From 8340b84a4752830feff6b77f3c23707dbd69f757 Mon Sep 17 00:00:00 2001 From: Jerome Glisse Date: Wed, 23 Sep 2009 18:48:56 +1000 Subject: [PATCH] drm/ttm: add pool wc/uc page allocator On AGP system we might allocate/free routinely uncached or wc memory, changing page from cached (wb) to uc or wc is very expensive and involves a lot of flushing. To improve performance this allocator use a pool of uc,wc pages. Pools are linked lists of pages. Ordered so that first is the latest adition to the pool and last is the oldest page. Old pages are periodicaly freed from pools to keep memory use at minimum. Pools are protected with spinlocks to allow multiple threads to allocate pages simultanously. Expensive operations must not hold the spinlock to maximise performance for multiple threads. V2: Fix ttm_page_pool_fill_locked. - max_cpages should be used for kmalloc instead of pages_to_alloc. - Set pool->npages to zero after taking all pages from the pool. Based on Jerome Glisse's and Dave Airlie's pool allocator. Signed-off-by: Jerome Glisse Signed-off-by: Dave Airlie Signed-off-by: Pauli Nieminen --- drivers/gpu/drm/ttm/Makefile | 2 +- drivers/gpu/drm/ttm/ttm_memory.c | 5 +- drivers/gpu/drm/ttm/ttm_page_alloc.c | 518 ++++++++++++++++++++++++++++++++++ drivers/gpu/drm/ttm/ttm_page_alloc.h | 38 +++ drivers/gpu/drm/ttm/ttm_tt.c | 47 ++-- 5 files changed, 584 insertions(+), 26 deletions(-) create mode 100644 drivers/gpu/drm/ttm/ttm_page_alloc.c create mode 100644 drivers/gpu/drm/ttm/ttm_page_alloc.h diff --git a/drivers/gpu/drm/ttm/Makefile b/drivers/gpu/drm/ttm/Makefile index 1e138f5..4256e20 100644 --- a/drivers/gpu/drm/ttm/Makefile +++ b/drivers/gpu/drm/ttm/Makefile @@ -4,6 +4,6 @@ ccflags-y := -Iinclude/drm ttm-y := ttm_agp_backend.o ttm_memory.o ttm_tt.o ttm_bo.o \ ttm_bo_util.o ttm_bo_vm.o ttm_module.o ttm_global.o \ - ttm_object.o ttm_lock.o ttm_execbuf_util.o + ttm_object.o ttm_lock.o ttm_execbuf_util.o ttm_page_alloc.o 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 f5245c0..d587fbe 100644 --- a/drivers/gpu/drm/ttm/ttm_memory.c +++ b/drivers/gpu/drm/ttm/ttm_memory.c @@ -32,6 +32,7 @@ #include #include #include +#include "ttm_page_alloc.h" #define TTM_MEMORY_ALLOC_RETRIES 4 @@ -394,6 +395,7 @@ int ttm_mem_global_init(struct ttm_mem_global *glob) "Zone %7s: Available graphics memory: %llu kiB.\n", zone->name, (unsigned long long) zone->max_mem >> 10); } + ttm_page_alloc_init(); return 0; out_no_zone: ttm_mem_global_release(glob); @@ -413,9 +415,10 @@ void ttm_mem_global_release(struct ttm_mem_global *glob) zone = glob->zones[i]; kobject_del(&zone->kobj); kobject_put(&zone->kobj); - } + } kobject_del(&glob->kobj); kobject_put(&glob->kobj); + ttm_page_alloc_fini(); } EXPORT_SYMBOL(ttm_mem_global_release); diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c new file mode 100644 index 0000000..e269a10 --- /dev/null +++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c @@ -0,0 +1,518 @@ +/* + * Copyright (c) Red Hat Inc. + + * 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. + * + * Authors: Dave Airlie + * Jerome Glisse + * Pauli Nieminen + */ + +/* simple list based uncached page pool + * - Pool collects resently freed pages for reuse + * - Use page->lru to keep a free list + * - doesn't track currently in use pages + */ +#include +#include +#include +#include +#include +#include + +#include +#include "ttm/ttm_bo_driver.h" +#include "ttm_page_alloc.h" + +/* Number of 4k pages to add at once */ +#define NUM_PAGES_TO_ADD 64 +/* times are in msecs */ +#define TIME_TO_KEEP_PAGE_IN_POOL 15000 +#define PAGE_FREE_INTERVAL 1000 + +/** + * list is a FILO stack. + * All pages pushed into it must go to begin while end of list must hold the + * least recently used pages. When pages in end of the list are not recently + * use they are freed. + * + * All expensive operations must be done outside of pool lock to prevent + * contention on lock. + */ +struct page_pool { + struct list_head list; + int gfp_flags; + unsigned npages; + spinlock_t lock; +}; + +# define MAX_NUM_POOLS 4 + +struct pool_manager { + bool page_alloc_inited; + unsigned long time_to_wait; + + struct timer_list timer; + + union { + struct page_pool pools[MAX_NUM_POOLS]; + struct { + struct page_pool wc_pool; + struct page_pool uc_pool; + struct page_pool wc_pool_dma32; + struct page_pool uc_pool_dma32; + } ; + }; +}; + + +static struct pool_manager _manager; + +#ifdef CONFIG_X86 +/* TODO: add this to x86 like _uc, this version here is inefficient */ +static int set_pages_array_wc(struct page **pages, int addrinarray) +{ + int i; + + for (i = 0; i < addrinarray; i++) + set_memory_wc((unsigned long)page_address(pages[i]), 1); + return 0; +} +#else +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 + + +static void ttm_page_pool_init_locked(struct page_pool *pool) +{ + spin_lock_init(&pool->lock); + INIT_LIST_HEAD(&pool->list); +} + +static struct page_pool *ttm_get_pool(int flags, + enum ttm_caching_state cstate) +{ + int pool_index; + + if (cstate == tt_cached) + return NULL; + + if (cstate == tt_wc) + pool_index = 0x0; + else + pool_index = 0x1; + + if (flags & TTM_PAGE_FLAG_DMA32) + pool_index |= 0x2; + + return &_manager.pools[pool_index]; +} + +static void ttm_pages_put(struct page *pages[], unsigned npages) +{ + unsigned i; + set_pages_array_wb(pages, npages); + for (i = 0; i < npages; ++i) + __free_page(pages[i]); +} + +/** + * Free pages from pool. + * If limit_with_time is set to true pages that wil lbe have to have been + * in pool for time specified by pool_manager.time_to_wait. + * + * Pages are freed in NUM_PAES_TO_ADD chunks to prevent freeing from + * blocking others operations on pool. + * + * This is called from the timer handler. + **/ +static void ttm_page_pool_free(struct page_pool *pool, + const bool limit_with_time) +{ + struct page *page; + struct page *pages_to_free[NUM_PAGES_TO_ADD]; + unsigned freed_pages; + unsigned long now = jiffies + _manager.time_to_wait; + unsigned long irq_flags; + +restart: + freed_pages = 0; + spin_lock_irqsave(&pool->lock, irq_flags); + + { + list_for_each_entry_reverse(page, &pool->list, lru) { + + if (limit_with_time + && time_after(now, page->private)) + break; + + pages_to_free[freed_pages++] = page; + if (freed_pages >= NUM_PAGES_TO_ADD) { + /* Limit the maximum time that we hold the lock + * not to block more important threads */ + __list_del(page->lru.prev, &pool->list); + pool->npages -= freed_pages; + spin_unlock_irqrestore(&pool->lock, irq_flags); + + ttm_pages_put(pages_to_free, freed_pages); + goto restart; + } + } + + pool->npages -= freed_pages; + if (freed_pages) + __list_del(&page->lru, &pool->list); + + spin_unlock_irqrestore(&pool->lock, irq_flags); + + if (freed_pages) + ttm_pages_put(pages_to_free, freed_pages); + } +} + +/* periodicaly free unused pages from pool */ +static void ttm_page_pool_timer_handler(unsigned long man_addr) +{ + int i; + struct pool_manager *m = (struct pool_manager *)man_addr; + + for (i = 0; i < MAX_NUM_POOLS; ++i) + ttm_page_pool_free(&_manager.pools[i], true); + + mod_timer(&m->timer, jiffies + msecs_to_jiffies(PAGE_FREE_INTERVAL)); +} + +static void ttm_page_pool_timer_create(struct pool_manager *m) +{ + /* Create a timer that is fired every second to clean + * the free list */ + init_timer(&m->timer); + m->timer.expires = jiffies + msecs_to_jiffies(PAGE_FREE_INTERVAL); + m->timer.function = ttm_page_pool_timer_handler; + m->timer.data = (unsigned long)m; +} + +static void ttm_page_pool_timer_free(struct pool_manager *m) +{ + del_timer_sync(&m->timer); +} + +static void ttm_set_pages_caching(struct page **pages, + enum ttm_caching_state cstate, unsigned cpages) +{ + /* Set page caching */ + switch (cstate) { + case tt_uncached: + set_pages_array_uc(pages, cpages); + break; + case tt_wc: + set_pages_array_wc(pages, cpages); + break; + default: + break; + } +} +/** + * Fill new pages to array. + */ +static int ttm_pages_pool_fill_locked(struct page_pool *pool, int flags, + enum ttm_caching_state cstate, unsigned count, + unsigned long *irq_flags) +{ + unsigned pages_to_alloc = count - pool->npages; + struct page **pages; + struct page *page; + struct list_head h; + unsigned i, cpages; + unsigned max_cpages = min(pages_to_alloc, + (unsigned)(PAGE_SIZE/sizeof(*page))); + + /* reserve allready allocated pages from pool */ + INIT_LIST_HEAD(&h); + list_splice_init(&pool->list, &h); + i = pool->npages; + pages_to_alloc = count; + pool->npages = 0; + + /* We are now safe to allow others users to modify the pool so unlock + * it. Allocating pages is expensive operation. */ + spin_unlock_irqrestore(&pool->lock, *irq_flags); + + pages = kmalloc(max_cpages*sizeof(*page), GFP_KERNEL); + + if (!pages) { + printk(KERN_ERR "[ttm] unable to allocate table for new pages."); + spin_lock_irqsave(&pool->lock, *irq_flags); + return -ENOMEM; + } + + /* i is set to number of pages already in pool */ + for (cpages = 0; i < pages_to_alloc; ++i) { + page = alloc_page(pool->gfp_flags); + if (!page) { + printk(KERN_ERR "[ttm] unable to get page %u\n", i); + + ttm_put_pages(&h, flags, cstate); + spin_lock_irqsave(&pool->lock, *irq_flags); + return -ENOMEM; + } + +#ifdef CONFIG_HIGHMEM + /* gfp flags of highmem page should never be dma32 so we + * we should be fine in such case + */ + if (!PageHighMem(page)) +#endif + { + pages[cpages++] = page; + if (cpages == max_cpages) { + + ttm_set_pages_caching(pages, cstate, cpages); + cpages = 0; + } + } + list_add(&page->lru, &h); + } + + if (cpages) + ttm_set_pages_caching(pages, cstate, cpages); + + spin_lock_irqsave(&pool->lock, *irq_flags); + + list_splice(&h, &pool->list); + + pool->npages += pages_to_alloc; + + return 0; +} + +/* + * 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) +{ + struct page_pool *pool = ttm_get_pool(flags, cstate); + struct page *page = NULL; + unsigned long irq_flags; + int gfp_flags = 0; + int retries = 1; + int r; + + /* No pool for cached pages */ + if (cstate == tt_cached) { + if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) + gfp_flags |= __GFP_ZERO; + + if (flags & TTM_PAGE_FLAG_DMA32) + gfp_flags |= GFP_DMA32; + else + gfp_flags |= __GFP_HIGHMEM; + + /* fallback in case of problems with pool allocation */ +alloc_pages: + for (r = 0; r < count; ++r) { + page = alloc_page(gfp_flags); + if (!page) { + + printk(KERN_ERR "[ttm] unable to allocate page."); + return -ENOMEM; + } + + list_add(&page->lru, pages); + } + return 0; + } + + gfp_flags = pool->gfp_flags; + + spin_lock_irqsave(&pool->lock, irq_flags); + do { + /* First we look if pool has the page we want */ + if (pool->npages >= count) { + + + /* Find the cut location */ + if (count <= pool->npages/2) { + unsigned c = 0; + list_for_each_entry(page, &pool->list, lru) { + if (++c >= count) + break; + } + } else { + /* We have to point to the last pages cut */ + unsigned c = 0; + unsigned rcount = pool->npages - count + 1; + + list_for_each_entry_reverse(page, &pool->list, lru) { + if (++c >= rcount) + break; + } + } + + list_cut_position(pages, &pool->list, &page->lru); + + pool->npages -= count; + spin_unlock_irqrestore(&pool->lock, irq_flags); + + if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) { + list_for_each_entry(page, pages, lru) { + clear_page(page_address(page)); + } + } + + return 0; + } + + if (retries == 0) + break; + + --retries; + + r = ttm_pages_pool_fill_locked(pool, flags, cstate, + count, &irq_flags); + if (r) { + spin_unlock_irqrestore(&pool->lock, irq_flags); + return r; + } + + } while (1); + + spin_unlock_irqrestore(&pool->lock, irq_flags); + + /* Pool allocator failed without errors. we try to allocate new pages + * instead. */ + if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) + gfp_flags |= __GFP_ZERO; + + goto alloc_pages; +} + +/* Put all pages in pages list to correct pool to wait for reuse */ +void ttm_put_pages(struct list_head *pages, int flags, + enum ttm_caching_state cstate) +{ + struct page_pool *pool = ttm_get_pool(flags, cstate); + struct page *p, *tmp; + unsigned page_count = 0; + unsigned long irq_flags; + unsigned long now = jiffies; + + if (pool == NULL) { + list_for_each_entry_safe(p, tmp, pages, lru) { + __free_page(p); + } + /* We took the ownership -> clear the pages list*/ + INIT_LIST_HEAD(pages); + return; + } + + list_for_each_entry_safe(p, tmp, pages, lru) { + +#ifdef CONFIG_HIGHMEM + /* we don't have pool for highmem -> free them */ + if (PageHighMem(p)) { + list_del(&p->lru); + __free_page(p); + } else +#endif + { + p->private = now; + ++page_count; + } + + } + + spin_lock_irqsave(&pool->lock, irq_flags); + list_splice(pages, &pool->list); + pool->npages += page_count; + spin_unlock_irqrestore(&pool->lock, irq_flags); + + /* We took the ownership -> clear the pages list*/ + INIT_LIST_HEAD(pages); +} + + +int ttm_page_alloc_init(void) +{ + if (_manager.page_alloc_inited) + return 0; + + ttm_page_pool_init_locked(&_manager.wc_pool); + _manager.wc_pool.gfp_flags = GFP_HIGHUSER; + ttm_page_pool_init_locked(&_manager.uc_pool); + _manager.uc_pool.gfp_flags = GFP_HIGHUSER; + ttm_page_pool_init_locked(&_manager.wc_pool_dma32); + _manager.wc_pool_dma32.gfp_flags = GFP_USER | GFP_DMA32; + ttm_page_pool_init_locked(&_manager.uc_pool_dma32); + _manager.uc_pool_dma32.gfp_flags = GFP_USER | GFP_DMA32; + + _manager.time_to_wait = msecs_to_jiffies(TIME_TO_KEEP_PAGE_IN_POOL); + ttm_page_pool_timer_create(&_manager); + _manager.page_alloc_inited = 1; + return 0; +} + +void ttm_page_alloc_fini(void) +{ + int i; + + if (!_manager.page_alloc_inited) + return; + + ttm_page_pool_timer_free(&_manager); + + for (i = 0; i < MAX_NUM_POOLS; ++i) + ttm_page_pool_free(&_manager.pools[i], false); + + _manager.page_alloc_inited = 0; +} diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.h b/drivers/gpu/drm/ttm/ttm_page_alloc.h new file mode 100644 index 0000000..eb0ddf4 --- /dev/null +++ b/drivers/gpu/drm/ttm/ttm_page_alloc.h @@ -0,0 +1,38 @@ +/* + * Copyright (c) Red Hat Inc. + + * 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. + * + * Authors: Dave Airlie + * Jerome Glisse + */ +#ifndef TTM_PAGE_ALLOC +#define TTM_PAGE_ALLOC + +#include "ttm/ttm_bo_driver.h" + +void ttm_put_pages(struct list_head *pages, int flags, + enum ttm_caching_state cstate); +int ttm_get_pages(struct list_head *pages, int flags, + enum ttm_caching_state cstate, unsigned count); +int ttm_page_alloc_init(void); +void ttm_page_alloc_fini(void); + +#endif diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c index 3d47a2c..d4e20f6 100644 --- a/drivers/gpu/drm/ttm/ttm_tt.c +++ b/drivers/gpu/drm/ttm/ttm_tt.c @@ -38,6 +38,7 @@ #include "ttm/ttm_module.h" #include "ttm/ttm_bo_driver.h" #include "ttm/ttm_placement.h" +#include "ttm_page_alloc.h" static int ttm_tt_swapin(struct ttm_tt *ttm); @@ -72,21 +73,6 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm) ttm->pages = NULL; } -static struct page *ttm_tt_alloc_page(unsigned page_flags) -{ - gfp_t gfp_flags = GFP_USER; - - if (page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) - gfp_flags |= __GFP_ZERO; - - if (page_flags & TTM_PAGE_FLAG_DMA32) - gfp_flags |= __GFP_DMA32; - else - gfp_flags |= __GFP_HIGHMEM; - - return alloc_page(gfp_flags); -} - static void ttm_tt_free_user_pages(struct ttm_tt *ttm) { int write; @@ -127,15 +113,21 @@ static void ttm_tt_free_user_pages(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; while (NULL == (p = ttm->pages[index])) { - p = ttm_tt_alloc_page(ttm->page_flags); - if (!p) + INIT_LIST_HEAD(&h); + + ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1); + + if (ret != 0) 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)) goto out_err; @@ -244,10 +236,10 @@ static int ttm_tt_set_caching(struct ttm_tt *ttm, if (ttm->caching_state == c_state) return 0; - if (c_state != tt_cached) { - ret = ttm_tt_populate(ttm); - if (unlikely(ret != 0)) - return ret; + if (ttm->state == tt_unpopulated) { + /* Change caching but don't populate */ + ttm->caching_state = c_state; + return 0; } if (ttm->caching_state == tt_cached) @@ -298,25 +290,32 @@ 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); + if (be) be->func->clear(be); - (void)ttm_tt_set_caching(ttm, tt_cached); 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"); + "Leaking pages (%d/%d).\n", + page_count(cur_page), count); ttm_mem_global_free_page(ttm->glob->mem_glob, cur_page); - __free_page(cur_page); + list_add(&cur_page->lru, &h); + count++; } } + ttm_put_pages(&h, ttm->page_flags, ttm->caching_state); ttm->state = tt_unpopulated; ttm->first_himem_page = ttm->num_pages; ttm->last_lomem_page = -1; -- 1.6.3.3