On Mon, Jun 14, 2021 at 09:45:37PM +0900, Sergey Senozhatsky wrote:
Hi,
We are observing some user-space crashes (sigabort, segfaults etc.) under moderate memory pressure (pretty far from severe pressure) which have one thing in common - restrictive GFP mask in setup_scratch_page().
For instance, (stable 4.19) drivers/gpu/drm/i915/i915_gem_gtt.c
(trimmed down version)
static int gen8_init_scratch(struct i915_address_space *vm) { setup_scratch_page(vm, __GFP_HIGHMEM);
vm->scratch_pt = alloc_pt(vm); vm->scratch_pd = alloc_pd(vm); if (use_4lvl(vm)) { vm->scratch_pdp = alloc_pdp(vm); }
}
gen8_init_scratch() function puts a rather inconsistent restrictions on mm.
Looking at it line by line:
setup_scratch_page() uses very restrictive gfp mask: __GFP_HIGHMEM | __GFP_ZERO | __GFP_RETRY_MAYFAIL
it doesn't try to reclaim anything and fails almost immediately.
alloc_pt() - uses more permissive gfp mask: GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN
alloc_pd() - likewise: GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN
alloc_pdp() - very permissive gfp mask: GFP_KERNEL
So can all allocations in gen8_init_scratch() use GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN
Yeah that looks all fairly broken tbh. The only thing I didn't know was that GFP_DMA32 wasn't a full gfp mask with reclaim bits set as needed. I guess it would be clearer if we use GFP_KERNEL | __GFP_DMA32 for these.
The commit that introduced a lot of this, including I915_GFP_ALLOW_FAIL seems to be
commit 1abb70f5955d1a9021f96359a2c6502ca569b68d Author: Chris Wilson chris@chris-wilson.co.uk Date: Tue May 22 09:36:43 2018 +0100
drm/i915/gtt: Allow pagedirectory allocations to fail
which used a selftest as justification, not real world workloads, so looks rather dubious.
Adding Matt Auld to this thread, maybe he has ideas.
Thanks, Daniel
?
E.g.
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c index a12430187108..e862680b9c93 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c @@ -792,7 +792,7 @@ alloc_pdp(struct i915_address_space *vm)
GEM_BUG_ON(!use_4lvl(vm));
pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
pdp = kzalloc(sizeof(*pdp), I915_GFP_ALLOW_FAIL); if (!pdp) return ERR_PTR(-ENOMEM);
@@ -1262,7 +1262,7 @@ static int gen8_init_scratch(struct i915_address_space *vm) { int ret;
ret = setup_scratch_page(vm, __GFP_HIGHMEM);
ret = setup_scratch_page(vm, GFP_KERNEL | __GFP_HIGHMEM); if (ret) return ret;
@@ -1972,7 +1972,7 @@ static int gen6_ppgtt_init_scratch(struct gen6_hw_ppgtt *ppgtt) u32 pde; int ret;
ret = setup_scratch_page(vm, __GFP_HIGHMEM);
ret = setup_scratch_page(vm, GFP_KERNEL | __GFP_HIGHMEM); if (ret) return ret;
@@ -3078,7 +3078,7 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return -ENOMEM; }
ret = setup_scratch_page(&ggtt->vm, GFP_DMA32);
ret = setup_scratch_page(&ggtt->vm, GFP_KERNEL | GFP_DMA32); if (ret) { DRM_ERROR("Scratch setup failed\n"); /* iounmap will also get called at remove, but meh */
It's quite similar on stable 5.4 - setup_scratch_page() uses restrictive gfp mask again.
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c index f614646ed3f9..99d78b1052df 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c @@ -1378,7 +1378,7 @@ static int gen8_init_scratch(struct i915_address_space *vm) return 0; }
ret = setup_scratch_page(vm, __GFP_HIGHMEM);
ret = setup_scratch_page(vm, GFP_KERNEL | __GFP_HIGHMEM); if (ret) return ret;
@@ -1753,7 +1753,7 @@ static int gen6_ppgtt_init_scratch(struct gen6_ppgtt *ppgtt) struct i915_page_directory * const pd = ppgtt->base.pd; int ret;
ret = setup_scratch_page(vm, __GFP_HIGHMEM);
ret = setup_scratch_page(vm, GFP_KERNEL | __GFP_HIGHMEM); if (ret) return ret;
@@ -2860,7 +2860,7 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return -ENOMEM; }
ret = setup_scratch_page(&ggtt->vm, GFP_DMA32);
ret = setup_scratch_page(&ggtt->vm, GFP_KERNEL | GFP_DMA32); if (ret) { DRM_ERROR("Scratch setup failed\n"); /* iounmap will also get called at remove, but meh */