* [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) @ 2021-07-21 20:13 Jason Ekstrand 2021-07-21 20:13 ` [PATCH 1/7] drm/i915/gem: Check object_can_migrate from object_migrate Jason Ekstrand ` (6 more replies) 0 siblings, 7 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Jason Ekstrand This patch series fixes an issue with discrete graphics on Intel where we allowed dma-buf import while leaving the object in local memory. This breaks down pretty badly if the import happened on a different physical device. v7: - Drop "drm/i915/gem/ttm: Place new BOs in the requested region" - Add a new "drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create()" - Misc. review feedback from Matthew Auld v8: - Misc. review feedback from Matthew Auld Jason Ekstrand (5): drm/i915/gem: Check object_can_migrate from object_migrate drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() drm/i915/gem: Unify user object creation (v3) drm/i915/gem/ttm: Respect the objection region in placement_from_obj Thomas Hellström (2): drm/i915/gem: Correct the locking and pin pattern for dma-buf (v8) drm/i915/gem: Migrate to system at dma-buf attach time (v7) drivers/gpu/drm/i915/gem/i915_gem_create.c | 177 ++++++++-------- drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 58 ++++-- drivers/gpu/drm/i915/gem/i915_gem_object.c | 13 +- drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 + drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 190 +++++++++++++++++- .../drm/i915/gem/selftests/i915_gem_migrate.c | 15 -- 7 files changed, 330 insertions(+), 130 deletions(-) -- 2.31.1 ^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH 1/7] drm/i915/gem: Check object_can_migrate from object_migrate 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 2/7] drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) Jason Ekstrand ` (5 subsequent siblings) 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Jason Ekstrand, Matthew Auld We don't roll them together entirely because there are still a couple cases where we want a separate can_migrate check. For instance, the display code checks that you can migrate a buffer to LMEM before it accepts it in fb_create. The dma-buf import code also uses it to do an early check and return a different error code if someone tries to attach a LMEM-only dma-buf to another driver. However, no one actually wants to call object_migrate when can_migrate has failed. The stated intention is for self-tests but none of those actually take advantage of this unsafe migration. Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Cc: Daniel Vetter <daniel@ffwll.ch> Reviewed-by: Matthew Auld <matthew.auld@intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_object.c | 13 ++----------- .../gpu/drm/i915/gem/selftests/i915_gem_migrate.c | 15 --------------- 2 files changed, 2 insertions(+), 26 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 9da7b288b7ede..f2244ae09a613 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -584,12 +584,6 @@ bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj, * completed yet, and to accomplish that, i915_gem_object_wait_migration() * must be called. * - * This function is a bit more permissive than i915_gem_object_can_migrate() - * to allow for migrating objects where the caller knows exactly what is - * happening. For example within selftests. More specifically this - * function allows migrating I915_BO_ALLOC_USER objects to regions - * that are not in the list of allowable regions. - * * Note: the @ww parameter is not used yet, but included to make sure * callers put some effort into obtaining a valid ww ctx if one is * available. @@ -616,11 +610,8 @@ int i915_gem_object_migrate(struct drm_i915_gem_object *obj, if (obj->mm.region == mr) return 0; - if (!i915_gem_object_evictable(obj)) - return -EBUSY; - - if (!obj->ops->migrate) - return -EOPNOTSUPP; + if (!i915_gem_object_can_migrate(obj, id)) + return -EINVAL; return obj->ops->migrate(obj, mr); } diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c index 0b7144d2991ca..28a700f08b49a 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c @@ -61,11 +61,6 @@ static int igt_create_migrate(struct intel_gt *gt, enum intel_region_id src, if (err) continue; - if (!i915_gem_object_can_migrate(obj, dst)) { - err = -EINVAL; - continue; - } - err = i915_gem_object_migrate(obj, &ww, dst); if (err) continue; @@ -114,11 +109,6 @@ static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww, return err; if (i915_gem_object_is_lmem(obj)) { - if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) { - pr_err("object can't migrate to smem.\n"); - return -EINVAL; - } - err = i915_gem_object_migrate(obj, ww, INTEL_REGION_SMEM); if (err) { pr_err("Object failed migration to smem\n"); @@ -137,11 +127,6 @@ static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww, } } else { - if (!i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) { - pr_err("object can't migrate to lmem.\n"); - return -EINVAL; - } - err = i915_gem_object_migrate(obj, ww, INTEL_REGION_LMEM); if (err) { pr_err("Object failed migration to lmem\n"); -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 2/7] drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 1/7] drm/i915/gem: Check object_can_migrate from object_migrate Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 3/7] drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() Jason Ekstrand ` (4 subsequent siblings) 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Matthew Auld, Jason Ekstrand Since we don't allow changing the set of regions after creation, we can make ext_set_placements() build up the region set directly in the create_ext and assign it to the object later. This is similar to what we did for contexts with the proto-context only simpler because there's no funny object shuffling. This will be used in the next patch to allow us to de-duplicate a bunch of code. Also, since we know the maximum number of regions up-front, we can use a fixed-size temporary array for the regions. This simplifies memory management a bit for this new delayed approach. v2 (Matthew Auld): - Get rid of MAX_N_PLACEMENTS - Drop kfree(placements) from set_placements() v3 (Matthew Auld): - Properly set ext_data->n_placements Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matthew Auld <matthew.auld@intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_create.c | 82 ++++++++++++---------- 1 file changed, 46 insertions(+), 36 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c b/drivers/gpu/drm/i915/gem/i915_gem_create.c index 51f92e4b1a69d..aa687b10dcd45 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_create.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c @@ -27,10 +27,13 @@ static u32 object_max_page_size(struct drm_i915_gem_object *obj) return max_page_size; } -static void object_set_placements(struct drm_i915_gem_object *obj, - struct intel_memory_region **placements, - unsigned int n_placements) +static int object_set_placements(struct drm_i915_gem_object *obj, + struct intel_memory_region **placements, + unsigned int n_placements) { + struct intel_memory_region **arr; + unsigned int i; + GEM_BUG_ON(!n_placements); /* @@ -44,9 +47,20 @@ static void object_set_placements(struct drm_i915_gem_object *obj, obj->mm.placements = &i915->mm.regions[mr->id]; obj->mm.n_placements = 1; } else { - obj->mm.placements = placements; + arr = kmalloc_array(n_placements, + sizeof(struct intel_memory_region *), + GFP_KERNEL); + if (!arr) + return -ENOMEM; + + for (i = 0; i < n_placements; i++) + arr[i] = placements[i]; + + obj->mm.placements = arr; obj->mm.n_placements = n_placements; } + + return 0; } static int i915_gem_publish(struct drm_i915_gem_object *obj, @@ -148,7 +162,9 @@ i915_gem_dumb_create(struct drm_file *file, return -ENOMEM; mr = intel_memory_region_by_type(to_i915(dev), mem_type); - object_set_placements(obj, &mr, 1); + ret = object_set_placements(obj, &mr, 1); + if (ret) + goto object_free; ret = i915_gem_setup(obj, args->size); if (ret) @@ -184,7 +200,9 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, return -ENOMEM; mr = intel_memory_region_by_type(i915, INTEL_MEMORY_SYSTEM); - object_set_placements(obj, &mr, 1); + ret = object_set_placements(obj, &mr, 1); + if (ret) + goto object_free; ret = i915_gem_setup(obj, args->size); if (ret) @@ -199,7 +217,8 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, struct create_ext { struct drm_i915_private *i915; - struct drm_i915_gem_object *vanilla_object; + struct intel_memory_region *placements[INTEL_REGION_UNKNOWN]; + unsigned int n_placements; }; static void repr_placements(char *buf, size_t size, @@ -230,8 +249,7 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, struct drm_i915_private *i915 = ext_data->i915; struct drm_i915_gem_memory_class_instance __user *uregions = u64_to_user_ptr(args->regions); - struct drm_i915_gem_object *obj = ext_data->vanilla_object; - struct intel_memory_region **placements; + struct intel_memory_region *placements[INTEL_REGION_UNKNOWN]; u32 mask; int i, ret = 0; @@ -245,6 +263,8 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, ret = -EINVAL; } + BUILD_BUG_ON(ARRAY_SIZE(i915->mm.regions) != ARRAY_SIZE(placements)); + BUILD_BUG_ON(ARRAY_SIZE(ext_data->placements) != ARRAY_SIZE(placements)); if (args->num_regions > ARRAY_SIZE(i915->mm.regions)) { drm_dbg(&i915->drm, "num_regions is too large\n"); ret = -EINVAL; @@ -253,21 +273,13 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, if (ret) return ret; - placements = kmalloc_array(args->num_regions, - sizeof(struct intel_memory_region *), - GFP_KERNEL); - if (!placements) - return -ENOMEM; - mask = 0; for (i = 0; i < args->num_regions; i++) { struct drm_i915_gem_memory_class_instance region; struct intel_memory_region *mr; - if (copy_from_user(®ion, uregions, sizeof(region))) { - ret = -EFAULT; - goto out_free; - } + if (copy_from_user(®ion, uregions, sizeof(region))) + return -EFAULT; mr = intel_memory_region_lookup(i915, region.memory_class, @@ -293,14 +305,14 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, ++uregions; } - if (obj->mm.placements) { + if (ext_data->n_placements) { ret = -EINVAL; goto out_dump; } - object_set_placements(obj, placements, args->num_regions); - if (args->num_regions == 1) - kfree(placements); + ext_data->n_placements = args->num_regions; + for (i = 0; i < args->num_regions; i++) + ext_data->placements[i] = placements[i]; return 0; @@ -308,11 +320,11 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, if (1) { char buf[256]; - if (obj->mm.placements) { + if (ext_data->n_placements) { repr_placements(buf, sizeof(buf), - obj->mm.placements, - obj->mm.n_placements); + ext_data->placements, + ext_data->n_placements); drm_dbg(&i915->drm, "Placements were already set in previous EXT. Existing placements: %s\n", buf); @@ -322,8 +334,6 @@ static int set_placements(struct drm_i915_gem_create_ext_memory_regions *args, drm_dbg(&i915->drm, "New placements(so far validated): %s\n", buf); } -out_free: - kfree(placements); return ret; } @@ -358,7 +368,6 @@ i915_gem_create_ext_ioctl(struct drm_device *dev, void *data, struct drm_i915_private *i915 = to_i915(dev); struct drm_i915_gem_create_ext *args = data; struct create_ext ext_data = { .i915 = i915 }; - struct intel_memory_region **placements_ext; struct drm_i915_gem_object *obj; int ret; @@ -371,21 +380,22 @@ i915_gem_create_ext_ioctl(struct drm_device *dev, void *data, if (!obj) return -ENOMEM; - ext_data.vanilla_object = obj; ret = i915_user_extensions(u64_to_user_ptr(args->extensions), create_extensions, ARRAY_SIZE(create_extensions), &ext_data); - placements_ext = obj->mm.placements; if (ret) goto object_free; - if (!placements_ext) { - struct intel_memory_region *mr = + if (!ext_data.n_placements) { + ext_data.placements[0] = intel_memory_region_by_type(i915, INTEL_MEMORY_SYSTEM); - - object_set_placements(obj, &mr, 1); + ext_data.n_placements = 1; } + ret = object_set_placements(obj, ext_data.placements, + ext_data.n_placements); + if (ret) + goto object_free; ret = i915_gem_setup(obj, args->size); if (ret) @@ -395,7 +405,7 @@ i915_gem_create_ext_ioctl(struct drm_device *dev, void *data, object_free: if (obj->mm.n_placements > 1) - kfree(placements_ext); + kfree(obj->mm.placements); i915_gem_object_free(obj); return ret; } -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 3/7] drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 1/7] drm/i915/gem: Check object_can_migrate from object_migrate Jason Ekstrand 2021-07-21 20:13 ` [PATCH 2/7] drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 4/7] drm/i915/gem: Unify user object creation (v3) Jason Ekstrand ` (3 subsequent siblings) 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Matthew Auld, Jason Ekstrand This doesn't really fix anything serious since the chances of a client creating and destroying a mass of dumb BOs is pretty low. However, it is called by the other two create IOCTLs to garbage collect old objects. Call it here too for consistency. Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matthew Auld <matthew.auld@intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_create.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c b/drivers/gpu/drm/i915/gem/i915_gem_create.c index aa687b10dcd45..adcce37c04b8d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_create.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c @@ -151,6 +151,8 @@ i915_gem_dumb_create(struct drm_file *file, if (args->pitch < args->width) return -EINVAL; + i915_gem_flush_free_objects(i915); + args->size = mul_u32_u32(args->pitch, args->height); mem_type = INTEL_MEMORY_SYSTEM; -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 4/7] drm/i915/gem: Unify user object creation (v3) 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand ` (2 preceding siblings ...) 2021-07-21 20:13 ` [PATCH 3/7] drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand ` (2 subsequent siblings) 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Matthew Auld, Jason Ekstrand Instead of hand-rolling the same three calls in each function, pull them into an i915_gem_object_create_user helper. Apart from re-ordering of the placements array ENOMEM check, there should be no functional change. v2 (Matthew Auld): - Add the call to i915_gem_flush_free_objects() from i915_gem_dumb_create() in a separate patch - Move i915_gem_object_alloc() below the simple error checks v3 (Matthew Auld): - Add __ to i915_gem_object_create_user and kerneldoc which warns the caller that it's not validating anything. Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Matthew Auld <matthew.auld@intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_create.c | 119 ++++++++++----------- drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 + 2 files changed, 58 insertions(+), 65 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c b/drivers/gpu/drm/i915/gem/i915_gem_create.c index adcce37c04b8d..23fee13a33844 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_create.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c @@ -11,13 +11,14 @@ #include "i915_trace.h" #include "i915_user_extensions.h" -static u32 object_max_page_size(struct drm_i915_gem_object *obj) +static u32 object_max_page_size(struct intel_memory_region **placements, + unsigned int n_placements) { u32 max_page_size = 0; int i; - for (i = 0; i < obj->mm.n_placements; i++) { - struct intel_memory_region *mr = obj->mm.placements[i]; + for (i = 0; i < n_placements; i++) { + struct intel_memory_region *mr = placements[i]; GEM_BUG_ON(!is_power_of_2(mr->min_page_size)); max_page_size = max_t(u32, max_page_size, mr->min_page_size); @@ -81,22 +82,46 @@ static int i915_gem_publish(struct drm_i915_gem_object *obj, return 0; } -static int -i915_gem_setup(struct drm_i915_gem_object *obj, u64 size) +/** + * Creates a new object using the same path as DRM_I915_GEM_CREATE_EXT + * @i915: i915 private + * @size: size of the buffer, in bytes + * @placements: possible placement regions, in priority order + * @n_placements: number of possible placement regions + * + * This function is exposed primarily for selftests and does very little + * error checking. It is assumed that the set of placement regions has + * already been verified to be valid. + */ +struct drm_i915_gem_object * +__i915_gem_object_create_user(struct drm_i915_private *i915, u64 size, + struct intel_memory_region **placements, + unsigned int n_placements) { - struct intel_memory_region *mr = obj->mm.placements[0]; + struct intel_memory_region *mr = placements[0]; + struct drm_i915_gem_object *obj; unsigned int flags; int ret; - size = round_up(size, object_max_page_size(obj)); + i915_gem_flush_free_objects(i915); + + size = round_up(size, object_max_page_size(placements, n_placements)); if (size == 0) - return -EINVAL; + return ERR_PTR(-EINVAL); /* For most of the ABI (e.g. mmap) we think in system pages */ GEM_BUG_ON(!IS_ALIGNED(size, PAGE_SIZE)); if (i915_gem_object_size_2big(size)) - return -E2BIG; + return ERR_PTR(-E2BIG); + + obj = i915_gem_object_alloc(); + if (!obj) + return ERR_PTR(-ENOMEM); + + ret = object_set_placements(obj, placements, n_placements); + if (ret) + goto object_free; /* * I915_BO_ALLOC_USER will make sure the object is cleared before @@ -106,12 +131,18 @@ i915_gem_setup(struct drm_i915_gem_object *obj, u64 size) ret = mr->ops->init_object(mr, obj, size, 0, flags); if (ret) - return ret; + goto object_free; GEM_BUG_ON(size != obj->base.size); trace_i915_gem_object_create(obj); - return 0; + return obj; + +object_free: + if (obj->mm.n_placements > 1) + kfree(obj->mm.placements); + i915_gem_object_free(obj); + return ERR_PTR(ret); } int @@ -124,7 +155,6 @@ i915_gem_dumb_create(struct drm_file *file, enum intel_memory_type mem_type; int cpp = DIV_ROUND_UP(args->bpp, 8); u32 format; - int ret; switch (cpp) { case 1: @@ -151,32 +181,19 @@ i915_gem_dumb_create(struct drm_file *file, if (args->pitch < args->width) return -EINVAL; - i915_gem_flush_free_objects(i915); - args->size = mul_u32_u32(args->pitch, args->height); mem_type = INTEL_MEMORY_SYSTEM; if (HAS_LMEM(to_i915(dev))) mem_type = INTEL_MEMORY_LOCAL; - obj = i915_gem_object_alloc(); - if (!obj) - return -ENOMEM; - mr = intel_memory_region_by_type(to_i915(dev), mem_type); - ret = object_set_placements(obj, &mr, 1); - if (ret) - goto object_free; - ret = i915_gem_setup(obj, args->size); - if (ret) - goto object_free; + obj = __i915_gem_object_create_user(to_i915(dev), args->size, &mr, 1); + if (IS_ERR(obj)) + return PTR_ERR(obj); return i915_gem_publish(obj, file, &args->size, &args->handle); - -object_free: - i915_gem_object_free(obj); - return ret; } /** @@ -193,28 +210,14 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, struct drm_i915_gem_create *args = data; struct drm_i915_gem_object *obj; struct intel_memory_region *mr; - int ret; - - i915_gem_flush_free_objects(i915); - - obj = i915_gem_object_alloc(); - if (!obj) - return -ENOMEM; mr = intel_memory_region_by_type(i915, INTEL_MEMORY_SYSTEM); - ret = object_set_placements(obj, &mr, 1); - if (ret) - goto object_free; - ret = i915_gem_setup(obj, args->size); - if (ret) - goto object_free; + obj = __i915_gem_object_create_user(i915, args->size, &mr, 1); + if (IS_ERR(obj)) + return PTR_ERR(obj); return i915_gem_publish(obj, file, &args->size, &args->handle); - -object_free: - i915_gem_object_free(obj); - return ret; } struct create_ext { @@ -376,38 +379,24 @@ i915_gem_create_ext_ioctl(struct drm_device *dev, void *data, if (args->flags) return -EINVAL; - i915_gem_flush_free_objects(i915); - - obj = i915_gem_object_alloc(); - if (!obj) - return -ENOMEM; - ret = i915_user_extensions(u64_to_user_ptr(args->extensions), create_extensions, ARRAY_SIZE(create_extensions), &ext_data); if (ret) - goto object_free; + return ret; if (!ext_data.n_placements) { ext_data.placements[0] = intel_memory_region_by_type(i915, INTEL_MEMORY_SYSTEM); ext_data.n_placements = 1; } - ret = object_set_placements(obj, ext_data.placements, - ext_data.n_placements); - if (ret) - goto object_free; - ret = i915_gem_setup(obj, args->size); - if (ret) - goto object_free; + obj = __i915_gem_object_create_user(i915, args->size, + ext_data.placements, + ext_data.n_placements); + if (IS_ERR(obj)) + return PTR_ERR(obj); return i915_gem_publish(obj, file, &args->size, &args->handle); - -object_free: - if (obj->mm.n_placements > 1) - kfree(obj->mm.placements); - i915_gem_object_free(obj); - return ret; } diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index f3ede43282dc6..0896ac532f5e5 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -61,6 +61,10 @@ i915_gem_object_create_shmem(struct drm_i915_private *i915, struct drm_i915_gem_object * i915_gem_object_create_shmem_from_data(struct drm_i915_private *i915, const void *data, resource_size_t size); +struct drm_i915_gem_object * +__i915_gem_object_create_user(struct drm_i915_private *i915, u64 size, + struct intel_memory_region **placements, + unsigned int n_placements); extern const struct drm_i915_gem_object_ops i915_gem_shmem_ops; -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand ` (3 preceding siblings ...) 2021-07-21 20:13 ` [PATCH 4/7] drm/i915/gem: Unify user object creation (v3) Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 6/7] drm/i915/gem: Correct the locking and pin pattern for dma-buf (v8) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 7/7] drm/i915/gem: Migrate to system at dma-buf attach time (v7) Jason Ekstrand 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Thomas Hellström, Matthew Auld, Jason Ekstrand Whenever we had a user object (n_placements > 0), we were ignoring obj->mm.region and always putting obj->placements[0] as the requested region. For LMEM+SMEM objects, this was causing them to get shoved into LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, i915_gem_object_migrate(). Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index f253b11e9e367..b76bdd978a5cc 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, unsigned int i; placement->num_placement = 1; - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : - obj->mm.region, requested, flags); + i915_ttm_place_from_region(obj->mm.region, requested, flags); /* Cache this on object? */ placement->num_busy_placement = num_allowed; -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 6/7] drm/i915/gem: Correct the locking and pin pattern for dma-buf (v8) 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand ` (4 preceding siblings ...) 2021-07-21 20:13 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 2021-07-21 20:13 ` [PATCH 7/7] drm/i915/gem: Migrate to system at dma-buf attach time (v7) Jason Ekstrand 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel Cc: Thomas Hellström, Jason Ekstrand, Michael J . Ruhl From: Thomas Hellström <thomas.hellstrom@linux.intel.com> If our exported dma-bufs are imported by another instance of our driver, that instance will typically have the imported dma-bufs locked during dma_buf_map_attachment(). But the exporter also locks the same reservation object in the map_dma_buf() callback, which leads to recursive locking. So taking the lock inside _pin_pages_unlocked() is incorrect. Additionally, the current pinning code path is contrary to the defined way that pinning should occur. Remove the explicit pin/unpin from the map/umap functions and move them to the attach/detach allowing correct locking to occur, and to match the static dma-buf drm_prime pattern. Add a live selftest to exercise both dynamic and non-dynamic exports. v2: - Extend the selftest with a fake dynamic importer. - Provide real pin and unpin callbacks to not abuse the interface. v3: (ruhl) - Remove the dynamic export support and move the pinning into the attach/detach path. v4: (ruhl) - Put pages does not need to assert on the dma-resv v5: (jason) - Lock around dma_buf_unmap_attachment() when emulating a dynamic importer in the subtests. - Use pin_pages_unlocked v6: (jason) - Use dma_buf_attach instead of dma_buf_attach_dynamic in the selftests v7: (mauld) - Use __i915_gem_object_get_pages (2 __underscores) instead of the 4 ____underscore version in the selftests v8: (mauld) - Drop the kernel doc from the static i915_gem_dmabuf_attach function - Add missing "err = PTR_ERR()" to a bunch of selftest error cases Reported-by: Michael J. Ruhl <michael.j.ruhl@intel.com> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com> Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com> Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> --- drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 37 ++++-- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 109 +++++++++++++++++- 2 files changed, 132 insertions(+), 14 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c index 616c3a2f1baf0..59dc56ae14d6b 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -12,6 +12,8 @@ #include "i915_gem_object.h" #include "i915_scatterlist.h" +I915_SELFTEST_DECLARE(static bool force_different_devices;) + static struct drm_i915_gem_object *dma_buf_to_obj(struct dma_buf *buf) { return to_intel_bo(buf->priv); @@ -25,15 +27,11 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme struct scatterlist *src, *dst; int ret, i; - ret = i915_gem_object_pin_pages_unlocked(obj); - if (ret) - goto err; - /* Copy sg so that we make an independent mapping */ st = kmalloc(sizeof(struct sg_table), GFP_KERNEL); if (st == NULL) { ret = -ENOMEM; - goto err_unpin_pages; + goto err; } ret = sg_alloc_table(st, obj->mm.pages->nents, GFP_KERNEL); @@ -58,8 +56,6 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme sg_free_table(st); err_free: kfree(st); -err_unpin_pages: - i915_gem_object_unpin_pages(obj); err: return ERR_PTR(ret); } @@ -68,13 +64,9 @@ static void i915_gem_unmap_dma_buf(struct dma_buf_attachment *attachment, struct sg_table *sg, enum dma_data_direction dir) { - struct drm_i915_gem_object *obj = dma_buf_to_obj(attachment->dmabuf); - dma_unmap_sgtable(attachment->dev, sg, dir, DMA_ATTR_SKIP_CPU_SYNC); sg_free_table(sg); kfree(sg); - - i915_gem_object_unpin_pages(obj); } static int i915_gem_dmabuf_vmap(struct dma_buf *dma_buf, struct dma_buf_map *map) @@ -168,7 +160,25 @@ static int i915_gem_end_cpu_access(struct dma_buf *dma_buf, enum dma_data_direct return err; } +static int i915_gem_dmabuf_attach(struct dma_buf *dmabuf, + struct dma_buf_attachment *attach) +{ + struct drm_i915_gem_object *obj = dma_buf_to_obj(dmabuf); + + return i915_gem_object_pin_pages_unlocked(obj); +} + +static void i915_gem_dmabuf_detach(struct dma_buf *dmabuf, + struct dma_buf_attachment *attach) +{ + struct drm_i915_gem_object *obj = dma_buf_to_obj(dmabuf); + + i915_gem_object_unpin_pages(obj); +} + static const struct dma_buf_ops i915_dmabuf_ops = { + .attach = i915_gem_dmabuf_attach, + .detach = i915_gem_dmabuf_detach, .map_dma_buf = i915_gem_map_dma_buf, .unmap_dma_buf = i915_gem_unmap_dma_buf, .release = drm_gem_dmabuf_release, @@ -204,6 +214,8 @@ static int i915_gem_object_get_pages_dmabuf(struct drm_i915_gem_object *obj) struct sg_table *pages; unsigned int sg_page_sizes; + assert_object_held(obj); + pages = dma_buf_map_attachment(obj->base.import_attach, DMA_BIDIRECTIONAL); if (IS_ERR(pages)) @@ -241,7 +253,8 @@ struct drm_gem_object *i915_gem_prime_import(struct drm_device *dev, if (dma_buf->ops == &i915_dmabuf_ops) { obj = dma_buf_to_obj(dma_buf); /* is it from our device? */ - if (obj->base.dev == dev) { + if (obj->base.dev == dev && + !I915_SELFTEST_ONLY(force_different_devices)) { /* * Importing dmabuf exported from out own gem increases * refcount on gem itself instead of f_count of dmabuf. diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c index dd74bc09ec88d..d4ce01e6ee854 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c @@ -35,7 +35,7 @@ static int igt_dmabuf_export(void *arg) static int igt_dmabuf_import_self(void *arg) { struct drm_i915_private *i915 = arg; - struct drm_i915_gem_object *obj; + struct drm_i915_gem_object *obj, *import_obj; struct drm_gem_object *import; struct dma_buf *dmabuf; int err; @@ -65,14 +65,118 @@ static int igt_dmabuf_import_self(void *arg) err = -EINVAL; goto out_import; } + import_obj = to_intel_bo(import); + + i915_gem_object_lock(import_obj, NULL); + err = __i915_gem_object_get_pages(import_obj); + i915_gem_object_unlock(import_obj); + if (err) { + pr_err("Same object dma-buf get_pages failed!\n"); + goto out_import; + } err = 0; out_import: - i915_gem_object_put(to_intel_bo(import)); + i915_gem_object_put(import_obj); +out_dmabuf: + dma_buf_put(dmabuf); +out: + i915_gem_object_put(obj); + return err; +} + +static int igt_dmabuf_import_same_driver(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct drm_i915_gem_object *obj, *import_obj; + struct drm_gem_object *import; + struct dma_buf *dmabuf; + struct dma_buf_attachment *import_attach; + struct sg_table *st; + long timeout; + int err; + + force_different_devices = true; + obj = i915_gem_object_create_shmem(i915, PAGE_SIZE); + if (IS_ERR(obj)) { + err = PTR_ERR(obj); + goto out_ret; + } + + dmabuf = i915_gem_prime_export(&obj->base, 0); + if (IS_ERR(dmabuf)) { + pr_err("i915_gem_prime_export failed with err=%ld\n", + PTR_ERR(dmabuf)); + err = PTR_ERR(dmabuf); + goto out; + } + + import = i915_gem_prime_import(&i915->drm, dmabuf); + if (IS_ERR(import)) { + pr_err("i915_gem_prime_import failed with err=%ld\n", + PTR_ERR(import)); + err = PTR_ERR(import); + goto out_dmabuf; + } + + if (import == &obj->base) { + pr_err("i915_gem_prime_import reused gem object!\n"); + err = -EINVAL; + goto out_import; + } + + import_obj = to_intel_bo(import); + + i915_gem_object_lock(import_obj, NULL); + err = __i915_gem_object_get_pages(import_obj); + if (err) { + pr_err("Different objects dma-buf get_pages failed!\n"); + i915_gem_object_unlock(import_obj); + goto out_import; + } + + /* + * If the exported object is not in system memory, something + * weird is going on. TODO: When p2p is supported, this is no + * longer considered weird. + */ + if (obj->mm.region != i915->mm.regions[INTEL_REGION_SMEM]) { + pr_err("Exported dma-buf is not in system memory\n"); + err = -EINVAL; + } + + i915_gem_object_unlock(import_obj); + + /* Now try a fake an importer */ + import_attach = dma_buf_attach(dmabuf, obj->base.dev->dev); + if (IS_ERR(import_attach)) { + err = PTR_ERR(import_attach); + goto out_import; + } + + st = dma_buf_map_attachment(import_attach, DMA_BIDIRECTIONAL); + if (IS_ERR(st)) { + err = PTR_ERR(st); + goto out_detach; + } + + timeout = dma_resv_wait_timeout(dmabuf->resv, false, true, 5 * HZ); + if (!timeout) { + pr_err("dmabuf wait for exclusive fence timed out.\n"); + timeout = -ETIME; + } + err = timeout > 0 ? 0 : timeout; + dma_buf_unmap_attachment(import_attach, st, DMA_BIDIRECTIONAL); +out_detach: + dma_buf_detach(dmabuf, import_attach); +out_import: + i915_gem_object_put(import_obj); out_dmabuf: dma_buf_put(dmabuf); out: i915_gem_object_put(obj); +out_ret: + force_different_devices = false; return err; } @@ -286,6 +390,7 @@ int i915_gem_dmabuf_live_selftests(struct drm_i915_private *i915) { static const struct i915_subtest tests[] = { SUBTEST(igt_dmabuf_export), + SUBTEST(igt_dmabuf_import_same_driver), }; return i915_subtests(tests, i915); -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 7/7] drm/i915/gem: Migrate to system at dma-buf attach time (v7) 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand ` (5 preceding siblings ...) 2021-07-21 20:13 ` [PATCH 6/7] drm/i915/gem: Correct the locking and pin pattern for dma-buf (v8) Jason Ekstrand @ 2021-07-21 20:13 ` Jason Ekstrand 6 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:13 UTC (permalink / raw) To: intel-gfx, dri-devel Cc: Thomas Hellström, kernel test robot, Jason Ekstrand, Michael J . Ruhl From: Thomas Hellström <thomas.hellstrom@linux.intel.com> Until we support p2p dma or as a complement to that, migrate data to system memory at dma-buf attach time if possible. v2: - Rebase on dynamic exporter. Update the igt_dmabuf_import_same_driver selftest to migrate if we are LMEM capable. v3: - Migrate also in the pin() callback. v4: - Migrate in attach v5: (jason) - Lock around the migration v6: (jason) - Move the can_migrate check outside the lock - Rework the selftests to test more migration conditions. In particular, SMEM, LMEM, and LMEM+SMEM are all checked. v7: (mauld) - Misc style nits Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com> Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com> Reported-by: kernel test robot <lkp@intel.com> Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> --- drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 23 ++++- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 87 ++++++++++++++++++- 2 files changed, 106 insertions(+), 4 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c index 59dc56ae14d6b..afa34111de02e 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -164,8 +164,29 @@ static int i915_gem_dmabuf_attach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach) { struct drm_i915_gem_object *obj = dma_buf_to_obj(dmabuf); + struct i915_gem_ww_ctx ww; + int err; + + if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) + return -EOPNOTSUPP; + + for_i915_gem_ww(&ww, err, true) { + err = i915_gem_object_lock(obj, &ww); + if (err) + continue; + + err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_SMEM); + if (err) + continue; - return i915_gem_object_pin_pages_unlocked(obj); + err = i915_gem_object_wait_migration(obj, 0); + if (err) + continue; + + err = i915_gem_object_pin_pages(obj); + } + + return err; } static void i915_gem_dmabuf_detach(struct dma_buf *dmabuf, diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c index d4ce01e6ee854..ffae7df5e4d7d 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c @@ -85,9 +85,63 @@ static int igt_dmabuf_import_self(void *arg) return err; } -static int igt_dmabuf_import_same_driver(void *arg) +static int igt_dmabuf_import_same_driver_lmem(void *arg) { struct drm_i915_private *i915 = arg; + struct intel_memory_region *lmem = i915->mm.regions[INTEL_REGION_LMEM]; + struct drm_i915_gem_object *obj; + struct drm_gem_object *import; + struct dma_buf *dmabuf; + int err; + + if (!lmem) + return 0; + + force_different_devices = true; + + obj = __i915_gem_object_create_user(i915, PAGE_SIZE, &lmem, 1); + if (IS_ERR(obj)) { + pr_err("__i915_gem_object_create_user failed with err=%ld\n", + PTR_ERR(dmabuf)); + err = PTR_ERR(obj); + goto out_ret; + } + + dmabuf = i915_gem_prime_export(&obj->base, 0); + if (IS_ERR(dmabuf)) { + pr_err("i915_gem_prime_export failed with err=%ld\n", + PTR_ERR(dmabuf)); + err = PTR_ERR(dmabuf); + goto out; + } + + /* + * We expect an import of an LMEM-only object to fail with + * -EOPNOTSUPP because it can't be migrated to SMEM. + */ + import = i915_gem_prime_import(&i915->drm, dmabuf); + if (!IS_ERR(import)) { + drm_gem_object_put(import); + pr_err("i915_gem_prime_import succeeded when it shouldn't have\n"); + err = -EINVAL; + } else if (PTR_ERR(import) != -EOPNOTSUPP) { + pr_err("i915_gem_prime_import failed with the wrong err=%ld\n", + PTR_ERR(import)); + err = PTR_ERR(import); + } + + dma_buf_put(dmabuf); +out: + i915_gem_object_put(obj); +out_ret: + force_different_devices = false; + return err; +} + +static int igt_dmabuf_import_same_driver(struct drm_i915_private *i915, + struct intel_memory_region **regions, + unsigned int num_regions) +{ struct drm_i915_gem_object *obj, *import_obj; struct drm_gem_object *import; struct dma_buf *dmabuf; @@ -97,8 +151,12 @@ static int igt_dmabuf_import_same_driver(void *arg) int err; force_different_devices = true; - obj = i915_gem_object_create_shmem(i915, PAGE_SIZE); + + obj = __i915_gem_object_create_user(i915, PAGE_SIZE, + regions, num_regions); if (IS_ERR(obj)) { + pr_err("__i915_gem_object_create_user failed with err=%ld\n", + PTR_ERR(dmabuf)); err = PTR_ERR(obj); goto out_ret; } @@ -180,6 +238,27 @@ static int igt_dmabuf_import_same_driver(void *arg) return err; } +static int igt_dmabuf_import_same_driver_smem(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct intel_memory_region *smem = i915->mm.regions[INTEL_REGION_SMEM]; + + return igt_dmabuf_import_same_driver(i915, &smem, 1); +} + +static int igt_dmabuf_import_same_driver_lmem_smem(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct intel_memory_region *regions[2]; + + if (!i915->mm.regions[INTEL_REGION_LMEM]) + return 0; + + regions[0] = i915->mm.regions[INTEL_REGION_LMEM]; + regions[1] = i915->mm.regions[INTEL_REGION_SMEM]; + return igt_dmabuf_import_same_driver(i915, regions, 2); +} + static int igt_dmabuf_import(void *arg) { struct drm_i915_private *i915 = arg; @@ -390,7 +469,9 @@ int i915_gem_dmabuf_live_selftests(struct drm_i915_private *i915) { static const struct i915_subtest tests[] = { SUBTEST(igt_dmabuf_export), - SUBTEST(igt_dmabuf_import_same_driver), + SUBTEST(igt_dmabuf_import_same_driver_lmem), + SUBTEST(igt_dmabuf_import_same_driver_smem), + SUBTEST(igt_dmabuf_import_same_driver_lmem_smem), }; return i915_subtests(tests, i915); -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v7) @ 2021-07-16 14:14 Jason Ekstrand 2021-07-16 14:14 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-16 14:14 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Jason Ekstrand This patch series fixes an issue with discrete graphics on Intel where we allowed dma-buf import while leaving the object in local memory. This breaks down pretty badly if the import happened on a different physical device. v7: - Drop "drm/i915/gem/ttm: Place new BOs in the requested region" - Add a new "drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create()" - Misc. review feedback from Matthew Auld Jason Ekstrand (5): drm/i915/gem: Check object_can_migrate from object_migrate drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() drm/i915/gem: Unify user object creation (v2) drm/i915/gem/ttm: Respect the objection region in placement_from_obj Thomas Hellström (2): drm/i915/gem: Correct the locking and pin pattern for dma-buf (v6) drm/i915/gem: Migrate to system at dma-buf attach time (v6) drivers/gpu/drm/i915/gem/i915_gem_create.c | 165 ++++++++-------- drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 64 ++++-- drivers/gpu/drm/i915/gem/i915_gem_object.c | 13 +- drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 + drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 184 +++++++++++++++++- .../drm/i915/gem/selftests/i915_gem_migrate.c | 15 -- 7 files changed, 318 insertions(+), 130 deletions(-) -- 2.31.1 ^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 14:14 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v7) Jason Ekstrand @ 2021-07-16 14:14 ` Jason Ekstrand 0 siblings, 0 replies; 23+ messages in thread From: Jason Ekstrand @ 2021-07-16 14:14 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Thomas Hellström, Matthew Auld, Jason Ekstrand Whenever we had a user object (n_placements > 0), we were ignoring obj->mm.region and always putting obj->placements[0] as the requested region. For LMEM+SMEM objects, this was causing them to get shoved into LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, i915_gem_object_migrate(). Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index 6589411396d3f..8eeb73c7c401c 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, unsigned int i; placement->num_placement = 1; - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : - obj->mm.region, requested, flags); + i915_ttm_place_from_region(obj->mm.region, requested, flags); /* Cache this on object? */ placement->num_busy_placement = num_allowed; -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device @ 2021-07-15 22:38 Jason Ekstrand 2021-07-15 22:38 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-15 22:38 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Jason Ekstrand This patch series fixes an issue with discrete graphics on Intel where we allowed dma-buf import while leaving the object in local memory. This breaks down pretty badly if the import happened on a different physical device. Jason Ekstrand (5): drm/i915/gem: Check object_can_migrate from object_migrate drm/i915/gem: Refactor placement setup for i915_gem_object_create* drm/i915/gem: Unify user object creation drm/i915/gem/ttm: Place new BOs in the requested region drm/i915/gem/ttm: Respect the objection region in placement_from_obj Thomas Hellström (2): drm/i915/gem: Correct the locking and pin pattern for dma-buf (v6) drm/i915/gem: Migrate to system at dma-buf attach time (v6) drivers/gpu/drm/i915/gem/i915_gem_create.c | 159 ++++++++------- drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 64 ++++-- drivers/gpu/drm/i915/gem/i915_gem_object.c | 13 +- drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 + drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 11 +- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 184 +++++++++++++++++- .../drm/i915/gem/selftests/i915_gem_migrate.c | 15 -- 7 files changed, 324 insertions(+), 126 deletions(-) -- 2.31.1 ^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-15 22:38 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device Jason Ekstrand @ 2021-07-15 22:38 ` Jason Ekstrand 2021-07-16 13:54 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-15 22:38 UTC (permalink / raw) To: intel-gfx, dri-devel; +Cc: Thomas Hellström, Matthew Auld, Jason Ekstrand Whenever we had a user object (n_placements > 0), we were ignoring obj->mm.region and always putting obj->placements[0] as the requested region. For LMEM+SMEM objects, this was causing them to get shoved into LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, i915_gem_object_migrate(). Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index d30f274c329c7..5985e994d56cf 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, unsigned int i; placement->num_placement = 1; - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : - obj->mm.region, requested, flags); + i915_ttm_place_from_region(obj->mm.region, requested, flags); /* Cache this on object? */ placement->num_busy_placement = num_allowed; -- 2.31.1 ^ permalink raw reply related [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-15 22:38 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand @ 2021-07-16 13:54 ` Matthew Auld 2021-07-16 14:10 ` Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-16 13:54 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > Whenever we had a user object (n_placements > 0), we were ignoring > obj->mm.region and always putting obj->placements[0] as the requested > region. For LMEM+SMEM objects, this was causing them to get shoved into > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > i915_gem_object_migrate(). i915_ttm_migrate calls i915_ttm_place_from_region() directly with the requested region, so there shouldn't be an issue with migration right? Do you have some more details? > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > Cc: Matthew Auld <matthew.auld@intel.com> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > --- > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > 1 file changed, 1 insertion(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > index d30f274c329c7..5985e994d56cf 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > unsigned int i; > > placement->num_placement = 1; > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > - obj->mm.region, requested, flags); > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > /* Cache this on object? */ > placement->num_busy_placement = num_allowed; > -- > 2.31.1 > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 13:54 ` Matthew Auld @ 2021-07-16 14:10 ` Jason Ekstrand 2021-07-16 15:52 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-16 14:10 UTC (permalink / raw) To: Matthew Auld Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > obj->mm.region and always putting obj->placements[0] as the requested > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > i915_gem_object_migrate(). > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > requested region, so there shouldn't be an issue with migration right? > Do you have some more details? With i915_ttm_migrate directly, no. But, in the last patch in the series, we're trying to migrate LMEM+SMEM buffers into SMEM on attach() and pin it there. This blows up in a very unexpected (IMO) way. The flow goes something like this: - Client attempts a dma-buf import from another device - In attach() we call i915_gem_object_migrate() which calls i915_ttm_migrate() which migrates as requested. - Once the migration is complete, we call i915_gem_object_pin_pages() which calls i915_ttm_get_pages() which depends on i915_ttm_placement_from_obj() and so migrates it right back to LMEM. Maybe the problem here is actually that our TTM code isn't respecting obj->mm.pages_pin_count? In case you can't tell, I really have no clue what I'm doing here. I'm really stumbling around in the dark finding things that make my bug go away. I'm happy for the feedback. --Jason > > > > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > > Cc: Matthew Auld <matthew.auld@intel.com> > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > > --- > > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > index d30f274c329c7..5985e994d56cf 100644 > > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > > unsigned int i; > > > > placement->num_placement = 1; > > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > > - obj->mm.region, requested, flags); > > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > > > /* Cache this on object? */ > > placement->num_busy_placement = num_allowed; > > -- > > 2.31.1 > > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 14:10 ` Jason Ekstrand @ 2021-07-16 15:52 ` Matthew Auld 2021-07-16 16:00 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-16 15:52 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > <matthew.william.auld@gmail.com> wrote: > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > obj->mm.region and always putting obj->placements[0] as the requested > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > i915_gem_object_migrate(). > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > requested region, so there shouldn't be an issue with migration right? > > Do you have some more details? > > With i915_ttm_migrate directly, no. But, in the last patch in the > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > attach() and pin it there. This blows up in a very unexpected (IMO) > way. The flow goes something like this: > > - Client attempts a dma-buf import from another device > - In attach() we call i915_gem_object_migrate() which calls > i915_ttm_migrate() which migrates as requested. > - Once the migration is complete, we call i915_gem_object_pin_pages() > which calls i915_ttm_get_pages() which depends on > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. The mm.pages must be NULL here, otherwise it would just increment the pages_pin_count? > > Maybe the problem here is actually that our TTM code isn't respecting > obj->mm.pages_pin_count? I think if the resource is moved, we always nuke the mm.pages after being notified of the move. Also TTM is also not allowed to move pinned buffers. I guess if we are evicted/swapped, so assuming we are not holding the object lock, and it's not pinned, the future call to get_pages() will see mm.pages = NULL, even though the ttm_resource is still there, and because we prioritise the placements[0], instead of mm.region we end up moving it for no good reason. But in your case you are holding the lock, or it's pinned? Also is this just with the selftest, or something real? > > In case you can't tell, I really have no clue what I'm doing here. > I'm really stumbling around in the dark finding things that make my > bug go away. I'm happy for the feedback. > > --Jason > > > > > > > > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > > > Cc: Matthew Auld <matthew.auld@intel.com> > > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > > > --- > > > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > index d30f274c329c7..5985e994d56cf 100644 > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > > > unsigned int i; > > > > > > placement->num_placement = 1; > > > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > > > - obj->mm.region, requested, flags); > > > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > > > > > /* Cache this on object? */ > > > placement->num_busy_placement = num_allowed; > > > -- > > > 2.31.1 > > > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 15:52 ` Matthew Auld @ 2021-07-16 16:00 ` Matthew Auld 2021-07-16 17:38 ` Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-16 16:00 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, 16 Jul 2021 at 16:52, Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > i915_gem_object_migrate(). > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > requested region, so there shouldn't be an issue with migration right? > > > Do you have some more details? > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > attach() and pin it there. This blows up in a very unexpected (IMO) > > way. The flow goes something like this: > > > > - Client attempts a dma-buf import from another device > > - In attach() we call i915_gem_object_migrate() which calls > > i915_ttm_migrate() which migrates as requested. > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > which calls i915_ttm_get_pages() which depends on > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > The mm.pages must be NULL here, otherwise it would just increment the > pages_pin_count? > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > obj->mm.pages_pin_count? > > I think if the resource is moved, we always nuke the mm.pages after > being notified of the move. Also TTM is also not allowed to move > pinned buffers. > > I guess if we are evicted/swapped, so assuming we are not holding the > object lock, and it's not pinned, the future call to get_pages() will > see mm.pages = NULL, even though the ttm_resource is still there, and > because we prioritise the placements[0], instead of mm.region we end > up moving it for no good reason. But in your case you are holding the > lock, or it's pinned? Also is this just with the selftest, or > something real? Or at least in the selftest I see ____i915_gem_object_get_pages() which doesn't even consider the mm.pages AFAIK. > > > > > In case you can't tell, I really have no clue what I'm doing here. > > I'm really stumbling around in the dark finding things that make my > > bug go away. I'm happy for the feedback. > > > > --Jason > > > > > > > > > > > > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > > > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > > > > Cc: Matthew Auld <matthew.auld@intel.com> > > > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > > > > --- > > > > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > > > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > index d30f274c329c7..5985e994d56cf 100644 > > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > > > > unsigned int i; > > > > > > > > placement->num_placement = 1; > > > > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > > > > - obj->mm.region, requested, flags); > > > > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > > > > > > > /* Cache this on object? */ > > > > placement->num_busy_placement = num_allowed; > > > > -- > > > > 2.31.1 > > > > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 16:00 ` Matthew Auld @ 2021-07-16 17:38 ` Jason Ekstrand 2021-07-16 18:44 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-16 17:38 UTC (permalink / raw) To: Matthew Auld Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > <matthew.william.auld@gmail.com> wrote: > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > i915_gem_object_migrate(). > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > requested region, so there shouldn't be an issue with migration right? > > > > Do you have some more details? > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > way. The flow goes something like this: > > > > > > - Client attempts a dma-buf import from another device > > > - In attach() we call i915_gem_object_migrate() which calls > > > i915_ttm_migrate() which migrates as requested. > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > which calls i915_ttm_get_pages() which depends on > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > The mm.pages must be NULL here, otherwise it would just increment the > > pages_pin_count? Given that the test is using the ____four_underscores version, it doesn't have that check. However, this executes after we've done the dma-buf import which pinned pages. So we should definitely have pages. > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > obj->mm.pages_pin_count? > > > > I think if the resource is moved, we always nuke the mm.pages after > > being notified of the move. Also TTM is also not allowed to move > > pinned buffers. > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > object lock, and it's not pinned, the future call to get_pages() will > > see mm.pages = NULL, even though the ttm_resource is still there, and > > because we prioritise the placements[0], instead of mm.region we end > > up moving it for no good reason. But in your case you are holding the > > lock, or it's pinned? Also is this just with the selftest, or > > something real? > > Or at least in the selftest I see ____i915_gem_object_get_pages() > which doesn't even consider the mm.pages AFAIK. The bogus migration is happening as part of the __i915_gem_object_get_pages() (2 __underscores) call in i915_gem_dmabuf_attach (see last patch). That code is attempting to migrate the BO to SMEM and then pin it there using the obvious calls to do so. However, in the pin_pages call, it gets implicitly migrated back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() migrating things at all? --Jason > > > > > > > > In case you can't tell, I really have no clue what I'm doing here. > > > I'm really stumbling around in the dark finding things that make my > > > bug go away. I'm happy for the feedback. > > > > > > --Jason > > > > > > > > > > > > > > > > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > > > > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > > > > > Cc: Matthew Auld <matthew.auld@intel.com> > > > > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > > > > > --- > > > > > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > > > > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > index d30f274c329c7..5985e994d56cf 100644 > > > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > > > > > unsigned int i; > > > > > > > > > > placement->num_placement = 1; > > > > > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > > > > > - obj->mm.region, requested, flags); > > > > > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > > > > > > > > > /* Cache this on object? */ > > > > > placement->num_busy_placement = num_allowed; > > > > > -- > > > > > 2.31.1 > > > > > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 17:38 ` Jason Ekstrand @ 2021-07-16 18:44 ` Matthew Auld 2021-07-16 19:49 ` Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-16 18:44 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > <matthew.william.auld@gmail.com> wrote: > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > i915_gem_object_migrate(). > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > Do you have some more details? > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > way. The flow goes something like this: > > > > > > > > - Client attempts a dma-buf import from another device > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > i915_ttm_migrate() which migrates as requested. > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > which calls i915_ttm_get_pages() which depends on > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > pages_pin_count? > > Given that the test is using the ____four_underscores version, it > doesn't have that check. However, this executes after we've done the > dma-buf import which pinned pages. So we should definitely have > pages. We shouldn't call ____four_underscores() if we might already have pages though. Under non-TTM that would leak the pages, and in TTM we might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for example nothing was moved. I take it we can't just call pin_pages()? Four scary underscores usually means "don't call this in normal code". > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > obj->mm.pages_pin_count? > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > being notified of the move. Also TTM is also not allowed to move > > > pinned buffers. > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > object lock, and it's not pinned, the future call to get_pages() will > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > because we prioritise the placements[0], instead of mm.region we end > > > up moving it for no good reason. But in your case you are holding the > > > lock, or it's pinned? Also is this just with the selftest, or > > > something real? > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > which doesn't even consider the mm.pages AFAIK. > > The bogus migration is happening as part of the > __i915_gem_object_get_pages() (2 __underscores) call in > i915_gem_dmabuf_attach (see last patch). That code is attempting to > migrate the BO to SMEM and then pin it there using the obvious calls > to do so. However, in the pin_pages call, it gets implicitly migrated > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > migrating things at all? Not sure yet, but __two_underscores() checks if i915_gem_object_has_pages() before actually calling into i915_ttm_get_pages(), so the mm.pages would have to be NULL here for some reason, so best guess is something to do with move_notify(). > > --Jason > > > > > > > > > > > > In case you can't tell, I really have no clue what I'm doing here. > > > > I'm really stumbling around in the dark finding things that make my > > > > bug go away. I'm happy for the feedback. > > > > > > > > --Jason > > > > > > > > > > > > > > > > > > > > > Signed-off-by: Jason Ekstrand <jason@jlekstrand.net> > > > > > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > > > > > > Cc: Matthew Auld <matthew.auld@intel.com> > > > > > > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > > > > > > --- > > > > > > drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 3 +-- > > > > > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > > > > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > > index d30f274c329c7..5985e994d56cf 100644 > > > > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c > > > > > > @@ -150,8 +150,7 @@ i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, > > > > > > unsigned int i; > > > > > > > > > > > > placement->num_placement = 1; > > > > > > - i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : > > > > > > - obj->mm.region, requested, flags); > > > > > > + i915_ttm_place_from_region(obj->mm.region, requested, flags); > > > > > > > > > > > > /* Cache this on object? */ > > > > > > placement->num_busy_placement = num_allowed; > > > > > > -- > > > > > > 2.31.1 > > > > > > ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 18:44 ` Matthew Auld @ 2021-07-16 19:49 ` Jason Ekstrand 2021-07-19 13:34 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-16 19:49 UTC (permalink / raw) To: Matthew Auld Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > > > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > > i915_gem_object_migrate(). > > > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > > Do you have some more details? > > > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > > way. The flow goes something like this: > > > > > > > > > > - Client attempts a dma-buf import from another device > > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > > i915_ttm_migrate() which migrates as requested. > > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > > which calls i915_ttm_get_pages() which depends on > > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > > pages_pin_count? > > > > Given that the test is using the ____four_underscores version, it > > doesn't have that check. However, this executes after we've done the > > dma-buf import which pinned pages. So we should definitely have > > pages. > > We shouldn't call ____four_underscores() if we might already have > pages though. Under non-TTM that would leak the pages, and in TTM we > might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > example nothing was moved. I take it we can't just call pin_pages()? > Four scary underscores usually means "don't call this in normal code". I've switched the ____four_underscores call to a __two_underscores in the selftests and it had no effect, good or bad. But, still, probably better to call that one. > > > > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > > obj->mm.pages_pin_count? > > > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > > being notified of the move. Also TTM is also not allowed to move > > > > pinned buffers. > > > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > > object lock, and it's not pinned, the future call to get_pages() will > > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > > because we prioritise the placements[0], instead of mm.region we end > > > > up moving it for no good reason. But in your case you are holding the > > > > lock, or it's pinned? Also is this just with the selftest, or > > > > something real? > > > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > > which doesn't even consider the mm.pages AFAIK. > > > > The bogus migration is happening as part of the > > __i915_gem_object_get_pages() (2 __underscores) call in > > i915_gem_dmabuf_attach (see last patch). That code is attempting to > > migrate the BO to SMEM and then pin it there using the obvious calls > > to do so. However, in the pin_pages call, it gets implicitly migrated > > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > > migrating things at all? > > Not sure yet, but __two_underscores() checks if > i915_gem_object_has_pages() before actually calling into > i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > some reason, so best guess is something to do with move_notify(). Did a bit of experimenting along those lines and added the following to the self-test BEFORE the export/import: i915_gem_object_lock(obj, NULL); err = __i915_gem_object_get_pages(obj); __i915_gem_object_unpin_pages(obj); i915_gem_object_unlock(obj); if (err) { pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); goto out_ret; } This seems to make the migration happen as expected without this patch. So it seems the problem only exists on buffers that haven't gotten any backing storage yet (if I'm understanding get_pages correctly). One potential work-around (not sure if this is a good idea or not!) would be to do this inside dmabuf_attach(). Is this reliable? Once it has pages will it always have pages? Or are there crazy races I need to be worried about here? --Jason ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-16 19:49 ` Jason Ekstrand @ 2021-07-19 13:34 ` Matthew Auld 2021-07-21 20:11 ` Jason Ekstrand 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-19 13:34 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: > > On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld > <matthew.william.auld@gmail.com> wrote: > > > > On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > > > i915_gem_object_migrate(). > > > > > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > > > Do you have some more details? > > > > > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > > > way. The flow goes something like this: > > > > > > > > > > > > - Client attempts a dma-buf import from another device > > > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > > > i915_ttm_migrate() which migrates as requested. > > > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > > > which calls i915_ttm_get_pages() which depends on > > > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > > > pages_pin_count? > > > > > > Given that the test is using the ____four_underscores version, it > > > doesn't have that check. However, this executes after we've done the > > > dma-buf import which pinned pages. So we should definitely have > > > pages. > > > > We shouldn't call ____four_underscores() if we might already have > > pages though. Under non-TTM that would leak the pages, and in TTM we > > might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > > example nothing was moved. I take it we can't just call pin_pages()? > > Four scary underscores usually means "don't call this in normal code". > > I've switched the ____four_underscores call to a __two_underscores in > the selftests and it had no effect, good or bad. But, still, probably > better to call that one. > > > > > > > > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > > > obj->mm.pages_pin_count? > > > > > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > > > being notified of the move. Also TTM is also not allowed to move > > > > > pinned buffers. > > > > > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > > > object lock, and it's not pinned, the future call to get_pages() will > > > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > > > because we prioritise the placements[0], instead of mm.region we end > > > > > up moving it for no good reason. But in your case you are holding the > > > > > lock, or it's pinned? Also is this just with the selftest, or > > > > > something real? > > > > > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > > > which doesn't even consider the mm.pages AFAIK. > > > > > > The bogus migration is happening as part of the > > > __i915_gem_object_get_pages() (2 __underscores) call in > > > i915_gem_dmabuf_attach (see last patch). That code is attempting to > > > migrate the BO to SMEM and then pin it there using the obvious calls > > > to do so. However, in the pin_pages call, it gets implicitly migrated > > > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > > > migrating things at all? > > > > Not sure yet, but __two_underscores() checks if > > i915_gem_object_has_pages() before actually calling into > > i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > > some reason, so best guess is something to do with move_notify(). > > Did a bit of experimenting along those lines and added the following > to the self-test BEFORE the export/import: > > i915_gem_object_lock(obj, NULL); > err = __i915_gem_object_get_pages(obj); > __i915_gem_object_unpin_pages(obj); > i915_gem_object_unlock(obj); > if (err) { > pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); > goto out_ret; > } > > This seems to make the migration happen as expected without this > patch. So it seems the problem only exists on buffers that haven't > gotten any backing storage yet (if I'm understanding get_pages > correctly). > > One potential work-around (not sure if this is a good idea or not!) > would be to do this inside dmabuf_attach(). Is this reliable? Once > it has pages will it always have pages? Or are there crazy races I > need to be worried about here? It turns out that the i915_ttm_adjust_gem_after_move() call in ttm_object_init will always update the mm.region to system memory(so that it matches the ttm resource), which seems reasonable given the default system placeholder thing, but does seem slightly iffy since we haven't actually moved/allocated anything. So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since mm.region == mr. Which ofc means when we actually call get_pages() all that happens is that we allocate the pages in system memory(or without this patch placements[0]). Also with this patch lmem+smem, will always be placed in smem first, regardless of the placements ordering. For now we could maybe just split i915_ttm_adjust_gem_after_move() so we skip the part which updates the mm.region here in the init portion, since that should only happen when we try to place the object for real? > > --Jason ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-19 13:34 ` Matthew Auld @ 2021-07-21 20:11 ` Jason Ekstrand 2021-07-22 9:49 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Jason Ekstrand @ 2021-07-21 20:11 UTC (permalink / raw) To: Matthew Auld Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Mon, Jul 19, 2021 at 8:35 AM Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > > > > > > On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > > > > i915_gem_object_migrate(). > > > > > > > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > > > > Do you have some more details? > > > > > > > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > > > > way. The flow goes something like this: > > > > > > > > > > > > > > - Client attempts a dma-buf import from another device > > > > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > > > > i915_ttm_migrate() which migrates as requested. > > > > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > > > > which calls i915_ttm_get_pages() which depends on > > > > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > > > > pages_pin_count? > > > > > > > > Given that the test is using the ____four_underscores version, it > > > > doesn't have that check. However, this executes after we've done the > > > > dma-buf import which pinned pages. So we should definitely have > > > > pages. > > > > > > We shouldn't call ____four_underscores() if we might already have > > > pages though. Under non-TTM that would leak the pages, and in TTM we > > > might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > > > example nothing was moved. I take it we can't just call pin_pages()? > > > Four scary underscores usually means "don't call this in normal code". > > > > I've switched the ____four_underscores call to a __two_underscores in > > the selftests and it had no effect, good or bad. But, still, probably > > better to call that one. > > > > > > > > > > > > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > > > > obj->mm.pages_pin_count? > > > > > > > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > > > > being notified of the move. Also TTM is also not allowed to move > > > > > > pinned buffers. > > > > > > > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > > > > object lock, and it's not pinned, the future call to get_pages() will > > > > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > > > > because we prioritise the placements[0], instead of mm.region we end > > > > > > up moving it for no good reason. But in your case you are holding the > > > > > > lock, or it's pinned? Also is this just with the selftest, or > > > > > > something real? > > > > > > > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > > > > which doesn't even consider the mm.pages AFAIK. > > > > > > > > The bogus migration is happening as part of the > > > > __i915_gem_object_get_pages() (2 __underscores) call in > > > > i915_gem_dmabuf_attach (see last patch). That code is attempting to > > > > migrate the BO to SMEM and then pin it there using the obvious calls > > > > to do so. However, in the pin_pages call, it gets implicitly migrated > > > > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > > > > migrating things at all? > > > > > > Not sure yet, but __two_underscores() checks if > > > i915_gem_object_has_pages() before actually calling into > > > i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > > > some reason, so best guess is something to do with move_notify(). > > > > Did a bit of experimenting along those lines and added the following > > to the self-test BEFORE the export/import: > > > > i915_gem_object_lock(obj, NULL); > > err = __i915_gem_object_get_pages(obj); > > __i915_gem_object_unpin_pages(obj); > > i915_gem_object_unlock(obj); > > if (err) { > > pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); > > goto out_ret; > > } > > > > This seems to make the migration happen as expected without this > > patch. So it seems the problem only exists on buffers that haven't > > gotten any backing storage yet (if I'm understanding get_pages > > correctly). > > > > One potential work-around (not sure if this is a good idea or not!) > > would be to do this inside dmabuf_attach(). Is this reliable? Once > > it has pages will it always have pages? Or are there crazy races I > > need to be worried about here? > > It turns out that the i915_ttm_adjust_gem_after_move() call in > ttm_object_init will always update the mm.region to system memory(so > that it matches the ttm resource), which seems reasonable given the > default system placeholder thing, but does seem slightly iffy since we > haven't actually moved/allocated anything. > > So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since > mm.region == mr. Which ofc means when we actually call get_pages() all > that happens is that we allocate the pages in system memory(or without > this patch placements[0]). Also with this patch lmem+smem, will always > be placed in smem first, regardless of the placements ordering. > > For now we could maybe just split i915_ttm_adjust_gem_after_move() so > we skip the part which updates the mm.region here in the init portion, > since that should only happen when we try to place the object for > real? Doesn't that mean we would end up with obj->mm.region and obj->mm.res->mem_type are out-of-sync? That seems bad. I would think we'd want the two in sync at all times. It seems like the fundamental problem here is that, when it's created, the object isn't really in any memory region at all. While I don't think obj->mm.region == NULL is allowed or a good idea, it does seem closer to the ground truth. Perhaps what we really want is for i915_gem_object_migrate to get_pages before it does the migration to ensure that pages exist. The only call to i915_gem_object_migrate in the code-base today is in the display code and it's immediately followed by pin_pages(). For that matter, maybe the call we actually want is i915_object_migrate_and_pin that does the whole lot. Thoughts? --Jason P.S. I'm going to go ahead and send another version with your other comments addressed. We can keep this discussion going here for now. ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-21 20:11 ` Jason Ekstrand @ 2021-07-22 9:49 ` Matthew Auld 2021-07-22 9:59 ` Matthew Auld 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-22 9:49 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Wed, 21 Jul 2021 at 21:11, Jason Ekstrand <jason@jlekstrand.net> wrote: > > On Mon, Jul 19, 2021 at 8:35 AM Matthew Auld > <matthew.william.auld@gmail.com> wrote: > > > > On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > > > > > i915_gem_object_migrate(). > > > > > > > > > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > > > > > Do you have some more details? > > > > > > > > > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > > > > > way. The flow goes something like this: > > > > > > > > > > > > > > > > - Client attempts a dma-buf import from another device > > > > > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > > > > > i915_ttm_migrate() which migrates as requested. > > > > > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > > > > > which calls i915_ttm_get_pages() which depends on > > > > > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > > > > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > > > > > pages_pin_count? > > > > > > > > > > Given that the test is using the ____four_underscores version, it > > > > > doesn't have that check. However, this executes after we've done the > > > > > dma-buf import which pinned pages. So we should definitely have > > > > > pages. > > > > > > > > We shouldn't call ____four_underscores() if we might already have > > > > pages though. Under non-TTM that would leak the pages, and in TTM we > > > > might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > > > > example nothing was moved. I take it we can't just call pin_pages()? > > > > Four scary underscores usually means "don't call this in normal code". > > > > > > I've switched the ____four_underscores call to a __two_underscores in > > > the selftests and it had no effect, good or bad. But, still, probably > > > better to call that one. > > > > > > > > > > > > > > > > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > > > > > obj->mm.pages_pin_count? > > > > > > > > > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > > > > > being notified of the move. Also TTM is also not allowed to move > > > > > > > pinned buffers. > > > > > > > > > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > > > > > object lock, and it's not pinned, the future call to get_pages() will > > > > > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > > > > > because we prioritise the placements[0], instead of mm.region we end > > > > > > > up moving it for no good reason. But in your case you are holding the > > > > > > > lock, or it's pinned? Also is this just with the selftest, or > > > > > > > something real? > > > > > > > > > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > > > > > which doesn't even consider the mm.pages AFAIK. > > > > > > > > > > The bogus migration is happening as part of the > > > > > __i915_gem_object_get_pages() (2 __underscores) call in > > > > > i915_gem_dmabuf_attach (see last patch). That code is attempting to > > > > > migrate the BO to SMEM and then pin it there using the obvious calls > > > > > to do so. However, in the pin_pages call, it gets implicitly migrated > > > > > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > > > > > migrating things at all? > > > > > > > > Not sure yet, but __two_underscores() checks if > > > > i915_gem_object_has_pages() before actually calling into > > > > i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > > > > some reason, so best guess is something to do with move_notify(). > > > > > > Did a bit of experimenting along those lines and added the following > > > to the self-test BEFORE the export/import: > > > > > > i915_gem_object_lock(obj, NULL); > > > err = __i915_gem_object_get_pages(obj); > > > __i915_gem_object_unpin_pages(obj); > > > i915_gem_object_unlock(obj); > > > if (err) { > > > pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); > > > goto out_ret; > > > } > > > > > > This seems to make the migration happen as expected without this > > > patch. So it seems the problem only exists on buffers that haven't > > > gotten any backing storage yet (if I'm understanding get_pages > > > correctly). > > > > > > One potential work-around (not sure if this is a good idea or not!) > > > would be to do this inside dmabuf_attach(). Is this reliable? Once > > > it has pages will it always have pages? Or are there crazy races I > > > need to be worried about here? > > > > It turns out that the i915_ttm_adjust_gem_after_move() call in > > ttm_object_init will always update the mm.region to system memory(so > > that it matches the ttm resource), which seems reasonable given the > > default system placeholder thing, but does seem slightly iffy since we > > haven't actually moved/allocated anything. > > > > So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since > > mm.region == mr. Which ofc means when we actually call get_pages() all > > that happens is that we allocate the pages in system memory(or without > > this patch placements[0]). Also with this patch lmem+smem, will always > > be placed in smem first, regardless of the placements ordering. > > > > For now we could maybe just split i915_ttm_adjust_gem_after_move() so > > we skip the part which updates the mm.region here in the init portion, > > since that should only happen when we try to place the object for > > real? > > Doesn't that mean we would end up with obj->mm.region and > obj->mm.res->mem_type are out-of-sync? That seems bad. I would think > we'd want the two in sync at all times. It likely doesn't matter since all roads lead to i915_ttm_get_pages() when we need to actually use the object? Also updating the mm.region in ttm_object_init() to reflect the dummy ttm resource seems a little scary, since any existing is_lmem() check now needs to happen after we place the object. Or at least the existing callers(for kernel internal objects) might not have expected that behaviour. Not sure if we checked all the callers. > > It seems like the fundamental problem here is that, when it's created, > the object isn't really in any memory region at all. While I don't > think obj->mm.region == NULL is allowed or a good idea, it does seem > closer to the ground truth. Yeah, seems reasonable, especially for create_user where we don't know the placement until we actually call get_pages(). I think for internal users like with create_lmem() setting the mm.region early still makes some sense? > > Perhaps what we really want is for i915_gem_object_migrate to > get_pages before it does the migration to ensure that pages exist. > The only call to i915_gem_object_migrate in the code-base today is in > the display code and it's immediately followed by pin_pages(). For > that matter, maybe the call we actually want is > i915_object_migrate_and_pin that does the whole lot. I guess the only downside is that we might end up doing a real migration, with mempy or the blitter vs just changing the preferred placement for later? I think just go with whatever you feel is the simplest for now. > > Thoughts? > > --Jason > > P.S. I'm going to go ahead and send another version with your other > comments addressed. We can keep this discussion going here for now. ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-22 9:49 ` Matthew Auld @ 2021-07-22 9:59 ` Matthew Auld 2021-08-04 8:00 ` Thomas Hellström 0 siblings, 1 reply; 23+ messages in thread From: Matthew Auld @ 2021-07-22 9:59 UTC (permalink / raw) To: Jason Ekstrand Cc: Thomas Hellström, Intel Graphics Development, Matthew Auld, ML dri-devel On Thu, 22 Jul 2021 at 10:49, Matthew Auld <matthew.william.auld@gmail.com> wrote: > > On Wed, 21 Jul 2021 at 21:11, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > On Mon, Jul 19, 2021 at 8:35 AM Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > > > > > > On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > On Fri, 16 Jul 2021 at 16:52, Matthew Auld > > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > > > On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > > > On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > > > > > > > > > <matthew.william.auld@gmail.com> wrote: > > > > > > > > > > > > > > > > > > > > On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > > > > > > > > > > > > > > > > > > > > > > Whenever we had a user object (n_placements > 0), we were ignoring > > > > > > > > > > > obj->mm.region and always putting obj->placements[0] as the requested > > > > > > > > > > > region. For LMEM+SMEM objects, this was causing them to get shoved into > > > > > > > > > > > LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > > > > > > > > > > > i915_gem_object_migrate(). > > > > > > > > > > > > > > > > > > > > i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > > > > > > > > > > requested region, so there shouldn't be an issue with migration right? > > > > > > > > > > Do you have some more details? > > > > > > > > > > > > > > > > > > With i915_ttm_migrate directly, no. But, in the last patch in the > > > > > > > > > series, we're trying to migrate LMEM+SMEM buffers into SMEM on > > > > > > > > > attach() and pin it there. This blows up in a very unexpected (IMO) > > > > > > > > > way. The flow goes something like this: > > > > > > > > > > > > > > > > > > - Client attempts a dma-buf import from another device > > > > > > > > > - In attach() we call i915_gem_object_migrate() which calls > > > > > > > > > i915_ttm_migrate() which migrates as requested. > > > > > > > > > - Once the migration is complete, we call i915_gem_object_pin_pages() > > > > > > > > > which calls i915_ttm_get_pages() which depends on > > > > > > > > > i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > > > > > > > > > > > > > > > > The mm.pages must be NULL here, otherwise it would just increment the > > > > > > > > pages_pin_count? > > > > > > > > > > > > Given that the test is using the ____four_underscores version, it > > > > > > doesn't have that check. However, this executes after we've done the > > > > > > dma-buf import which pinned pages. So we should definitely have > > > > > > pages. > > > > > > > > > > We shouldn't call ____four_underscores() if we might already have > > > > > pages though. Under non-TTM that would leak the pages, and in TTM we > > > > > might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > > > > > example nothing was moved. I take it we can't just call pin_pages()? > > > > > Four scary underscores usually means "don't call this in normal code". > > > > > > > > I've switched the ____four_underscores call to a __two_underscores in > > > > the selftests and it had no effect, good or bad. But, still, probably > > > > better to call that one. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Maybe the problem here is actually that our TTM code isn't respecting > > > > > > > > > obj->mm.pages_pin_count? > > > > > > > > > > > > > > > > I think if the resource is moved, we always nuke the mm.pages after > > > > > > > > being notified of the move. Also TTM is also not allowed to move > > > > > > > > pinned buffers. > > > > > > > > > > > > > > > > I guess if we are evicted/swapped, so assuming we are not holding the > > > > > > > > object lock, and it's not pinned, the future call to get_pages() will > > > > > > > > see mm.pages = NULL, even though the ttm_resource is still there, and > > > > > > > > because we prioritise the placements[0], instead of mm.region we end > > > > > > > > up moving it for no good reason. But in your case you are holding the > > > > > > > > lock, or it's pinned? Also is this just with the selftest, or > > > > > > > > something real? > > > > > > > > > > > > > > Or at least in the selftest I see ____i915_gem_object_get_pages() > > > > > > > which doesn't even consider the mm.pages AFAIK. > > > > > > > > > > > > The bogus migration is happening as part of the > > > > > > __i915_gem_object_get_pages() (2 __underscores) call in > > > > > > i915_gem_dmabuf_attach (see last patch). That code is attempting to > > > > > > migrate the BO to SMEM and then pin it there using the obvious calls > > > > > > to do so. However, in the pin_pages call, it gets implicitly migrated > > > > > > back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > > > > > > migrating things at all? > > > > > > > > > > Not sure yet, but __two_underscores() checks if > > > > > i915_gem_object_has_pages() before actually calling into > > > > > i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > > > > > some reason, so best guess is something to do with move_notify(). > > > > > > > > Did a bit of experimenting along those lines and added the following > > > > to the self-test BEFORE the export/import: > > > > > > > > i915_gem_object_lock(obj, NULL); > > > > err = __i915_gem_object_get_pages(obj); > > > > __i915_gem_object_unpin_pages(obj); > > > > i915_gem_object_unlock(obj); > > > > if (err) { > > > > pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); > > > > goto out_ret; > > > > } > > > > > > > > This seems to make the migration happen as expected without this > > > > patch. So it seems the problem only exists on buffers that haven't > > > > gotten any backing storage yet (if I'm understanding get_pages > > > > correctly). > > > > > > > > One potential work-around (not sure if this is a good idea or not!) > > > > would be to do this inside dmabuf_attach(). Is this reliable? Once > > > > it has pages will it always have pages? Or are there crazy races I > > > > need to be worried about here? > > > > > > It turns out that the i915_ttm_adjust_gem_after_move() call in > > > ttm_object_init will always update the mm.region to system memory(so > > > that it matches the ttm resource), which seems reasonable given the > > > default system placeholder thing, but does seem slightly iffy since we > > > haven't actually moved/allocated anything. > > > > > > So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since > > > mm.region == mr. Which ofc means when we actually call get_pages() all > > > that happens is that we allocate the pages in system memory(or without > > > this patch placements[0]). Also with this patch lmem+smem, will always > > > be placed in smem first, regardless of the placements ordering. > > > > > > For now we could maybe just split i915_ttm_adjust_gem_after_move() so > > > we skip the part which updates the mm.region here in the init portion, > > > since that should only happen when we try to place the object for > > > real? > > > > Doesn't that mean we would end up with obj->mm.region and > > obj->mm.res->mem_type are out-of-sync? That seems bad. I would think > > we'd want the two in sync at all times. > > It likely doesn't matter since all roads lead to i915_ttm_get_pages() > when we need to actually use the object? > > Also updating the mm.region in ttm_object_init() to reflect the dummy > ttm resource seems a little scary, since any existing is_lmem() check > now needs to happen after we place the object. Or at least the > existing callers(for kernel internal objects) might not have expected > that behaviour. Not sure if we checked all the callers. > > > > > It seems like the fundamental problem here is that, when it's created, > > the object isn't really in any memory region at all. While I don't > > think obj->mm.region == NULL is allowed or a good idea, it does seem > > closer to the ground truth. > > Yeah, seems reasonable, especially for create_user where we don't know > the placement until we actually call get_pages(). I think for internal > users like with create_lmem() setting the mm.region early still makes > some sense? > > > > > Perhaps what we really want is for i915_gem_object_migrate to > > get_pages before it does the migration to ensure that pages exist. > > The only call to i915_gem_object_migrate in the code-base today is in > > the display code and it's immediately followed by pin_pages(). For > > that matter, maybe the call we actually want is > > i915_object_migrate_and_pin that does the whole lot. > > I guess the only downside is that we might end up doing a real > migration, with mempy or the blitter vs just changing the preferred > placement for later? I think just go with whatever you feel is the > simplest for now. Another cheapo could be to drop the mr == mm.region noop, and just try to place the object at mr anyway? > > > > > Thoughts? > > > > --Jason > > > > P.S. I'm going to go ahead and send another version with your other > > comments addressed. We can keep this discussion going here for now. ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-07-22 9:59 ` Matthew Auld @ 2021-08-04 8:00 ` Thomas Hellström 2021-08-04 14:35 ` Daniel Vetter 0 siblings, 1 reply; 23+ messages in thread From: Thomas Hellström @ 2021-08-04 8:00 UTC (permalink / raw) To: Matthew Auld, Jason Ekstrand, Daniel Vetter Cc: Intel Graphics Development, ML dri-devel, Matthew Auld Hi, On 7/22/21 11:59 AM, Matthew Auld wrote: > On Thu, 22 Jul 2021 at 10:49, Matthew Auld > <matthew.william.auld@gmail.com> wrote: >> On Wed, 21 Jul 2021 at 21:11, Jason Ekstrand <jason@jlekstrand.net> wrote: >>> On Mon, Jul 19, 2021 at 8:35 AM Matthew Auld >>> <matthew.william.auld@gmail.com> wrote: >>>> On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: >>>>> On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld >>>>> <matthew.william.auld@gmail.com> wrote: >>>>>> On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: >>>>>>> On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld >>>>>>> <matthew.william.auld@gmail.com> wrote: >>>>>>>> On Fri, 16 Jul 2021 at 16:52, Matthew Auld >>>>>>>> <matthew.william.auld@gmail.com> wrote: >>>>>>>>> On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: >>>>>>>>>> On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld >>>>>>>>>> <matthew.william.auld@gmail.com> wrote: >>>>>>>>>>> On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: >>>>>>>>>>>> Whenever we had a user object (n_placements > 0), we were ignoring >>>>>>>>>>>> obj->mm.region and always putting obj->placements[0] as the requested >>>>>>>>>>>> region. For LMEM+SMEM objects, this was causing them to get shoved into >>>>>>>>>>>> LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, >>>>>>>>>>>> i915_gem_object_migrate(). >>>>>>>>>>> i915_ttm_migrate calls i915_ttm_place_from_region() directly with the >>>>>>>>>>> requested region, so there shouldn't be an issue with migration right? >>>>>>>>>>> Do you have some more details? >>>>>>>>>> With i915_ttm_migrate directly, no. But, in the last patch in the >>>>>>>>>> series, we're trying to migrate LMEM+SMEM buffers into SMEM on >>>>>>>>>> attach() and pin it there. This blows up in a very unexpected (IMO) >>>>>>>>>> way. The flow goes something like this: >>>>>>>>>> >>>>>>>>>> - Client attempts a dma-buf import from another device >>>>>>>>>> - In attach() we call i915_gem_object_migrate() which calls >>>>>>>>>> i915_ttm_migrate() which migrates as requested. >>>>>>>>>> - Once the migration is complete, we call i915_gem_object_pin_pages() >>>>>>>>>> which calls i915_ttm_get_pages() which depends on >>>>>>>>>> i915_ttm_placement_from_obj() and so migrates it right back to LMEM. >>>>>>>>> The mm.pages must be NULL here, otherwise it would just increment the >>>>>>>>> pages_pin_count? >>>>>>> Given that the test is using the ____four_underscores version, it >>>>>>> doesn't have that check. However, this executes after we've done the >>>>>>> dma-buf import which pinned pages. So we should definitely have >>>>>>> pages. >>>>>> We shouldn't call ____four_underscores() if we might already have >>>>>> pages though. Under non-TTM that would leak the pages, and in TTM we >>>>>> might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for >>>>>> example nothing was moved. I take it we can't just call pin_pages()? >>>>>> Four scary underscores usually means "don't call this in normal code". >>>>> I've switched the ____four_underscores call to a __two_underscores in >>>>> the selftests and it had no effect, good or bad. But, still, probably >>>>> better to call that one. >>>>> >>>>>>>>>> Maybe the problem here is actually that our TTM code isn't respecting >>>>>>>>>> obj->mm.pages_pin_count? >>>>>>>>> I think if the resource is moved, we always nuke the mm.pages after >>>>>>>>> being notified of the move. Also TTM is also not allowed to move >>>>>>>>> pinned buffers. >>>>>>>>> >>>>>>>>> I guess if we are evicted/swapped, so assuming we are not holding the >>>>>>>>> object lock, and it's not pinned, the future call to get_pages() will >>>>>>>>> see mm.pages = NULL, even though the ttm_resource is still there, and >>>>>>>>> because we prioritise the placements[0], instead of mm.region we end >>>>>>>>> up moving it for no good reason. But in your case you are holding the >>>>>>>>> lock, or it's pinned? Also is this just with the selftest, or >>>>>>>>> something real? >>>>>>>> Or at least in the selftest I see ____i915_gem_object_get_pages() >>>>>>>> which doesn't even consider the mm.pages AFAIK. >>>>>>> The bogus migration is happening as part of the >>>>>>> __i915_gem_object_get_pages() (2 __underscores) call in >>>>>>> i915_gem_dmabuf_attach (see last patch). That code is attempting to >>>>>>> migrate the BO to SMEM and then pin it there using the obvious calls >>>>>>> to do so. However, in the pin_pages call, it gets implicitly migrated >>>>>>> back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() >>>>>>> migrating things at all? >>>>>> Not sure yet, but __two_underscores() checks if >>>>>> i915_gem_object_has_pages() before actually calling into >>>>>> i915_ttm_get_pages(), so the mm.pages would have to be NULL here for >>>>>> some reason, so best guess is something to do with move_notify(). >>>>> Did a bit of experimenting along those lines and added the following >>>>> to the self-test BEFORE the export/import: >>>>> >>>>> i915_gem_object_lock(obj, NULL); >>>>> err = __i915_gem_object_get_pages(obj); >>>>> __i915_gem_object_unpin_pages(obj); >>>>> i915_gem_object_unlock(obj); >>>>> if (err) { >>>>> pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); >>>>> goto out_ret; >>>>> } >>>>> >>>>> This seems to make the migration happen as expected without this >>>>> patch. So it seems the problem only exists on buffers that haven't >>>>> gotten any backing storage yet (if I'm understanding get_pages >>>>> correctly). >>>>> >>>>> One potential work-around (not sure if this is a good idea or not!) >>>>> would be to do this inside dmabuf_attach(). Is this reliable? Once >>>>> it has pages will it always have pages? Or are there crazy races I >>>>> need to be worried about here? >>>> It turns out that the i915_ttm_adjust_gem_after_move() call in >>>> ttm_object_init will always update the mm.region to system memory(so >>>> that it matches the ttm resource), which seems reasonable given the >>>> default system placeholder thing, but does seem slightly iffy since we >>>> haven't actually moved/allocated anything. >>>> >>>> So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since >>>> mm.region == mr. Which ofc means when we actually call get_pages() all >>>> that happens is that we allocate the pages in system memory(or without >>>> this patch placements[0]). Also with this patch lmem+smem, will always >>>> be placed in smem first, regardless of the placements ordering. >>>> >>>> For now we could maybe just split i915_ttm_adjust_gem_after_move() so >>>> we skip the part which updates the mm.region here in the init portion, >>>> since that should only happen when we try to place the object for >>>> real? >>> Doesn't that mean we would end up with obj->mm.region and >>> obj->mm.res->mem_type are out-of-sync? That seems bad. I would think >>> we'd want the two in sync at all times. >> It likely doesn't matter since all roads lead to i915_ttm_get_pages() >> when we need to actually use the object? >> >> Also updating the mm.region in ttm_object_init() to reflect the dummy >> ttm resource seems a little scary, since any existing is_lmem() check >> now needs to happen after we place the object. Or at least the >> existing callers(for kernel internal objects) might not have expected >> that behaviour. Not sure if we checked all the callers. >> >>> It seems like the fundamental problem here is that, when it's created, >>> the object isn't really in any memory region at all. While I don't >>> think obj->mm.region == NULL is allowed or a good idea, it does seem >>> closer to the ground truth. >> Yeah, seems reasonable, especially for create_user where we don't know >> the placement until we actually call get_pages(). I think for internal >> users like with create_lmem() setting the mm.region early still makes >> some sense? >> >>> Perhaps what we really want is for i915_gem_object_migrate to >>> get_pages before it does the migration to ensure that pages exist. >>> The only call to i915_gem_object_migrate in the code-base today is in >>> the display code and it's immediately followed by pin_pages(). For >>> that matter, maybe the call we actually want is >>> i915_object_migrate_and_pin that does the whole lot. >> I guess the only downside is that we might end up doing a real >> migration, with mempy or the blitter vs just changing the preferred >> placement for later? I think just go with whatever you feel is the >> simplest for now. > Another cheapo could be to drop the mr == mm.region noop, and just try > to place the object at mr anyway? > There are a number of things to consider here, First, as Jason found out what's keeping thing from working as intended is that we actually call into TTM get_pages() after migration, since the object isn't populated with pages yet. That's indeed a bug. We should probably have migrate be migrate_and_populate(): Whatever kernel code decides to migrate needs to hold the object lock over the operation where data needs to be migrated or in the worst case call pin() under the lock which currently needs to be the case for dma-buf and display. If we blindly just look at obj->mm.region() in get_pages() then if an object with allowable placements in lmem and smem initially gets placed in lmem, and then evicted to smem it will never migrate back to lmem unless if there is an explicit i915_gem_object_migrate(), but again, that's perhaps what we want? I guess we need to more clearly define the migration policies; for example should we attempt to migrate evicted buffers back to lmem on each execbuf where they are referenced, even if they haven't lost their pages? On region dicrepance between gem and TTM there is a short DOC: section in i915_gem_ttm.c /Thomas >>> Thoughts? >>> >>> --Jason >>> >>> P.S. I'm going to go ahead and send another version with your other >>> comments addressed. We can keep this discussion going here for now. ^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj 2021-08-04 8:00 ` Thomas Hellström @ 2021-08-04 14:35 ` Daniel Vetter 0 siblings, 0 replies; 23+ messages in thread From: Daniel Vetter @ 2021-08-04 14:35 UTC (permalink / raw) To: Thomas Hellström Cc: Matthew Auld, Jason Ekstrand, Intel Graphics Development, ML dri-devel, Matthew Auld On Wed, Aug 4, 2021 at 10:00 AM Thomas Hellström <thomas.hellstrom@linux.intel.com> wrote: > > Hi, > > On 7/22/21 11:59 AM, Matthew Auld wrote: > > On Thu, 22 Jul 2021 at 10:49, Matthew Auld > > <matthew.william.auld@gmail.com> wrote: > >> On Wed, 21 Jul 2021 at 21:11, Jason Ekstrand <jason@jlekstrand.net> wrote: > >>> On Mon, Jul 19, 2021 at 8:35 AM Matthew Auld > >>> <matthew.william.auld@gmail.com> wrote: > >>>> On Fri, 16 Jul 2021 at 20:49, Jason Ekstrand <jason@jlekstrand.net> wrote: > >>>>> On Fri, Jul 16, 2021 at 1:45 PM Matthew Auld > >>>>> <matthew.william.auld@gmail.com> wrote: > >>>>>> On Fri, 16 Jul 2021 at 18:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > >>>>>>> On Fri, Jul 16, 2021 at 11:00 AM Matthew Auld > >>>>>>> <matthew.william.auld@gmail.com> wrote: > >>>>>>>> On Fri, 16 Jul 2021 at 16:52, Matthew Auld > >>>>>>>> <matthew.william.auld@gmail.com> wrote: > >>>>>>>>> On Fri, 16 Jul 2021 at 15:10, Jason Ekstrand <jason@jlekstrand.net> wrote: > >>>>>>>>>> On Fri, Jul 16, 2021 at 8:54 AM Matthew Auld > >>>>>>>>>> <matthew.william.auld@gmail.com> wrote: > >>>>>>>>>>> On Thu, 15 Jul 2021 at 23:39, Jason Ekstrand <jason@jlekstrand.net> wrote: > >>>>>>>>>>>> Whenever we had a user object (n_placements > 0), we were ignoring > >>>>>>>>>>>> obj->mm.region and always putting obj->placements[0] as the requested > >>>>>>>>>>>> region. For LMEM+SMEM objects, this was causing them to get shoved into > >>>>>>>>>>>> LMEM on every i915_ttm_get_pages() even when SMEM was requested by, say, > >>>>>>>>>>>> i915_gem_object_migrate(). > >>>>>>>>>>> i915_ttm_migrate calls i915_ttm_place_from_region() directly with the > >>>>>>>>>>> requested region, so there shouldn't be an issue with migration right? > >>>>>>>>>>> Do you have some more details? > >>>>>>>>>> With i915_ttm_migrate directly, no. But, in the last patch in the > >>>>>>>>>> series, we're trying to migrate LMEM+SMEM buffers into SMEM on > >>>>>>>>>> attach() and pin it there. This blows up in a very unexpected (IMO) > >>>>>>>>>> way. The flow goes something like this: > >>>>>>>>>> > >>>>>>>>>> - Client attempts a dma-buf import from another device > >>>>>>>>>> - In attach() we call i915_gem_object_migrate() which calls > >>>>>>>>>> i915_ttm_migrate() which migrates as requested. > >>>>>>>>>> - Once the migration is complete, we call i915_gem_object_pin_pages() > >>>>>>>>>> which calls i915_ttm_get_pages() which depends on > >>>>>>>>>> i915_ttm_placement_from_obj() and so migrates it right back to LMEM. > >>>>>>>>> The mm.pages must be NULL here, otherwise it would just increment the > >>>>>>>>> pages_pin_count? > >>>>>>> Given that the test is using the ____four_underscores version, it > >>>>>>> doesn't have that check. However, this executes after we've done the > >>>>>>> dma-buf import which pinned pages. So we should definitely have > >>>>>>> pages. > >>>>>> We shouldn't call ____four_underscores() if we might already have > >>>>>> pages though. Under non-TTM that would leak the pages, and in TTM we > >>>>>> might hit the WARN_ON(mm->pages) in __i915_ttm_get_pages(), if for > >>>>>> example nothing was moved. I take it we can't just call pin_pages()? > >>>>>> Four scary underscores usually means "don't call this in normal code". > >>>>> I've switched the ____four_underscores call to a __two_underscores in > >>>>> the selftests and it had no effect, good or bad. But, still, probably > >>>>> better to call that one. > >>>>> > >>>>>>>>>> Maybe the problem here is actually that our TTM code isn't respecting > >>>>>>>>>> obj->mm.pages_pin_count? > >>>>>>>>> I think if the resource is moved, we always nuke the mm.pages after > >>>>>>>>> being notified of the move. Also TTM is also not allowed to move > >>>>>>>>> pinned buffers. > >>>>>>>>> > >>>>>>>>> I guess if we are evicted/swapped, so assuming we are not holding the > >>>>>>>>> object lock, and it's not pinned, the future call to get_pages() will > >>>>>>>>> see mm.pages = NULL, even though the ttm_resource is still there, and > >>>>>>>>> because we prioritise the placements[0], instead of mm.region we end > >>>>>>>>> up moving it for no good reason. But in your case you are holding the > >>>>>>>>> lock, or it's pinned? Also is this just with the selftest, or > >>>>>>>>> something real? > >>>>>>>> Or at least in the selftest I see ____i915_gem_object_get_pages() > >>>>>>>> which doesn't even consider the mm.pages AFAIK. > >>>>>>> The bogus migration is happening as part of the > >>>>>>> __i915_gem_object_get_pages() (2 __underscores) call in > >>>>>>> i915_gem_dmabuf_attach (see last patch). That code is attempting to > >>>>>>> migrate the BO to SMEM and then pin it there using the obvious calls > >>>>>>> to do so. However, in the pin_pages call, it gets implicitly migrated > >>>>>>> back to LMEM thanks to i915_ttm_get_pages(). Why is _get_pages() > >>>>>>> migrating things at all? > >>>>>> Not sure yet, but __two_underscores() checks if > >>>>>> i915_gem_object_has_pages() before actually calling into > >>>>>> i915_ttm_get_pages(), so the mm.pages would have to be NULL here for > >>>>>> some reason, so best guess is something to do with move_notify(). > >>>>> Did a bit of experimenting along those lines and added the following > >>>>> to the self-test BEFORE the export/import: > >>>>> > >>>>> i915_gem_object_lock(obj, NULL); > >>>>> err = __i915_gem_object_get_pages(obj); > >>>>> __i915_gem_object_unpin_pages(obj); > >>>>> i915_gem_object_unlock(obj); > >>>>> if (err) { > >>>>> pr_err("__i915_gem_object_get_pages failed with err=%d\n", err); > >>>>> goto out_ret; > >>>>> } > >>>>> > >>>>> This seems to make the migration happen as expected without this > >>>>> patch. So it seems the problem only exists on buffers that haven't > >>>>> gotten any backing storage yet (if I'm understanding get_pages > >>>>> correctly). > >>>>> > >>>>> One potential work-around (not sure if this is a good idea or not!) > >>>>> would be to do this inside dmabuf_attach(). Is this reliable? Once > >>>>> it has pages will it always have pages? Or are there crazy races I > >>>>> need to be worried about here? > >>>> It turns out that the i915_ttm_adjust_gem_after_move() call in > >>>> ttm_object_init will always update the mm.region to system memory(so > >>>> that it matches the ttm resource), which seems reasonable given the > >>>> default system placeholder thing, but does seem slightly iffy since we > >>>> haven't actually moved/allocated anything. > >>>> > >>>> So effectively i915_ttm_migrate(SYSTEM) becomes a noop here since > >>>> mm.region == mr. Which ofc means when we actually call get_pages() all > >>>> that happens is that we allocate the pages in system memory(or without > >>>> this patch placements[0]). Also with this patch lmem+smem, will always > >>>> be placed in smem first, regardless of the placements ordering. > >>>> > >>>> For now we could maybe just split i915_ttm_adjust_gem_after_move() so > >>>> we skip the part which updates the mm.region here in the init portion, > >>>> since that should only happen when we try to place the object for > >>>> real? > >>> Doesn't that mean we would end up with obj->mm.region and > >>> obj->mm.res->mem_type are out-of-sync? That seems bad. I would think > >>> we'd want the two in sync at all times. > >> It likely doesn't matter since all roads lead to i915_ttm_get_pages() > >> when we need to actually use the object? > >> > >> Also updating the mm.region in ttm_object_init() to reflect the dummy > >> ttm resource seems a little scary, since any existing is_lmem() check > >> now needs to happen after we place the object. Or at least the > >> existing callers(for kernel internal objects) might not have expected > >> that behaviour. Not sure if we checked all the callers. > >> > >>> It seems like the fundamental problem here is that, when it's created, > >>> the object isn't really in any memory region at all. While I don't > >>> think obj->mm.region == NULL is allowed or a good idea, it does seem > >>> closer to the ground truth. > >> Yeah, seems reasonable, especially for create_user where we don't know > >> the placement until we actually call get_pages(). I think for internal > >> users like with create_lmem() setting the mm.region early still makes > >> some sense? > >> > >>> Perhaps what we really want is for i915_gem_object_migrate to > >>> get_pages before it does the migration to ensure that pages exist. > >>> The only call to i915_gem_object_migrate in the code-base today is in > >>> the display code and it's immediately followed by pin_pages(). For > >>> that matter, maybe the call we actually want is > >>> i915_object_migrate_and_pin that does the whole lot. > >> I guess the only downside is that we might end up doing a real > >> migration, with mempy or the blitter vs just changing the preferred > >> placement for later? I think just go with whatever you feel is the > >> simplest for now. > > Another cheapo could be to drop the mr == mm.region noop, and just try > > to place the object at mr anyway? > > > There are a number of things to consider here, > > First, as Jason found out what's keeping thing from working as intended > is that we actually call into TTM get_pages() after migration, since the > object isn't populated with pages yet. That's indeed a bug. > > We should probably have migrate be migrate_and_populate(): Whatever > kernel code decides to migrate needs to hold the object lock over the > operation where data needs to be migrated or in the worst case call > pin() under the lock which currently needs to be the case for dma-buf > and display. > > If we blindly just look at obj->mm.region() in get_pages() then if an > object with allowable placements in lmem and smem initially gets placed > in lmem, and then evicted to smem it will never migrate back to lmem > unless if there is an explicit i915_gem_object_migrate(), but again, > that's perhaps what we want? I guess we need to more clearly define the > migration policies; for example should we attempt to migrate evicted > buffers back to lmem on each execbuf where they are referenced, even if > they haven't lost their pages? Looking at amdgpu things are indeed complicated: - mmap adds some hints that cpu access is preferred (iirc at least) so that the unmappable vram problems aren't too awful - execbuf adds vram to the non-evict placement list whenever that makes sense (i.e. preferred place and no inferred hint like mmap access countering that) - for eviction there's a ratelimit, to make sure we're not thrashing terribly and spending all the gpu time moving buffers around with the copy engine Maybe another interim strategy would be to only evict non-busy buffers, not sure ttm supports that already. We definitely don't want to unconditionally force all buffers into lmem on every execbuf. -Daniel > On region dicrepance between gem and TTM there is a short DOC: section > in i915_gem_ttm.c > > /Thomas > > > >>> Thoughts? > >>> > >>> --Jason > >>> > >>> P.S. I'm going to go ahead and send another version with your other > >>> comments addressed. We can keep this discussion going here for now. -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch ^ permalink raw reply [flat|nested] 23+ messages in thread
end of thread, other threads:[~2021-08-04 14:35 UTC | newest] Thread overview: 23+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2021-07-21 20:13 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v8) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 1/7] drm/i915/gem: Check object_can_migrate from object_migrate Jason Ekstrand 2021-07-21 20:13 ` [PATCH 2/7] drm/i915/gem: Refactor placement setup for i915_gem_object_create* (v2) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 3/7] drm/i915/gem: Call i915_gem_flush_free_objects() in i915_gem_dumb_create() Jason Ekstrand 2021-07-21 20:13 ` [PATCH 4/7] drm/i915/gem: Unify user object creation (v3) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand 2021-07-21 20:13 ` [PATCH 6/7] drm/i915/gem: Correct the locking and pin pattern for dma-buf (v8) Jason Ekstrand 2021-07-21 20:13 ` [PATCH 7/7] drm/i915/gem: Migrate to system at dma-buf attach time (v7) Jason Ekstrand -- strict thread matches above, loose matches on Subject: below -- 2021-07-16 14:14 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device (v7) Jason Ekstrand 2021-07-16 14:14 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand 2021-07-15 22:38 [PATCH 0/7] drm/i915: Migrate memory to SMEM when imported cross-device Jason Ekstrand 2021-07-15 22:38 ` [PATCH 5/7] drm/i915/gem/ttm: Respect the objection region in placement_from_obj Jason Ekstrand 2021-07-16 13:54 ` Matthew Auld 2021-07-16 14:10 ` Jason Ekstrand 2021-07-16 15:52 ` Matthew Auld 2021-07-16 16:00 ` Matthew Auld 2021-07-16 17:38 ` Jason Ekstrand 2021-07-16 18:44 ` Matthew Auld 2021-07-16 19:49 ` Jason Ekstrand 2021-07-19 13:34 ` Matthew Auld 2021-07-21 20:11 ` Jason Ekstrand 2021-07-22 9:49 ` Matthew Auld 2021-07-22 9:59 ` Matthew Auld 2021-08-04 8:00 ` Thomas Hellström 2021-08-04 14:35 ` Daniel Vetter
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).