| Message ID | 1393559258-1405-1-git-send-email-benjamin.widawsky@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, 2014-02-27 at 19:47 -0800, Ben Widawsky wrote: > We normally clear the page tables as one of the first things during > initialization. They are however wired up (and potentially valid) before > we clear them. I might be missing something, but afaics the page directories/tables are not in use until after ppgtt->enable()/mm_switch() is called on them, which is after the clear_range() call. I'd understand if it's about leaving uninitialized stuff _after_ clear_range() is called. But I think because of the 1G size alignment for ppgtt that's not possible either. --Imre > To prevent the GPU from doing anything we might later regret, simply get > zeroed pages, which always mean invalid on all GENs. > > NOTE: that a similar paranoia could be applied to GGTT via making sure > all entries are invalid ASAP. I think the extra work required to fix > such a BIOS bug is unwarranted until proven necessary. > > v2: Remove useless GFP_ZERO in the kcallocs > > Signed-off-by: Ben Widawsky <ben@bwidawsk.net> > --- > drivers/gpu/drm/i915/i915_gem_gtt.c | 7 ++++--- > 1 file changed, 4 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c > index 0c27d8a..5e3957e 100644 > --- a/drivers/gpu/drm/i915/i915_gem_gtt.c > +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c > @@ -359,7 +359,7 @@ static struct page **__gen8_alloc_page_tables(void) > return ERR_PTR(-ENOMEM); > > for (i = 0; i < GEN8_PDES_PER_PAGE; i++) { > - pt_pages[i] = alloc_page(GFP_KERNEL); > + pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); > if (!pt_pages[i]) > goto bail; > } > @@ -421,7 +421,8 @@ static int gen8_ppgtt_allocate_dma(struct i915_hw_ppgtt *ppgtt) > static int gen8_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt, > const int max_pdp) > { > - ppgtt->pd_pages = alloc_pages(GFP_KERNEL, get_order(max_pdp << PAGE_SHIFT)); > + ppgtt->pd_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, > + get_order(max_pdp << PAGE_SHIFT)); > if (!ppgtt->pd_pages) > return -ENOMEM; > > @@ -1021,7 +1022,7 @@ static int gen6_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt) > return -ENOMEM; > > for (i = 0; i < ppgtt->num_pd_entries; i++) { > - ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL); > + ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); > if (!ppgtt->pt_pages[i]) { > gen6_ppgtt_free(ppgtt); > return -ENOMEM;
On Wed, Mar 05, 2014 at 06:32:06PM +0200, Imre Deak wrote: > On Thu, 2014-02-27 at 19:47 -0800, Ben Widawsky wrote: > > We normally clear the page tables as one of the first things during > > initialization. They are however wired up (and potentially valid) before > > we clear them. > > I might be missing something, but afaics the page directories/tables are > not in use until after ppgtt->enable()/mm_switch() is called on them, > which is after the clear_range() call. > > I'd understand if it's about leaving uninitialized stuff _after_ > clear_range() is called. But I think because of the 1G size alignment > for ppgtt that's not possible either. > > --Imre The only case I was able to fathom was if we accidentally connect a PDE before we populate the page table. I felt it was a rather harmless patch though. I do agree with the IRC conversation that it shouldn't happen. It was in lines with the same reason of why we never BUG_ON. > > > To prevent the GPU from doing anything we might later regret, simply get > > zeroed pages, which always mean invalid on all GENs. > > > > NOTE: that a similar paranoia could be applied to GGTT via making sure > > all entries are invalid ASAP. I think the extra work required to fix > > such a BIOS bug is unwarranted until proven necessary. > > > > v2: Remove useless GFP_ZERO in the kcallocs > > > > Signed-off-by: Ben Widawsky <ben@bwidawsk.net> > > --- > > drivers/gpu/drm/i915/i915_gem_gtt.c | 7 ++++--- > > 1 file changed, 4 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c > > index 0c27d8a..5e3957e 100644 > > --- a/drivers/gpu/drm/i915/i915_gem_gtt.c > > +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c > > @@ -359,7 +359,7 @@ static struct page **__gen8_alloc_page_tables(void) > > return ERR_PTR(-ENOMEM); > > > > for (i = 0; i < GEN8_PDES_PER_PAGE; i++) { > > - pt_pages[i] = alloc_page(GFP_KERNEL); > > + pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); > > if (!pt_pages[i]) > > goto bail; > > } > > @@ -421,7 +421,8 @@ static int gen8_ppgtt_allocate_dma(struct i915_hw_ppgtt *ppgtt) > > static int gen8_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt, > > const int max_pdp) > > { > > - ppgtt->pd_pages = alloc_pages(GFP_KERNEL, get_order(max_pdp << PAGE_SHIFT)); > > + ppgtt->pd_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, > > + get_order(max_pdp << PAGE_SHIFT)); > > if (!ppgtt->pd_pages) > > return -ENOMEM; > > > > @@ -1021,7 +1022,7 @@ static int gen6_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt) > > return -ENOMEM; > > > > for (i = 0; i < ppgtt->num_pd_entries; i++) { > > - ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL); > > + ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); > > if (!ppgtt->pt_pages[i]) { > > gen6_ppgtt_free(ppgtt); > > return -ENOMEM; >
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c index 0c27d8a..5e3957e 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c @@ -359,7 +359,7 @@ static struct page **__gen8_alloc_page_tables(void) return ERR_PTR(-ENOMEM); for (i = 0; i < GEN8_PDES_PER_PAGE; i++) { - pt_pages[i] = alloc_page(GFP_KERNEL); + pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); if (!pt_pages[i]) goto bail; } @@ -421,7 +421,8 @@ static int gen8_ppgtt_allocate_dma(struct i915_hw_ppgtt *ppgtt) static int gen8_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt, const int max_pdp) { - ppgtt->pd_pages = alloc_pages(GFP_KERNEL, get_order(max_pdp << PAGE_SHIFT)); + ppgtt->pd_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, + get_order(max_pdp << PAGE_SHIFT)); if (!ppgtt->pd_pages) return -ENOMEM; @@ -1021,7 +1022,7 @@ static int gen6_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt) return -ENOMEM; for (i = 0; i < ppgtt->num_pd_entries; i++) { - ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL); + ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO); if (!ppgtt->pt_pages[i]) { gen6_ppgtt_free(ppgtt); return -ENOMEM;
We normally clear the page tables as one of the first things during initialization. They are however wired up (and potentially valid) before we clear them. To prevent the GPU from doing anything we might later regret, simply get zeroed pages, which always mean invalid on all GENs. NOTE: that a similar paranoia could be applied to GGTT via making sure all entries are invalid ASAP. I think the extra work required to fix such a BIOS bug is unwarranted until proven necessary. v2: Remove useless GFP_ZERO in the kcallocs Signed-off-by: Ben Widawsky <ben@bwidawsk.net> --- drivers/gpu/drm/i915/i915_gem_gtt.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)