct *work)
ttm_shrink(glob, true, 0ULL);
}
-static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- uint64_t mem;
- int ret;
-
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram - si->totalhigh;
- mem *= si->mem_unit;
-
- zone->name = "kernel";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_kernel = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-
-#ifdef CONFIG_HIGHMEM
-static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone;
- uint64_t mem;
- int ret;
-
- if (si->totalhigh == 0)
- return 0;
-
- zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram;
- mem *= si->mem_unit;
-
- zone->name = "highmem";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_highmem = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-#else
-static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- uint64_t mem;
- int ret;
-
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram;
- mem *= si->mem_unit;
-
- /**
- * No special dma32 zone needed.
- */
-
- if (mem <= ((uint64_t) 1ULL << 32)) {
- kfree(zone);
- return 0;
- }
-
- /*
- * Limit max dma32 memory to 4GB for now
- * until we can figure out how big this
- * zone really is.
- */
-
- mem = ((uint64_t) 1ULL << 32);
- zone->name = "dma32";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_dma32 = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-#endif
-
int ttm_mem_global_init(struct ttm_mem_global *glob)
{
struct sysinfo si;
int ret;
- int i;
- struct ttm_mem_zone *zone;
spin_lock_init(&glob->lock);
glob->swap_queue = create_singlethread_workqueue("ttm_swap");
@@ -374,50 +247,29 @@ int ttm_mem_global_init(struct ttm_mem_global *glob)
return ret;
}
+ /* compute limit */
si_meminfo(&si);
-
- ret = ttm_mem_init_kernel_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#ifdef CONFIG_HIGHMEM
- ret = ttm_mem_init_highmem_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#else
- ret = ttm_mem_init_dma32_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#endif
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- printk(KERN_INFO TTM_PFX
- "Zone %7s: Available graphics memory: %llu kiB.\n",
- zone->name, (unsigned long long) zone->max_mem >> 10);
- }
- ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
+ glob->mem = si.totalram;
+ glob->mem *= si.mem_unit;
+ glob->used_mem = 0;
+ glob->used_dma32_mem = 0;
+ glob->max_mem = glob->mem >> 1;
+ glob->emer_mem = (glob->mem >> 1) + (glob->mem >> 2);
+ glob->swap_limit = glob->max_mem - (glob->mem >> 3);
+ glob->swap_dma32_limit = ((1ULL << 32) >> 1) - (((1ULL << 32) >> 3));
+ ttm_page_alloc_init(glob, glob->max_mem/(2*PAGE_SIZE));
return 0;
-out_no_zone:
- ttm_mem_global_release(glob);
- return ret;
}
EXPORT_SYMBOL(ttm_mem_global_init);
void ttm_mem_global_release(struct ttm_mem_global *glob)
{
- unsigned int i;
- struct ttm_mem_zone *zone;
-
/* let the page allocator first stop the shrink work. */
ttm_page_alloc_fini();
flush_workqueue(glob->swap_queue);
destroy_workqueue(glob->swap_queue);
glob->swap_queue = NULL;
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- kobject_del(&zone->kobj);
- kobject_put(&zone->kobj);
- }
kobject_del(&glob->kobj);
kobject_put(&glob->kobj);
}
@@ -426,18 +278,14 @@ EXPORT_SYMBOL(ttm_mem_global_release);
static void ttm_check_swapping(struct ttm_mem_global *glob)
{
bool needs_swapping = false;
- unsigned int i;
- struct ttm_mem_zone *zone;
spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (zone->used_mem > zone->swap_limit) {
- needs_swapping = true;
- break;
- }
+ if ((glob->used_mem + glob->used_dma32_mem) > glob->swap_limit) {
+ needs_swapping = true;
+ }
+ if (glob->used_dma32_mem > glob->swap_dma32_limit) {
+ needs_swapping = true;
}
-
spin_unlock(&glob->lock);
if (unlikely(needs_swapping))
@@ -445,142 +293,69 @@ static void ttm_check_swapping(struct ttm_mem_global *glob)
}
-static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t amount)
-{
- unsigned int i;
- struct ttm_mem_zone *zone;
-
- spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
- zone->used_mem -= amount;
- }
- spin_unlock(&glob->lock);
-}
-
void ttm_mem_global_free(struct ttm_mem_global *glob,
uint64_t amount)
{
- return ttm_mem_global_free_zone(glob, NULL, amount);
-}
-EXPORT_SYMBOL(ttm_mem_global_free);
-
-static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t amount, bool reserve)
-{
- uint64_t limit;
- int ret = -ENOMEM;
- unsigned int i;
- struct ttm_mem_zone *zone;
-
spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
-
- limit = (capable(CAP_SYS_ADMIN)) ?
- zone->emer_mem : zone->max_mem;
-
- if (zone->used_mem > limit)
- goto out_unlock;
- }
-
- if (reserve) {
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
- zone->used_mem += amount;
- }
- }
-
- ret = 0;
-out_unlock:
+ glob->used_mem -= amount;
spin_unlock(&glob->lock);
- ttm_check_swapping(glob);
-
- return ret;
}
+EXPORT_SYMBOL(ttm_mem_global_free);
-
-static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t memory,
- bool no_wait, bool interruptible)
+int ttm_mem_global_alloc(struct ttm_mem_global *glob,
+ uint64_t memory,
+ bool no_wait)
{
- int count = TTM_MEMORY_ALLOC_RETRIES;
+ unsigned long limit;
+ int i;
- while (unlikely(ttm_mem_global_reserve(glob,
- single_zone,
- memory, true)
- != 0)) {
+ for (i = 0; i < TTM_MEMORY_RETRIES; i++) {
+ spin_lock(&glob->lock);
+ limit = (capable(CAP_SYS_ADMIN)) ? glob->emer_mem : glob->max_mem;
+ if ((glob->used_mem + glob->used_dma32_mem + memory) < limit) {
+ glob->used_mem += memory;
+ spin_unlock(&glob->lock);
+ return 0;
+ }
+ spin_unlock(&glob->lock);
if (no_wait)
return -ENOMEM;
- if (unlikely(count-- == 0))
- return -ENOMEM;
ttm_shrink(glob, false, memory + (memory >> 2) + 16);
}
-
- return 0;
-}
-
-int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
- bool no_wait, bool interruptible)
-{
- /**
- * Normal allocations of kernel memory are registered in
- * all zones.
- */
-
- return ttm_mem_global_alloc_zone(glob, NULL, memory, no_wait,
- interruptible);
+ return -ENOMEM;
}
EXPORT_SYMBOL(ttm_mem_global_alloc);
int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
struct page *page,
- bool no_wait, bool interruptible)
+ bool no_wait)
{
- struct ttm_mem_zone *zone = NULL;
-
- /**
- * Page allocations may be registed in a single zone
- * only if highmem or !dma32.
- */
-
-#ifdef CONFIG_HIGHMEM
- if (PageHighMem(page) && glob->zone_highmem != NULL)
- zone = glob->zone_highmem;
-#else
- if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
- zone = glob->zone_kernel;
-#endif
- return ttm_mem_global_alloc_zone(glob, zone, PAGE_SIZE, no_wait,
- interruptible);
+ if (ttm_mem_global_alloc(glob, PAGE_SIZE, no_wait))
+ return -ENOMEM;
+
+ /* check if page is dma32 */
+ if (page_to_pfn(page) > 0x00100000UL) {
+ spin_lock(&glob->lock);
+ glob->used_mem -= PAGE_SIZE;
+ glob->used_dma32_mem += PAGE_SIZE;
+ spin_unlock(&glob->lock);
+ }
+ ttm_check_swapping(glob);
+ return 0;
}
void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page)
{
- struct ttm_mem_zone *zone = NULL;
-
-#ifdef CONFIG_HIGHMEM
- if (PageHighMem(page) && glob->zone_highmem != NULL)
- zone = glob->zone_highmem;
-#else
- if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
- zone = glob->zone_kernel;
-#endif
- ttm_mem_global_free_zone(glob, zone, PAGE_SIZE);
+ spin_lock(&glob->lock);
+ if (page_to_pfn(page) > 0x00100000UL) {
+ glob->used_dma32_mem -= PAGE_SIZE;
+ } else {
+ glob->used_mem -= PAGE_SIZE;
+ }
+ spin_unlock(&glob->lock);
}
-
size_t ttm_round_pot(size_t size)
{
if ((size & (size - 1)) == 0)
diff --git a/drivers/gpu/drm/ttm/ttm_object.c b/drivers/gpu/drm/ttm/ttm_object.c
index 93577f2..662c6fc 100644
--- a/drivers/gpu/drm/ttm/ttm_object.c
+++ b/drivers/gpu/drm/ttm/ttm_object.c
@@ -267,8 +267,7 @@ int ttm_ref_object_add(struct ttm_object_file *tfile,
}
read_unlock(&tfile->lock);
- ret = ttm_mem_global_alloc(mem_glob, sizeof(*ref),
- false, false);
+ ret = ttm_mem_global_alloc(mem_glob, sizeof(*ref), false);
if (unlikely(ret != 0))
return ret;
ref = kmalloc(sizeof(*ref), GFP_KERNEL);
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 3fb4c6d..e5432efa 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -81,7 +81,7 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
p = list_first_entry(&h, struct page, lru);
- ret = ttm_mem_global_alloc_page(mem_glob, p, false, false);
+ ret = ttm_mem_global_alloc_page(mem_glob, p, false);
if (unlikely(ret != 0))
goto out_err;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 15fb260..8fd1e6b 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -503,8 +503,7 @@ int vmw_fence_create(struct vmw_fence_manager *fman,
struct vmw_fence_obj *fence;
int ret;
- ret = ttm_mem_global_alloc(mem_glob, fman->fence_size,
- false, false);
+ ret = ttm_mem_global_alloc(mem_glob, fman->fence_size, false);
if (unlikely(ret != 0))
return ret;
@@ -573,8 +572,7 @@ int vmw_user_fence_create(struct drm_file *file_priv,
* be created by a user-space request.
*/
- ret = ttm_mem_global_alloc(mem_glob, fman->user_fence_size,
- false, false);
+ ret = ttm_mem_global_alloc(mem_glob, fman->user_fence_size, false);
if (unlikely(ret != 0))
return ret;
@@ -966,7 +964,7 @@ int vmw_event_fence_action_create(struct drm_file *file_priv,
* event size itself.
*/
- ret = ttm_mem_global_alloc(mem_glob, size, false, interruptible);
+ ret = ttm_mem_global_alloc(mem_glob, size, false);
if (unlikely(ret != 0))
return ret;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 86c5e4c..3c43ac0 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -430,7 +430,7 @@ int vmw_context_define_ioctl(struct drm_device *dev, void *data,
ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
vmw_user_context_size,
- false, true);
+ false);
if (unlikely(ret != 0)) {
if (ret != -ERESTARTSYS)
DRM_ERROR("Out of graphics memory for context"
@@ -1298,7 +1298,7 @@ int vmw_surface_define_ioctl(struct drm_device *dev, void *data,
return ret;
ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
- size, false, true);
+ size, false);
if (unlikely(ret != 0)) {
if (ret != -ERESTARTSYS)
DRM_ERROR("Out of graphics memory for surface"
@@ -1560,7 +1560,7 @@ int vmw_dmabuf_init(struct vmw_private *dev_priv,
vmw_dmabuf_acc_size(bdev->glob,
(size + PAGE_SIZE - 1) >> PAGE_SHIFT);
- ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
+ ret = ttm_mem_global_alloc(mem_glob, acc_size, false);
if (unlikely(ret != 0)) {
/* we must free the bo here as
* ttm_buffer_object_init does so as well */
@@ -1851,7 +1851,7 @@ int vmw_stream_claim_ioctl(struct drm_device *dev, void *data,
ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
vmw_user_stream_size,
- false, true);
+ false);
if (unlikely(ret != 0)) {
if (ret != -ERESTARTSYS)
DRM_ERROR("Out of graphics memory for stream"
diff --git a/include/drm/ttm/ttm_memory.h b/include/drm/ttm/ttm_memory.h
index 26c1f78..d24792b 100644
--- a/include/drm/ttm/ttm_memory.h
+++ b/include/drm/ttm/ttm_memory.h
@@ -81,14 +81,13 @@ struct ttm_mem_global {
struct work_struct work;
wait_queue_head_t queue;
spinlock_t lock;
- struct ttm_mem_zone *zones[TTM_MEM_MAX_ZONES];
- unsigned int num_zones;
- struct ttm_mem_zone *zone_kernel;
-#ifdef CONFIG_HIGHMEM
- struct ttm_mem_zone *zone_highmem;
-#else
- struct ttm_mem_zone *zone_dma32;
-#endif
+ unsigned long mem;
+ unsigned long max_mem;
+ unsigned long emer_mem;
+ unsigned long used_mem;
+ unsigned long used_dma32_mem;
+ unsigned long swap_limit;
+ unsigned long swap_dma32_limit;
};
/**
@@ -146,14 +145,16 @@ static inline void ttm_mem_unregister_shrink(struct ttm_mem_global *glob,
extern int ttm_mem_global_init(struct ttm_mem_global *glob);
extern void ttm_mem_global_release(struct ttm_mem_global *glob);
-extern int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
- bool no_wait, bool interruptible);
+extern int ttm_mem_global_alloc(struct ttm_mem_global *glob,
+ uint64_t memory,
+ bool no_wait);
extern void ttm_mem_global_free(struct ttm_mem_global *glob,
uint64_t amount);
extern int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
struct page *page,
- bool no_wait, bool interruptible);
+ bool no_wait);
extern void ttm_mem_global_free_page(struct ttm_mem_global *glob,
struct page *page);
extern size_t ttm_round_pot(size_t size);
+
#endif
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 06/13] drm/ttm: convert page allocation to use page ptr array instead of list V4
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (4 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 05/13] drm/ttm: overhaul memory accounting j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 07/13] drm/ttm: test for dma_address array allocation failure j.glisse
` (7 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
Use the ttm_tt page ptr array for page allocation, move the list to
array unwinding into the page allocation functions.
V2 split the fix to use ttm put page as a separate fix
properly fill pages array when TTM_PAGE_FLAG_ZERO_ALLOC is not
set
V3 Added back page_count()==1 check when freeing page
V4 Rebase on top of memory accounting overhaul
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
---
drivers/gpu/drm/ttm/ttm_memory.c | 47 +++++++++++------
drivers/gpu/drm/ttm/ttm_page_alloc.c | 90 ++++++++++++++++++++--------------
drivers/gpu/drm/ttm/ttm_tt.c | 68 ++++++++------------------
include/drm/ttm/ttm_memory.h | 13 +++--
include/drm/ttm/ttm_page_alloc.h | 17 +++---
5 files changed, 120 insertions(+), 115 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index b550baf..98f6899 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -326,32 +326,45 @@ int ttm_mem_global_alloc(struct ttm_mem_global *glob,
}
EXPORT_SYMBOL(ttm_mem_global_alloc);
-int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
- struct page *page,
- bool no_wait)
+int ttm_mem_global_alloc_pages(struct ttm_mem_global *glob,
+ unsigned npages,
+ bool no_wait)
{
-
- if (ttm_mem_global_alloc(glob, PAGE_SIZE, no_wait))
+ if (ttm_mem_global_alloc(glob, PAGE_SIZE * npages, no_wait))
return -ENOMEM;
+ ttm_check_swapping(glob);
+ return 0;
+}
+
+void ttm_mem_global_account_pages(struct ttm_mem_global *glob,
+ struct page **pages,
+ unsigned npages)
+{
+ unsigned i;
/* check if page is dma32 */
- if (page_to_pfn(page) > 0x00100000UL) {
- spin_lock(&glob->lock);
- glob->used_mem -= PAGE_SIZE;
- glob->used_dma32_mem += PAGE_SIZE;
- spin_unlock(&glob->lock);
+ spin_lock(&glob->lock);
+ for (i = 0; i < npages; i++) {
+ if (page_to_pfn(pages[i]) > 0x00100000UL) {
+ glob->used_mem -= PAGE_SIZE;
+ glob->used_dma32_mem += PAGE_SIZE;
+ }
}
- ttm_check_swapping(glob);
- return 0;
+ spin_unlock(&glob->lock);
}
-void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page)
+void ttm_mem_global_free_pages(struct ttm_mem_global *glob,
+ struct page **pages, unsigned npages)
{
+ unsigned i;
+
spin_lock(&glob->lock);
- if (page_to_pfn(page) > 0x00100000UL) {
- glob->used_dma32_mem -= PAGE_SIZE;
- } else {
- glob->used_mem -= PAGE_SIZE;
+ for (i = 0; i < npages; i++) {
+ if (page_to_pfn(pages[i]) > 0x00100000UL) {
+ glob->used_dma32_mem -= PAGE_SIZE;
+ } else {
+ glob->used_mem -= PAGE_SIZE;
+ }
}
spin_unlock(&glob->lock);
}
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index 727e93d..c4f18b9 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -619,8 +619,10 @@ static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
* @return count of pages still required to fulfill the request.
*/
static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
- struct list_head *pages, int ttm_flags,
- enum ttm_caching_state cstate, unsigned count)
+ struct list_head *pages,
+ int ttm_flags,
+ enum ttm_caching_state cstate,
+ unsigned count)
{
unsigned long irq_flags;
struct list_head *p;
@@ -664,13 +666,14 @@ out:
* On success pages list will hold count number of correctly
* cached pages.
*/
-int ttm_get_pages(struct list_head *pages, int flags,
- enum ttm_caching_state cstate, unsigned count,
- dma_addr_t *dma_address)
+int ttm_get_pages(struct page **pages, unsigned npages, int flags,
+ enum ttm_caching_state cstate, dma_addr_t *dma_address)
{
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
struct page *p = NULL;
+ struct list_head plist;
gfp_t gfp_flags = GFP_USER;
+ unsigned count = 0;
int r;
/* set zero flag for page allocation if required */
@@ -684,94 +687,107 @@ int ttm_get_pages(struct list_head *pages, int flags,
else
gfp_flags |= GFP_HIGHUSER;
- for (r = 0; r < count; ++r) {
- p = alloc_page(gfp_flags);
- if (!p) {
-
+ for (count = 0; count < npages; ++count) {
+ pages[count] = alloc_page(gfp_flags);
+ if (pages[count] == NULL) {
printk(KERN_ERR TTM_PFX
"Unable to allocate page.");
return -ENOMEM;
}
-
- list_add(&p->lru, pages);
}
return 0;
}
-
/* combine zero flag to pool flags */
gfp_flags |= pool->gfp_flags;
/* First we take pages from the pool */
- count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count);
+ INIT_LIST_HEAD(&plist);
+ npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
/* clear the pages coming from the pool if requested */
+ count = 0;
+ list_for_each_entry(p, &plist, lru) {
+ pages[count++] = p;
+ }
if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
- list_for_each_entry(p, pages, lru) {
+ list_for_each_entry(p, &plist, lru) {
clear_page(page_address(p));
}
}
/* If pool didn't have enough pages allocate new one. */
- if (count > 0) {
+ if (npages > 0) {
/* ttm_alloc_new_pages doesn't reference pool so we can run
* multiple requests in parallel.
**/
- r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count);
+ INIT_LIST_HEAD(&plist);
+ r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate, npages);
+ list_for_each_entry(p, &plist, lru) {
+ pages[count++] = p;
+ }
if (r) {
/* If there is any pages in the list put them back to
* the pool. */
printk(KERN_ERR TTM_PFX
"Failed to allocate extra pages "
"for large request.");
- ttm_put_pages(pages, 0, flags, cstate, NULL);
+ ttm_put_pages(pages, count, flags, cstate, NULL);
return r;
}
}
-
return 0;
}
/* Put all pages in pages list to correct pool to wait for reuse */
-void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
+void ttm_put_pages(struct page **pages, unsigned npages, int flags,
enum ttm_caching_state cstate, dma_addr_t *dma_address)
{
unsigned long irq_flags;
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
- struct page *p, *tmp;
+ unsigned i;
if (pool == NULL) {
/* No pool for this memory type so free the pages */
- list_for_each_entry_safe(p, tmp, pages, lru) {
- __free_page(p);
+ for (i = 0; i < npages; i++) {
+ if (pages[i]) {
+ if (page_count(pages[i]) != 1)
+ printk(KERN_ERR TTM_PFX
+ "Erroneous page count. "
+ "Leaking pages.\n");
+ __free_page(pages[i]);
+ pages[i] = NULL;
+ }
}
- /* Make the pages list empty */
- INIT_LIST_HEAD(pages);
return;
}
- if (page_count == 0) {
- list_for_each_entry_safe(p, tmp, pages, lru) {
- ++page_count;
- }
- }
spin_lock_irqsave(&pool->lock, irq_flags);
- list_splice_init(pages, &pool->list);
- pool->npages += page_count;
+ for (i = 0; i < npages; i++) {
+ if (pages[i]) {
+ if (page_count(pages[i]) != 1)
+ printk(KERN_ERR TTM_PFX
+ "Erroneous page count. "
+ "Leaking pages.\n");
+ list_add_tail(&pages[i]->lru, &pool->list);
+ pages[i] = NULL;
+ pool->npages++;
+ }
+ }
/* Check that we don't go over the pool limit */
- page_count = 0;
+ npages = 0;
if (pool->npages > _manager->options.max_size) {
- page_count = pool->npages - _manager->options.max_size;
+ npages = pool->npages - _manager->options.max_size;
/* free at least NUM_PAGES_TO_ALLOC number of pages
* to reduce calls to set_memory_wb */
- if (page_count < NUM_PAGES_TO_ALLOC)
- page_count = NUM_PAGES_TO_ALLOC;
+ if (npages < NUM_PAGES_TO_ALLOC)
+ npages = NUM_PAGES_TO_ALLOC;
}
spin_unlock_irqrestore(&pool->lock, irq_flags);
- if (page_count)
- ttm_page_pool_free(pool, page_count);
+ if (npages)
+ ttm_page_pool_free(pool, npages);
}
static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index e5432efa..303bbba 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -65,35 +65,23 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm)
static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
{
struct page *p;
- struct list_head h;
struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
int ret;
if (NULL == (p = ttm->pages[index])) {
+ ret = ttm_mem_global_alloc_pages(mem_glob, 1, false);
+ if (unlikely(ret != 0))
+ return NULL;
- INIT_LIST_HEAD(&h);
-
- ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1,
+ ret = ttm_get_pages(&ttm->pages[index], 1, ttm->page_flags,
+ ttm->caching_state,
&ttm->dma_address[index]);
-
if (ret != 0)
return NULL;
- p = list_first_entry(&h, struct page, lru);
-
- ret = ttm_mem_global_alloc_page(mem_glob, p, false);
- if (unlikely(ret != 0))
- goto out_err;
-
- ttm->pages[index] = p;
+ ttm_mem_global_account_pages(mem_glob, &ttm->pages[index], 1);
}
return p;
-out_err:
- INIT_LIST_HEAD(&h);
- list_add(&p->lru, &h);
- ttm_put_pages(&h, 1, ttm->page_flags,
- ttm->caching_state, &ttm->dma_address[index]);
- return NULL;
}
struct page *ttm_tt_get_page(struct ttm_tt *ttm, int index)
@@ -110,8 +98,7 @@ struct page *ttm_tt_get_page(struct ttm_tt *ttm, int index)
int ttm_tt_populate(struct ttm_tt *ttm)
{
- struct page *page;
- unsigned long i;
+ struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
struct ttm_backend *be;
int ret;
@@ -126,11 +113,16 @@ int ttm_tt_populate(struct ttm_tt *ttm)
be = ttm->be;
- for (i = 0; i < ttm->num_pages; ++i) {
- page = __ttm_tt_get_page(ttm, i);
- if (!page)
- return -ENOMEM;
- }
+ ret = ttm_mem_global_alloc_pages(mem_glob, ttm->num_pages, false);
+ if (unlikely(ret != 0))
+ return -ENOMEM;
+
+ ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
+ ttm->caching_state, ttm->dma_address);
+ if (ret != 0)
+ return -ENOMEM;
+
+ ttm_mem_global_account_pages(mem_glob, ttm->pages, ttm->num_pages);
be->func->populate(be, ttm->num_pages, ttm->pages,
ttm->dummy_read_page, ttm->dma_address);
@@ -242,33 +234,15 @@ EXPORT_SYMBOL(ttm_tt_set_placement_caching);
static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
{
- int i;
- unsigned count = 0;
- struct list_head h;
- struct page *cur_page;
struct ttm_backend *be = ttm->be;
-
- INIT_LIST_HEAD(&h);
+ struct ttm_mem_global *glob = ttm->glob->mem_glob;
if (be)
be->func->clear(be);
- for (i = 0; i < ttm->num_pages; ++i) {
- cur_page = ttm->pages[i];
- ttm->pages[i] = NULL;
- if (cur_page) {
- if (page_count(cur_page) != 1)
- printk(KERN_ERR TTM_PFX
- "Erroneous page count. "
- "Leaking pages.\n");
- ttm_mem_global_free_page(ttm->glob->mem_glob,
- cur_page);
- list_add(&cur_page->lru, &h);
- count++;
- }
- }
- ttm_put_pages(&h, count, ttm->page_flags, ttm->caching_state,
- ttm->dma_address);
+ ttm_mem_global_free_pages(glob, ttm->pages, ttm->num_pages);
+ ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
+ ttm->caching_state, ttm->dma_address);
ttm->state = tt_unpopulated;
}
diff --git a/include/drm/ttm/ttm_memory.h b/include/drm/ttm/ttm_memory.h
index d24792b..c0c732d 100644
--- a/include/drm/ttm/ttm_memory.h
+++ b/include/drm/ttm/ttm_memory.h
@@ -150,11 +150,14 @@ extern int ttm_mem_global_alloc(struct ttm_mem_global *glob,
bool no_wait);
extern void ttm_mem_global_free(struct ttm_mem_global *glob,
uint64_t amount);
-extern int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
- struct page *page,
- bool no_wait);
-extern void ttm_mem_global_free_page(struct ttm_mem_global *glob,
- struct page *page);
+extern int ttm_mem_global_alloc_pages(struct ttm_mem_global *glob,
+ unsigned npages,
+ bool no_wait);
+extern void ttm_mem_global_free_pages(struct ttm_mem_global *glob,
+ struct page **pages, unsigned npages);
+extern void ttm_mem_global_account_pages(struct ttm_mem_global *glob,
+ struct page **pages,
+ unsigned npages);
extern size_t ttm_round_pot(size_t size);
#endif
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 129de12..fffb3bd 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -32,29 +32,28 @@
/**
* Get count number of pages from pool to pages list.
*
- * @pages: head of empty linked list where pages are filled.
+ * @pages: array of pages ptr
+ * @npages: number of pages to allocate.
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state for the page.
- * @count: number of pages to allocate.
* @dma_address: The DMA (bus) address of pages (if TTM_PAGE_FLAG_DMA32 set).
*/
-int ttm_get_pages(struct list_head *pages,
+int ttm_get_pages(struct page **pages,
+ unsigned npages,
int flags,
enum ttm_caching_state cstate,
- unsigned count,
dma_addr_t *dma_address);
/**
* Put linked list of pages to pool.
*
- * @pages: list of pages to free.
- * @page_count: number of pages in the list. Zero can be passed for unknown
- * count.
+ * @pages: array of pages ptr
+ * @npages: number of pages to free.
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
* @dma_address: The DMA (bus) address of pages (if TTM_PAGE_FLAG_DMA32 set).
*/
-void ttm_put_pages(struct list_head *pages,
- unsigned page_count,
+void ttm_put_pages(struct page **pages,
+ unsigned npages,
int flags,
enum ttm_caching_state cstate,
dma_addr_t *dma_address);
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 07/13] drm/ttm: test for dma_address array allocation failure
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (5 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 06/13] drm/ttm: convert page allocation to use page ptr array instead of list V4 j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 08/13] drm/ttm: merge ttm_backend and ttm_tt V2 j.glisse
` (6 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
---
drivers/gpu/drm/ttm/ttm_tt.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 303bbba..2dab08b 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -293,7 +293,7 @@ struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
ttm->dummy_read_page = dummy_read_page;
ttm_tt_alloc_page_directory(ttm);
- if (!ttm->pages) {
+ if (!ttm->pages || !ttm->dma_address) {
ttm_tt_destroy(ttm);
printk(KERN_ERR TTM_PFX "Failed allocating page table\n");
return NULL;
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 08/13] drm/ttm: merge ttm_backend and ttm_tt V2
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (6 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 07/13] drm/ttm: test for dma_address array allocation failure j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 09/13] drm/ttm: introduce callback for ttm_tt populate & unpopulate V2 j.glisse
` (5 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
ttm_backend will exist only and only with a ttm_tt, and ttm_tt
will be of interesting use only when bind to a backend. Thus to
avoid code & data duplication btw the two merge them.
V2 Rebase on top of memory accountign overhaul
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/nouveau/nouveau_bo.c | 14 ++-
drivers/gpu/drm/nouveau/nouveau_drv.h | 5 +-
drivers/gpu/drm/nouveau/nouveau_sgdma.c | 188 ++++++++++++--------------
drivers/gpu/drm/radeon/radeon_ttm.c | 222 ++++++++++++-------------------
drivers/gpu/drm/ttm/ttm_agp_backend.c | 88 +++++--------
drivers/gpu/drm/ttm/ttm_bo.c | 9 +-
drivers/gpu/drm/ttm/ttm_tt.c | 60 ++-------
drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c | 66 +++------
include/drm/ttm/ttm_bo_driver.h | 104 ++++++---------
9 files changed, 295 insertions(+), 461 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 7226f41..b060fa4 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -343,8 +343,10 @@ nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
*mem = val;
}
-static struct ttm_backend *
-nouveau_bo_create_ttm_backend_entry(struct ttm_bo_device *bdev)
+static struct ttm_tt *
+nouveau_ttm_tt_create(struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
{
struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
struct drm_device *dev = dev_priv->dev;
@@ -352,11 +354,13 @@ nouveau_bo_create_ttm_backend_entry(struct ttm_bo_device *bdev)
switch (dev_priv->gart_info.type) {
#if __OS_HAS_AGP
case NOUVEAU_GART_AGP:
- return ttm_agp_backend_init(bdev, dev->agp->bridge);
+ return ttm_agp_tt_create(bdev, dev->agp->bridge,
+ size, page_flags, dummy_read_page);
#endif
case NOUVEAU_GART_PDMA:
case NOUVEAU_GART_HW:
- return nouveau_sgdma_init_ttm(dev);
+ return nouveau_sgdma_create_ttm(bdev, size, page_flags,
+ dummy_read_page);
default:
NV_ERROR(dev, "Unknown GART type %d\n",
dev_priv->gart_info.type);
@@ -1045,7 +1049,7 @@ nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
}
struct ttm_bo_driver nouveau_bo_driver = {
- .create_ttm_backend_entry = nouveau_bo_create_ttm_backend_entry,
+ .ttm_tt_create = &nouveau_ttm_tt_create,
.invalidate_caches = nouveau_bo_invalidate_caches,
.init_mem_type = nouveau_bo_init_mem_type,
.evict_flags = nouveau_bo_evict_flags,
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 29837da..0c53e39 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -1000,7 +1000,10 @@ extern int nouveau_sgdma_init(struct drm_device *);
extern void nouveau_sgdma_takedown(struct drm_device *);
extern uint32_t nouveau_sgdma_get_physical(struct drm_device *,
uint32_t offset);
-extern struct ttm_backend *nouveau_sgdma_init_ttm(struct drm_device *);
+extern struct ttm_tt *nouveau_sgdma_create_ttm(struct ttm_bo_device *bdev,
+ unsigned long size,
+ uint32_t page_flags,
+ struct page *dummy_read_page);
/* nouveau_debugfs.c */
#if defined(CONFIG_DRM_NOUVEAU_DEBUG)
diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
index b75258a..bc2ab90 100644
--- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c
+++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
@@ -8,44 +8,23 @@
#define NV_CTXDMA_PAGE_MASK (NV_CTXDMA_PAGE_SIZE - 1)
struct nouveau_sgdma_be {
- struct ttm_backend backend;
+ struct ttm_tt ttm;
struct drm_device *dev;
-
- dma_addr_t *pages;
- unsigned nr_pages;
- bool unmap_pages;
-
u64 offset;
- bool bound;
};
static int
-nouveau_sgdma_populate(struct ttm_backend *be, unsigned long num_pages,
- struct page **pages, struct page *dummy_read_page,
- dma_addr_t *dma_addrs)
+nouveau_sgdma_dma_map(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_device *dev = nvbe->dev;
int i;
- NV_DEBUG(nvbe->dev, "num_pages = %ld\n", num_pages);
-
- nvbe->pages = dma_addrs;
- nvbe->nr_pages = num_pages;
- nvbe->unmap_pages = true;
-
- /* this code path isn't called and is incorrect anyways */
- if (0) { /* dma_addrs[0] != DMA_ERROR_CODE) { */
- nvbe->unmap_pages = false;
- return 0;
- }
-
- for (i = 0; i < num_pages; i++) {
- nvbe->pages[i] = pci_map_page(dev->pdev, pages[i], 0,
- PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
- if (pci_dma_mapping_error(dev->pdev, nvbe->pages[i])) {
- nvbe->nr_pages = --i;
- be->func->clear(be);
+ for (i = 0; i < ttm->num_pages; i++) {
+ ttm->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
+ 0, PAGE_SIZE,
+ PCI_DMA_BIDIRECTIONAL);
+ if (pci_dma_mapping_error(dev->pdev, ttm->dma_address[i])) {
return -EFAULT;
}
}
@@ -54,53 +33,52 @@ nouveau_sgdma_populate(struct ttm_backend *be, unsigned long num_pages,
}
static void
-nouveau_sgdma_clear(struct ttm_backend *be)
+nouveau_sgdma_dma_unmap(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_device *dev = nvbe->dev;
+ int i;
- if (nvbe->bound)
- be->func->unbind(be);
-
- if (nvbe->unmap_pages) {
- while (nvbe->nr_pages--) {
- pci_unmap_page(dev->pdev, nvbe->pages[nvbe->nr_pages],
+ for (i = 0; i < ttm->num_pages; i++) {
+ if (ttm->dma_address[i]) {
+ pci_unmap_page(dev->pdev, ttm->dma_address[i],
PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
}
+ ttm->dma_address[i] = 0;
}
}
static void
-nouveau_sgdma_destroy(struct ttm_backend *be)
+nouveau_sgdma_destroy(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
- if (be) {
+ if (ttm) {
NV_DEBUG(nvbe->dev, "\n");
-
- if (nvbe) {
- if (nvbe->pages)
- be->func->clear(be);
- kfree(nvbe);
- }
+ kfree(nvbe);
}
}
static int
-nv04_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
+nv04_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_device *dev = nvbe->dev;
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
unsigned i, j, pte;
+ int r;
NV_DEBUG(dev, "pg=0x%lx\n", mem->start);
+ r = nouveau_sgdma_dma_map(ttm);
+ if (r) {
+ return r;
+ }
nvbe->offset = mem->start << PAGE_SHIFT;
pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
- for (i = 0; i < nvbe->nr_pages; i++) {
- dma_addr_t dma_offset = nvbe->pages[i];
+ for (i = 0; i < ttm->num_pages; i++) {
+ dma_addr_t dma_offset = ttm->dma_address[i];
uint32_t offset_l = lower_32_bits(dma_offset);
for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++) {
@@ -109,14 +87,13 @@ nv04_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
}
}
- nvbe->bound = true;
return 0;
}
static int
-nv04_sgdma_unbind(struct ttm_backend *be)
+nv04_sgdma_unbind(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_device *dev = nvbe->dev;
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
@@ -124,22 +101,20 @@ nv04_sgdma_unbind(struct ttm_backend *be)
NV_DEBUG(dev, "\n");
- if (!nvbe->bound)
+ if (ttm->state != tt_bound)
return 0;
pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
- for (i = 0; i < nvbe->nr_pages; i++) {
+ for (i = 0; i < ttm->num_pages; i++) {
for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++)
nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
}
- nvbe->bound = false;
+ nouveau_sgdma_dma_unmap(ttm);
return 0;
}
static struct ttm_backend_func nv04_sgdma_backend = {
- .populate = nouveau_sgdma_populate,
- .clear = nouveau_sgdma_clear,
.bind = nv04_sgdma_bind,
.unbind = nv04_sgdma_unbind,
.destroy = nouveau_sgdma_destroy
@@ -158,16 +133,21 @@ nv41_sgdma_flush(struct nouveau_sgdma_be *nvbe)
}
static int
-nv41_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
+nv41_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
- dma_addr_t *list = nvbe->pages;
+ dma_addr_t *list = ttm->dma_address;
u32 pte = mem->start << 2;
- u32 cnt = nvbe->nr_pages;
+ u32 cnt = ttm->num_pages;
+ int r;
nvbe->offset = mem->start << PAGE_SHIFT;
+ r = nouveau_sgdma_dma_map(ttm);
+ if (r) {
+ return r;
+ }
while (cnt--) {
nv_wo32(pgt, pte, (*list++ >> 7) | 1);
@@ -175,18 +155,17 @@ nv41_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
}
nv41_sgdma_flush(nvbe);
- nvbe->bound = true;
return 0;
}
static int
-nv41_sgdma_unbind(struct ttm_backend *be)
+nv41_sgdma_unbind(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
u32 pte = (nvbe->offset >> 12) << 2;
- u32 cnt = nvbe->nr_pages;
+ u32 cnt = ttm->num_pages;
while (cnt--) {
nv_wo32(pgt, pte, 0x00000000);
@@ -194,24 +173,23 @@ nv41_sgdma_unbind(struct ttm_backend *be)
}
nv41_sgdma_flush(nvbe);
- nvbe->bound = false;
+ nouveau_sgdma_dma_unmap(ttm);
return 0;
}
static struct ttm_backend_func nv41_sgdma_backend = {
- .populate = nouveau_sgdma_populate,
- .clear = nouveau_sgdma_clear,
.bind = nv41_sgdma_bind,
.unbind = nv41_sgdma_unbind,
.destroy = nouveau_sgdma_destroy
};
static void
-nv44_sgdma_flush(struct nouveau_sgdma_be *nvbe)
+nv44_sgdma_flush(struct ttm_tt *ttm)
{
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_device *dev = nvbe->dev;
- nv_wr32(dev, 0x100814, (nvbe->nr_pages - 1) << 12);
+ nv_wr32(dev, 0x100814, (ttm->num_pages - 1) << 12);
nv_wr32(dev, 0x100808, nvbe->offset | 0x20);
if (!nv_wait(dev, 0x100808, 0x00000001, 0x00000001))
NV_ERROR(dev, "gart flush timeout: 0x%08x\n",
@@ -270,17 +248,21 @@ nv44_sgdma_fill(struct nouveau_gpuobj *pgt, dma_addr_t *list, u32 base, u32 cnt)
}
static int
-nv44_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
+nv44_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
- dma_addr_t *list = nvbe->pages;
+ dma_addr_t *list = ttm->dma_address;
u32 pte = mem->start << 2, tmp[4];
- u32 cnt = nvbe->nr_pages;
- int i;
+ u32 cnt = ttm->num_pages;
+ int i, r;
nvbe->offset = mem->start << PAGE_SHIFT;
+ r = nouveau_sgdma_dma_map(ttm);
+ if (r) {
+ return r;
+ }
if (pte & 0x0000000c) {
u32 max = 4 - ((pte >> 2) & 0x3);
@@ -305,19 +287,18 @@ nv44_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
if (cnt)
nv44_sgdma_fill(pgt, list, pte, cnt);
- nv44_sgdma_flush(nvbe);
- nvbe->bound = true;
+ nv44_sgdma_flush(ttm);
return 0;
}
static int
-nv44_sgdma_unbind(struct ttm_backend *be)
+nv44_sgdma_unbind(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+ struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
u32 pte = (nvbe->offset >> 12) << 2;
- u32 cnt = nvbe->nr_pages;
+ u32 cnt = ttm->num_pages;
if (pte & 0x0000000c) {
u32 max = 4 - ((pte >> 2) & 0x3);
@@ -339,55 +320,53 @@ nv44_sgdma_unbind(struct ttm_backend *be)
if (cnt)
nv44_sgdma_fill(pgt, NULL, pte, cnt);
- nv44_sgdma_flush(nvbe);
- nvbe->bound = false;
+ nv44_sgdma_flush(ttm);
+ nouveau_sgdma_dma_unmap(ttm);
return 0;
}
static struct ttm_backend_func nv44_sgdma_backend = {
- .populate = nouveau_sgdma_populate,
- .clear = nouveau_sgdma_clear,
.bind = nv44_sgdma_bind,
.unbind = nv44_sgdma_unbind,
.destroy = nouveau_sgdma_destroy
};
static int
-nv50_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
+nv50_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
struct nouveau_mem *node = mem->mm_node;
+ int r;
+
/* noop: bound in move_notify() */
- node->pages = nvbe->pages;
- nvbe->pages = (dma_addr_t *)node;
- nvbe->bound = true;
+ r = nouveau_sgdma_dma_map(ttm);
+ if (r) {
+ return r;
+ }
+ node->pages = ttm->dma_address;
return 0;
}
static int
-nv50_sgdma_unbind(struct ttm_backend *be)
+nv50_sgdma_unbind(struct ttm_tt *ttm)
{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
- struct nouveau_mem *node = (struct nouveau_mem *)nvbe->pages;
/* noop: unbound in move_notify() */
- nvbe->pages = node->pages;
- node->pages = NULL;
- nvbe->bound = false;
+ nouveau_sgdma_dma_unmap(ttm);
return 0;
}
static struct ttm_backend_func nv50_sgdma_backend = {
- .populate = nouveau_sgdma_populate,
- .clear = nouveau_sgdma_clear,
.bind = nv50_sgdma_bind,
.unbind = nv50_sgdma_unbind,
.destroy = nouveau_sgdma_destroy
};
-struct ttm_backend *
-nouveau_sgdma_init_ttm(struct drm_device *dev)
+struct ttm_tt *
+nouveau_sgdma_create_ttm(struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
+ struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
+ struct drm_device *dev = dev_priv->dev;
struct nouveau_sgdma_be *nvbe;
nvbe = kzalloc(sizeof(*nvbe), GFP_KERNEL);
@@ -395,9 +374,12 @@ nouveau_sgdma_init_ttm(struct drm_device *dev)
return NULL;
nvbe->dev = dev;
+ nvbe->ttm.func = dev_priv->gart_info.func;
- nvbe->backend.func = dev_priv->gart_info.func;
- return &nvbe->backend;
+ if (ttm_tt_init(&nvbe->ttm, bdev, size, page_flags, dummy_read_page)) {
+ return NULL;
+ }
+ return &nvbe->ttm;
}
int
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 97c76ae..53ff62b 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -114,24 +114,6 @@ static void radeon_ttm_global_fini(struct radeon_device *rdev)
}
}
-struct ttm_backend *radeon_ttm_backend_create(struct radeon_device *rdev);
-
-static struct ttm_backend*
-radeon_create_ttm_backend_entry(struct ttm_bo_device *bdev)
-{
- struct radeon_device *rdev;
-
- rdev = radeon_get_rdev(bdev);
-#if __OS_HAS_AGP
- if (rdev->flags & RADEON_IS_AGP) {
- return ttm_agp_backend_init(bdev, rdev->ddev->agp->bridge);
- } else
-#endif
- {
- return radeon_ttm_backend_create(rdev);
- }
-}
-
static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
{
return 0;
@@ -515,8 +497,93 @@ static bool radeon_sync_obj_signaled(void *sync_obj, void *sync_arg)
return radeon_fence_signaled((struct radeon_fence *)sync_obj);
}
+/*
+ * TTM backend functions.
+ */
+struct radeon_ttm_tt {
+ struct ttm_tt ttm;
+ struct radeon_device *rdev;
+ u64 offset;
+};
+
+static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
+ struct ttm_mem_reg *bo_mem)
+{
+ struct radeon_ttm_tt *gtt;
+ int r;
+
+ gtt = container_of(ttm, struct radeon_ttm_tt, ttm);
+ gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
+ if (!ttm->num_pages) {
+ WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
+ ttm->num_pages, bo_mem, ttm);
+ }
+ r = radeon_gart_bind(gtt->rdev, gtt->offset,
+ ttm->num_pages, ttm->pages, ttm->dma_address);
+ if (r) {
+ DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
+ ttm->num_pages, (unsigned)gtt->offset);
+ return r;
+ }
+ return 0;
+}
+
+static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
+{
+ struct radeon_ttm_tt *gtt;
+
+ gtt = container_of(ttm, struct radeon_ttm_tt, ttm);
+ radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
+ return 0;
+}
+
+static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
+{
+ struct radeon_ttm_tt *gtt;
+
+ gtt = container_of(ttm, struct radeon_ttm_tt, ttm);
+ if (ttm->state == tt_bound) {
+ radeon_ttm_backend_unbind(ttm);
+ }
+ kfree(gtt);
+}
+
+static struct ttm_backend_func radeon_backend_func = {
+ .bind = &radeon_ttm_backend_bind,
+ .unbind = &radeon_ttm_backend_unbind,
+ .destroy = &radeon_ttm_backend_destroy,
+};
+
+struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
+{
+ struct radeon_device *rdev;
+ struct radeon_ttm_tt *gtt;
+
+ rdev = radeon_get_rdev(bdev);
+#if __OS_HAS_AGP
+ if (rdev->flags & RADEON_IS_AGP) {
+ return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
+ size, page_flags, dummy_read_page);
+ }
+#endif
+
+ gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
+ if (gtt == NULL) {
+ return NULL;
+ }
+ gtt->ttm.func = &radeon_backend_func;
+ gtt->rdev = rdev;
+ if (ttm_tt_init(>t->ttm, bdev, size, page_flags, dummy_read_page)) {
+ return NULL;
+ }
+ return >t->ttm;
+}
+
+
static struct ttm_bo_driver radeon_bo_driver = {
- .create_ttm_backend_entry = &radeon_create_ttm_backend_entry,
+ .ttm_tt_create = &radeon_ttm_tt_create,
.invalidate_caches = &radeon_invalidate_caches,
.init_mem_type = &radeon_init_mem_type,
.evict_flags = &radeon_evict_flags,
@@ -680,123 +747,6 @@ int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
}
-/*
- * TTM backend functions.
- */
-struct radeon_ttm_backend {
- struct ttm_backend backend;
- struct radeon_device *rdev;
- unsigned long num_pages;
- struct page **pages;
- struct page *dummy_read_page;
- dma_addr_t *dma_addrs;
- bool populated;
- bool bound;
- unsigned offset;
-};
-
-static int radeon_ttm_backend_populate(struct ttm_backend *backend,
- unsigned long num_pages,
- struct page **pages,
- struct page *dummy_read_page,
- dma_addr_t *dma_addrs)
-{
- struct radeon_ttm_backend *gtt;
-
- gtt = container_of(backend, struct radeon_ttm_backend, backend);
- gtt->pages = pages;
- gtt->dma_addrs = dma_addrs;
- gtt->num_pages = num_pages;
- gtt->dummy_read_page = dummy_read_page;
- gtt->populated = true;
- return 0;
-}
-
-static void radeon_ttm_backend_clear(struct ttm_backend *backend)
-{
- struct radeon_ttm_backend *gtt;
-
- gtt = container_of(backend, struct radeon_ttm_backend, backend);
- gtt->pages = NULL;
- gtt->dma_addrs = NULL;
- gtt->num_pages = 0;
- gtt->dummy_read_page = NULL;
- gtt->populated = false;
- gtt->bound = false;
-}
-
-
-static int radeon_ttm_backend_bind(struct ttm_backend *backend,
- struct ttm_mem_reg *bo_mem)
-{
- struct radeon_ttm_backend *gtt;
- int r;
-
- gtt = container_of(backend, struct radeon_ttm_backend, backend);
- gtt->offset = bo_mem->start << PAGE_SHIFT;
- if (!gtt->num_pages) {
- WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
- gtt->num_pages, bo_mem, backend);
- }
- r = radeon_gart_bind(gtt->rdev, gtt->offset,
- gtt->num_pages, gtt->pages, gtt->dma_addrs);
- if (r) {
- DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
- gtt->num_pages, gtt->offset);
- return r;
- }
- gtt->bound = true;
- return 0;
-}
-
-static int radeon_ttm_backend_unbind(struct ttm_backend *backend)
-{
- struct radeon_ttm_backend *gtt;
-
- gtt = container_of(backend, struct radeon_ttm_backend, backend);
- radeon_gart_unbind(gtt->rdev, gtt->offset, gtt->num_pages);
- gtt->bound = false;
- return 0;
-}
-
-static void radeon_ttm_backend_destroy(struct ttm_backend *backend)
-{
- struct radeon_ttm_backend *gtt;
-
- gtt = container_of(backend, struct radeon_ttm_backend, backend);
- if (gtt->bound) {
- radeon_ttm_backend_unbind(backend);
- }
- kfree(gtt);
-}
-
-static struct ttm_backend_func radeon_backend_func = {
- .populate = &radeon_ttm_backend_populate,
- .clear = &radeon_ttm_backend_clear,
- .bind = &radeon_ttm_backend_bind,
- .unbind = &radeon_ttm_backend_unbind,
- .destroy = &radeon_ttm_backend_destroy,
-};
-
-struct ttm_backend *radeon_ttm_backend_create(struct radeon_device *rdev)
-{
- struct radeon_ttm_backend *gtt;
-
- gtt = kzalloc(sizeof(struct radeon_ttm_backend), GFP_KERNEL);
- if (gtt == NULL) {
- return NULL;
- }
- gtt->backend.bdev = &rdev->mman.bdev;
- gtt->backend.func = &radeon_backend_func;
- gtt->rdev = rdev;
- gtt->pages = NULL;
- gtt->num_pages = 0;
- gtt->dummy_read_page = NULL;
- gtt->populated = false;
- gtt->bound = false;
- return >t->backend;
-}
-
#define RADEON_DEBUGFS_MEM_TYPES 2
#if defined(CONFIG_DEBUG_FS)
diff --git a/drivers/gpu/drm/ttm/ttm_agp_backend.c b/drivers/gpu/drm/ttm/ttm_agp_backend.c
index 1c4a72f..14ebd36 100644
--- a/drivers/gpu/drm/ttm/ttm_agp_backend.c
+++ b/drivers/gpu/drm/ttm/ttm_agp_backend.c
@@ -40,45 +40,33 @@
#include <asm/agp.h>
struct ttm_agp_backend {
- struct ttm_backend backend;
+ struct ttm_tt ttm;
struct agp_memory *mem;
struct agp_bridge_data *bridge;
};
-static int ttm_agp_populate(struct ttm_backend *backend,
- unsigned long num_pages, struct page **pages,
- struct page *dummy_read_page,
- dma_addr_t *dma_addrs)
+static int ttm_agp_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
{
- struct ttm_agp_backend *agp_be =
- container_of(backend, struct ttm_agp_backend, backend);
- struct page **cur_page, **last_page = pages + num_pages;
+ struct ttm_agp_backend *agp_be = container_of(ttm, struct ttm_agp_backend, ttm);
+ struct drm_mm_node *node = bo_mem->mm_node;
struct agp_memory *mem;
+ int ret, cached = (bo_mem->placement & TTM_PL_FLAG_CACHED);
+ unsigned i;
- mem = agp_allocate_memory(agp_be->bridge, num_pages, AGP_USER_MEMORY);
+ mem = agp_allocate_memory(agp_be->bridge, ttm->num_pages, AGP_USER_MEMORY);
if (unlikely(mem == NULL))
return -ENOMEM;
mem->page_count = 0;
- for (cur_page = pages; cur_page < last_page; ++cur_page) {
- struct page *page = *cur_page;
+ for (i = 0; i < ttm->num_pages; i++) {
+ struct page *page = ttm->pages[i];
+
if (!page)
- page = dummy_read_page;
+ page = ttm->dummy_read_page;
mem->pages[mem->page_count++] = page;
}
agp_be->mem = mem;
- return 0;
-}
-
-static int ttm_agp_bind(struct ttm_backend *backend, struct ttm_mem_reg *bo_mem)
-{
- struct ttm_agp_backend *agp_be =
- container_of(backend, struct ttm_agp_backend, backend);
- struct drm_mm_node *node = bo_mem->mm_node;
- struct agp_memory *mem = agp_be->mem;
- int cached = (bo_mem->placement & TTM_PL_FLAG_CACHED);
- int ret;
mem->is_flushed = 1;
mem->type = (cached) ? AGP_USER_CACHED_MEMORY : AGP_USER_MEMORY;
@@ -90,50 +78,38 @@ static int ttm_agp_bind(struct ttm_backend *backend, struct ttm_mem_reg *bo_mem)
return ret;
}
-static int ttm_agp_unbind(struct ttm_backend *backend)
+static int ttm_agp_unbind(struct ttm_tt *ttm)
{
- struct ttm_agp_backend *agp_be =
- container_of(backend, struct ttm_agp_backend, backend);
-
- if (agp_be->mem->is_bound)
- return agp_unbind_memory(agp_be->mem);
- else
- return 0;
-}
+ struct ttm_agp_backend *agp_be = container_of(ttm, struct ttm_agp_backend, ttm);
-static void ttm_agp_clear(struct ttm_backend *backend)
-{
- struct ttm_agp_backend *agp_be =
- container_of(backend, struct ttm_agp_backend, backend);
- struct agp_memory *mem = agp_be->mem;
-
- if (mem) {
- ttm_agp_unbind(backend);
- agp_free_memory(mem);
+ if (agp_be->mem) {
+ if (agp_be->mem->is_bound)
+ return agp_unbind_memory(agp_be->mem);
+ agp_free_memory(agp_be->mem);
+ agp_be->mem = NULL;
}
- agp_be->mem = NULL;
+ return 0;
}
-static void ttm_agp_destroy(struct ttm_backend *backend)
+static void ttm_agp_destroy(struct ttm_tt *ttm)
{
- struct ttm_agp_backend *agp_be =
- container_of(backend, struct ttm_agp_backend, backend);
+ struct ttm_agp_backend *agp_be = container_of(ttm, struct ttm_agp_backend, ttm);
if (agp_be->mem)
- ttm_agp_clear(backend);
+ ttm_agp_unbind(ttm);
kfree(agp_be);
}
static struct ttm_backend_func ttm_agp_func = {
- .populate = ttm_agp_populate,
- .clear = ttm_agp_clear,
.bind = ttm_agp_bind,
.unbind = ttm_agp_unbind,
.destroy = ttm_agp_destroy,
};
-struct ttm_backend *ttm_agp_backend_init(struct ttm_bo_device *bdev,
- struct agp_bridge_data *bridge)
+struct ttm_tt *ttm_agp_tt_create(struct ttm_bo_device *bdev,
+ struct agp_bridge_data *bridge,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
{
struct ttm_agp_backend *agp_be;
@@ -143,10 +119,14 @@ struct ttm_backend *ttm_agp_backend_init(struct ttm_bo_device *bdev,
agp_be->mem = NULL;
agp_be->bridge = bridge;
- agp_be->backend.func = &ttm_agp_func;
- agp_be->backend.bdev = bdev;
- return &agp_be->backend;
+ agp_be->ttm.func = &ttm_agp_func;
+
+ if (ttm_tt_init(&agp_be->ttm, bdev, size, page_flags, dummy_read_page)) {
+ return NULL;
+ }
+
+ return &agp_be->ttm;
}
-EXPORT_SYMBOL(ttm_agp_backend_init);
+EXPORT_SYMBOL(ttm_agp_tt_create);
#endif
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 92712798..c31f40f 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -337,8 +337,8 @@ static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
if (zero_alloc)
page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
case ttm_bo_type_kernel:
- bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
- page_flags, glob->dummy_read_page);
+ bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
+ page_flags, glob->dummy_read_page);
if (unlikely(bo->ttm == NULL))
ret = -ENOMEM;
break;
@@ -1437,10 +1437,7 @@ int ttm_bo_global_init(struct drm_global_reference *ref)
goto out_no_shrink;
}
- glob->ttm_bo_extra_size =
- ttm_round_pot(sizeof(struct ttm_tt)) +
- ttm_round_pot(sizeof(struct ttm_backend));
-
+ glob->ttm_bo_extra_size = ttm_round_pot(sizeof(struct ttm_tt));
glob->ttm_bo_size = glob->ttm_bo_extra_size +
ttm_round_pot(sizeof(struct ttm_buffer_object));
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 2dab08b..8c2d0df 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -99,7 +99,6 @@ struct page *ttm_tt_get_page(struct ttm_tt *ttm, int index)
int ttm_tt_populate(struct ttm_tt *ttm)
{
struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
- struct ttm_backend *be;
int ret;
if (ttm->state != tt_unpopulated)
@@ -111,8 +110,6 @@ int ttm_tt_populate(struct ttm_tt *ttm)
return ret;
}
- be = ttm->be;
-
ret = ttm_mem_global_alloc_pages(mem_glob, ttm->num_pages, false);
if (unlikely(ret != 0))
return -ENOMEM;
@@ -123,9 +120,6 @@ int ttm_tt_populate(struct ttm_tt *ttm)
return -ENOMEM;
ttm_mem_global_account_pages(mem_glob, ttm->pages, ttm->num_pages);
-
- be->func->populate(be, ttm->num_pages, ttm->pages,
- ttm->dummy_read_page, ttm->dma_address);
ttm->state = tt_unbound;
return 0;
}
@@ -234,12 +228,8 @@ EXPORT_SYMBOL(ttm_tt_set_placement_caching);
static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
{
- struct ttm_backend *be = ttm->be;
struct ttm_mem_global *glob = ttm->glob->mem_glob;
- if (be)
- be->func->clear(be);
-
ttm_mem_global_free_pages(glob, ttm->pages, ttm->num_pages);
ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
ttm->caching_state, ttm->dma_address);
@@ -248,20 +238,11 @@ static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
void ttm_tt_destroy(struct ttm_tt *ttm)
{
- struct ttm_backend *be;
-
if (unlikely(ttm == NULL))
return;
- be = ttm->be;
- if (likely(be != NULL)) {
- be->func->destroy(be);
- ttm->be = NULL;
- }
-
if (likely(ttm->pages != NULL)) {
ttm_tt_free_alloced_pages(ttm);
-
ttm_tt_free_page_directory(ttm);
}
@@ -269,52 +250,38 @@ void ttm_tt_destroy(struct ttm_tt *ttm)
ttm->swap_storage)
fput(ttm->swap_storage);
- kfree(ttm);
+ ttm->swap_storage = NULL;
+ ttm->func->destroy(ttm);
}
-struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
- uint32_t page_flags, struct page *dummy_read_page)
+int ttm_tt_init(struct ttm_tt *ttm, struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
{
- struct ttm_bo_driver *bo_driver = bdev->driver;
- struct ttm_tt *ttm;
-
- if (!bo_driver)
- return NULL;
-
- ttm = kzalloc(sizeof(*ttm), GFP_KERNEL);
- if (!ttm)
- return NULL;
-
+ ttm->bdev = bdev;
ttm->glob = bdev->glob;
ttm->num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
ttm->caching_state = tt_cached;
ttm->page_flags = page_flags;
-
ttm->dummy_read_page = dummy_read_page;
+ ttm->state = tt_unpopulated;
ttm_tt_alloc_page_directory(ttm);
if (!ttm->pages || !ttm->dma_address) {
ttm_tt_destroy(ttm);
printk(KERN_ERR TTM_PFX "Failed allocating page table\n");
- return NULL;
- }
- ttm->be = bo_driver->create_ttm_backend_entry(bdev);
- if (!ttm->be) {
- ttm_tt_destroy(ttm);
- printk(KERN_ERR TTM_PFX "Failed creating ttm backend entry\n");
- return NULL;
+ return -ENOMEM;
}
- ttm->state = tt_unpopulated;
- return ttm;
+ return 0;
}
+EXPORT_SYMBOL(ttm_tt_init);
void ttm_tt_unbind(struct ttm_tt *ttm)
{
int ret;
- struct ttm_backend *be = ttm->be;
if (ttm->state == tt_bound) {
- ret = be->func->unbind(be);
+ ret = ttm->func->unbind(ttm);
BUG_ON(ret);
ttm->state = tt_unbound;
}
@@ -323,7 +290,6 @@ void ttm_tt_unbind(struct ttm_tt *ttm)
int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
{
int ret = 0;
- struct ttm_backend *be;
if (!ttm)
return -EINVAL;
@@ -331,13 +297,11 @@ int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
if (ttm->state == tt_bound)
return 0;
- be = ttm->be;
-
ret = ttm_tt_populate(ttm);
if (ret)
return ret;
- ret = be->func->bind(be, bo_mem);
+ ret = ttm->func->bind(ttm, bo_mem);
if (unlikely(ret != 0))
return ret;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
index 5a72ed9..cc72435 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
@@ -139,85 +139,61 @@ struct ttm_placement vmw_srf_placement = {
.busy_placement = gmr_vram_placement_flags
};
-struct vmw_ttm_backend {
- struct ttm_backend backend;
- struct page **pages;
- unsigned long num_pages;
+struct vmw_ttm_tt {
+ struct ttm_tt ttm;
struct vmw_private *dev_priv;
int gmr_id;
};
-static int vmw_ttm_populate(struct ttm_backend *backend,
- unsigned long num_pages, struct page **pages,
- struct page *dummy_read_page,
- dma_addr_t *dma_addrs)
+static int vmw_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
{
- struct vmw_ttm_backend *vmw_be =
- container_of(backend, struct vmw_ttm_backend, backend);
-
- vmw_be->pages = pages;
- vmw_be->num_pages = num_pages;
-
- return 0;
-}
-
-static int vmw_ttm_bind(struct ttm_backend *backend, struct ttm_mem_reg *bo_mem)
-{
- struct vmw_ttm_backend *vmw_be =
- container_of(backend, struct vmw_ttm_backend, backend);
+ struct vmw_ttm_tt *vmw_be = container_of(ttm, struct vmw_ttm_tt, ttm);
vmw_be->gmr_id = bo_mem->start;
- return vmw_gmr_bind(vmw_be->dev_priv, vmw_be->pages,
- vmw_be->num_pages, vmw_be->gmr_id);
+ return vmw_gmr_bind(vmw_be->dev_priv, ttm->pages,
+ ttm->num_pages, vmw_be->gmr_id);
}
-static int vmw_ttm_unbind(struct ttm_backend *backend)
+static int vmw_ttm_unbind(struct ttm_tt *ttm)
{
- struct vmw_ttm_backend *vmw_be =
- container_of(backend, struct vmw_ttm_backend, backend);
+ struct vmw_ttm_tt *vmw_be = container_of(ttm, struct vmw_ttm_tt, ttm);
vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
return 0;
}
-static void vmw_ttm_clear(struct ttm_backend *backend)
+static void vmw_ttm_destroy(struct ttm_tt *ttm)
{
- struct vmw_ttm_backend *vmw_be =
- container_of(backend, struct vmw_ttm_backend, backend);
-
- vmw_be->pages = NULL;
- vmw_be->num_pages = 0;
-}
-
-static void vmw_ttm_destroy(struct ttm_backend *backend)
-{
- struct vmw_ttm_backend *vmw_be =
- container_of(backend, struct vmw_ttm_backend, backend);
+ struct vmw_ttm_tt *vmw_be = container_of(ttm, struct vmw_ttm_tt, ttm);
kfree(vmw_be);
}
static struct ttm_backend_func vmw_ttm_func = {
- .populate = vmw_ttm_populate,
- .clear = vmw_ttm_clear,
.bind = vmw_ttm_bind,
.unbind = vmw_ttm_unbind,
.destroy = vmw_ttm_destroy,
};
-struct ttm_backend *vmw_ttm_backend_init(struct ttm_bo_device *bdev)
+struct ttm_tt *vmw_ttm_tt_create(struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page)
{
- struct vmw_ttm_backend *vmw_be;
+ struct vmw_ttm_tt *vmw_be;
vmw_be = kmalloc(sizeof(*vmw_be), GFP_KERNEL);
if (!vmw_be)
return NULL;
- vmw_be->backend.func = &vmw_ttm_func;
+ vmw_be->ttm.func = &vmw_ttm_func;
vmw_be->dev_priv = container_of(bdev, struct vmw_private, bdev);
- return &vmw_be->backend;
+ if (ttm_tt_init(&vmw_be->ttm, bdev, size, page_flags, dummy_read_page)) {
+ return NULL;
+ }
+
+ return &vmw_be->ttm;
}
int vmw_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
@@ -357,7 +333,7 @@ static int vmw_sync_obj_wait(void *sync_obj, void *sync_arg,
}
struct ttm_bo_driver vmw_bo_driver = {
- .create_ttm_backend_entry = vmw_ttm_backend_init,
+ .ttm_tt_create = &vmw_ttm_tt_create,
.invalidate_caches = vmw_invalidate_caches,
.init_mem_type = vmw_init_mem_type,
.evict_flags = vmw_evict_flags,
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 6d17140..6b8c5cd 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -43,36 +43,9 @@ struct ttm_backend;
struct ttm_backend_func {
/**
- * struct ttm_backend_func member populate
- *
- * @backend: Pointer to a struct ttm_backend.
- * @num_pages: Number of pages to populate.
- * @pages: Array of pointers to ttm pages.
- * @dummy_read_page: Page to be used instead of NULL pages in the
- * array @pages.
- * @dma_addrs: Array of DMA (bus) address of the ttm pages.
- *
- * Populate the backend with ttm pages. Depending on the backend,
- * it may or may not copy the @pages array.
- */
- int (*populate) (struct ttm_backend *backend,
- unsigned long num_pages, struct page **pages,
- struct page *dummy_read_page,
- dma_addr_t *dma_addrs);
- /**
- * struct ttm_backend_func member clear
- *
- * @backend: Pointer to a struct ttm_backend.
- *
- * This is an "unpopulate" function. Release all resources
- * allocated with populate.
- */
- void (*clear) (struct ttm_backend *backend);
-
- /**
* struct ttm_backend_func member bind
*
- * @backend: Pointer to a struct ttm_backend.
+ * @ttm: Pointer to a struct ttm_tt.
* @bo_mem: Pointer to a struct ttm_mem_reg describing the
* memory type and location for binding.
*
@@ -80,40 +53,27 @@ struct ttm_backend_func {
* indicated by @bo_mem. This function should be able to handle
* differences between aperture and system page sizes.
*/
- int (*bind) (struct ttm_backend *backend, struct ttm_mem_reg *bo_mem);
+ int (*bind) (struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem);
/**
* struct ttm_backend_func member unbind
*
- * @backend: Pointer to a struct ttm_backend.
+ * @ttm: Pointer to a struct ttm_tt.
*
* Unbind previously bound backend pages. This function should be
* able to handle differences between aperture and system page sizes.
*/
- int (*unbind) (struct ttm_backend *backend);
+ int (*unbind) (struct ttm_tt *ttm);
/**
* struct ttm_backend_func member destroy
*
- * @backend: Pointer to a struct ttm_backend.
+ * @ttm: Pointer to a struct ttm_tt.
*
- * Destroy the backend.
+ * Destroy the backend. This will be call back from ttm_tt_destroy so
+ * don't call ttm_tt_destroy from the callback or infinite loop.
*/
- void (*destroy) (struct ttm_backend *backend);
-};
-
-/**
- * struct ttm_backend
- *
- * @bdev: Pointer to a struct ttm_bo_device.
- * @func: Pointer to a struct ttm_backend_func that describes
- * the backend methods.
- *
- */
-
-struct ttm_backend {
- struct ttm_bo_device *bdev;
- struct ttm_backend_func *func;
+ void (*destroy) (struct ttm_tt *ttm);
};
#define TTM_PAGE_FLAG_WRITE (1 << 3)
@@ -131,6 +91,9 @@ enum ttm_caching_state {
/**
* struct ttm_tt
*
+ * @bdev: Pointer to a struct ttm_bo_device.
+ * @func: Pointer to a struct ttm_backend_func that describes
+ * the backend methods.
* @dummy_read_page: Page to map where the ttm_tt page array contains a NULL
* pointer.
* @pages: Array of pages backing the data.
@@ -148,6 +111,8 @@ enum ttm_caching_state {
*/
struct ttm_tt {
+ struct ttm_bo_device *bdev;
+ struct ttm_backend_func *func;
struct page *dummy_read_page;
struct page **pages;
uint32_t page_flags;
@@ -336,15 +301,22 @@ struct ttm_mem_type_manager {
struct ttm_bo_driver {
/**
- * struct ttm_bo_driver member create_ttm_backend_entry
+ * ttm_tt_create
*
- * @bdev: The buffer object device.
+ * @bdev: pointer to a struct ttm_bo_device:
+ * @size: Size of the data needed backing.
+ * @page_flags: Page flags as identified by TTM_PAGE_FLAG_XX flags.
+ * @dummy_read_page: See struct ttm_bo_device.
*
- * Create a driver specific struct ttm_backend.
+ * Create a struct ttm_tt to back data with system memory pages.
+ * No pages are actually allocated.
+ * Returns:
+ * NULL: Out of memory.
*/
-
- struct ttm_backend *(*create_ttm_backend_entry)
- (struct ttm_bo_device *bdev);
+ struct ttm_tt *(*ttm_tt_create)(struct ttm_bo_device *bdev,
+ unsigned long size,
+ uint32_t page_flags,
+ struct page *dummy_read_page);
/**
* struct ttm_bo_driver member invalidate_caches
@@ -585,8 +557,9 @@ ttm_flag_masked(uint32_t *old, uint32_t new, uint32_t mask)
}
/**
- * ttm_tt_create
+ * ttm_tt_init
*
+ * @ttm: The struct ttm_tt.
* @bdev: pointer to a struct ttm_bo_device:
* @size: Size of the data needed backing.
* @page_flags: Page flags as identified by TTM_PAGE_FLAG_XX flags.
@@ -597,10 +570,9 @@ ttm_flag_masked(uint32_t *old, uint32_t new, uint32_t mask)
* Returns:
* NULL: Out of memory.
*/
-extern struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev,
- unsigned long size,
- uint32_t page_flags,
- struct page *dummy_read_page);
+extern int ttm_tt_init(struct ttm_tt *ttm, struct ttm_bo_device *bdev,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page);
/**
* ttm_ttm_bind:
@@ -626,7 +598,7 @@ extern int ttm_tt_populate(struct ttm_tt *ttm);
*
* @ttm: The struct ttm_tt.
*
- * Unbind, unpopulate and destroy a struct ttm_tt.
+ * Unbind, unpopulate and destroy common struct ttm_tt.
*/
extern void ttm_tt_destroy(struct ttm_tt *ttm);
@@ -1013,17 +985,23 @@ extern const struct ttm_mem_type_manager_func ttm_bo_manager_func;
#include <linux/agp_backend.h>
/**
- * ttm_agp_backend_init
+ * ttm_agp_tt_create
*
* @bdev: Pointer to a struct ttm_bo_device.
* @bridge: The agp bridge this device is sitting on.
+ * @size: Size of the data needed backing.
+ * @page_flags: Page flags as identified by TTM_PAGE_FLAG_XX flags.
+ * @dummy_read_page: See struct ttm_bo_device.
+ *
*
* Create a TTM backend that uses the indicated AGP bridge as an aperture
* for TT memory. This function uses the linux agpgart interface to
* bind and unbind memory backing a ttm_tt.
*/
-extern struct ttm_backend *ttm_agp_backend_init(struct ttm_bo_device *bdev,
- struct agp_bridge_data *bridge);
+extern struct ttm_tt *ttm_agp_tt_create(struct ttm_bo_device *bdev,
+ struct agp_bridge_data *bridge,
+ unsigned long size, uint32_t page_flags,
+ struct page *dummy_read_page);
#endif
#endif
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 09/13] drm/ttm: introduce callback for ttm_tt populate & unpopulate V2
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (7 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 08/13] drm/ttm: merge ttm_backend and ttm_tt V2 j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 10/13] ttm: Provide DMA aware TTM page pool code. V5 j.glisse
` (4 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Jerome Glisse
From: Jerome Glisse <jglisse@redhat.com>
Move the page allocation and freeing to driver callback and
provide ttm code helper function for those.
Most intrusive change, is the fact that we now only fully
populate an object this simplify some of code designed around
the page fault design.
V2 Rebase on top of memory accounting overhaul
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/nouveau/nouveau_bo.c | 3 +
drivers/gpu/drm/radeon/radeon_ttm.c | 2 +
drivers/gpu/drm/ttm/ttm_bo_util.c | 31 ++++++-----
drivers/gpu/drm/ttm/ttm_bo_vm.c | 13 +++--
drivers/gpu/drm/ttm/ttm_page_alloc.c | 45 +++++++++++++++++
drivers/gpu/drm/ttm/ttm_tt.c | 86 ++------------------------------
drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c | 3 +
include/drm/ttm/ttm_bo_driver.h | 41 +++++++++-------
include/drm/ttm/ttm_page_alloc.h | 18 +++++++
9 files changed, 123 insertions(+), 119 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index b060fa4..7e5ca3f 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -28,6 +28,7 @@
*/
#include "drmP.h"
+#include "ttm/ttm_page_alloc.h"
#include "nouveau_drm.h"
#include "nouveau_drv.h"
@@ -1050,6 +1051,8 @@ nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
struct ttm_bo_driver nouveau_bo_driver = {
.ttm_tt_create = &nouveau_ttm_tt_create,
+ .ttm_tt_populate = &ttm_page_alloc_ttm_tt_populate,
+ .ttm_tt_unpopulate = &ttm_page_alloc_ttm_tt_unpopulate,
.invalidate_caches = nouveau_bo_invalidate_caches,
.init_mem_type = nouveau_bo_init_mem_type,
.evict_flags = nouveau_bo_evict_flags,
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 53ff62b..490afce 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -584,6 +584,8 @@ struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
static struct ttm_bo_driver radeon_bo_driver = {
.ttm_tt_create = &radeon_ttm_tt_create,
+ .ttm_tt_populate = &ttm_page_alloc_ttm_tt_populate,
+ .ttm_tt_unpopulate = &ttm_page_alloc_ttm_tt_unpopulate,
.invalidate_caches = &radeon_invalidate_caches,
.init_mem_type = &radeon_init_mem_type,
.evict_flags = &radeon_evict_flags,
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index 082fcae..60f204d 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -244,7 +244,7 @@ static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
unsigned long page,
pgprot_t prot)
{
- struct page *d = ttm_tt_get_page(ttm, page);
+ struct page *d = ttm->pages[page];
void *dst;
if (!d)
@@ -281,7 +281,7 @@ static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
unsigned long page,
pgprot_t prot)
{
- struct page *s = ttm_tt_get_page(ttm, page);
+ struct page *s = ttm->pages[page];
void *src;
if (!s)
@@ -342,6 +342,12 @@ int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
if (old_iomap == NULL && ttm == NULL)
goto out2;
+ if (ttm->state == tt_unpopulated) {
+ ret = ttm->bdev->driver->ttm_tt_populate(ttm);
+ if (ret)
+ goto out1;
+ }
+
add = 0;
dir = 1;
@@ -502,10 +508,16 @@ static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
{
struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
struct ttm_tt *ttm = bo->ttm;
- struct page *d;
- int i;
+ int ret;
BUG_ON(!ttm);
+
+ if (ttm->state == tt_unpopulated) {
+ ret = ttm->bdev->driver->ttm_tt_populate(ttm);
+ if (ret)
+ return ret;
+ }
+
if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
/*
* We're mapping a single page, and the desired
@@ -513,18 +525,9 @@ static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
*/
map->bo_kmap_type = ttm_bo_map_kmap;
- map->page = ttm_tt_get_page(ttm, start_page);
+ map->page = ttm->pages[start_page];
map->virtual = kmap(map->page);
} else {
- /*
- * Populate the part we're mapping;
- */
- for (i = start_page; i < start_page + num_pages; ++i) {
- d = ttm_tt_get_page(ttm, i);
- if (!d)
- return -ENOMEM;
- }
-
/*
* We need to use vmap to get the desired page protection
* or to make the buffer object look contiguous.
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
index 221b924..bc1d751 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
@@ -174,18 +174,19 @@ static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
vma->vm_page_prot = (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
vm_get_page_prot(vma->vm_flags) :
ttm_io_prot(bo->mem.placement, vma->vm_page_prot);
- }
- /*
- * Speculatively prefault a number of pages. Only error on
- * first page.
- */
+ /* Allocate all page at once, most common usage */
+ if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
+ retval = VM_FAULT_OOM;
+ goto out_io_unlock;
+ }
+ }
for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
if (bo->mem.bus.is_iomem)
pfn = ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT) + page_offset;
else {
- page = ttm_tt_get_page(ttm, page_offset);
+ page = ttm->pages[page_offset];
if (unlikely(!page && i == 0)) {
retval = VM_FAULT_OOM;
goto out_io_unlock;
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index c4f18b9..ada2cad 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -852,6 +852,51 @@ void ttm_page_alloc_fini(void)
_manager = NULL;
}
+int ttm_page_alloc_ttm_tt_populate(struct ttm_tt *ttm)
+{
+ struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
+ int ret;
+
+ if (ttm->state != tt_unpopulated)
+ return 0;
+
+ ret = ttm_mem_global_alloc_pages(mem_glob, ttm->num_pages, false);
+ if (unlikely(ret != 0))
+ return -ENOMEM;
+
+ ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
+ ttm->caching_state, ttm->dma_address);
+ if (ret != 0) {
+ ttm_mem_global_free(mem_glob, ttm->num_pages * PAGE_SIZE);
+ return -ENOMEM;
+ }
+
+ ttm_mem_global_account_pages(mem_glob, ttm->pages, ttm->num_pages);
+
+ if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
+ ret = ttm_tt_swapin(ttm);
+ if (unlikely(ret != 0)) {
+ ttm_page_alloc_ttm_tt_unpopulate(ttm);
+ return ret;
+ }
+ }
+
+ ttm->state = tt_unbound;
+ return 0;
+}
+EXPORT_SYMBOL(ttm_page_alloc_ttm_tt_populate);
+
+void ttm_page_alloc_ttm_tt_unpopulate(struct ttm_tt *ttm)
+{
+ struct ttm_mem_global *glob = ttm->glob->mem_glob;
+
+ ttm_mem_global_free_pages(glob, ttm->pages, ttm->num_pages);
+ ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
+ ttm->caching_state, ttm->dma_address);
+ ttm->state = tt_unpopulated;
+}
+EXPORT_SYMBOL(ttm_page_alloc_ttm_tt_unpopulate);
+
int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
{
struct ttm_page_pool *p;
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 8c2d0df..d1be1c6 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -42,8 +42,6 @@
#include "ttm/ttm_placement.h"
#include "ttm/ttm_page_alloc.h"
-static int ttm_tt_swapin(struct ttm_tt *ttm);
-
/**
* Allocates storage for pointers to the pages that back the ttm.
*/
@@ -62,69 +60,6 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm)
ttm->dma_address = NULL;
}
-static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
-{
- struct page *p;
- struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
- int ret;
-
- if (NULL == (p = ttm->pages[index])) {
- ret = ttm_mem_global_alloc_pages(mem_glob, 1, false);
- if (unlikely(ret != 0))
- return NULL;
-
- ret = ttm_get_pages(&ttm->pages[index], 1, ttm->page_flags,
- ttm->caching_state,
- &ttm->dma_address[index]);
- if (ret != 0)
- return NULL;
-
- ttm_mem_global_account_pages(mem_glob, &ttm->pages[index], 1);
- }
- return p;
-}
-
-struct page *ttm_tt_get_page(struct ttm_tt *ttm, int index)
-{
- int ret;
-
- if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
- ret = ttm_tt_swapin(ttm);
- if (unlikely(ret != 0))
- return NULL;
- }
- return __ttm_tt_get_page(ttm, index);
-}
-
-int ttm_tt_populate(struct ttm_tt *ttm)
-{
- struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
- int ret;
-
- if (ttm->state != tt_unpopulated)
- return 0;
-
- if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
- ret = ttm_tt_swapin(ttm);
- if (unlikely(ret != 0))
- return ret;
- }
-
- ret = ttm_mem_global_alloc_pages(mem_glob, ttm->num_pages, false);
- if (unlikely(ret != 0))
- return -ENOMEM;
-
- ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
- ttm->caching_state, ttm->dma_address);
- if (ret != 0)
- return -ENOMEM;
-
- ttm_mem_global_account_pages(mem_glob, ttm->pages, ttm->num_pages);
- ttm->state = tt_unbound;
- return 0;
-}
-EXPORT_SYMBOL(ttm_tt_populate);
-
#ifdef CONFIG_X86
static inline int ttm_tt_set_page_caching(struct page *p,
enum ttm_caching_state c_old,
@@ -226,23 +161,13 @@ int ttm_tt_set_placement_caching(struct ttm_tt *ttm, uint32_t placement)
}
EXPORT_SYMBOL(ttm_tt_set_placement_caching);
-static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
-{
- struct ttm_mem_global *glob = ttm->glob->mem_glob;
-
- ttm_mem_global_free_pages(glob, ttm->pages, ttm->num_pages);
- ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
- ttm->caching_state, ttm->dma_address);
- ttm->state = tt_unpopulated;
-}
-
void ttm_tt_destroy(struct ttm_tt *ttm)
{
if (unlikely(ttm == NULL))
return;
if (likely(ttm->pages != NULL)) {
- ttm_tt_free_alloced_pages(ttm);
+ ttm->bdev->driver->ttm_tt_unpopulate(ttm);
ttm_tt_free_page_directory(ttm);
}
@@ -297,7 +222,7 @@ int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
if (ttm->state == tt_bound)
return 0;
- ret = ttm_tt_populate(ttm);
+ ret = ttm->bdev->driver->ttm_tt_populate(ttm);
if (ret)
return ret;
@@ -311,7 +236,7 @@ int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
}
EXPORT_SYMBOL(ttm_tt_bind);
-static int ttm_tt_swapin(struct ttm_tt *ttm)
+int ttm_tt_swapin(struct ttm_tt *ttm)
{
struct address_space *swap_space;
struct file *swap_storage;
@@ -333,7 +258,7 @@ static int ttm_tt_swapin(struct ttm_tt *ttm)
ret = PTR_ERR(from_page);
goto out_err;
}
- to_page = __ttm_tt_get_page(ttm, i);
+ to_page = ttm->pages[i];
if (unlikely(to_page == NULL))
goto out_err;
@@ -354,7 +279,6 @@ static int ttm_tt_swapin(struct ttm_tt *ttm)
return 0;
out_err:
- ttm_tt_free_alloced_pages(ttm);
return ret;
}
@@ -406,7 +330,7 @@ int ttm_tt_swapout(struct ttm_tt *ttm, struct file *persistent_swap_storage)
page_cache_release(to_page);
}
- ttm_tt_free_alloced_pages(ttm);
+ ttm->bdev->driver->ttm_tt_unpopulate(ttm);
ttm->swap_storage = swap_storage;
ttm->page_flags |= TTM_PAGE_FLAG_SWAPPED;
if (persistent_swap_storage)
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
index cc72435..383993b 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
@@ -28,6 +28,7 @@
#include "vmwgfx_drv.h"
#include "ttm/ttm_bo_driver.h"
#include "ttm/ttm_placement.h"
+#include "ttm/ttm_page_alloc.h"
static uint32_t vram_placement_flags = TTM_PL_FLAG_VRAM |
TTM_PL_FLAG_CACHED;
@@ -334,6 +335,8 @@ static int vmw_sync_obj_wait(void *sync_obj, void *sync_arg,
struct ttm_bo_driver vmw_bo_driver = {
.ttm_tt_create = &vmw_ttm_tt_create,
+ .ttm_tt_populate = &ttm_page_alloc_ttm_tt_populate,
+ .ttm_tt_unpopulate = &ttm_page_alloc_ttm_tt_unpopulate,
.invalidate_caches = vmw_invalidate_caches,
.init_mem_type = vmw_init_mem_type,
.evict_flags = vmw_evict_flags,
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 6b8c5cd..ae06e42 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -319,6 +319,26 @@ struct ttm_bo_driver {
struct page *dummy_read_page);
/**
+ * ttm_tt_populate
+ *
+ * @ttm: The struct ttm_tt to contain the backing pages.
+ *
+ * Allocate all backing pages
+ * Returns:
+ * -ENOMEM: Out of memory.
+ */
+ int (*ttm_tt_populate)(struct ttm_tt *ttm);
+
+ /**
+ * ttm_tt_unpopulate
+ *
+ * @ttm: The struct ttm_tt to contain the backing pages.
+ *
+ * Free all backing page
+ */
+ void (*ttm_tt_unpopulate)(struct ttm_tt *ttm);
+
+ /**
* struct ttm_bo_driver member invalidate_caches
*
* @bdev: the buffer object device.
@@ -585,15 +605,6 @@ extern int ttm_tt_init(struct ttm_tt *ttm, struct ttm_bo_device *bdev,
extern int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem);
/**
- * ttm_tt_populate:
- *
- * @ttm: The struct ttm_tt to contain the backing pages.
- *
- * Add backing pages to all of @ttm
- */
-extern int ttm_tt_populate(struct ttm_tt *ttm);
-
-/**
* ttm_ttm_destroy:
*
* @ttm: The struct ttm_tt.
@@ -612,19 +623,13 @@ extern void ttm_tt_destroy(struct ttm_tt *ttm);
extern void ttm_tt_unbind(struct ttm_tt *ttm);
/**
- * ttm_ttm_destroy:
+ * ttm_tt_swapin:
*
* @ttm: The struct ttm_tt.
- * @index: Index of the desired page.
*
- * Return a pointer to the struct page backing @ttm at page
- * index @index. If the page is unpopulated, one will be allocated to
- * populate that index.
- *
- * Returns:
- * NULL on OOM.
+ * Swap in a previously swap out ttm_tt.
*/
-extern struct page *ttm_tt_get_page(struct ttm_tt *ttm, int index);
+extern int ttm_tt_swapin(struct ttm_tt *ttm);
/**
* ttm_tt_cache_flush:
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index fffb3bd..18a2d6b 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -67,6 +67,24 @@ int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages);
void ttm_page_alloc_fini(void);
/**
+ * ttm_page_alloc_ttm_tt_populate:
+ *
+ * @ttm: The struct ttm_tt to contain the backing pages.
+ *
+ * Add backing pages to all of @ttm
+ */
+extern int ttm_page_alloc_ttm_tt_populate(struct ttm_tt *ttm);
+
+/**
+ * ttm_page_alloc_ttm_tt_unpopulate:
+ *
+ * @ttm: The struct ttm_tt which to free backing pages.
+ *
+ * Free all pages of @ttm
+ */
+extern void ttm_page_alloc_ttm_tt_unpopulate(struct ttm_tt *ttm);
+
+/**
* Output the state of pools to debugfs file
*/
extern int ttm_page_alloc_debugfs(struct seq_file *m, void *data);
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 10/13] ttm: Provide DMA aware TTM page pool code. V5
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (8 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 09/13] drm/ttm: introduce callback for ttm_tt populate & unpopulate V2 j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 11/13] swiotlb: Expose swiotlb_nr_tlb function to modules j.glisse
` (3 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom
From: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
In TTM world the pages for the graphic drivers are kept in three different
pools: write combined, uncached, and cached (write-back). When the pages
are used by the graphic driver the graphic adapter via its built in MMU
(or AGP) programs these pages in. The programming requires the virtual address
(from the graphic adapter perspective) and the physical address (either System RAM
or the memory on the card) which is obtained using the pci_map_* calls (which does the
virtual to physical - or bus address translation). During the graphic application's
"life" those pages can be shuffled around, swapped out to disk, moved from the
VRAM to System RAM or vice-versa. This all works with the existing TTM pool code
- except when we want to use the software IOTLB (SWIOTLB) code to "map" the physical
addresses to the graphic adapter MMU. We end up programming the bounce buffer's
physical address instead of the TTM pool memory's and get a non-worky driver.
There are two solutions:
1) using the DMA API to allocate pages that are screened by the DMA API, or
2) using the pci_sync_* calls to copy the pages from the bounce-buffer and back.
This patch fixes the issue by allocating pages using the DMA API. The second
is a viable option - but it has performance drawbacks and potential correctness
issues - think of the write cache page being bounced (SWIOTLB->TTM), the
WC is set on the TTM page and the copy from SWIOTLB not making it to the TTM
page until the page has been recycled in the pool (and used by another application).
The bounce buffer does not get activated often - only in cases where we have
a 32-bit capable card and we want to use a page that is allocated above the
4GB limit. The bounce buffer offers the solution of copying the contents
of that 4GB page to an location below 4GB and then back when the operation has been
completed (or vice-versa). This is done by using the 'pci_sync_*' calls.
Note: If you look carefully enough in the existing TTM page pool code you will
notice the GFP_DMA32 flag is used - which should guarantee that the provided page
is under 4GB. It certainly is the case, except this gets ignored in two cases:
- If user specifies 'swiotlb=force' which bounces _every_ page.
- If user is using a Xen's PV Linux guest (which uses the SWIOTLB and the
underlaying PFN's aren't necessarily under 4GB).
To not have this extra copying done the other option is to allocate the pages
using the DMA API so that there is not need to map the page and perform the
expensive 'pci_sync_*' calls.
This DMA API capable TTM pool requires for this the 'struct device' to
properly call the DMA API. It also has to track the virtual and bus address of
the page being handed out in case it ends up being swapped out or de-allocated -
to make sure it is de-allocated using the proper's 'struct device'.
Implementation wise the code keeps two lists: one that is attached to the
'struct device' (via the dev->dma_pools list) and a global one to be used when
the 'struct device' is unavailable (think shrinker code). The global list can
iterate over all of the 'struct device' and its associated dma_pool. The list
in dev->dma_pools can only iterate the device's dma_pool.
/[struct device_pool]\
/---------------------------------------------------| dev |
/ +-------| dma_pool |
/-----+------\ / \--------------------/
|struct device| /-->[struct dma_pool for WC]</ /[struct device_pool]\
| dma_pools +----+ /-| dev |
| ... | \--->[struct dma_pool for uncached]<-/--| dma_pool |
\-----+------/ / \--------------------/
\----------------------------------------------/
[Two pools associated with the device (WC and UC), and the parallel list
containing the 'struct dev' and 'struct dma_pool' entries]
The maximum amount of dma pools a device can have is six: write-combined,
uncached, and cached; then there are the DMA32 variants which are:
write-combined dma32, uncached dma32, and cached dma32.
Currently this code only gets activated when any variant of the SWIOTLB IOMMU
code is running (Intel without VT-d, AMD without GART, IBM Calgary and Xen PV
with PCI devices).
Tested-by: Michel Dänzer <michel@daenzer.net>
[v1: Using swiotlb_nr_tbl instead of swiotlb_enabled]
[v2: Major overhaul - added 'inuse_list' to seperate used from inuse and reorder
the order of lists to get better performance.]
[v3: Added comments/and some logic based on review, Added Jerome tag]
[v4: rebase on top of ttm_tt & ttm_backend merge]
[v5: rebase on top of ttm memory accounting overhaul]
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/Makefile | 4 +
drivers/gpu/drm/ttm/ttm_memory.c | 2 +
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 1235 ++++++++++++++++++++++++++++++
include/drm/ttm/ttm_bo_driver.h | 2 +
include/drm/ttm/ttm_page_alloc.h | 36 +
5 files changed, 1279 insertions(+), 0 deletions(-)
create mode 100644 drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
diff --git a/drivers/gpu/drm/ttm/Makefile b/drivers/gpu/drm/ttm/Makefile
index f3cf6f0..b2b33dd 100644
--- a/drivers/gpu/drm/ttm/Makefile
+++ b/drivers/gpu/drm/ttm/Makefile
@@ -7,4 +7,8 @@ ttm-y := ttm_agp_backend.o ttm_memory.o ttm_tt.o ttm_bo.o \
ttm_object.o ttm_lock.o ttm_execbuf_util.o ttm_page_alloc.o \
ttm_bo_manager.o
+ifeq ($(CONFIG_SWIOTLB),y)
+ttm-y += ttm_page_alloc_dma.o
+endif
+
obj-$(CONFIG_DRM_TTM) += ttm.o
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index 98f6899..3b12f05 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -258,6 +258,7 @@ int ttm_mem_global_init(struct ttm_mem_global *glob)
glob->swap_limit = glob->max_mem - (glob->mem >> 3);
glob->swap_dma32_limit = ((1ULL << 32) >> 1) - (((1ULL << 32) >> 3));
ttm_page_alloc_init(glob, glob->max_mem/(2*PAGE_SIZE));
+ ttm_dma_page_alloc_init(glob, glob->max_mem/(2*PAGE_SIZE));
return 0;
}
EXPORT_SYMBOL(ttm_mem_global_init);
@@ -266,6 +267,7 @@ void ttm_mem_global_release(struct ttm_mem_global *glob)
{
/* let the page allocator first stop the shrink work. */
ttm_page_alloc_fini();
+ ttm_dma_page_alloc_fini();
flush_workqueue(glob->swap_queue);
destroy_workqueue(glob->swap_queue);
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
new file mode 100644
index 0000000..0850836
--- /dev/null
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
@@ -0,0 +1,1235 @@
+/*
+ * Copyright 2011 (c) Oracle Corp.
+
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sub license,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ *
+ * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
+ */
+
+/*
+ * A simple DMA pool losely based on dmapool.c. It has certain advantages
+ * over the DMA pools:
+ * - Pool collects resently freed pages for reuse (and hooks up to
+ * the shrinker).
+ * - Tracks currently in use pages
+ * - Tracks whether the page is UC, WB or cached (and reverts to WB
+ * when freed).
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/list.h>
+#include <linux/seq_file.h> /* for seq_printf */
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/highmem.h>
+#include <linux/mm_types.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+#include <linux/atomic.h>
+#include <linux/device.h>
+#include <linux/kthread.h>
+#include "ttm/ttm_bo_driver.h"
+#include "ttm/ttm_page_alloc.h"
+#ifdef TTM_HAS_AGP
+#include <asm/agp.h>
+#endif
+
+#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
+#define SMALL_ALLOCATION 16
+#define FREE_ALL_PAGES (~0U)
+/* times are in msecs */
+#define IS_UNDEFINED (0)
+#define IS_WC (1<<1)
+#define IS_UC (1<<2)
+#define IS_CACHED (1<<3)
+#define IS_DMA32 (1<<4)
+
+enum pool_type {
+ POOL_IS_UNDEFINED,
+ POOL_IS_WC = IS_WC,
+ POOL_IS_UC = IS_UC,
+ POOL_IS_CACHED = IS_CACHED,
+ POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
+ POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
+ POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
+};
+/*
+ * The pool structure. There are usually six pools:
+ * - generic (not restricted to DMA32):
+ * - write combined, uncached, cached.
+ * - dma32 (up to 2^32 - so up 4GB):
+ * - write combined, uncached, cached.
+ * for each 'struct device'. The 'cached' is for pages that are actively used.
+ * The other ones can be shrunk by the shrinker API if neccessary.
+ * @pools: The 'struct device->dma_pools' link.
+ * @type: Type of the pool
+ * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
+ * used with irqsave/irqrestore variants because pool allocator maybe called
+ * from delayed work.
+ * @inuse_list: Pool of pages that are in use. The order is very important and
+ * it is in the order that the TTM pages that are put back are in.
+ * @free_list: Pool of pages that are free to be used. No order requirements.
+ * @dev: The device that is associated with these pools.
+ * @size: Size used during DMA allocation.
+ * @npages_free: Count of available pages for re-use.
+ * @npages_in_use: Count of pages that are in use.
+ * @nfrees: Stats when pool is shrinking.
+ * @nrefills: Stats when the pool is grown.
+ * @gfp_flags: Flags to pass for alloc_page.
+ * @fill_lock: Allows only one pool fill operation at time.
+ * @name: Name of the pool.
+ * @dev_name: Name derieved from dev - similar to how dev_info works.
+ * Used during shutdown as the dev_info during release is unavailable.
+ */
+struct dma_pool {
+ struct list_head pools; /* The 'struct device->dma_pools link */
+ enum pool_type type;
+ spinlock_t lock;
+ struct list_head inuse_list;
+ struct list_head free_list;
+ struct device *dev;
+ unsigned size;
+ unsigned npages_free;
+ unsigned npages_in_use;
+ unsigned long nfrees; /* Stats when shrunk. */
+ unsigned long nrefills; /* Stats when grown. */
+ gfp_t gfp_flags;
+ bool fill_lock;
+ char name[13]; /* "cached dma32" */
+ char dev_name[64]; /* Constructed from dev */
+};
+
+/*
+ * The accounting page keeping track of the allocated page along with
+ * the DMA address.
+ * @page_list: The link to the 'page_list' in 'struct dma_pool'.
+ * @vaddr: The virtual address of the page
+ * @dma: The bus address of the page. If the page is not allocated
+ * via the DMA API, it will be -1.
+ */
+struct dma_page {
+ struct list_head page_list;
+ void *vaddr;
+ struct page *p;
+ dma_addr_t dma;
+};
+
+/*
+ * Limits for the pool. They are handled without locks because only place where
+ * they may change is in sysfs store. They won't have immediate effect anyway
+ * so forcing serialization to access them is pointless.
+ */
+
+struct ttm_pool_opts {
+ unsigned alloc_size;
+ unsigned max_size;
+ unsigned small;
+};
+
+/*
+ * Contains the list of all of the 'struct device' and their corresponding
+ * DMA pools. Guarded by _mutex->lock.
+ * @pools: The link to 'struct ttm_pool_manager->pools'
+ * @dev: The 'struct device' associated with the 'pool'
+ * @pool: The 'struct dma_pool' associated with the 'dev'
+ */
+struct device_pools {
+ struct list_head pools;
+ struct device *dev;
+ struct dma_pool *pool;
+};
+
+/*
+ * struct ttm_pool_manager - Holds memory pools for fast allocation
+ *
+ * @lock: Lock used when adding/removing from pools
+ * @pools: List of 'struct device' and 'struct dma_pool' tuples.
+ * @options: Limits for the pool.
+ * @npools: Total amount of pools in existence.
+ * @shrinker: The structure used by [un|]register_shrinker
+ */
+struct ttm_pool_manager {
+ struct mutex lock;
+ struct list_head pools;
+ struct ttm_pool_opts options;
+ unsigned npools;
+ struct shrinker mm_shrink;
+ struct kobject kobj;
+};
+
+static struct ttm_pool_manager *_manager;
+
+static struct attribute ttm_page_pool_max = {
+ .name = "pool_max_size",
+ .mode = S_IRUGO | S_IWUSR
+};
+static struct attribute ttm_page_pool_small = {
+ .name = "pool_small_allocation",
+ .mode = S_IRUGO | S_IWUSR
+};
+static struct attribute ttm_page_pool_alloc_size = {
+ .name = "pool_allocation_size",
+ .mode = S_IRUGO | S_IWUSR
+};
+
+static struct attribute *ttm_pool_attrs[] = {
+ &ttm_page_pool_max,
+ &ttm_page_pool_small,
+ &ttm_page_pool_alloc_size,
+ NULL
+};
+
+static void ttm_pool_kobj_release(struct kobject *kobj)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ kfree(m);
+}
+
+static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
+ const char *buffer, size_t size)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ int chars;
+ unsigned val;
+ chars = sscanf(buffer, "%u", &val);
+ if (chars == 0)
+ return size;
+
+ /* Convert kb to number of pages */
+ val = val / (PAGE_SIZE >> 10);
+
+ if (attr == &ttm_page_pool_max)
+ m->options.max_size = val;
+ else if (attr == &ttm_page_pool_small)
+ m->options.small = val;
+ else if (attr == &ttm_page_pool_alloc_size) {
+ if (val > NUM_PAGES_TO_ALLOC*8) {
+ printk(KERN_ERR TTM_PFX
+ "Setting allocation size to %lu "
+ "is not allowed. Recommended size is "
+ "%lu\n",
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
+ return size;
+ } else if (val > NUM_PAGES_TO_ALLOC) {
+ printk(KERN_WARNING TTM_PFX
+ "Setting allocation size to "
+ "larger than %lu is not recommended.\n",
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
+ }
+ m->options.alloc_size = val;
+ }
+
+ return size;
+}
+
+static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
+ char *buffer)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ unsigned val = 0;
+
+ if (attr == &ttm_page_pool_max)
+ val = m->options.max_size;
+ else if (attr == &ttm_page_pool_small)
+ val = m->options.small;
+ else if (attr == &ttm_page_pool_alloc_size)
+ val = m->options.alloc_size;
+
+ val = val * (PAGE_SIZE >> 10);
+
+ return snprintf(buffer, PAGE_SIZE, "%u\n", val);
+}
+
+static const struct sysfs_ops ttm_pool_sysfs_ops = {
+ .show = &ttm_pool_show,
+ .store = &ttm_pool_store,
+};
+
+static struct kobj_type ttm_pool_kobj_type = {
+ .release = &ttm_pool_kobj_release,
+ .sysfs_ops = &ttm_pool_sysfs_ops,
+ .default_attrs = ttm_pool_attrs,
+};
+
+#ifndef CONFIG_X86
+static int set_pages_array_wb(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ unmap_page_from_agp(pages[i]);
+#endif
+ return 0;
+}
+
+static int set_pages_array_wc(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ map_page_into_agp(pages[i]);
+#endif
+ return 0;
+}
+
+static int set_pages_array_uc(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ map_page_into_agp(pages[i]);
+#endif
+ return 0;
+}
+#endif /* for !CONFIG_X86 */
+
+static int ttm_set_pages_caching(struct dma_pool *pool,
+ struct page **pages, unsigned cpages)
+{
+ int r = 0;
+ /* Set page caching */
+ if (pool->type & IS_UC) {
+ r = set_pages_array_uc(pages, cpages);
+ if (r)
+ pr_err(TTM_PFX
+ "%s: Failed to set %d pages to uc!\n",
+ pool->dev_name, cpages);
+ }
+ if (pool->type & IS_WC) {
+ r = set_pages_array_wc(pages, cpages);
+ if (r)
+ pr_err(TTM_PFX
+ "%s: Failed to set %d pages to wc!\n",
+ pool->dev_name, cpages);
+ }
+ return r;
+}
+
+static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
+{
+ dma_addr_t dma = d_page->dma;
+ dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
+
+ kfree(d_page);
+ d_page = NULL;
+}
+static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
+{
+ struct dma_page *d_page;
+
+ d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
+ if (!d_page)
+ return NULL;
+
+ d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
+ &d_page->dma,
+ pool->gfp_flags);
+ if (d_page->vaddr)
+ d_page->p = virt_to_page(d_page->vaddr);
+ else {
+ kfree(d_page);
+ d_page = NULL;
+ }
+ return d_page;
+}
+static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
+{
+ enum pool_type type = IS_UNDEFINED;
+
+ if (flags & TTM_PAGE_FLAG_DMA32)
+ type |= IS_DMA32;
+ if (cstate == tt_cached)
+ type |= IS_CACHED;
+ else if (cstate == tt_uncached)
+ type |= IS_UC;
+ else
+ type |= IS_WC;
+
+ return type;
+}
+
+static void ttm_pool_update_free_locked(struct dma_pool *pool,
+ unsigned freed_pages)
+{
+ pool->npages_free -= freed_pages;
+ pool->nfrees += freed_pages;
+
+}
+
+/* set memory back to wb and free the pages. */
+static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
+ struct page *pages[], unsigned npages)
+{
+ struct dma_page *d_page, *tmp;
+
+ if (npages && set_pages_array_wb(pages, npages))
+ pr_err(TTM_PFX "%s: Failed to set %d pages to wb!\n",
+ pool->dev_name, npages);
+
+ if (npages > 1) {
+ pr_debug("%s: (%s:%d) Freeing %d pages at once (lockless).\n",
+ pool->dev_name, pool->name, current->pid, npages);
+ }
+
+ list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
+ list_del(&d_page->page_list);
+ __ttm_dma_free_page(pool, d_page);
+ }
+}
+
+/*
+ * Free pages from pool.
+ *
+ * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
+ * number of pages in one go.
+ *
+ * @pool: to free the pages from
+ * @nr_free: If set to true will free all pages in pool
+ **/
+static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
+{
+ unsigned long irq_flags;
+ struct dma_page *dma_p, *tmp;
+ struct page **pages_to_free;
+ struct list_head d_pages;
+ unsigned freed_pages = 0,
+ npages_to_free = nr_free;
+
+ if (NUM_PAGES_TO_ALLOC < nr_free)
+ npages_to_free = NUM_PAGES_TO_ALLOC;
+#if 0
+ if (nr_free > 1) {
+ pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
+ pool->dev_name, pool->name, current->pid,
+ npages_to_free, nr_free);
+ }
+#endif
+ pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
+ GFP_KERNEL);
+
+ if (!pages_to_free) {
+ pr_err(TTM_PFX
+ "%s: Failed to allocate memory for pool free operation.\n",
+ pool->dev_name);
+ return 0;
+ }
+ INIT_LIST_HEAD(&d_pages);
+restart:
+ spin_lock_irqsave(&pool->lock, irq_flags);
+
+ /* We picking the oldest ones off the list */
+ list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
+ page_list) {
+ if (freed_pages >= npages_to_free)
+ break;
+
+ /* Move the dma_page from one list to another. */
+ list_move(&dma_p->page_list, &d_pages);
+
+ pages_to_free[freed_pages++] = dma_p->p;
+ /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
+ if (freed_pages >= NUM_PAGES_TO_ALLOC) {
+
+ ttm_pool_update_free_locked(pool, freed_pages);
+ /**
+ * Because changing page caching is costly
+ * we unlock the pool to prevent stalling.
+ */
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ ttm_dma_pages_put(pool, &d_pages, pages_to_free,
+ freed_pages);
+
+ INIT_LIST_HEAD(&d_pages);
+
+ if (likely(nr_free != FREE_ALL_PAGES))
+ nr_free -= freed_pages;
+
+ if (NUM_PAGES_TO_ALLOC >= nr_free)
+ npages_to_free = nr_free;
+ else
+ npages_to_free = NUM_PAGES_TO_ALLOC;
+
+ freed_pages = 0;
+
+ /* free all so restart the processing */
+ if (nr_free)
+ goto restart;
+
+ /* Not allowed to fall through or break because
+ * following context is inside spinlock while we are
+ * outside here.
+ */
+ goto out;
+
+ }
+ }
+
+ /* remove range of pages from the pool */
+ if (freed_pages) {
+ ttm_pool_update_free_locked(pool, freed_pages);
+ nr_free -= freed_pages;
+ }
+
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ if (freed_pages)
+ ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
+out:
+ kfree(pages_to_free);
+ return nr_free;
+}
+
+static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
+{
+ struct device_pools *p;
+ struct dma_pool *pool;
+
+ if (!dev)
+ return;
+
+ mutex_lock(&_manager->lock);
+ list_for_each_entry_reverse(p, &_manager->pools, pools) {
+ if (p->dev != dev)
+ continue;
+ pool = p->pool;
+ if (pool->type != type)
+ continue;
+
+ list_del(&p->pools);
+ kfree(p);
+ _manager->npools--;
+ break;
+ }
+ list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
+ if (pool->type != type)
+ continue;
+ /* Takes a spinlock.. */
+ ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
+ WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
+ /* This code path is called after _all_ references to the
+ * struct device has been dropped - so nobody should be
+ * touching it. In case somebody is trying to _add_ we are
+ * guarded by the mutex. */
+ list_del(&pool->pools);
+ kfree(pool);
+ break;
+ }
+ mutex_unlock(&_manager->lock);
+}
+
+/*
+ * On free-ing of the 'struct device' this deconstructor is run.
+ * Albeit the pool might have already been freed earlier.
+ */
+static void ttm_dma_pool_release(struct device *dev, void *res)
+{
+ struct dma_pool *pool = *(struct dma_pool **)res;
+
+ if (pool)
+ ttm_dma_free_pool(dev, pool->type);
+}
+
+static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
+{
+ return *(struct dma_pool **)res == match_data;
+}
+
+static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
+ enum pool_type type)
+{
+ char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
+ enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
+ struct device_pools *sec_pool = NULL;
+ struct dma_pool *pool = NULL, **ptr;
+ unsigned i;
+ int ret = -ENODEV;
+ char *p;
+
+ if (!dev)
+ return NULL;
+
+ ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
+ if (!ptr)
+ return NULL;
+
+ ret = -ENOMEM;
+
+ pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
+ dev_to_node(dev));
+ if (!pool)
+ goto err_mem;
+
+ sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
+ dev_to_node(dev));
+ if (!sec_pool)
+ goto err_mem;
+
+ INIT_LIST_HEAD(&sec_pool->pools);
+ sec_pool->dev = dev;
+ sec_pool->pool = pool;
+
+ INIT_LIST_HEAD(&pool->free_list);
+ INIT_LIST_HEAD(&pool->inuse_list);
+ INIT_LIST_HEAD(&pool->pools);
+ spin_lock_init(&pool->lock);
+ pool->dev = dev;
+ pool->npages_free = pool->npages_in_use = 0;
+ pool->nfrees = 0;
+ pool->gfp_flags = flags;
+ pool->size = PAGE_SIZE;
+ pool->type = type;
+ pool->nrefills = 0;
+ pool->fill_lock = false;
+ p = pool->name;
+ for (i = 0; i < 5; i++) {
+ if (type & t[i]) {
+ p += snprintf(p, sizeof(pool->name) - (p - pool->name),
+ "%s", n[i]);
+ }
+ }
+ *p = 0;
+ /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
+ * - the kobj->name has already been deallocated.*/
+ snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
+ dev_driver_string(dev), dev_name(dev));
+ mutex_lock(&_manager->lock);
+ /* You can get the dma_pool from either the global: */
+ list_add(&sec_pool->pools, &_manager->pools);
+ _manager->npools++;
+ /* or from 'struct device': */
+ list_add(&pool->pools, &dev->dma_pools);
+ mutex_unlock(&_manager->lock);
+
+ *ptr = pool;
+ devres_add(dev, ptr);
+
+ return pool;
+err_mem:
+ devres_free(ptr);
+ kfree(sec_pool);
+ kfree(pool);
+ return ERR_PTR(ret);
+}
+
+static struct dma_pool *ttm_dma_find_pool(struct device *dev,
+ enum pool_type type)
+{
+ struct dma_pool *pool, *tmp, *found = NULL;
+
+ if (type == IS_UNDEFINED)
+ return found;
+
+ /* NB: We iterate on the 'struct dev' which has no spinlock, but
+ * it does have a kref which we have taken. The kref is taken during
+ * graphic driver loading - in the drm_pci_init it calls either
+ * pci_dev_get or pci_register_driver which both end up taking a kref
+ * on 'struct device'.
+ *
+ * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
+ * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
+ * thing is at that point of time there are no pages associated with the
+ * driver so this function will not be called.
+ */
+ list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
+ if (pool->type != type)
+ continue;
+ found = pool;
+ break;
+ }
+ return found;
+}
+
+/*
+ * Free pages the pages that failed to change the caching state. If there
+ * are pages that have changed their caching state already put them to the
+ * pool.
+ */
+static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
+ struct list_head *d_pages,
+ struct page **failed_pages,
+ unsigned cpages)
+{
+ struct dma_page *d_page, *tmp;
+ struct page *p;
+ unsigned i = 0;
+
+ p = failed_pages[0];
+ if (!p)
+ return;
+ /* Find the failed page. */
+ list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
+ if (d_page->p != p)
+ continue;
+ /* .. and then progress over the full list. */
+ list_del(&d_page->page_list);
+ __ttm_dma_free_page(pool, d_page);
+ if (++i < cpages)
+ p = failed_pages[i];
+ else
+ break;
+ }
+
+}
+
+/*
+ * Allocate 'count' pages, and put 'need' number of them on the
+ * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
+ * The full list of pages should also be on 'd_pages'.
+ * We return zero for success, and negative numbers as errors.
+ */
+static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
+ struct list_head *d_pages,
+ unsigned count)
+{
+ struct page **caching_array;
+ struct dma_page *dma_p;
+ struct page *p;
+ int r = 0;
+ unsigned i, cpages;
+ unsigned max_cpages = min(count,
+ (unsigned)(PAGE_SIZE/sizeof(struct page *)));
+
+ /* allocate array for page caching change */
+ caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
+
+ if (!caching_array) {
+ pr_err(TTM_PFX
+ "%s: Unable to allocate table for new pages.",
+ pool->dev_name);
+ return -ENOMEM;
+ }
+
+ if (count > 1) {
+ pr_debug("%s: (%s:%d) Getting %d pages\n",
+ pool->dev_name, pool->name, current->pid,
+ count);
+ }
+
+ for (i = 0, cpages = 0; i < count; ++i) {
+ dma_p = __ttm_dma_alloc_page(pool);
+ if (!dma_p) {
+ pr_err(TTM_PFX "%s: Unable to get page %u.\n",
+ pool->dev_name, i);
+
+ /* store already allocated pages in the pool after
+ * setting the caching state */
+ if (cpages) {
+ r = ttm_set_pages_caching(pool, caching_array,
+ cpages);
+ if (r)
+ ttm_dma_handle_caching_state_failure(
+ pool, d_pages, caching_array,
+ cpages);
+ }
+ r = -ENOMEM;
+ goto out;
+ }
+ p = dma_p->p;
+#ifdef CONFIG_HIGHMEM
+ /* gfp flags of highmem page should never be dma32 so we
+ * we should be fine in such case
+ */
+ if (!PageHighMem(p))
+#endif
+ {
+ caching_array[cpages++] = p;
+ if (cpages == max_cpages) {
+ /* Note: Cannot hold the spinlock */
+ r = ttm_set_pages_caching(pool, caching_array,
+ cpages);
+ if (r) {
+ ttm_dma_handle_caching_state_failure(
+ pool, d_pages, caching_array,
+ cpages);
+ goto out;
+ }
+ cpages = 0;
+ }
+ }
+ list_add(&dma_p->page_list, d_pages);
+ }
+
+ if (cpages) {
+ r = ttm_set_pages_caching(pool, caching_array, cpages);
+ if (r)
+ ttm_dma_handle_caching_state_failure(pool, d_pages,
+ caching_array, cpages);
+ }
+out:
+ kfree(caching_array);
+ return r;
+}
+
+static int ttm_dma_check_alloc_list(struct ttm_tt *ttm)
+{
+ struct dma_page *d_page;
+ unsigned i = 0;
+
+ list_for_each_entry(d_page, &ttm->alloc_list, page_list) {
+ if (d_page->p != ttm->pages[i++]) {
+ pr_err(TTM_PFX "dma page list & array mismatch %d", i);
+ return -EINVAL;
+ }
+ }
+ return 0;
+}
+
+/*
+ * @return count of pages still required to fulfill the request.
+*/
+static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
+ unsigned count,
+ unsigned long *irq_flags)
+{
+ int r = count;
+
+ if (pool->fill_lock)
+ return r;
+
+ pool->fill_lock = true;
+ if (count < _manager->options.small &&
+ count > pool->npages_free) {
+ struct list_head d_pages;
+ unsigned alloc_size = _manager->options.alloc_size;
+
+ INIT_LIST_HEAD(&d_pages);
+
+ spin_unlock_irqrestore(&pool->lock, *irq_flags);
+
+ /* Returns how many more are neccessary to fulfill the
+ * request. */
+ r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, alloc_size);
+
+ spin_lock_irqsave(&pool->lock, *irq_flags);
+ if (!r) {
+ /* Add the fresh to the end.. */
+ list_splice(&d_pages, &pool->free_list);
+ ++pool->nrefills;
+ pool->npages_free += alloc_size;
+ } else {
+ struct dma_page *d_page;
+ unsigned cpages = 0;
+
+ pr_err(TTM_PFX "%s: Failed to fill %s pool (r:%d)!\n",
+ pool->dev_name, pool->name, r);
+
+ list_for_each_entry(d_page, &d_pages, page_list) {
+ cpages++;
+ }
+ list_splice_tail(&d_pages, &pool->free_list);
+ pool->npages_free += cpages;
+ }
+ }
+ pool->fill_lock = false;
+ return r;
+
+}
+
+/*
+ * @return count of pages still required to fulfill the request.
+ * The populate list is actually a stack (not that is matters as TTM
+ * allocates one page at a time.
+ */
+static int ttm_dma_pool_get_pages(struct dma_pool *pool,
+ struct ttm_tt *ttm)
+{
+ struct dma_page *d_page, *tmp;
+ unsigned long irq_flags;
+ unsigned i, count = ttm->num_pages;
+ int r;
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ r = ttm_dma_page_pool_fill_locked(pool, count, &irq_flags);
+ if (r < 0) {
+ pr_debug("%s: (%s:%d) Asked for %d, got %d %s.\n",
+ pool->dev_name, pool->name, current->pid, count, r,
+ (r < 0) ? "err:" : "pages");
+ goto out;
+ }
+ if (!pool->npages_free)
+ goto out;
+ if (count > 1) {
+ pr_debug("%s: (%s:%d) Looking in free list for %d pages. "\
+ "(have %d pages free)\n",
+ pool->dev_name, pool->name, current->pid, count,
+ pool->npages_free);
+ }
+ i = 0;
+ /* We are holding the spinlock.. */
+ /* Note: The the 'pages' (and inuse_list) is expected to be a stack,
+ * so we put the entries in the right order (and on the inuse list
+ * in the reverse order to compenstate for freeing - which inverts the
+ * 'pages' order).
+ */
+ list_for_each_entry_safe(d_page, tmp, &pool->free_list, page_list) {
+ ttm->pages[i] = d_page->p;
+ ttm->dma_address[i] = d_page->dma;
+ i++;
+ list_move_tail(&d_page->page_list, &ttm->alloc_list);
+ if (i == count)
+ break;
+ }
+ count -= i;
+ pool->npages_in_use += i;
+ pool->npages_free -= i;
+out:
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+ if (count)
+ pr_debug("%s: (%s:%d) Need %d more.\n",
+ pool->dev_name, pool->name, current->pid, count);
+ return count;
+}
+
+/*
+ * On success pages list will hold count number of correctly
+ * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
+ */
+int ttm_dma_populate(struct ttm_tt *ttm, struct device *dev)
+{
+ struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
+ struct dma_pool *pool;
+ gfp_t gfp_flags;
+ enum pool_type type;
+ unsigned i;
+ int r;
+
+ r = ttm_mem_global_alloc_pages(mem_glob, ttm->num_pages, false);
+ if (r)
+ return -ENOMEM;
+
+ type = ttm_to_type(ttm->page_flags, ttm->caching_state);
+
+ if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
+ gfp_flags = GFP_USER | GFP_DMA32;
+ else
+ gfp_flags = GFP_HIGHUSER;
+
+ if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
+ gfp_flags |= __GFP_ZERO;
+
+ pool = ttm_dma_find_pool(dev, type);
+ if (!pool) {
+ pool = ttm_dma_pool_init(dev, gfp_flags, type);
+ if (IS_ERR_OR_NULL(pool)) {
+ ttm_mem_global_free(mem_glob, ttm->num_pages * PAGE_SIZE);
+ return -ENOMEM;
+ }
+ }
+
+ /* ?? we might want to check see if pages are already
+ * alloced, for the moment just trust the caller
+ */
+ INIT_LIST_HEAD(&ttm->alloc_list);
+ /* Take pages out of a pool (if applicable) */
+ r = ttm_dma_pool_get_pages(pool, ttm);
+ /* clear the pages coming from the pool if requested */
+ if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
+ for (i = 0; i < (ttm->num_pages - r); i++) {
+ clear_page(page_address(ttm->pages[i]));
+ }
+ }
+ /* If pool didn't have enough pages allocate new one. */
+ if (r > 0) {
+ struct list_head d_pages;
+ unsigned long irq_flags;
+ unsigned pages_need = r;
+
+ INIT_LIST_HEAD(&d_pages);
+ i = ttm->num_pages - pages_need;
+
+ /* Note, we are running without locking here..
+ * and we have to manually add the stack to the inuse pool. */
+ r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, pages_need);
+
+ if (r == 0) {
+ struct dma_page *d_page;
+
+ /* Since the pages are directly going to the inuse_list
+ * which is stack based, lets treat it as a stack.
+ */
+ list_for_each_entry(d_page, &d_pages, page_list) {
+ ttm->pages[i] = d_page->p;
+ ttm->dma_address[i] = d_page->dma;
+ i++;
+ }
+ list_splice_tail(&d_pages, &ttm->alloc_list);
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ pool->npages_in_use += pages_need;
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+ } else {
+ struct dma_page *d_page;
+ unsigned cpages = 0;
+
+ /* If there is any pages in the list put them back to
+ * the pool. */
+ pr_err(TTM_PFX
+ "%s: Failed to allocate extra pages "
+ "for large request.",
+ pool->dev_name);
+
+ list_for_each_entry(d_page, &d_pages, page_list) {
+ cpages++;
+ }
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ pool->npages_free += cpages;
+ /* We don't care about ordering on the free_list. */
+ list_splice(&d_pages, &pool->free_list);
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+ ttm_mem_global_free(mem_glob, ttm->num_pages * PAGE_SIZE);
+ return -ENOMEM;
+ }
+ }
+ ttm_mem_global_account_pages(mem_glob, ttm->pages, ttm->num_pages);
+
+ if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
+ r = ttm_tt_swapin(ttm);
+ if (unlikely(r!= 0)) {
+ ttm_dma_unpopulate(ttm, dev);
+ return r;
+ }
+ }
+
+ ttm->state = tt_unbound;
+ return 0;
+}
+EXPORT_SYMBOL_GPL(ttm_dma_populate);
+
+/* Get good estimation how many pages are free in pools */
+static int ttm_dma_pool_get_num_unused_pages(void)
+{
+ struct device_pools *p;
+ unsigned total = 0;
+
+ mutex_lock(&_manager->lock);
+ list_for_each_entry(p, &_manager->pools, pools) {
+ if (p)
+ total += p->pool->npages_free;
+ }
+ mutex_unlock(&_manager->lock);
+ return total;
+}
+
+/* Put all pages in pages list to correct pool to wait for reuse */
+void ttm_dma_unpopulate(struct ttm_tt *ttm, struct device *dev)
+{
+ struct ttm_mem_global *glob = ttm->glob->mem_glob;
+ struct dma_pool *pool;
+ enum pool_type type;
+ bool is_cached = false;
+ unsigned count = 0, i;
+ unsigned long irq_flags;
+
+ ttm_mem_global_free_pages(glob, ttm->pages, ttm->num_pages);
+
+ type = ttm_to_type(ttm->page_flags, ttm->caching_state);
+ pool = ttm_dma_find_pool(dev, type);
+ if (!pool) {
+ WARN_ON(!pool);
+ return;
+ }
+ is_cached = (ttm_dma_find_pool(pool->dev,
+ ttm_to_type(ttm->page_flags, tt_cached)) == pool);
+
+ if (ttm->num_pages > 1) {
+ dev_dbg(pool->dev, "(%s:%d) Attempting to %s %ld pages.\n",
+ pool->name, current->pid,
+ (is_cached) ? "destroy" : "recycle", ttm->num_pages);
+ }
+
+ /* this should never happen, should we panic here ?*/
+ if (ttm_dma_check_alloc_list(ttm)) {
+ pr_err(TTM_PFX "Could not free the other ttm tt pages %ld!\n",
+ ttm->num_pages);
+ return;
+ }
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ count = 0;
+ pool->npages_in_use -= ttm->num_pages;
+ if (is_cached) {
+ pool->nfrees += ttm->num_pages;
+ } else {
+ pool->npages_free += ttm->num_pages;
+ list_splice(&ttm->alloc_list, &pool->free_list);
+ if (pool->npages_free > _manager->options.max_size) {
+ count = pool->npages_free - _manager->options.max_size;
+ }
+ }
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ if (is_cached) {
+ ttm_dma_pages_put(pool, &ttm->alloc_list,
+ ttm->pages, ttm->num_pages);
+ }
+
+ INIT_LIST_HEAD(&ttm->alloc_list);
+ for (i = 0; i < ttm->num_pages; i++) {
+ ttm->pages[i] = NULL;
+ ttm->dma_address[i] = 0;
+ }
+
+ /* shrink pool if necessary */
+ if (count)
+ ttm_dma_page_pool_free(pool, count);
+ ttm->state = tt_unpopulated;
+}
+EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
+
+/**
+ * Callback for mm to request pool to reduce number of page held.
+ */
+static int ttm_dma_pool_mm_shrink(struct shrinker *shrink,
+ struct shrink_control *sc)
+{
+ static atomic_t start_pool = ATOMIC_INIT(0);
+ unsigned idx = 0;
+ unsigned pool_offset = atomic_add_return(1, &start_pool);
+ unsigned shrink_pages = sc->nr_to_scan;
+ struct device_pools *p;
+
+ if (list_empty(&_manager->pools))
+ return 0;
+
+ mutex_lock(&_manager->lock);
+ pool_offset = pool_offset % _manager->npools;
+ list_for_each_entry(p, &_manager->pools, pools) {
+ unsigned nr_free;
+
+ if (!p && !p->dev)
+ continue;
+ if (shrink_pages == 0)
+ break;
+ /* Do it in round-robin fashion. */
+ if (++idx < pool_offset)
+ continue;
+ nr_free = shrink_pages;
+ shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free);
+ pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
+ p->pool->dev_name, p->pool->name, current->pid, nr_free,
+ shrink_pages);
+ }
+ mutex_unlock(&_manager->lock);
+ /* return estimated number of unused pages in pool */
+ return ttm_dma_pool_get_num_unused_pages();
+}
+
+static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
+{
+ manager->mm_shrink.shrink = &ttm_dma_pool_mm_shrink;
+ manager->mm_shrink.seeks = 1;
+ register_shrinker(&manager->mm_shrink);
+}
+
+static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
+{
+ unregister_shrinker(&manager->mm_shrink);
+}
+
+int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
+{
+ int ret = -ENOMEM;
+
+ WARN_ON(_manager);
+
+ printk(KERN_INFO TTM_PFX "Initializing DMA pool allocator.\n");
+
+ _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
+ if (!_manager)
+ goto err_manager;
+
+ mutex_init(&_manager->lock);
+ INIT_LIST_HEAD(&_manager->pools);
+
+ _manager->options.max_size = max_pages;
+ _manager->options.small = SMALL_ALLOCATION;
+ _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
+
+ /* This takes care of auto-freeing the _manager */
+ ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
+ &glob->kobj, "dma_pool");
+ if (unlikely(ret != 0)) {
+ kobject_put(&_manager->kobj);
+ goto err;
+ }
+ ttm_dma_pool_mm_shrink_init(_manager);
+ return 0;
+err_manager:
+ kfree(_manager);
+ _manager = NULL;
+err:
+ return ret;
+}
+
+void ttm_dma_page_alloc_fini(void)
+{
+ struct device_pools *p, *t;
+
+ printk(KERN_INFO TTM_PFX "Finalizing DMA pool allocator.\n");
+ ttm_dma_pool_mm_shrink_fini(_manager);
+
+ list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
+ dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
+ current->pid);
+ WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
+ ttm_dma_pool_match, p->pool));
+ ttm_dma_free_pool(p->dev, p->pool->type);
+ }
+ kobject_put(&_manager->kobj);
+ _manager = NULL;
+}
+
+int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
+{
+ struct device_pools *p;
+ struct dma_pool *pool = NULL;
+ char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
+ "name", "virt", "busaddr"};
+
+ if (!_manager) {
+ seq_printf(m, "No pool allocator running.\n");
+ return 0;
+ }
+ seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
+ h[0], h[1], h[2], h[3], h[4], h[5]);
+ mutex_lock(&_manager->lock);
+ list_for_each_entry(p, &_manager->pools, pools) {
+ struct device *dev = p->dev;
+ if (!dev)
+ continue;
+ pool = p->pool;
+ seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
+ pool->name, pool->nrefills,
+ pool->nfrees, pool->npages_in_use,
+ pool->npages_free,
+ pool->dev_name);
+ }
+ mutex_unlock(&_manager->lock);
+ return 0;
+}
+EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index ae06e42..beef9ab 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -104,6 +104,7 @@ enum ttm_caching_state {
* @caching_state: The current caching state of the pages.
* @state: The current binding state of the pages.
* @dma_address: The DMA (bus) addresses of the pages (if TTM_PAGE_FLAG_DMA32)
+ * @alloc_list: used by some page allocation backend
*
* This is a structure holding the pages, caching- and aperture binding
* status for a buffer object that isn't backed by fixed (VRAM / AGP)
@@ -127,6 +128,7 @@ struct ttm_tt {
tt_unpopulated,
} state;
dma_addr_t *dma_address;
+ struct list_head alloc_list;
};
#define TTM_MEMTYPE_FLAG_FIXED (1 << 0) /* Fixed (on-card) PCI memory */
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 18a2d6b..f127635 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -88,4 +88,40 @@ extern void ttm_page_alloc_ttm_tt_unpopulate(struct ttm_tt *ttm);
* Output the state of pools to debugfs file
*/
extern int ttm_page_alloc_debugfs(struct seq_file *m, void *data);
+
+
+#ifdef CONFIG_SWIOTLB
+/**
+ * Initialize pool allocator.
+ */
+int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages);
+
+/**
+ * Free pool allocator.
+ */
+void ttm_dma_page_alloc_fini(void);
+
+/**
+ * Output the state of pools to debugfs file
+ */
+extern int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data);
+
+int ttm_dma_populate(struct ttm_tt *ttm, struct device *dev);
+extern void ttm_dma_unpopulate(struct ttm_tt *ttm, struct device *dev);
+
+#else
+static inline int ttm_dma_page_alloc_init(struct ttm_mem_global *glob,
+ unsigned max_pages)
+{
+ return -ENODEV;
+}
+
+static inline void ttm_dma_page_alloc_fini(void) { return; }
+
+static inline int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
+{
+ return 0;
+}
+#endif
+
#endif
--
1.7.7.1
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 11/13] swiotlb: Expose swiotlb_nr_tlb function to modules
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (9 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 10/13] ttm: Provide DMA aware TTM page pool code. V5 j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 12/13] drm/radeon/kms: Enable the TTM DMA pool if swiotlb is on V2 j.glisse
` (2 subsequent siblings)
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: FUJITA Tomonori, thellstrom
From: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
As a mechanism to detect whether SWIOTLB is enabled or not.
We also fix the spelling - it was swioltb instead of
swiotlb.
CC: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
[v1: Ripped out swiotlb_enabled]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/xen/swiotlb-xen.c | 2 +-
include/linux/swiotlb.h | 2 +-
lib/swiotlb.c | 5 +++--
3 files changed, 5 insertions(+), 4 deletions(-)
diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
index c984768..c50fb0b 100644
--- a/drivers/xen/swiotlb-xen.c
+++ b/drivers/xen/swiotlb-xen.c
@@ -152,7 +152,7 @@ void __init xen_swiotlb_init(int verbose)
char *m = NULL;
unsigned int repeat = 3;
- nr_tbl = swioltb_nr_tbl();
+ nr_tbl = swiotlb_nr_tbl();
if (nr_tbl)
xen_io_tlb_nslabs = nr_tbl;
else {
diff --git a/include/linux/swiotlb.h b/include/linux/swiotlb.h
index 445702c..e872526 100644
--- a/include/linux/swiotlb.h
+++ b/include/linux/swiotlb.h
@@ -24,7 +24,7 @@ extern int swiotlb_force;
extern void swiotlb_init(int verbose);
extern void swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose);
-extern unsigned long swioltb_nr_tbl(void);
+extern unsigned long swiotlb_nr_tbl(void);
/*
* Enumeration for sync targets
diff --git a/lib/swiotlb.c b/lib/swiotlb.c
index 99093b3..058935e 100644
--- a/lib/swiotlb.c
+++ b/lib/swiotlb.c
@@ -110,11 +110,11 @@ setup_io_tlb_npages(char *str)
__setup("swiotlb=", setup_io_tlb_npages);
/* make io_tlb_overflow tunable too? */
-unsigned long swioltb_nr_tbl(void)
+unsigned long swiotlb_nr_tbl(void)
{
return io_tlb_nslabs;
}
-
+EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
/* Note that this doesn't work with highmem page */
static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
volatile void *address)
@@ -321,6 +321,7 @@ void __init swiotlb_free(void)
free_bootmem_late(__pa(io_tlb_start),
PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
}
+ io_tlb_nslabs = 0;
}
static int is_swiotlb_buffer(phys_addr_t paddr)
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 12/13] drm/radeon/kms: Enable the TTM DMA pool if swiotlb is on V2
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (10 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 11/13] swiotlb: Expose swiotlb_nr_tlb function to modules j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:22 ` [PATCH 13/13] drm/nouveau: enable the TTM DMA pool on 32-bit DMA only device V2 j.glisse
2011-11-09 20:25 ` ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator Thomas Hellstrom
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Dave Airlie
From: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
With the exception that we do not handle the AGP case. We only
deal with PCIe cards such as ATI ES1000 or HD3200 that have been
detected to only do DMA up to 32-bits.
V2 force dma32 if we fail to set bigger dma mask
CC: Dave Airlie <airlied@redhat.com>
CC: Alex Deucher <alexdeucher@gmail.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
---
drivers/gpu/drm/radeon/radeon.h | 1 -
drivers/gpu/drm/radeon/radeon_device.c | 6 ++
drivers/gpu/drm/radeon/radeon_gart.c | 29 +-----------
drivers/gpu/drm/radeon/radeon_ttm.c | 83 +++++++++++++++++++++++++++++--
4 files changed, 84 insertions(+), 35 deletions(-)
diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index e3170c7..63257ba 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -332,7 +332,6 @@ struct radeon_gart {
union radeon_gart_table table;
struct page **pages;
dma_addr_t *pages_addr;
- bool *ttm_alloced;
bool ready;
};
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index c33bc91..7c31321 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -765,8 +765,14 @@ int radeon_device_init(struct radeon_device *rdev,
r = pci_set_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
if (r) {
rdev->need_dma32 = true;
+ dma_bits = 32;
printk(KERN_WARNING "radeon: No suitable DMA available.\n");
}
+ r = pci_set_consistent_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
+ if (r) {
+ pci_set_consistent_dma_mask(rdev->pdev, DMA_BIT_MASK(32));
+ printk(KERN_WARNING "radeon: No coherent DMA available.\n");
+ }
/* Registers mapping */
/* TODO: block userspace mapping of io register */
diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
index fdc3a9a..18f496c 100644
--- a/drivers/gpu/drm/radeon/radeon_gart.c
+++ b/drivers/gpu/drm/radeon/radeon_gart.c
@@ -149,9 +149,6 @@ void radeon_gart_unbind(struct radeon_device *rdev, unsigned offset,
p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
for (i = 0; i < pages; i++, p++) {
if (rdev->gart.pages[p]) {
- if (!rdev->gart.ttm_alloced[p])
- pci_unmap_page(rdev->pdev, rdev->gart.pages_addr[p],
- PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
rdev->gart.pages[p] = NULL;
rdev->gart.pages_addr[p] = rdev->dummy_page.addr;
page_base = rdev->gart.pages_addr[p];
@@ -181,23 +178,7 @@ int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
for (i = 0; i < pages; i++, p++) {
- /* we reverted the patch using dma_addr in TTM for now but this
- * code stops building on alpha so just comment it out for now */
- if (0) { /*dma_addr[i] != DMA_ERROR_CODE) */
- rdev->gart.ttm_alloced[p] = true;
- rdev->gart.pages_addr[p] = dma_addr[i];
- } else {
- /* we need to support large memory configurations */
- /* assume that unbind have already been call on the range */
- rdev->gart.pages_addr[p] = pci_map_page(rdev->pdev, pagelist[i],
- 0, PAGE_SIZE,
- PCI_DMA_BIDIRECTIONAL);
- if (pci_dma_mapping_error(rdev->pdev, rdev->gart.pages_addr[p])) {
- /* FIXME: failed to map page (return -ENOMEM?) */
- radeon_gart_unbind(rdev, offset, pages);
- return -ENOMEM;
- }
- }
+ rdev->gart.pages_addr[p] = dma_addr[i];
rdev->gart.pages[p] = pagelist[i];
page_base = rdev->gart.pages_addr[p];
for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
@@ -259,12 +240,6 @@ int radeon_gart_init(struct radeon_device *rdev)
radeon_gart_fini(rdev);
return -ENOMEM;
}
- rdev->gart.ttm_alloced = kzalloc(sizeof(bool) *
- rdev->gart.num_cpu_pages, GFP_KERNEL);
- if (rdev->gart.ttm_alloced == NULL) {
- radeon_gart_fini(rdev);
- return -ENOMEM;
- }
/* set GART entry to point to the dummy page by default */
for (i = 0; i < rdev->gart.num_cpu_pages; i++) {
rdev->gart.pages_addr[i] = rdev->dummy_page.addr;
@@ -281,10 +256,8 @@ void radeon_gart_fini(struct radeon_device *rdev)
rdev->gart.ready = false;
kfree(rdev->gart.pages);
kfree(rdev->gart.pages_addr);
- kfree(rdev->gart.ttm_alloced);
rdev->gart.pages = NULL;
rdev->gart.pages_addr = NULL;
- rdev->gart.ttm_alloced = NULL;
radeon_dummy_page_fini(rdev);
}
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 490afce..82119a5 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -581,11 +581,73 @@ struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
return >t->ttm;
}
+static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
+{
+ struct radeon_device *rdev;
+ unsigned i;
+ int r;
+
+ if (ttm->state != tt_unpopulated)
+ return 0;
+
+ rdev = radeon_get_rdev(ttm->bdev);
+
+#ifdef CONFIG_SWIOTLB
+ if (rdev->need_dma32 && swiotlb_nr_tbl()) {
+ return ttm_dma_populate(ttm, rdev->dev);
+ }
+#endif
+
+ r = ttm_page_alloc_ttm_tt_populate(ttm);
+ if (r) {
+ return r;
+ }
+
+ for (i = 0; i < ttm->num_pages; i++) {
+ ttm->dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
+ 0, PAGE_SIZE,
+ PCI_DMA_BIDIRECTIONAL);
+ if (pci_dma_mapping_error(rdev->pdev, ttm->dma_address[i])) {
+ while (--i) {
+ pci_unmap_page(rdev->pdev, ttm->dma_address[i],
+ PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+ ttm->dma_address[i] = 0;
+ }
+ ttm_page_alloc_ttm_tt_unpopulate(ttm);
+ return -EFAULT;
+ }
+ }
+ return 0;
+}
+
+static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
+{
+ struct radeon_device *rdev;
+ unsigned i;
+
+ rdev = radeon_get_rdev(ttm->bdev);
+
+#ifdef CONFIG_SWIOTLB
+ if (rdev->need_dma32 && swiotlb_nr_tbl()) {
+ ttm_dma_unpopulate(ttm, rdev->dev);
+ return;
+ }
+#endif
+
+ for (i = 0; i < ttm->num_pages; i++) {
+ if (ttm->dma_address[i]) {
+ pci_unmap_page(rdev->pdev, ttm->dma_address[i],
+ PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+ }
+ }
+
+ ttm_page_alloc_ttm_tt_unpopulate(ttm);
+}
static struct ttm_bo_driver radeon_bo_driver = {
.ttm_tt_create = &radeon_ttm_tt_create,
- .ttm_tt_populate = &ttm_page_alloc_ttm_tt_populate,
- .ttm_tt_unpopulate = &ttm_page_alloc_ttm_tt_unpopulate,
+ .ttm_tt_populate = &radeon_ttm_tt_populate,
+ .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
.invalidate_caches = &radeon_invalidate_caches,
.init_mem_type = &radeon_init_mem_type,
.evict_flags = &radeon_evict_flags,
@@ -771,8 +833,8 @@ static int radeon_mm_dump_table(struct seq_file *m, void *data)
static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
{
#if defined(CONFIG_DEBUG_FS)
- static struct drm_info_list radeon_mem_types_list[RADEON_DEBUGFS_MEM_TYPES+1];
- static char radeon_mem_types_names[RADEON_DEBUGFS_MEM_TYPES+1][32];
+ static struct drm_info_list radeon_mem_types_list[RADEON_DEBUGFS_MEM_TYPES+2];
+ static char radeon_mem_types_names[RADEON_DEBUGFS_MEM_TYPES+2][32];
unsigned i;
for (i = 0; i < RADEON_DEBUGFS_MEM_TYPES; i++) {
@@ -794,8 +856,17 @@ static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
radeon_mem_types_list[i].name = radeon_mem_types_names[i];
radeon_mem_types_list[i].show = &ttm_page_alloc_debugfs;
radeon_mem_types_list[i].driver_features = 0;
- radeon_mem_types_list[i].data = NULL;
- return radeon_debugfs_add_files(rdev, radeon_mem_types_list, RADEON_DEBUGFS_MEM_TYPES+1);
+ radeon_mem_types_list[i++].data = NULL;
+#ifdef CONFIG_SWIOTLB
+ if (rdev->need_dma32 && swiotlb_nr_tbl()) {
+ sprintf(radeon_mem_types_names[i], "ttm_dma_page_pool");
+ radeon_mem_types_list[i].name = radeon_mem_types_names[i];
+ radeon_mem_types_list[i].show = &ttm_dma_page_alloc_debugfs;
+ radeon_mem_types_list[i].driver_features = 0;
+ radeon_mem_types_list[i++].data = NULL;
+ }
+#endif
+ return radeon_debugfs_add_files(rdev, radeon_mem_types_list, i);
#endif
return 0;
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* [PATCH 13/13] drm/nouveau: enable the TTM DMA pool on 32-bit DMA only device V2
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (11 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 12/13] drm/radeon/kms: Enable the TTM DMA pool if swiotlb is on V2 j.glisse
@ 2011-11-09 20:22 ` j.glisse
2011-11-09 20:25 ` ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator Thomas Hellstrom
13 siblings, 0 replies; 28+ messages in thread
From: j.glisse @ 2011-11-09 20:22 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom, Ben Skeggs, Dave Airlie
From: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
If the card is capable of more than 32-bit, then use the default
TTM page pool code which allocates from anywhere in the memory.
Note: If the 'ttm.no_dma' parameter is set, the override is ignored
and the default TTM pool is used.
V2 use pci_set_consistent_dma_mask
CC: Ben Skeggs <bskeggs@redhat.com>
CC: Francisco Jerez <currojerez@riseup.net>
CC: Dave Airlie <airlied@redhat.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
---
drivers/gpu/drm/nouveau/nouveau_bo.c | 73 ++++++++++++++++++++++++++++-
drivers/gpu/drm/nouveau/nouveau_debugfs.c | 1 +
drivers/gpu/drm/nouveau/nouveau_mem.c | 6 ++
drivers/gpu/drm/nouveau/nouveau_sgdma.c | 60 +-----------------------
4 files changed, 79 insertions(+), 61 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 7e5ca3f..36234a7 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -1049,10 +1049,79 @@ nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
nouveau_fence_unref(&old_fence);
}
+static int
+nouveau_ttm_tt_populate(struct ttm_tt *ttm)
+{
+ struct drm_nouveau_private *dev_priv;
+ struct drm_device *dev;
+ unsigned i;
+ int r;
+
+ if (ttm->state != tt_unpopulated)
+ return 0;
+
+ dev_priv = nouveau_bdev(ttm->bdev);
+ dev = dev_priv->dev;
+
+#ifdef CONFIG_SWIOTLB
+ if ((dma_get_mask(dev->dev) <= DMA_BIT_MASK(32)) && swiotlb_nr_tbl()) {
+ return ttm_dma_populate(ttm, dev->dev);
+ }
+#endif
+
+ r = ttm_page_alloc_ttm_tt_populate(ttm);
+ if (r) {
+ return r;
+ }
+
+ for (i = 0; i < ttm->num_pages; i++) {
+ ttm->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
+ 0, PAGE_SIZE,
+ PCI_DMA_BIDIRECTIONAL);
+ if (pci_dma_mapping_error(dev->pdev, ttm->dma_address[i])) {
+ while (--i) {
+ pci_unmap_page(dev->pdev, ttm->dma_address[i],
+ PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+ ttm->dma_address[i] = 0;
+ }
+ ttm_page_alloc_ttm_tt_unpopulate(ttm);
+ return -EFAULT;
+ }
+ }
+ return 0;
+}
+
+static void
+nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
+{
+ struct drm_nouveau_private *dev_priv;
+ struct drm_device *dev;
+ unsigned i;
+
+ dev_priv = nouveau_bdev(ttm->bdev);
+ dev = dev_priv->dev;
+
+#ifdef CONFIG_SWIOTLB
+ if ((dma_get_mask(dev->dev) <= DMA_BIT_MASK(32)) && swiotlb_nr_tbl()) {
+ ttm_dma_unpopulate(ttm, dev->dev);
+ return;
+ }
+#endif
+
+ for (i = 0; i < ttm->num_pages; i++) {
+ if (ttm->dma_address[i]) {
+ pci_unmap_page(dev->pdev, ttm->dma_address[i],
+ PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+ }
+ }
+
+ ttm_page_alloc_ttm_tt_unpopulate(ttm);
+}
+
struct ttm_bo_driver nouveau_bo_driver = {
.ttm_tt_create = &nouveau_ttm_tt_create,
- .ttm_tt_populate = &ttm_page_alloc_ttm_tt_populate,
- .ttm_tt_unpopulate = &ttm_page_alloc_ttm_tt_unpopulate,
+ .ttm_tt_populate = &nouveau_ttm_tt_populate,
+ .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
.invalidate_caches = nouveau_bo_invalidate_caches,
.init_mem_type = nouveau_bo_init_mem_type,
.evict_flags = nouveau_bo_evict_flags,
diff --git a/drivers/gpu/drm/nouveau/nouveau_debugfs.c b/drivers/gpu/drm/nouveau/nouveau_debugfs.c
index 8e15923..f52c2db 100644
--- a/drivers/gpu/drm/nouveau/nouveau_debugfs.c
+++ b/drivers/gpu/drm/nouveau/nouveau_debugfs.c
@@ -178,6 +178,7 @@ static struct drm_info_list nouveau_debugfs_list[] = {
{ "memory", nouveau_debugfs_memory_info, 0, NULL },
{ "vbios.rom", nouveau_debugfs_vbios_image, 0, NULL },
{ "ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL },
+ { "ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL },
};
#define NOUVEAU_DEBUGFS_ENTRIES ARRAY_SIZE(nouveau_debugfs_list)
diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index 36bec48..37fcaa2 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -407,6 +407,12 @@ nouveau_mem_vram_init(struct drm_device *dev)
ret = pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(dma_bits));
if (ret)
return ret;
+ ret = pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(dma_bits));
+ if (ret) {
+ /* Reset to default value. */
+ pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(32));
+ }
+
ret = nouveau_ttm_global_init(dev_priv);
if (ret)
diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
index bc2ab90..ee1eb7c 100644
--- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c
+++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
@@ -13,41 +13,6 @@ struct nouveau_sgdma_be {
u64 offset;
};
-static int
-nouveau_sgdma_dma_map(struct ttm_tt *ttm)
-{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
- struct drm_device *dev = nvbe->dev;
- int i;
-
- for (i = 0; i < ttm->num_pages; i++) {
- ttm->dma_address[i] = pci_map_page(dev->pdev, ttm->pages[i],
- 0, PAGE_SIZE,
- PCI_DMA_BIDIRECTIONAL);
- if (pci_dma_mapping_error(dev->pdev, ttm->dma_address[i])) {
- return -EFAULT;
- }
- }
-
- return 0;
-}
-
-static void
-nouveau_sgdma_dma_unmap(struct ttm_tt *ttm)
-{
- struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
- struct drm_device *dev = nvbe->dev;
- int i;
-
- for (i = 0; i < ttm->num_pages; i++) {
- if (ttm->dma_address[i]) {
- pci_unmap_page(dev->pdev, ttm->dma_address[i],
- PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
- }
- ttm->dma_address[i] = 0;
- }
-}
-
static void
nouveau_sgdma_destroy(struct ttm_tt *ttm)
{
@@ -67,13 +32,8 @@ nv04_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
unsigned i, j, pte;
- int r;
NV_DEBUG(dev, "pg=0x%lx\n", mem->start);
- r = nouveau_sgdma_dma_map(ttm);
- if (r) {
- return r;
- }
nvbe->offset = mem->start << PAGE_SHIFT;
pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
@@ -110,7 +70,6 @@ nv04_sgdma_unbind(struct ttm_tt *ttm)
nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
}
- nouveau_sgdma_dma_unmap(ttm);
return 0;
}
@@ -141,13 +100,8 @@ nv41_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
dma_addr_t *list = ttm->dma_address;
u32 pte = mem->start << 2;
u32 cnt = ttm->num_pages;
- int r;
nvbe->offset = mem->start << PAGE_SHIFT;
- r = nouveau_sgdma_dma_map(ttm);
- if (r) {
- return r;
- }
while (cnt--) {
nv_wo32(pgt, pte, (*list++ >> 7) | 1);
@@ -173,7 +127,6 @@ nv41_sgdma_unbind(struct ttm_tt *ttm)
}
nv41_sgdma_flush(nvbe);
- nouveau_sgdma_dma_unmap(ttm);
return 0;
}
@@ -256,13 +209,9 @@ nv44_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
dma_addr_t *list = ttm->dma_address;
u32 pte = mem->start << 2, tmp[4];
u32 cnt = ttm->num_pages;
- int i, r;
+ int i;
nvbe->offset = mem->start << PAGE_SHIFT;
- r = nouveau_sgdma_dma_map(ttm);
- if (r) {
- return r;
- }
if (pte & 0x0000000c) {
u32 max = 4 - ((pte >> 2) & 0x3);
@@ -321,7 +270,6 @@ nv44_sgdma_unbind(struct ttm_tt *ttm)
nv44_sgdma_fill(pgt, NULL, pte, cnt);
nv44_sgdma_flush(ttm);
- nouveau_sgdma_dma_unmap(ttm);
return 0;
}
@@ -335,13 +283,8 @@ static int
nv50_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
{
struct nouveau_mem *node = mem->mm_node;
- int r;
/* noop: bound in move_notify() */
- r = nouveau_sgdma_dma_map(ttm);
- if (r) {
- return r;
- }
node->pages = ttm->dma_address;
return 0;
}
@@ -350,7 +293,6 @@ static int
nv50_sgdma_unbind(struct ttm_tt *ttm)
{
/* noop: unbound in move_notify() */
- nouveau_sgdma_dma_unmap(ttm);
return 0;
}
--
1.7.7.1
^ permalink raw reply related [flat|nested] 28+ messages in thread
* Re: ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
` (12 preceding siblings ...)
2011-11-09 20:22 ` [PATCH 13/13] drm/nouveau: enable the TTM DMA pool on 32-bit DMA only device V2 j.glisse
@ 2011-11-09 20:25 ` Thomas Hellstrom
2011-11-09 21:03 ` Jerome Glisse
13 siblings, 1 reply; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-09 20:25 UTC (permalink / raw)
To: j.glisse; +Cc: dri-devel
On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> So i did an overhaul of ttm_memory, i believe the simplification i did
> make sense. See patch 5 for a longer explanation.
>
>
> Thomas with the ttm_memory change the allocation of pages won't happen
> if the accounting report that we are going over the limit and bo shrinker
> failed to free any memory to make room.
>
> The handling of dma32 zone is done as post pass of ttm memory accounting.
>
OK. I'll take a deeper look into this.
> Regarding the pagefault comment i removed, it doesn't make sense anymore
> because now we populate the whole page table in one shot. So there is
> no more prefaulting few pages but a full prefaulting. Thought i can
> add a comment stating that if you like.
>
It's important that we distinguish between populating, which populates
pages,
and faulting, which add ptes pointing to those pages.
Previously populating happened as a side effect of faulting, but now
that populating is done
in a single step, faulting (adding ptes) is still not. Usually a fault()
handler adds a single pte,
but TTM is different and tries to prefault more, but it is important
that we only error on the first
pte, so that's why the comment should stay.
> For the ttm_tt_dma struct to hold page allocator specific informations
> i think it can be done as an followup patch but if you prefer to have
> that in this patchset let me know i will respin with such changes.
>
>
I'm fine with having it as a separate patchset as long as it gets done :).
> I am in the process of retesting this whole serie and especialy the
> while memory accounting.
>
> Cheers,
> Jerome
>
/Thomas
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator
2011-11-09 20:25 ` ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator Thomas Hellstrom
@ 2011-11-09 21:03 ` Jerome Glisse
0 siblings, 0 replies; 28+ messages in thread
From: Jerome Glisse @ 2011-11-09 21:03 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: dri-devel
On Wed, Nov 09, 2011 at 09:25:20PM +0100, Thomas Hellstrom wrote:
> On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> >So i did an overhaul of ttm_memory, i believe the simplification i did
> >make sense. See patch 5 for a longer explanation.
> >
> >Thomas with the ttm_memory change the allocation of pages won't happen
> >if the accounting report that we are going over the limit and bo shrinker
> >failed to free any memory to make room.
> >
> >The handling of dma32 zone is done as post pass of ttm memory accounting.
>
> OK. I'll take a deeper look into this.
>
> >Regarding the pagefault comment i removed, it doesn't make sense anymore
> >because now we populate the whole page table in one shot. So there is
> >no more prefaulting few pages but a full prefaulting. Thought i can
> >add a comment stating that if you like.
> It's important that we distinguish between populating, which
> populates pages,
> and faulting, which add ptes pointing to those pages.
>
> Previously populating happened as a side effect of faulting, but now
> that populating is done
> in a single step, faulting (adding ptes) is still not. Usually a
> fault() handler adds a single pte,
> but TTM is different and tries to prefault more, but it is important
> that we only error on the first
> pte, so that's why the comment should stay.
>
Well yes it only fill numprefault pte, but no error can happen in the
prefault loop except for vm_insert_mixed failure, it's why i kept the
report error only if it fails on the first page. I actually did a full
pte populate at one point while working on that, dunno if that would
make sense.
> >For the ttm_tt_dma struct to hold page allocator specific informations
> >i think it can be done as an followup patch but if you prefer to have
> >that in this patchset let me know i will respin with such changes.
> >
>
> I'm fine with having it as a separate patchset as long as it gets done :).
>
I will spin a patch for that on top of the patchset.
Cheers,
Jerome
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-09 20:22 ` [PATCH 05/13] drm/ttm: overhaul memory accounting j.glisse
@ 2011-11-10 10:27 ` Thomas Hellstrom
2011-11-10 18:05 ` Jerome Glisse
0 siblings, 1 reply; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-10 10:27 UTC (permalink / raw)
To: j.glisse; +Cc: thellstrom, Jerome Glisse, dri-devel
On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> From: Jerome Glisse<jglisse@redhat.com>
>
> This is an overhaul of the ttm memory accounting. This tries to keep
> the same global behavior while removing the whole zone concept. It
> keeps a distrinction for dma32 so that we make sure that ttm don't
> starve the dma32 zone.
>
> There is 3 threshold for memory allocation :
> - max_mem is the maximum memory the whole ttm infrastructure is
> going to allow allocation for (exception of system process see
> below)
> - emer_mem is the maximum memory allowed for system process, this
> limit is> to max_mem
> - swap_limit is the threshold at which point ttm will start to
> try to swap object because ttm is getting close the max_mem
> limit
> - swap_dma32_limit is the threshold at which point ttm will start
> swap object to try to reduce the pressure on the dma32 zone. Note
> that we don't specificly target object to swap to it might very
> well free more memory from highmem rather than from dma32
>
> Accounting is done through used_mem& used_dma32_mem, which sum give
> the total amount of memory actually accounted by ttm.
>
> Idea is that allocation will fail if (used_mem + used_dma32_mem)>
> max_mem and if swapping fail to make enough room.
>
> The used_dma32_mem can be updated as a later stage, allowing to
> perform accounting test before allocating a whole batch of pages.
>
>
Jerome, you're removing a fair amount of functionality here, without
justifying
why it could be removed.
Consider a low-end system with 1G of kernel memory and 10G of highmem.
How do we avoid putting stress on the kernel memory? I also wouldn't be
too surprised if DMA32 zones appear in HIGHMEM systems in the future
making the current zone concept good to keep.
Also, in effect you move the DOS from *all* zones into the DMA32 zone
and create a race in that multiple simultaneous allocators can first
pre-allocate out of the global zone, and then update the DMA32 zone
without synchronization. In this way you might theoretically end up with
more DMA32 pages allocated than present in the zone.
With the proposed code there's also a theoretical problem in that a
potentially huge number of pages are unaccounted before they are
actually freed.
A possible way around all this is to pre-allocate out of *all* zones,
and after the big allocation release back memory to relevant zones. If
such a big allocation fails, one needs to revert back to a page-by-page
scheme.
/Thomas
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-10 10:27 ` Thomas Hellstrom
@ 2011-11-10 18:05 ` Jerome Glisse
2011-11-10 20:05 ` Thomas Hellstrom
0 siblings, 1 reply; 28+ messages in thread
From: Jerome Glisse @ 2011-11-10 18:05 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: Jerome Glisse, dri-devel
[-- Attachment #1: Type: text/plain, Size: 4306 bytes --]
On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
> On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> >From: Jerome Glisse<jglisse@redhat.com>
> >
> >This is an overhaul of the ttm memory accounting. This tries to keep
> >the same global behavior while removing the whole zone concept. It
> >keeps a distrinction for dma32 so that we make sure that ttm don't
> >starve the dma32 zone.
> >
> >There is 3 threshold for memory allocation :
> >- max_mem is the maximum memory the whole ttm infrastructure is
> > going to allow allocation for (exception of system process see
> > below)
> >- emer_mem is the maximum memory allowed for system process, this
> > limit is> to max_mem
> >- swap_limit is the threshold at which point ttm will start to
> > try to swap object because ttm is getting close the max_mem
> > limit
> >- swap_dma32_limit is the threshold at which point ttm will start
> > swap object to try to reduce the pressure on the dma32 zone. Note
> > that we don't specificly target object to swap to it might very
> > well free more memory from highmem rather than from dma32
> >
> >Accounting is done through used_mem& used_dma32_mem, which sum give
> >the total amount of memory actually accounted by ttm.
> >
> >Idea is that allocation will fail if (used_mem + used_dma32_mem)>
> >max_mem and if swapping fail to make enough room.
> >
> >The used_dma32_mem can be updated as a later stage, allowing to
> >perform accounting test before allocating a whole batch of pages.
> >
>
> Jerome, you're removing a fair amount of functionality here, without
> justifying
> why it could be removed.
All this code was overkill.
> Consider a low-end system with 1G of kernel memory and 10G of
> highmem. How do we avoid putting stress on the kernel memory? I also
> wouldn't be too surprised if DMA32 zones appear in HIGHMEM systems
> in the future making the current zone concept good to keep.
Right now kernel memory is accounted against all zone so it decrease
not only the kernel zone but also the dma32 & highmem if present.
Note also that kernel zone in current code == dma32 zone.
When it comes to future it looks a lot simpler, it seems everyone
is moving toward more capable and more advanced iommu that can remove
all the restriction on memory from the device pov. I mean even arm
are getting more and more advance iommu. I don't see any architecture
worse supporting not going down that road.
> Also, in effect you move the DOS from *all* zones into the DMA32
> zone and create a race in that multiple simultaneous allocators can
> first pre-allocate out of the global zone, and then update the DMA32
> zone without synchronization. In this way you might theoretically
> end up with more DMA32 pages allocated than present in the zone.
Ok a respin attached with a simple change, things will be
accounted against dma32 zone and only when we get page we will
decrease the dma32 zone usage that way no DOS on dma32.
It also deals with the case where there is still lot of highmem
but no more dma32.
> With the proposed code there's also a theoretical problem in that a
> potentially huge number of pages are unaccounted before they are
> actually freed.
What you mean unaccounted ? The way it works is :
r = global_memory_alloc(size)
if (r) fail
alloc pages
update memory accounting according to what page where allocated
So memory is always accounted before even being allocated (exception
are the kernel object for vmwgfx & ttm_bo but we can move accounting
there too if you want, those are small allocation and i didn't think
it was worse changing that).
> A possible way around all this is to pre-allocate out of *all*
> zones, and after the big allocation release back memory to relevant
> zones. If such a big allocation fails, one needs to revert back to a
> page-by-page scheme.
>
> /Thomas
Really i believe this make the whole accounting a lot simpler. The
whole zone business was overkill as especialy kernelzone == dma32zone
and highmem zone is a superset of this.
With my change what happen is that only the dma32 distinction is kept
around because there is whole batch of device that can only do dma32
and we need to make sure we don't starve those.
I believe my code is lot easier and straighforward to understand.
Cheers,
Jerome
[-- Attachment #2: 0005-drm-ttm-overhaul-memory-accounting-V2.patch --]
[-- Type: text/plain, Size: 31690 bytes --]
>From 03ec0abcfa4bc3060fd0d7a797758fc50088ec29 Mon Sep 17 00:00:00 2001
From: Jerome Glisse <jglisse@redhat.com>
Date: Wed, 9 Nov 2011 08:11:00 -0500
Subject: [PATCH] drm/ttm: overhaul memory accounting V2
This is an overhaul of the ttm memory accounting. This tries to keep
the same global behavior while removing the whole zone concept. It
keeps a distrinction for dma32 so that we make sure that ttm don't
starve the dma32 zone.
There is 3 threshold for memory allocation :
- max_mem is the maximum memory the whole ttm infrastructure is
going to allow allocation for (exception of system process see
below)
- emer_mem is the maximum memory allowed for system process, this
limit is > to max_mem
- swap_limit is the threshold at which point ttm will start to
try to swap object because ttm is getting close the max_mem
limit
- swap_dma32_limit is the threshold at which point ttm will start
swap object to try to reduce the pressure on the dma32 zone. Note
that we don't specificly target object to swap to it might very
well free more memory from highmem rather than from dma32
Accounting is done through used_mem & used_dma32_mem, which sum give
the total amount of memory actually accounted by ttm.
Idea is that allocation will fail if (used_mem + used_dma32_mem) >
max_mem and if swapping fail to make enough room.
The used_dma32_mem can be updated as a later stage, allowing to
perform accounting test before allocating a whole batch of pages.
V2 Move accouting before page allocation instead of leaving that
to a later patch. Always account first against dma32 zone and
only as a post pass rebalance accouting to highmem if pages
allow it.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
---
drivers/gpu/drm/ttm/ttm_bo.c | 9 +-
drivers/gpu/drm/ttm/ttm_memory.c | 554 ++++++++++--------------------
drivers/gpu/drm/ttm/ttm_object.c | 10 +-
drivers/gpu/drm/ttm/ttm_tt.c | 14 +-
drivers/gpu/drm/vmwgfx/vmwgfx_fence.c | 30 +-
drivers/gpu/drm/vmwgfx/vmwgfx_resource.c | 51 ++--
include/drm/ttm/ttm_memory.h | 27 +-
7 files changed, 263 insertions(+), 432 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 4bde335..b3674ba 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -1248,13 +1248,16 @@ int ttm_bo_create(struct ttm_bo_device *bdev,
{
struct ttm_buffer_object *bo;
struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
+ unsigned long available;
int ret;
size_t acc_size =
ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
- ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
- if (unlikely(ret != 0))
- return ret;
+ available = ttm_mem_global_alloc(mem_glob, acc_size, false);
+ if (available != acc_size) {
+ ttm_mem_global_free(mem_glob, available);
+ return -ENOMEM;
+ }
bo = kzalloc(sizeof(*bo), GFP_KERNEL);
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index e70ddd8..2aa08a7 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -35,21 +35,10 @@
#include <linux/module.h>
#include <linux/slab.h>
-#define TTM_MEMORY_ALLOC_RETRIES 4
-
-struct ttm_mem_zone {
- struct kobject kobj;
- struct ttm_mem_global *glob;
- const char *name;
- uint64_t zone_mem;
- uint64_t emer_mem;
- uint64_t max_mem;
- uint64_t swap_limit;
- uint64_t used_mem;
-};
+#define TTM_MEMORY_RETRIES 4
static struct attribute ttm_mem_sys = {
- .name = "zone_memory",
+ .name = "memory",
.mode = S_IRUGO
};
static struct attribute ttm_mem_emer = {
@@ -60,144 +49,156 @@ static struct attribute ttm_mem_max = {
.name = "available_memory",
.mode = S_IRUGO | S_IWUSR
};
+static struct attribute ttm_mem_dma32_max = {
+ .name = "available_dma32_memory",
+ .mode = S_IRUGO | S_IWUSR
+};
static struct attribute ttm_mem_swap = {
.name = "swap_limit",
.mode = S_IRUGO | S_IWUSR
};
+static struct attribute ttm_mem_dma32_swap = {
+ .name = "swap_dma32_limit",
+ .mode = S_IRUGO | S_IWUSR
+};
static struct attribute ttm_mem_used = {
.name = "used_memory",
.mode = S_IRUGO
};
+static struct attribute ttm_mem_dma32_used = {
+ .name = "used_dma32_memory",
+ .mode = S_IRUGO
+};
-static void ttm_mem_zone_kobj_release(struct kobject *kobj)
-{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
-
- printk(KERN_INFO TTM_PFX
- "Zone %7s: Used memory at exit: %llu kiB.\n",
- zone->name, (unsigned long long) zone->used_mem >> 10);
- kfree(zone);
-}
-
-static ssize_t ttm_mem_zone_show(struct kobject *kobj,
- struct attribute *attr,
- char *buffer)
+static ssize_t ttm_mem_global_show(struct kobject *kobj,
+ struct attribute *attr,
+ char *buffer)
{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
- uint64_t val = 0;
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
+ unsigned long val = 0;
- spin_lock(&zone->glob->lock);
+ spin_lock(&glob->lock);
if (attr == &ttm_mem_sys)
- val = zone->zone_mem;
+ val = glob->mem;
else if (attr == &ttm_mem_emer)
- val = zone->emer_mem;
+ val = glob->emer_mem;
else if (attr == &ttm_mem_max)
- val = zone->max_mem;
+ val = glob->max_mem;
+ else if (attr == &ttm_mem_dma32_max)
+ val = glob->max_dma32_mem;
else if (attr == &ttm_mem_swap)
- val = zone->swap_limit;
+ val = glob->swap_limit;
else if (attr == &ttm_mem_used)
- val = zone->used_mem;
- spin_unlock(&zone->glob->lock);
+ val = glob->used_mem;
+ else if (attr == &ttm_mem_dma32_used)
+ val = glob->used_dma32_mem;
+ else if (attr == &ttm_mem_dma32_swap)
+ val = glob->swap_dma32_limit;
+ spin_unlock(&glob->lock);
- return snprintf(buffer, PAGE_SIZE, "%llu\n",
- (unsigned long long) val >> 10);
+ return snprintf(buffer, PAGE_SIZE, "%lu\n", val >> 10);
}
static void ttm_check_swapping(struct ttm_mem_global *glob);
-static ssize_t ttm_mem_zone_store(struct kobject *kobj,
- struct attribute *attr,
- const char *buffer,
- size_t size)
+static ssize_t ttm_mem_global_store(struct kobject *kobj,
+ struct attribute *attr,
+ const char *buffer,
+ size_t size)
{
- struct ttm_mem_zone *zone =
- container_of(kobj, struct ttm_mem_zone, kobj);
- int chars;
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
unsigned long val;
- uint64_t val64;
+ int chars;
chars = sscanf(buffer, "%lu", &val);
if (chars == 0)
return size;
- val64 = val;
- val64 <<= 10;
+ val <<= 10;
+
+ spin_lock(&glob->lock);
+ /* limit to maximum memory */
+ if (val > glob->mem)
+ val = glob->mem;
- spin_lock(&zone->glob->lock);
- if (val64 > zone->zone_mem)
- val64 = zone->zone_mem;
if (attr == &ttm_mem_emer) {
- zone->emer_mem = val64;
- if (zone->max_mem > val64)
- zone->max_mem = val64;
+ glob->emer_mem = val;
+ if (glob->max_dma32_mem > val)
+ glob->max_dma32_mem = val;
} else if (attr == &ttm_mem_max) {
- zone->max_mem = val64;
- if (zone->emer_mem < val64)
- zone->emer_mem = val64;
- } else if (attr == &ttm_mem_swap)
- zone->swap_limit = val64;
- spin_unlock(&zone->glob->lock);
+ glob->max_mem = val;
+ } else if (attr == &ttm_mem_dma32_max) {
+ glob->max_dma32_mem = val;
+ if (glob->emer_mem < val)
+ glob->emer_mem = val;
+ } else if (attr == &ttm_mem_swap) {
+ glob->swap_limit = val;
+ } else if (attr == &ttm_mem_dma32_swap) {
+ glob->swap_dma32_limit = val;
+ }
+ spin_unlock(&glob->lock);
+ ttm_check_swapping(glob);
+ return size;
+}
- ttm_check_swapping(zone->glob);
+static void ttm_mem_global_kobj_release(struct kobject *kobj)
+{
+ struct ttm_mem_global *glob =
+ container_of(kobj, struct ttm_mem_global, kobj);
- return size;
+ kfree(glob);
}
-static struct attribute *ttm_mem_zone_attrs[] = {
+static struct attribute *ttm_mem_global_attrs[] = {
&ttm_mem_sys,
&ttm_mem_emer,
&ttm_mem_max,
&ttm_mem_swap,
+ &ttm_mem_dma32_swap,
&ttm_mem_used,
+ &ttm_mem_dma32_used,
NULL
};
-static const struct sysfs_ops ttm_mem_zone_ops = {
- .show = &ttm_mem_zone_show,
- .store = &ttm_mem_zone_store
-};
-
-static struct kobj_type ttm_mem_zone_kobj_type = {
- .release = &ttm_mem_zone_kobj_release,
- .sysfs_ops = &ttm_mem_zone_ops,
- .default_attrs = ttm_mem_zone_attrs,
+static const struct sysfs_ops ttm_mem_global_ops = {
+ .show = &ttm_mem_global_show,
+ .store = &ttm_mem_global_store
};
-static void ttm_mem_global_kobj_release(struct kobject *kobj)
-{
- struct ttm_mem_global *glob =
- container_of(kobj, struct ttm_mem_global, kobj);
-
- kfree(glob);
-}
-
static struct kobj_type ttm_mem_glob_kobj_type = {
.release = &ttm_mem_global_kobj_release,
+ .sysfs_ops = &ttm_mem_global_ops,
+ .default_attrs = ttm_mem_global_attrs,
};
-static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
- bool from_wq, uint64_t extra)
+static bool ttm_above_swap_target(struct ttm_mem_global *glob,
+ bool from_wq, uint64_t extra)
{
- unsigned int i;
- struct ttm_mem_zone *zone;
- uint64_t target;
-
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
+ unsigned long target;
- if (from_wq)
- target = zone->swap_limit;
- else if (capable(CAP_SYS_ADMIN))
- target = zone->emer_mem;
+ if (from_wq) {
+ if (glob->used_mem > glob->swap_limit) {
+ return true;
+ }
+ if (glob->used_dma32_mem > glob->swap_dma32_limit) {
+ return true;
+ }
+ } else {
+ if (capable(CAP_SYS_ADMIN))
+ target = glob->emer_mem;
else
- target = zone->max_mem;
-
- target = (extra > target) ? 0ULL : target;
-
- if (zone->used_mem > target)
+ target = glob->max_dma32_mem;
+ if (extra > target) {
+ return true;
+ }
+ if ((glob->used_dma32_mem + extra) > target) {
+ return true;
+ }
+ if ((glob->used_mem + glob->used_dma32_mem + extra) > glob->max_mem) {
return true;
+ }
}
return false;
}
@@ -208,31 +209,31 @@ static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
* Note that this function is reentrant:
* many threads may try to swap out at any given time.
*/
-
-static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
+static void ttm_shrink(struct ttm_mem_global *glob,
+ bool from_wq,
uint64_t extra)
{
- int ret;
struct ttm_mem_shrink *shrink;
+ int ret, nretries = TTM_MEMORY_RETRIES;
spin_lock(&glob->lock);
if (glob->shrink == NULL)
goto out;
- while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
+ while (ttm_above_swap_target(glob, from_wq, extra)) {
shrink = glob->shrink;
spin_unlock(&glob->lock);
ret = shrink->do_shrink(shrink);
spin_lock(&glob->lock);
if (unlikely(ret != 0))
goto out;
+ if (--nretries < 0)
+ goto out;
}
out:
spin_unlock(&glob->lock);
}
-
-
static void ttm_shrink_work(struct work_struct *work)
{
struct ttm_mem_global *glob =
@@ -241,127 +242,10 @@ static void ttm_shrink_work(struct work_struct *work)
ttm_shrink(glob, true, 0ULL);
}
-static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- uint64_t mem;
- int ret;
-
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram - si->totalhigh;
- mem *= si->mem_unit;
-
- zone->name = "kernel";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_kernel = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-
-#ifdef CONFIG_HIGHMEM
-static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone;
- uint64_t mem;
- int ret;
-
- if (si->totalhigh == 0)
- return 0;
-
- zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram;
- mem *= si->mem_unit;
-
- zone->name = "highmem";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_highmem = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-#else
-static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
- const struct sysinfo *si)
-{
- struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
- uint64_t mem;
- int ret;
-
- if (unlikely(!zone))
- return -ENOMEM;
-
- mem = si->totalram;
- mem *= si->mem_unit;
-
- /**
- * No special dma32 zone needed.
- */
-
- if (mem <= ((uint64_t) 1ULL << 32)) {
- kfree(zone);
- return 0;
- }
-
- /*
- * Limit max dma32 memory to 4GB for now
- * until we can figure out how big this
- * zone really is.
- */
-
- mem = ((uint64_t) 1ULL << 32);
- zone->name = "dma32";
- zone->zone_mem = mem;
- zone->max_mem = mem >> 1;
- zone->emer_mem = (mem >> 1) + (mem >> 2);
- zone->swap_limit = zone->max_mem - (mem >> 3);
- zone->used_mem = 0;
- zone->glob = glob;
- glob->zone_dma32 = zone;
- ret = kobject_init_and_add(
- &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
- if (unlikely(ret != 0)) {
- kobject_put(&zone->kobj);
- return ret;
- }
- glob->zones[glob->num_zones++] = zone;
- return 0;
-}
-#endif
-
int ttm_mem_global_init(struct ttm_mem_global *glob)
{
struct sysinfo si;
int ret;
- int i;
- struct ttm_mem_zone *zone;
spin_lock_init(&glob->lock);
glob->swap_queue = create_singlethread_workqueue("ttm_swap");
@@ -374,50 +258,36 @@ int ttm_mem_global_init(struct ttm_mem_global *glob)
return ret;
}
+ /* compute limit */
si_meminfo(&si);
-
- ret = ttm_mem_init_kernel_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#ifdef CONFIG_HIGHMEM
- ret = ttm_mem_init_highmem_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#else
- ret = ttm_mem_init_dma32_zone(glob, &si);
- if (unlikely(ret != 0))
- goto out_no_zone;
-#endif
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- printk(KERN_INFO TTM_PFX
- "Zone %7s: Available graphics memory: %llu kiB.\n",
- zone->name, (unsigned long long) zone->max_mem >> 10);
+ glob->mem = si.totalram;
+ glob->mem *= si.mem_unit;
+ glob->used_mem = 0;
+ glob->used_dma32_mem = 0;
+ glob->max_mem = glob->mem >> 1;
+ glob->max_dma32_mem = (1ULL << 32) >> 1;
+ if (glob->max_dma32_mem > glob->max_mem) {
+ glob->max_dma32_mem = glob->max_mem;
+ }
+ glob->emer_mem = glob->max_dma32_mem + (glob->max_dma32_mem >> 1);
+ glob->swap_limit = glob->max_mem - (glob->mem >> 3);
+ glob->swap_dma32_limit = ((1ULL << 32) >> 1) - (((1ULL << 32) >> 3));
+ if (glob->swap_dma32_limit > glob->swap_limit) {
+ glob->swap_dma32_limit = glob->swap_limit;
}
- ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
+ ttm_page_alloc_init(glob, glob->max_mem/(2*PAGE_SIZE));
return 0;
-out_no_zone:
- ttm_mem_global_release(glob);
- return ret;
}
EXPORT_SYMBOL(ttm_mem_global_init);
void ttm_mem_global_release(struct ttm_mem_global *glob)
{
- unsigned int i;
- struct ttm_mem_zone *zone;
-
/* let the page allocator first stop the shrink work. */
ttm_page_alloc_fini();
flush_workqueue(glob->swap_queue);
destroy_workqueue(glob->swap_queue);
glob->swap_queue = NULL;
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- kobject_del(&zone->kobj);
- kobject_put(&zone->kobj);
- }
kobject_del(&glob->kobj);
kobject_put(&glob->kobj);
}
@@ -426,18 +296,14 @@ EXPORT_SYMBOL(ttm_mem_global_release);
static void ttm_check_swapping(struct ttm_mem_global *glob)
{
bool needs_swapping = false;
- unsigned int i;
- struct ttm_mem_zone *zone;
spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (zone->used_mem > zone->swap_limit) {
- needs_swapping = true;
- break;
- }
+ if ((glob->used_mem + glob->used_dma32_mem) > glob->swap_limit) {
+ needs_swapping = true;
+ }
+ if (glob->used_dma32_mem > glob->swap_dma32_limit) {
+ needs_swapping = true;
}
-
spin_unlock(&glob->lock);
if (unlikely(needs_swapping))
@@ -445,142 +311,84 @@ static void ttm_check_swapping(struct ttm_mem_global *glob)
}
-static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t amount)
-{
- unsigned int i;
- struct ttm_mem_zone *zone;
-
- spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
- zone->used_mem -= amount;
- }
- spin_unlock(&glob->lock);
-}
-
void ttm_mem_global_free(struct ttm_mem_global *glob,
uint64_t amount)
{
- return ttm_mem_global_free_zone(glob, NULL, amount);
+ spin_lock(&glob->lock);
+ glob->used_dma32_mem -= amount;
+ spin_unlock(&glob->lock);
}
EXPORT_SYMBOL(ttm_mem_global_free);
-static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t amount, bool reserve)
+unsigned long ttm_mem_global_alloc(struct ttm_mem_global *glob,
+ uint64_t memory,
+ bool no_wait)
{
- uint64_t limit;
- int ret = -ENOMEM;
- unsigned int i;
- struct ttm_mem_zone *zone;
+ unsigned long limit, available = 0;
+ int i;
+ for (i = 0; i < TTM_MEMORY_RETRIES; i++) {
+ spin_lock(&glob->lock);
+ available = 0;
+ limit = (capable(CAP_SYS_ADMIN)) ? glob->emer_mem : glob->max_dma32_mem;
+ if (glob->used_dma32_mem > limit) {
+ available = limit - glob->used_dma32_mem;
+ }
+ if (memory <= available) {
+ glob->used_dma32_mem += memory;
+ spin_unlock(&glob->lock);
+ return memory;
+ }
+ if (no_wait) {
+ glob->used_dma32_mem += available;
+ spin_unlock(&glob->lock);
+ return available;
+ }
+ spin_unlock(&glob->lock);
+ ttm_shrink(glob, false, memory + (memory >> 2) + 16);
+ }
spin_lock(&glob->lock);
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
-
- limit = (capable(CAP_SYS_ADMIN)) ?
- zone->emer_mem : zone->max_mem;
-
- if (zone->used_mem > limit)
- goto out_unlock;
+ available = 0;
+ limit = (capable(CAP_SYS_ADMIN)) ? glob->emer_mem : glob->max_dma32_mem;
+ if (glob->used_dma32_mem > limit) {
+ available = limit - glob->used_dma32_mem;
}
-
- if (reserve) {
- for (i = 0; i < glob->num_zones; ++i) {
- zone = glob->zones[i];
- if (single_zone && zone != single_zone)
- continue;
- zone->used_mem += amount;
- }
+ if (memory <= available) {
+ glob->used_dma32_mem += memory;
+ spin_unlock(&glob->lock);
+ return memory;
}
-
- ret = 0;
-out_unlock:
+ glob->used_dma32_mem += available;
spin_unlock(&glob->lock);
- ttm_check_swapping(glob);
-
- return ret;
+ return available;
}
+EXPORT_SYMBOL(ttm_mem_global_alloc);
-
-static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
- struct ttm_mem_zone *single_zone,
- uint64_t memory,
- bool no_wait, bool interruptible)
+void ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
+ struct page *page)
{
- int count = TTM_MEMORY_ALLOC_RETRIES;
-
- while (unlikely(ttm_mem_global_reserve(glob,
- single_zone,
- memory, true)
- != 0)) {
- if (no_wait)
- return -ENOMEM;
- if (unlikely(count-- == 0))
- return -ENOMEM;
- ttm_shrink(glob, false, memory + (memory >> 2) + 16);
+ /* check if page is dma32 */
+ if (page_to_pfn(page) > 0x00100000UL) {
+ spin_lock(&glob->lock);
+ glob->used_dma32_mem -= PAGE_SIZE;
+ glob->used_mem += PAGE_SIZE;
+ spin_unlock(&glob->lock);
}
-
+ ttm_check_swapping(glob);
return 0;
}
-int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
- bool no_wait, bool interruptible)
-{
- /**
- * Normal allocations of kernel memory are registered in
- * all zones.
- */
-
- return ttm_mem_global_alloc_zone(glob, NULL, memory, no_wait,
- interruptible);
-}
-EXPORT_SYMBOL(ttm_mem_global_alloc);
-
-int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
- struct page *page,
- bool no_wait, bool interruptible)
-{
-
- struct ttm_mem_zone *zone = NULL;
-
- /**
- * Page allocations may be registed in a single zone
- * only if highmem or !dma32.
- */
-
-#ifdef CONFIG_HIGHMEM
- if (PageHighMem(page) && glob->zone_highmem != NULL)
- zone = glob->zone_highmem;
-#else
- if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
- zone = glob->zone_kernel;
-#endif
- return ttm_mem_global_alloc_zone(glob, zone, PAGE_SIZE, no_wait,
- interruptible);
-}
-
void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page)
{
- struct ttm_mem_zone *zone = NULL;
-
-#ifdef CONFIG_HIGHMEM
- if (PageHighMem(page) && glob->zone_highmem != NULL)
- zone = glob->zone_highmem;
-#else
- if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
- zone = glob->zone_kernel;
-#endif
- ttm_mem_global_free_zone(glob, zone, PAGE_SIZE);
+ spin_lock(&glob->lock);
+ if (page_to_pfn(page) > 0x00100000UL) {
+ glob->used_mem -= PAGE_SIZE;
+ } else {
+ glob->used_dma32_mem -= PAGE_SIZE;
+ }
+ spin_unlock(&glob->lock);
}
-
size_t ttm_round_pot(size_t size)
{
if ((size & (size - 1)) == 0)
diff --git a/drivers/gpu/drm/ttm/ttm_object.c b/drivers/gpu/drm/ttm/ttm_object.c
index 93577f2..c3b9e2c 100644
--- a/drivers/gpu/drm/ttm/ttm_object.c
+++ b/drivers/gpu/drm/ttm/ttm_object.c
@@ -250,6 +250,7 @@ int ttm_ref_object_add(struct ttm_object_file *tfile,
struct ttm_ref_object *ref;
struct drm_hash_item *hash;
struct ttm_mem_global *mem_glob = tfile->tdev->mem_glob;
+ unsigned long available;
int ret = -EINVAL;
if (existed != NULL)
@@ -267,10 +268,11 @@ int ttm_ref_object_add(struct ttm_object_file *tfile,
}
read_unlock(&tfile->lock);
- ret = ttm_mem_global_alloc(mem_glob, sizeof(*ref),
- false, false);
- if (unlikely(ret != 0))
- return ret;
+ available = ttm_mem_global_alloc(mem_glob, sizeof(*ref), false);
+ if (available != sizeof(*ref)) {
+ ttm_mem_global_free(mem_glob, available);
+ return -ENOMEM;
+ }
ref = kmalloc(sizeof(*ref), GFP_KERNEL);
if (unlikely(ref == NULL)) {
ttm_mem_global_free(mem_glob, sizeof(*ref));
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 3fb4c6d..d317bd4 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -67,24 +67,30 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
struct page *p;
struct list_head h;
struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
+ unsigned long available;
int ret;
if (NULL == (p = ttm->pages[index])) {
INIT_LIST_HEAD(&h);
+ available = ttm_mem_global_alloc(mem_glob, PAGE_SIZE, false);
+
ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1,
&ttm->dma_address[index]);
-
- if (ret != 0)
+ if (ret != 0) {
+ ttm_mem_global_free(mem_glob, available);
return NULL;
+ }
p = list_first_entry(&h, struct page, lru);
- ret = ttm_mem_global_alloc_page(mem_glob, p, false, false);
- if (unlikely(ret != 0))
+ if (available < PAGE_SIZE && page_to_pfn(p) < 0x00100000UL) {
+ ttm_mem_global_free(mem_glob, available);
goto out_err;
+ }
+ ttm_mem_global_alloc_page(mem_glob, p);
ttm->pages[index] = p;
}
return p;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 15fb260..f1d4b31 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -501,12 +501,14 @@ int vmw_fence_create(struct vmw_fence_manager *fman,
{
struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
struct vmw_fence_obj *fence;
+ unsigned long available;
int ret;
- ret = ttm_mem_global_alloc(mem_glob, fman->fence_size,
- false, false);
- if (unlikely(ret != 0))
- return ret;
+ available = ttm_mem_global_alloc(mem_glob, fman->fence_size, false);
+ if (available != fman->fence_size) {
+ ttm_mem_global_free(mem_glob, available);
+ return -ENOMEM;
+ }
fence = kzalloc(sizeof(*fence), GFP_KERNEL);
if (unlikely(fence == NULL)) {
@@ -566,6 +568,7 @@ int vmw_user_fence_create(struct drm_file *file_priv,
struct vmw_user_fence *ufence;
struct vmw_fence_obj *tmp;
struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
+ unsigned long available;
int ret;
/*
@@ -573,10 +576,11 @@ int vmw_user_fence_create(struct drm_file *file_priv,
* be created by a user-space request.
*/
- ret = ttm_mem_global_alloc(mem_glob, fman->user_fence_size,
- false, false);
- if (unlikely(ret != 0))
- return ret;
+ available = ttm_mem_global_alloc(mem_glob, fman->user_fence_size, false);
+ if (available != fman->user_fence_size) {
+ ttm_mem_global_free(mem_glob, available);
+ return -ENOMEM;
+ }
ufence = kzalloc(sizeof(*ufence), GFP_KERNEL);
if (unlikely(ufence == NULL)) {
@@ -959,16 +963,18 @@ int vmw_event_fence_action_create(struct drm_file *file_priv,
struct vmw_fence_manager *fman = fence->fman;
uint32_t size = fman->event_fence_action_size +
ttm_round_pot(event->length);
- int ret;
+ unsigned long available;
/*
* Account for internal structure size as well as the
* event size itself.
*/
- ret = ttm_mem_global_alloc(mem_glob, size, false, interruptible);
- if (unlikely(ret != 0))
- return ret;
+ available = ttm_mem_global_alloc(mem_glob, size, false);
+ if (available != size) {
+ ttm_mem_global_free(mem_glob, available);
+ return -ENOMEM;
+ }
eaction = kzalloc(sizeof(*eaction), GFP_KERNEL);
if (unlikely(eaction == NULL)) {
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 86c5e4c..9f2fa46d 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -413,6 +413,7 @@ int vmw_context_define_ioctl(struct drm_device *dev, void *data,
struct drm_vmw_context_arg *arg = (struct drm_vmw_context_arg *)data;
struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
struct vmw_master *vmaster = vmw_master(file_priv->master);
+ unsigned long available;
int ret;
@@ -428,13 +429,13 @@ int vmw_context_define_ioctl(struct drm_device *dev, void *data,
if (unlikely(ret != 0))
return ret;
- ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
- vmw_user_context_size,
- false, true);
- if (unlikely(ret != 0)) {
- if (ret != -ERESTARTSYS)
- DRM_ERROR("Out of graphics memory for context"
- " creation.\n");
+ available = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
+ vmw_user_context_size,
+ false);
+ if (available != vmw_user_context_size) {
+ ttm_mem_global_free(vmw_mem_glob(dev_priv), available);
+ ret = -ENOMEM;
+ DRM_ERROR("Out of graphics memory for context creation.\n");
goto out_unlock;
}
@@ -1275,6 +1276,7 @@ int vmw_surface_define_ioctl(struct drm_device *dev, void *data,
uint32_t num_sizes;
uint32_t size;
struct vmw_master *vmaster = vmw_master(file_priv->master);
+ unsigned long available;
if (unlikely(vmw_user_surface_size == 0))
vmw_user_surface_size = ttm_round_pot(sizeof(*user_srf)) +
@@ -1297,12 +1299,12 @@ int vmw_surface_define_ioctl(struct drm_device *dev, void *data,
if (unlikely(ret != 0))
return ret;
- ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
- size, false, true);
- if (unlikely(ret != 0)) {
- if (ret != -ERESTARTSYS)
- DRM_ERROR("Out of graphics memory for surface"
- " creation.\n");
+ available = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
+ size, false);
+ if (available != size) {
+ ttm_mem_global_free(vmw_mem_glob(dev_priv), available);
+ ret = -ENOMEM;
+ DRM_ERROR("Out of graphics memory for surface creation.\n");
goto out_unlock;
}
@@ -1552,6 +1554,7 @@ int vmw_dmabuf_init(struct vmw_private *dev_priv,
struct ttm_bo_device *bdev = &dev_priv->bdev;
struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
size_t acc_size;
+ unsigned long available;
int ret;
BUG_ON(!bo_free);
@@ -1560,12 +1563,13 @@ int vmw_dmabuf_init(struct vmw_private *dev_priv,
vmw_dmabuf_acc_size(bdev->glob,
(size + PAGE_SIZE - 1) >> PAGE_SHIFT);
- ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
- if (unlikely(ret != 0)) {
+ available = ttm_mem_global_alloc(mem_glob, acc_size, false);
+ if (available != acc_size) {
+ ttm_mem_global_free(vmw_mem_glob(dev_priv), available);
/* we must free the bo here as
* ttm_buffer_object_init does so as well */
bo_free(&vmw_bo->base);
- return ret;
+ return -ENOMEM;
}
memset(vmw_bo, 0, sizeof(*vmw_bo));
@@ -1835,6 +1839,7 @@ int vmw_stream_claim_ioctl(struct drm_device *dev, void *data,
struct drm_vmw_stream_arg *arg = (struct drm_vmw_stream_arg *)data;
struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
struct vmw_master *vmaster = vmw_master(file_priv->master);
+ unsigned long available;
int ret;
/*
@@ -1849,13 +1854,13 @@ int vmw_stream_claim_ioctl(struct drm_device *dev, void *data,
if (unlikely(ret != 0))
return ret;
- ret = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
- vmw_user_stream_size,
- false, true);
- if (unlikely(ret != 0)) {
- if (ret != -ERESTARTSYS)
- DRM_ERROR("Out of graphics memory for stream"
- " creation.\n");
+ available = ttm_mem_global_alloc(vmw_mem_glob(dev_priv),
+ vmw_user_stream_size,
+ false);
+ if (available != vmw_user_stream_size) {
+ ttm_mem_global_free(vmw_mem_glob(dev_priv), available);
+ ret = -ENOMEM;
+ DRM_ERROR("Out of graphics memory for stream creation.\n");
goto out_unlock;
}
diff --git a/include/drm/ttm/ttm_memory.h b/include/drm/ttm/ttm_memory.h
index 26c1f78..e001ff4 100644
--- a/include/drm/ttm/ttm_memory.h
+++ b/include/drm/ttm/ttm_memory.h
@@ -81,14 +81,14 @@ struct ttm_mem_global {
struct work_struct work;
wait_queue_head_t queue;
spinlock_t lock;
- struct ttm_mem_zone *zones[TTM_MEM_MAX_ZONES];
- unsigned int num_zones;
- struct ttm_mem_zone *zone_kernel;
-#ifdef CONFIG_HIGHMEM
- struct ttm_mem_zone *zone_highmem;
-#else
- struct ttm_mem_zone *zone_dma32;
-#endif
+ unsigned long mem;
+ unsigned long max_mem;
+ unsigned long max_dma32_mem;
+ unsigned long emer_mem;
+ unsigned long used_mem;
+ unsigned long used_dma32_mem;
+ unsigned long swap_limit;
+ unsigned long swap_dma32_limit;
};
/**
@@ -146,14 +146,15 @@ static inline void ttm_mem_unregister_shrink(struct ttm_mem_global *glob,
extern int ttm_mem_global_init(struct ttm_mem_global *glob);
extern void ttm_mem_global_release(struct ttm_mem_global *glob);
-extern int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
- bool no_wait, bool interruptible);
+extern unsigned long ttm_mem_global_alloc(struct ttm_mem_global *glob,
+ uint64_t memory,
+ bool no_wait);
extern void ttm_mem_global_free(struct ttm_mem_global *glob,
uint64_t amount);
-extern int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
- struct page *page,
- bool no_wait, bool interruptible);
+extern void ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
+ struct page *page);
extern void ttm_mem_global_free_page(struct ttm_mem_global *glob,
struct page *page);
extern size_t ttm_round_pot(size_t size);
+
#endif
--
1.7.7.1
[-- Attachment #3: Type: text/plain, Size: 159 bytes --]
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel
^ permalink raw reply related [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-10 18:05 ` Jerome Glisse
@ 2011-11-10 20:05 ` Thomas Hellstrom
2011-11-10 23:33 ` Jerome Glisse
0 siblings, 1 reply; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-10 20:05 UTC (permalink / raw)
To: Jerome Glisse; +Cc: Jerome Glisse, dri-devel
On 11/10/2011 07:05 PM, Jerome Glisse wrote:
> On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
>
>> On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
>>
>>> From: Jerome Glisse<jglisse@redhat.com>
>>>
>>> This is an overhaul of the ttm memory accounting. This tries to keep
>>> the same global behavior while removing the whole zone concept. It
>>> keeps a distrinction for dma32 so that we make sure that ttm don't
>>> starve the dma32 zone.
>>>
>>> There is 3 threshold for memory allocation :
>>> - max_mem is the maximum memory the whole ttm infrastructure is
>>> going to allow allocation for (exception of system process see
>>> below)
>>> - emer_mem is the maximum memory allowed for system process, this
>>> limit is> to max_mem
>>> - swap_limit is the threshold at which point ttm will start to
>>> try to swap object because ttm is getting close the max_mem
>>> limit
>>> - swap_dma32_limit is the threshold at which point ttm will start
>>> swap object to try to reduce the pressure on the dma32 zone. Note
>>> that we don't specificly target object to swap to it might very
>>> well free more memory from highmem rather than from dma32
>>>
>>> Accounting is done through used_mem& used_dma32_mem, which sum give
>>> the total amount of memory actually accounted by ttm.
>>>
>>> Idea is that allocation will fail if (used_mem + used_dma32_mem)>
>>> max_mem and if swapping fail to make enough room.
>>>
>>> The used_dma32_mem can be updated as a later stage, allowing to
>>> perform accounting test before allocating a whole batch of pages.
>>>
>>>
>> Jerome, you're removing a fair amount of functionality here, without
>> justifying
>> why it could be removed.
>>
> All this code was overkill.
>
[1] I don't agree, and since it's well tested, thought throught and
working, I see no obvious reason to alter it,
within the context of this patch series unless it's absolutely required
for the functionality.
>
>
>> Consider a low-end system with 1G of kernel memory and 10G of
>> highmem. How do we avoid putting stress on the kernel memory? I also
>> wouldn't be too surprised if DMA32 zones appear in HIGHMEM systems
>> in the future making the current zone concept good to keep.
>>
> Right now kernel memory is accounted against all zone so it decrease
> not only the kernel zone but also the dma32& highmem if present.
>
Do you mean that the code is incorrect? In that case, did you consider
the fact
that zones may overlap? (Although I admit the name "highmem" might be
misleading. Should be "total").
> Note also that kernel zone in current code == dma32 zone.
>
Last time I looked, the highmem split was typically at slightly less
than 1GB, depending on the hardware and desired setup. I admit that was
some time ago, but has that really changed? On all archs?
Furthermore, on !Highmem systems, All pages are in the kernel zone.
> When it comes to future it looks a lot simpler, it seems everyone
> is moving toward more capable and more advanced iommu that can remove
> all the restriction on memory from the device pov. I mean even arm
> are getting more and more advance iommu. I don't see any architecture
> worse supporting not going down that road.
>
While the proposed change is probably possible, with different low <->
high splits depending on whether HIGHMEM is defined or not, I think [1]
is a good reason for not pushing it through.
>
>> Also, in effect you move the DOS from *all* zones into the DMA32
>> zone and create a race in that multiple simultaneous allocators can
>> first pre-allocate out of the global zone, and then update the DMA32
>> zone without synchronization. In this way you might theoretically
>> end up with more DMA32 pages allocated than present in the zone.
>>
> Ok a respin attached with a simple change, things will be
> accounted against dma32 zone and only when we get page we will
> decrease the dma32 zone usage that way no DOS on dma32.
>
> It also deals with the case where there is still lot of highmem
> but no more dma32.
>
So why not just do a ttm_mem_global_alloc() for the pages you want to
allocate,
and add a proper adjustment function if memory turns out to be either
HIGHMEM or !DMA32
>
>> With the proposed code there's also a theoretical problem in that a
>> potentially huge number of pages are unaccounted before they are
>> actually freed.
>>
> What you mean unaccounted ? The way it works is :
> r = global_memory_alloc(size)
> if (r) fail
> alloc pages
> update memory accounting according to what page where allocated
>
> So memory is always accounted before even being allocated (exception
> are the kernel object for vmwgfx& ttm_bo but we can move accounting
> there too if you want, those are small allocation and i didn't think
> it was worse changing that).
>
No, I mean the sequence
unaccount_page_array()
---Race--
free_page_array()
/Thomas
> Cheers,
> Jerome
>
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-10 20:05 ` Thomas Hellstrom
@ 2011-11-10 23:33 ` Jerome Glisse
2011-11-11 7:49 ` Thomas Hellstrom
0 siblings, 1 reply; 28+ messages in thread
From: Jerome Glisse @ 2011-11-10 23:33 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: Jerome Glisse, dri-devel
On Thu, Nov 10, 2011 at 09:05:22PM +0100, Thomas Hellstrom wrote:
> On 11/10/2011 07:05 PM, Jerome Glisse wrote:
> >On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
> >>On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> >>>From: Jerome Glisse<jglisse@redhat.com>
> >>>
> >>>This is an overhaul of the ttm memory accounting. This tries to keep
> >>>the same global behavior while removing the whole zone concept. It
> >>>keeps a distrinction for dma32 so that we make sure that ttm don't
> >>>starve the dma32 zone.
> >>>
> >>>There is 3 threshold for memory allocation :
> >>>- max_mem is the maximum memory the whole ttm infrastructure is
> >>> going to allow allocation for (exception of system process see
> >>> below)
> >>>- emer_mem is the maximum memory allowed for system process, this
> >>> limit is> to max_mem
> >>>- swap_limit is the threshold at which point ttm will start to
> >>> try to swap object because ttm is getting close the max_mem
> >>> limit
> >>>- swap_dma32_limit is the threshold at which point ttm will start
> >>> swap object to try to reduce the pressure on the dma32 zone. Note
> >>> that we don't specificly target object to swap to it might very
> >>> well free more memory from highmem rather than from dma32
> >>>
> >>>Accounting is done through used_mem& used_dma32_mem, which sum give
> >>>the total amount of memory actually accounted by ttm.
> >>>
> >>>Idea is that allocation will fail if (used_mem + used_dma32_mem)>
> >>>max_mem and if swapping fail to make enough room.
> >>>
> >>>The used_dma32_mem can be updated as a later stage, allowing to
> >>>perform accounting test before allocating a whole batch of pages.
> >>>
> >>Jerome, you're removing a fair amount of functionality here, without
> >>justifying
> >>why it could be removed.
> >All this code was overkill.
>
> [1] I don't agree, and since it's well tested, thought throught and
> working, I see no obvious reason to alter it,
> within the context of this patch series unless it's absolutely
> required for the functionality.
Well one thing i can tell is that it doesn't work on radeon, i pushed
a test to libdrm and here it's the oom that starts doing its beating.
Anyway i won't alter it. Was just trying to make it works, ie be useful
while also being simpler.
> >>Consider a low-end system with 1G of kernel memory and 10G of
> >>highmem. How do we avoid putting stress on the kernel memory? I also
> >>wouldn't be too surprised if DMA32 zones appear in HIGHMEM systems
> >>in the future making the current zone concept good to keep.
> >Right now kernel memory is accounted against all zone so it decrease
> >not only the kernel zone but also the dma32& highmem if present.
>
> Do you mean that the code is incorrect? In that case, did you
> consider the fact
> that zones may overlap? (Although I admit the name "highmem" might
> be misleading. Should be "total").
Yeah i am well aware that zone overlap :)
> >Note also that kernel zone in current code == dma32 zone.
>
> Last time I looked, the highmem split was typically at slightly less
> than 1GB, depending on the hardware and desired setup. I admit that
> was some time ago, but has that really changed? On all archs?
> Furthermore, on !Highmem systems, All pages are in the kernel zone.
I was bit too much focused on my system where 1G of ram is wonderland
and 512M is the average. Thx to AMD i got a system with 8G i should
use it more.
Cheers,
Jerome
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-10 23:33 ` Jerome Glisse
@ 2011-11-11 7:49 ` Thomas Hellstrom
2011-11-11 8:47 ` Dave Airlie
2011-11-11 15:47 ` Jerome Glisse
0 siblings, 2 replies; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-11 7:49 UTC (permalink / raw)
To: Jerome Glisse; +Cc: Jerome Glisse, dri-devel
[-- Attachment #1.1: Type: text/plain, Size: 3763 bytes --]
On 11/11/2011 12:33 AM, Jerome Glisse wrote:
> On Thu, Nov 10, 2011 at 09:05:22PM +0100, Thomas Hellstrom wrote:
>
>> On 11/10/2011 07:05 PM, Jerome Glisse wrote:
>>
>>> On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
>>>
>>>> On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
>>>>
>>>>> From: Jerome Glisse<jglisse@redhat.com>
>>>>>
>>>>> This is an overhaul of the ttm memory accounting. This tries to keep
>>>>> the same global behavior while removing the whole zone concept. It
>>>>> keeps a distrinction for dma32 so that we make sure that ttm don't
>>>>> starve the dma32 zone.
>>>>>
>>>>> There is 3 threshold for memory allocation :
>>>>> - max_mem is the maximum memory the whole ttm infrastructure is
>>>>> going to allow allocation for (exception of system process see
>>>>> below)
>>>>> - emer_mem is the maximum memory allowed for system process, this
>>>>> limit is> to max_mem
>>>>> - swap_limit is the threshold at which point ttm will start to
>>>>> try to swap object because ttm is getting close the max_mem
>>>>> limit
>>>>> - swap_dma32_limit is the threshold at which point ttm will start
>>>>> swap object to try to reduce the pressure on the dma32 zone. Note
>>>>> that we don't specificly target object to swap to it might very
>>>>> well free more memory from highmem rather than from dma32
>>>>>
>>>>> Accounting is done through used_mem& used_dma32_mem, which sum give
>>>>> the total amount of memory actually accounted by ttm.
>>>>>
>>>>> Idea is that allocation will fail if (used_mem + used_dma32_mem)>
>>>>> max_mem and if swapping fail to make enough room.
>>>>>
>>>>> The used_dma32_mem can be updated as a later stage, allowing to
>>>>> perform accounting test before allocating a whole batch of pages.
>>>>>
>>>>>
>>>> Jerome, you're removing a fair amount of functionality here, without
>>>> justifying
>>>> why it could be removed.
>>>>
>>> All this code was overkill.
>>>
>> [1] I don't agree, and since it's well tested, thought throught and
>> working, I see no obvious reason to alter it,
>> within the context of this patch series unless it's absolutely
>> required for the functionality.
>>
> Well one thing i can tell is that it doesn't work on radeon, i pushed
> a test to libdrm and here it's the oom that starts doing its beating.
> Anyway i won't alter it. Was just trying to make it works, ie be useful
> while also being simpler.
>
Well if it doesn't work it should of course be fixed.
I'm not against fixing it nor making it simpler, but I think that
requires a detailed understanding of what's going wrong and how it needs
to be fixed. Not as part of a patch series that really tries to
accomplish something else.
The current code was tested extensively with psb and unichrome.
One good test for drivers with bo-backed textures is to continously
create fairly large texture images. The end result should be the swap
space starting to fill up and once there is no more swap space, the OOM
killer should kill your app, and kmalloc failures should be avoided. It
should be tricky to get a failure from the global alloc system, but a
huge amount of small buffer objects or fence objects should probably do it.
Naturally, that requires that all persistent drm objects created from
user-space are registered with their correct sizes, or at least a really
good size approximation. That includes things like gem flinks, that
could otherwise easily be exploited to bring a system down, simply by
guessing a gem name and create flinks to that name in an infinite loop.
What are the symptoms of the failure you're seeing with Radeon? Any
suggestions on why it happens?
Thanks,
Thomas
[-- Attachment #1.2: Type: text/html, Size: 4446 bytes --]
[-- Attachment #2: Type: text/plain, Size: 159 bytes --]
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-11 7:49 ` Thomas Hellstrom
@ 2011-11-11 8:47 ` Dave Airlie
2011-11-11 9:08 ` Thomas Hellstrom
2011-11-11 15:47 ` Jerome Glisse
1 sibling, 1 reply; 28+ messages in thread
From: Dave Airlie @ 2011-11-11 8:47 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: Jerome Glisse, dri-devel
> What are the symptoms of the failure you're seeing with Radeon? Any
> suggestions on why it happens?
I suspect radeon swapping might have some issues, I really think we
should put the effort into a test suite like the intel-gpu-tools,
instead of moving code around with no idea what is better or worse.
(Jerome you can take this as a strong hint to work on some tests :-)
Dave.
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-11 8:47 ` Dave Airlie
@ 2011-11-11 9:08 ` Thomas Hellstrom
0 siblings, 0 replies; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-11 9:08 UTC (permalink / raw)
To: dri-devel
On 11/11/2011 09:47 AM, Dave Airlie wrote:
>> What are the symptoms of the failure you're seeing with Radeon? Any
>> suggestions on why it happens?
>>
> I suspect radeon swapping might have some issues, I really think we
> should put the effort into a test suite like the intel-gpu-tools,
> instead of moving code around with no idea what is better or worse.
> (Jerome you can take this as a strong hint to work on some tests :-)
>
> Dave.
>
I think many tests like this can be written using OpenGL (using some
assumptions about where drivers store things like texture images etc.
That could be useful for all drivers).
But for the specific radeon bo case, I had a quick look at the code:
1) Gem stuff is not accounted anywhere AFAICT. Gem idr memory? flink memory?
2) Radeon fences doesn't seem to be accounted anywhere.
3) Radeon bo accounting is completely incorrect. The accounting size
given to ttm_bo_init() should be the size of the kmalloc'ed object,
including size of page arrays and dma_addr arrays and gem stuff, but not
the pages themselves. kmalloced memory sizes should also be rounded up
to the nearest power of two to account for allocator granularity.
Before this is fixed, we shouldn't IMHO really blame TTM and try to
rewrite it.
/Thomas
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-11 7:49 ` Thomas Hellstrom
2011-11-11 8:47 ` Dave Airlie
@ 2011-11-11 15:47 ` Jerome Glisse
2011-11-11 16:22 ` Thomas Hellstrom
1 sibling, 1 reply; 28+ messages in thread
From: Jerome Glisse @ 2011-11-11 15:47 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: Jerome Glisse, dri-devel
On Fri, Nov 11, 2011 at 08:49:39AM +0100, Thomas Hellstrom wrote:
> On 11/11/2011 12:33 AM, Jerome Glisse wrote:
> >On Thu, Nov 10, 2011 at 09:05:22PM +0100, Thomas Hellstrom wrote:
> >>On 11/10/2011 07:05 PM, Jerome Glisse wrote:
> >>>On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
> >>>>On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> >>>>>From: Jerome Glisse<jglisse@redhat.com>
> >>>>>
> >>>>>This is an overhaul of the ttm memory accounting. This tries to keep
> >>>>>the same global behavior while removing the whole zone concept. It
> >>>>>keeps a distrinction for dma32 so that we make sure that ttm don't
> >>>>>starve the dma32 zone.
> >>>>>
> >>>>>There is 3 threshold for memory allocation :
> >>>>>- max_mem is the maximum memory the whole ttm infrastructure is
> >>>>> going to allow allocation for (exception of system process see
> >>>>> below)
> >>>>>- emer_mem is the maximum memory allowed for system process, this
> >>>>> limit is> to max_mem
> >>>>>- swap_limit is the threshold at which point ttm will start to
> >>>>> try to swap object because ttm is getting close the max_mem
> >>>>> limit
> >>>>>- swap_dma32_limit is the threshold at which point ttm will start
> >>>>> swap object to try to reduce the pressure on the dma32 zone. Note
> >>>>> that we don't specificly target object to swap to it might very
> >>>>> well free more memory from highmem rather than from dma32
> >>>>>
> >>>>>Accounting is done through used_mem& used_dma32_mem, which sum give
> >>>>>the total amount of memory actually accounted by ttm.
> >>>>>
> >>>>>Idea is that allocation will fail if (used_mem + used_dma32_mem)>
> >>>>>max_mem and if swapping fail to make enough room.
> >>>>>
> >>>>>The used_dma32_mem can be updated as a later stage, allowing to
> >>>>>perform accounting test before allocating a whole batch of pages.
> >>>>>
> >>>>Jerome, you're removing a fair amount of functionality here, without
> >>>>justifying
> >>>>why it could be removed.
> >>>All this code was overkill.
> >>[1] I don't agree, and since it's well tested, thought throught and
> >>working, I see no obvious reason to alter it,
> >>within the context of this patch series unless it's absolutely
> >>required for the functionality.
> >Well one thing i can tell is that it doesn't work on radeon, i pushed
> >a test to libdrm and here it's the oom that starts doing its beating.
> >Anyway i won't alter it. Was just trying to make it works, ie be useful
> >while also being simpler.
>
> Well if it doesn't work it should of course be fixed.
>
> I'm not against fixing it nor making it simpler, but I think that
> requires a detailed understanding of what's going wrong and how it
> needs to be fixed. Not as part of a patch series that really tries
> to accomplish something else.
>
> The current code was tested extensively with psb and unichrome.
> One good test for drivers with bo-backed textures is to continously
> create fairly large texture images. The end result should be the
> swap space starting to fill up and once there is no more swap space,
> the OOM killer should kill your app, and kmalloc failures should be
> avoided. It should be tricky to get a failure from the global alloc
> system, but a huge amount of small buffer objects or fence objects
> should probably do it.
>
> Naturally, that requires that all persistent drm objects created
> from user-space are registered with their correct sizes, or at least
> a really good size approximation. That includes things like gem
> flinks, that could otherwise easily be exploited to bring a system
> down, simply by guessing a gem name and create flinks to that name
> in an infinite loop.
>
> What are the symptoms of the failure you're seeing with Radeon? Any
> suggestions on why it happens?
>
> Thanks,
> Thomas
I pushed my test case to libdrm yesterday, i basicly alloc ttm object
of 1 page in a loop and expect it to fail. I modified the kernel to
account 2 page for the ttm_buffer_object struct size so that the kernel
area should be exhausted long before i run out of memory on a 8G
config. What happen is that the oom start killing everythings except
my app, even the kernel logger daemon got kill before my app ...
I think the ttm_memory accounting for kernel object is not the right
way. What we want here is trying to avoid to starve kernel memory.
Thing is we are uniq in the kernel afaik (i take a look at bunch
of others driver to check that in net, audio, ...). in that we allocate
a lot of kernel object on the behalf of userspace.
I believe a better solution would be either to put a limit on the number
of objects. For instance 16 millions objects (wether fence of bo struct).
Or start talking with kernel folks about a dev_alloc function that could
be use to allocate memory and account for this memory against a device.
Kernel would be able to decide for a policy for this. There is already
a netdev_alloc_page but it's not use in any meaningfull way afaict.
Then when it comes to memory of buffer object themself, it would be a
lot better if we could somehow make it accountable against user process.
Like hooking into memory group (memcg). This would also allow us to
directly reuse kernel infrastructure, things like watermark on when
to start either swapin or reclaiming memory or simply refusing to
allocate more memory.
Cheers,
Jerome
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-11 15:47 ` Jerome Glisse
@ 2011-11-11 16:22 ` Thomas Hellstrom
2011-11-11 17:15 ` Jerome Glisse
0 siblings, 1 reply; 28+ messages in thread
From: Thomas Hellstrom @ 2011-11-11 16:22 UTC (permalink / raw)
To: Jerome Glisse; +Cc: Thomas Hellstrom, Jerome Glisse, dri-devel
On 11/11/2011 04:47 PM, Jerome Glisse wrote:
> On Fri, Nov 11, 2011 at 08:49:39AM +0100, Thomas Hellstrom wrote:
>
>> On 11/11/2011 12:33 AM, Jerome Glisse wrote:
>>
>>> On Thu, Nov 10, 2011 at 09:05:22PM +0100, Thomas Hellstrom wrote:
>>>
>>>> On 11/10/2011 07:05 PM, Jerome Glisse wrote:
>>>>
>>>>> On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
>>>>>
>>>>>> On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
>>>>>>
>>>>>>> From: Jerome Glisse<jglisse@redhat.com>
>>>>>>>
>>>>>>> This is an overhaul of the ttm memory accounting. This tries to keep
>>>>>>> the same global behavior while removing the whole zone concept. It
>>>>>>> keeps a distrinction for dma32 so that we make sure that ttm don't
>>>>>>> starve the dma32 zone.
>>>>>>>
>>>>>>> There is 3 threshold for memory allocation :
>>>>>>> - max_mem is the maximum memory the whole ttm infrastructure is
>>>>>>> going to allow allocation for (exception of system process see
>>>>>>> below)
>>>>>>> - emer_mem is the maximum memory allowed for system process, this
>>>>>>> limit is> to max_mem
>>>>>>> - swap_limit is the threshold at which point ttm will start to
>>>>>>> try to swap object because ttm is getting close the max_mem
>>>>>>> limit
>>>>>>> - swap_dma32_limit is the threshold at which point ttm will start
>>>>>>> swap object to try to reduce the pressure on the dma32 zone. Note
>>>>>>> that we don't specificly target object to swap to it might very
>>>>>>> well free more memory from highmem rather than from dma32
>>>>>>>
>>>>>>> Accounting is done through used_mem& used_dma32_mem, which sum give
>>>>>>> the total amount of memory actually accounted by ttm.
>>>>>>>
>>>>>>> Idea is that allocation will fail if (used_mem + used_dma32_mem)>
>>>>>>> max_mem and if swapping fail to make enough room.
>>>>>>>
>>>>>>> The used_dma32_mem can be updated as a later stage, allowing to
>>>>>>> perform accounting test before allocating a whole batch of pages.
>>>>>>>
>>>>>>>
>>>>>> Jerome, you're removing a fair amount of functionality here, without
>>>>>> justifying
>>>>>> why it could be removed.
>>>>>>
>>>>> All this code was overkill.
>>>>>
>>>> [1] I don't agree, and since it's well tested, thought throught and
>>>> working, I see no obvious reason to alter it,
>>>> within the context of this patch series unless it's absolutely
>>>> required for the functionality.
>>>>
>>> Well one thing i can tell is that it doesn't work on radeon, i pushed
>>> a test to libdrm and here it's the oom that starts doing its beating.
>>> Anyway i won't alter it. Was just trying to make it works, ie be useful
>>> while also being simpler.
>>>
>> Well if it doesn't work it should of course be fixed.
>>
>> I'm not against fixing it nor making it simpler, but I think that
>> requires a detailed understanding of what's going wrong and how it
>> needs to be fixed. Not as part of a patch series that really tries
>> to accomplish something else.
>>
>> The current code was tested extensively with psb and unichrome.
>> One good test for drivers with bo-backed textures is to continously
>> create fairly large texture images. The end result should be the
>> swap space starting to fill up and once there is no more swap space,
>> the OOM killer should kill your app, and kmalloc failures should be
>> avoided. It should be tricky to get a failure from the global alloc
>> system, but a huge amount of small buffer objects or fence objects
>> should probably do it.
>>
>> Naturally, that requires that all persistent drm objects created
>> from user-space are registered with their correct sizes, or at least
>> a really good size approximation. That includes things like gem
>> flinks, that could otherwise easily be exploited to bring a system
>> down, simply by guessing a gem name and create flinks to that name
>> in an infinite loop.
>>
>> What are the symptoms of the failure you're seeing with Radeon? Any
>> suggestions on why it happens?
>>
>> Thanks,
>> Thomas
>>
> I pushed my test case to libdrm yesterday, i basicly alloc ttm object
> of 1 page in a loop and expect it to fail. I modified the kernel to
> account 2 page for the ttm_buffer_object struct size so that the kernel
> area should be exhausted long before i run out of memory on a 8G
> config. What happen is that the oom start killing everythings except
> my app, even the kernel logger daemon got kill before my app ...
>
> I think the ttm_memory accounting for kernel object is not the right
> way.
>
....
So, yet again, TTM gets incorrectly blamed when things are not
working as expected.
The TTM memory accounting is designed to avoid pinning too much memory
for graphics, so that it can't be
used by the rest of the system. It's working well doing exactly that.
However, it can't stop your app from wanting to store too much data. It
just shuffles that data to swap. If too many apps want to store too much
data, eventually the computer runs out of swap space and the OOM killer
kicks in, and
tries to guess what app to kill. That's not TTM's business. Nor is it
DRM's business.
The only time the TTM memory accounting system blocks an allocation is
if there is too much pinned memory allocated (kmalloc, vmalloc) that it
can't release to swap space. It protects against kmalloc failures, but
it makes
no attempt to stop your app from wanting to store too much data.
/Thomas
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: [PATCH 05/13] drm/ttm: overhaul memory accounting
2011-11-11 16:22 ` Thomas Hellstrom
@ 2011-11-11 17:15 ` Jerome Glisse
0 siblings, 0 replies; 28+ messages in thread
From: Jerome Glisse @ 2011-11-11 17:15 UTC (permalink / raw)
To: Thomas Hellstrom; +Cc: Jerome Glisse, dri-devel
On Fri, Nov 11, 2011 at 05:22:26PM +0100, Thomas Hellstrom wrote:
> On 11/11/2011 04:47 PM, Jerome Glisse wrote:
> >On Fri, Nov 11, 2011 at 08:49:39AM +0100, Thomas Hellstrom wrote:
> >>On 11/11/2011 12:33 AM, Jerome Glisse wrote:
> >>>On Thu, Nov 10, 2011 at 09:05:22PM +0100, Thomas Hellstrom wrote:
> >>>>On 11/10/2011 07:05 PM, Jerome Glisse wrote:
> >>>>>On Thu, Nov 10, 2011 at 11:27:33AM +0100, Thomas Hellstrom wrote:
> >>>>>>On 11/09/2011 09:22 PM, j.glisse@gmail.com wrote:
> >>>>>>>From: Jerome Glisse<jglisse@redhat.com>
> >>>>>>>
> >>>>>>>This is an overhaul of the ttm memory accounting. This tries to keep
> >>>>>>>the same global behavior while removing the whole zone concept. It
> >>>>>>>keeps a distrinction for dma32 so that we make sure that ttm don't
> >>>>>>>starve the dma32 zone.
> >>>>>>>
> >>>>>>>There is 3 threshold for memory allocation :
> >>>>>>>- max_mem is the maximum memory the whole ttm infrastructure is
> >>>>>>> going to allow allocation for (exception of system process see
> >>>>>>> below)
> >>>>>>>- emer_mem is the maximum memory allowed for system process, this
> >>>>>>> limit is> to max_mem
> >>>>>>>- swap_limit is the threshold at which point ttm will start to
> >>>>>>> try to swap object because ttm is getting close the max_mem
> >>>>>>> limit
> >>>>>>>- swap_dma32_limit is the threshold at which point ttm will start
> >>>>>>> swap object to try to reduce the pressure on the dma32 zone. Note
> >>>>>>> that we don't specificly target object to swap to it might very
> >>>>>>> well free more memory from highmem rather than from dma32
> >>>>>>>
> >>>>>>>Accounting is done through used_mem& used_dma32_mem, which sum give
> >>>>>>>the total amount of memory actually accounted by ttm.
> >>>>>>>
> >>>>>>>Idea is that allocation will fail if (used_mem + used_dma32_mem)>
> >>>>>>>max_mem and if swapping fail to make enough room.
> >>>>>>>
> >>>>>>>The used_dma32_mem can be updated as a later stage, allowing to
> >>>>>>>perform accounting test before allocating a whole batch of pages.
> >>>>>>>
> >>>>>>Jerome, you're removing a fair amount of functionality here, without
> >>>>>>justifying
> >>>>>>why it could be removed.
> >>>>>All this code was overkill.
> >>>>[1] I don't agree, and since it's well tested, thought throught and
> >>>>working, I see no obvious reason to alter it,
> >>>>within the context of this patch series unless it's absolutely
> >>>>required for the functionality.
> >>>Well one thing i can tell is that it doesn't work on radeon, i pushed
> >>>a test to libdrm and here it's the oom that starts doing its beating.
> >>>Anyway i won't alter it. Was just trying to make it works, ie be useful
> >>>while also being simpler.
> >>Well if it doesn't work it should of course be fixed.
> >>
> >>I'm not against fixing it nor making it simpler, but I think that
> >>requires a detailed understanding of what's going wrong and how it
> >>needs to be fixed. Not as part of a patch series that really tries
> >>to accomplish something else.
> >>
> >>The current code was tested extensively with psb and unichrome.
> >>One good test for drivers with bo-backed textures is to continously
> >>create fairly large texture images. The end result should be the
> >>swap space starting to fill up and once there is no more swap space,
> >>the OOM killer should kill your app, and kmalloc failures should be
> >>avoided. It should be tricky to get a failure from the global alloc
> >>system, but a huge amount of small buffer objects or fence objects
> >>should probably do it.
> >>
> >>Naturally, that requires that all persistent drm objects created
> >>from user-space are registered with their correct sizes, or at least
> >>a really good size approximation. That includes things like gem
> >>flinks, that could otherwise easily be exploited to bring a system
> >>down, simply by guessing a gem name and create flinks to that name
> >>in an infinite loop.
> >>
> >>What are the symptoms of the failure you're seeing with Radeon? Any
> >>suggestions on why it happens?
> >>
> >>Thanks,
> >>Thomas
> >I pushed my test case to libdrm yesterday, i basicly alloc ttm object
> >of 1 page in a loop and expect it to fail. I modified the kernel to
> >account 2 page for the ttm_buffer_object struct size so that the kernel
> >area should be exhausted long before i run out of memory on a 8G
> >config. What happen is that the oom start killing everythings except
> >my app, even the kernel logger daemon got kill before my app ...
> >
> >I think the ttm_memory accounting for kernel object is not the right
> >way.
> ....
>
> So, yet again, TTM gets incorrectly blamed when things are not
> working as expected.
>
> The TTM memory accounting is designed to avoid pinning too much
> memory for graphics, so that it can't be
> used by the rest of the system. It's working well doing exactly that.
>
> However, it can't stop your app from wanting to store too much data.
> It just shuffles that data to swap. If too many apps want to store
> too much data, eventually the computer runs out of swap space and
> the OOM killer kicks in, and
> tries to guess what app to kill. That's not TTM's business. Nor is
> it DRM's business.
>
> The only time the TTM memory accounting system blocks an allocation
> is if there is too much pinned memory allocated (kmalloc, vmalloc)
> that it can't release to swap space. It protects against kmalloc
> failures, but it makes
> no attempt to stop your app from wanting to store too much data.
>
My test only try to allocate kernel memory, so i have 8G of ram
i allocate 1 page for each object (highmem forced) and then for
the accounting i account 2 page per each kernel kmalloc of
bo struct. So ttm memory should stop allowing any more allocation
as soon as i allocated for 400M of kernel memory. Which should
happen with only 200M of highmem allocate so all well below the
limit.
I also setup 4G of swap so in the end the oom shouldn't kick in
before we exhauste our kernel memory.
I think here the issue is that the kernel memory pool got decrease
wrongly when swaping object as the ttm buffer struct is not swaped
and still there.
I am chasing a bug in swapout + dma page alloc first, i will get
back to why ttm memory accouting fails after that.
Cheers,
Jerome
^ permalink raw reply [flat|nested] 28+ messages in thread
* Re: ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator
2011-11-03 23:39 j.glisse
@ 2011-11-07 17:50 ` Konrad Rzeszutek Wilk
0 siblings, 0 replies; 28+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-11-07 17:50 UTC (permalink / raw)
To: j.glisse; +Cc: thellstrom, dri-devel
On Thu, Nov 03, 2011 at 07:39:04PM -0400, j.glisse@gmail.com wrote:
> Hi,
>
> So updated patchset, only patch 5 seen change since last set.
> Last 3 patch are from your patchset, modified on top of mine.
Yup, and I've tested it on my radeon box. Going to test it
on the nvidia and some of the 32-bit troublesome cards.
Will need a patch to enable the TTM DMA under the nouveau
driver too, but I can spin that up.
>
> Konrad so i added you dma pool allocator on top of that
> and added support for it to radeon. All in all it's slightly
> smaller than your patchset.
It is! Thank you.
>
> Biggest change is use of a list_head in ttm_tt to keep the
> dma_page list inside the ttm_tt object allowing faster and
> lot simpler deallocation of page.
Yup.
>
> I only briefly test this code, it seems ok so far. Did you
> tested booting kernel with swiotlb=force and with your patchset ?
> Because here it doesn't work. I still don't understand why
> swiotlb want to create a bounce page when the page supplied
> fit the constraint. Need to dig into kernel history to see if
> there is any good reasons for that.
I think you found out why.. the force, does force every page.
>
> Otherwise i believe this whole patchset make things cleaner
> and simpler for ttm.
<nods> So do I.
Let me run it throught some of the machines this week to get
that extra comfortable feeling.
>
> Cheers,
> Jerome Glisse
^ permalink raw reply [flat|nested] 28+ messages in thread
* ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator
@ 2011-11-03 23:39 j.glisse
2011-11-07 17:50 ` Konrad Rzeszutek Wilk
0 siblings, 1 reply; 28+ messages in thread
From: j.glisse @ 2011-11-03 23:39 UTC (permalink / raw)
To: dri-devel; +Cc: thellstrom
Hi,
So updated patchset, only patch 5 seen change since last set.
Last 3 patch are from your patchset, modified on top of mine.
Konrad so i added you dma pool allocator on top of that
and added support for it to radeon. All in all it's slightly
smaller than your patchset.
Biggest change is use of a list_head in ttm_tt to keep the
dma_page list inside the ttm_tt object allowing faster and
lot simpler deallocation of page.
I only briefly test this code, it seems ok so far. Did you
tested booting kernel with swiotlb=force and with your patchset ?
Because here it doesn't work. I still don't understand why
swiotlb want to create a bounce page when the page supplied
fit the constraint. Need to dig into kernel history to see if
there is any good reasons for that.
Otherwise i believe this whole patchset make things cleaner
and simpler for ttm.
Cheers,
Jerome Glisse
^ permalink raw reply [flat|nested] 28+ messages in thread
end of thread, other threads:[~2011-11-11 17:11 UTC | newest]
Thread overview: 28+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2011-11-09 20:22 ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator j.glisse
2011-11-09 20:22 ` [PATCH 01/13] drm/ttm: remove userspace backed ttm object support j.glisse
2011-11-09 20:22 ` [PATCH 02/13] drm/ttm: remove split btw highmen and lowmem page j.glisse
2011-11-09 20:22 ` [PATCH 03/13] drm/ttm: remove unused backend flags field j.glisse
2011-11-09 20:22 ` [PATCH 04/13] drm/ttm: use ttm put pages function to properly restore cache attribute j.glisse
2011-11-09 20:22 ` [PATCH 05/13] drm/ttm: overhaul memory accounting j.glisse
2011-11-10 10:27 ` Thomas Hellstrom
2011-11-10 18:05 ` Jerome Glisse
2011-11-10 20:05 ` Thomas Hellstrom
2011-11-10 23:33 ` Jerome Glisse
2011-11-11 7:49 ` Thomas Hellstrom
2011-11-11 8:47 ` Dave Airlie
2011-11-11 9:08 ` Thomas Hellstrom
2011-11-11 15:47 ` Jerome Glisse
2011-11-11 16:22 ` Thomas Hellstrom
2011-11-11 17:15 ` Jerome Glisse
2011-11-09 20:22 ` [PATCH 06/13] drm/ttm: convert page allocation to use page ptr array instead of list V4 j.glisse
2011-11-09 20:22 ` [PATCH 07/13] drm/ttm: test for dma_address array allocation failure j.glisse
2011-11-09 20:22 ` [PATCH 08/13] drm/ttm: merge ttm_backend and ttm_tt V2 j.glisse
2011-11-09 20:22 ` [PATCH 09/13] drm/ttm: introduce callback for ttm_tt populate & unpopulate V2 j.glisse
2011-11-09 20:22 ` [PATCH 10/13] ttm: Provide DMA aware TTM page pool code. V5 j.glisse
2011-11-09 20:22 ` [PATCH 11/13] swiotlb: Expose swiotlb_nr_tlb function to modules j.glisse
2011-11-09 20:22 ` [PATCH 12/13] drm/radeon/kms: Enable the TTM DMA pool if swiotlb is on V2 j.glisse
2011-11-09 20:22 ` [PATCH 13/13] drm/nouveau: enable the TTM DMA pool on 32-bit DMA only device V2 j.glisse
2011-11-09 20:25 ` ttm: merge ttm_backend & ttm_tt, introduce ttm dma allocator Thomas Hellstrom
2011-11-09 21:03 ` Jerome Glisse
-- strict thread matches above, loose matches on Subject: below --
2011-11-03 23:39 j.glisse
2011-11-07 17:50 ` Konrad Rzeszutek Wilk
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.