[Intel-gfx] [PATCH 5/9] drm/i915/bdw: Reorganize PT allocations
Imre Deak
imre.deak at intel.com
Wed Feb 19 22:20:51 CET 2014
On Wed, 2014-02-19 at 13:06 -0800, Ben Widawsky wrote:
> On Wed, Feb 19, 2014 at 09:11:46PM +0200, Imre Deak wrote:
> > On Wed, 2014-02-12 at 14:28 -0800, Ben Widawsky wrote:
> > > The previous allocation mechanism would get 2 contiguous allocations,
> > > one for the page directories, and one for the page tables. As each page
> > > table is 1 page, and there are 512 of these per page directory, this
> > > goes to 1MB. An unfriendly request at best. Worse still, our HW now
> > ---^
> > Fwiw, 2MB.
>
> Thanks.
>
> >
> > > supports 4 page directories, and a 2MB allocation is not allowed.
> > >
> > > In order to fix this, this patch attempts to split up each page table
> > > allocation into a single, discrete allocation. There is nothing really
> > > fancy about the patch itself, it just has to manage an extra pointer
> > > indirection, and have a fancier bit of logic to free up the pages.
> > >
> > > To accommodate some of the added complexity, two new helpers are
> > > introduced to allocate, and free the page table pages.
> > >
> > > NOTE: I really wanted to split the way we do allocations, and the way in
> > > which we identify the page table/page directory being used. I found
> > > splitting this functionality up to be too unwieldy. I apologize in
> > > advance to the reviewer. I'd recommend looking at the result, rather
> > > than the diff.
> > >
> > > v2/NOTE2: This patch predated commit:
> > > 6f1cc993518462ccf039e195fabd47e7aa5bfd13
> > > Author: Chris Wilson <chris at chris-wilson.co.uk>
> > > Date: Tue Dec 31 15:50:31 2013 +0000
> > >
> > > drm/i915: Avoid dereference past end of page arr
> > >
> > > It fixed the same issue as that patch, but because of the limbo state of
> > > PPGTT, Chris patch was merged instead. The excess churn is a result of
> > > my using my original patch, which has my preferred naming. Primarily
> > > act_* is changed to which_*, but it's mostly the same otherwise. I've
> > > kept the convention Chris used for the pte wrap (I had something
> > > slightly different, and broken - but fixable)
> > >
> > > Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> > > ---
> > > drivers/gpu/drm/i915/i915_drv.h | 5 +-
> > > drivers/gpu/drm/i915/i915_gem_gtt.c | 127 ++++++++++++++++++++++++++++--------
> > > 2 files changed, 103 insertions(+), 29 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > > index 2ebad96..d9a6327 100644
> > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > @@ -691,6 +691,7 @@ struct i915_gtt {
> > > };
> > > #define gtt_total_entries(gtt) ((gtt).base.total >> PAGE_SHIFT)
> > >
> > > +#define GEN8_LEGACY_PDPS 4
> > > struct i915_hw_ppgtt {
> > > struct i915_address_space base;
> > > struct kref ref;
> > > @@ -698,14 +699,14 @@ struct i915_hw_ppgtt {
> > > unsigned num_pd_entries;
> > > union {
> > > struct page **pt_pages;
> > > - struct page *gen8_pt_pages;
> > > + struct page **gen8_pt_pages[GEN8_LEGACY_PDPS];
> > > };
> > > struct page *pd_pages;
> > > int num_pd_pages;
> > > int num_pt_pages;
> > > union {
> > > uint32_t pd_offset;
> > > - dma_addr_t pd_dma_addr[4];
> > > + dma_addr_t pd_dma_addr[GEN8_LEGACY_PDPS];
> > > };
> > > union {
> > > dma_addr_t *pt_dma_addr;
> > > diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> > > index 5bfc6ff..5299acc 100644
> > > --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> > > +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> > > @@ -64,7 +64,19 @@ typedef gen8_gtt_pte_t gen8_ppgtt_pde_t;
> > >
> > > #define GEN8_PTES_PER_PAGE (PAGE_SIZE / sizeof(gen8_gtt_pte_t))
> > > #define GEN8_PDES_PER_PAGE (PAGE_SIZE / sizeof(gen8_ppgtt_pde_t))
> > > -#define GEN8_LEGACY_PDPS 4
> > > +
> > > +/* GEN8 legacy style addressis defined as a 3 level page table:
> > > + * 31:30 | 29:21 | 20:12 | 11:0
> > > + * PDPE | PDE | PTE | offset
> > > + * The difference as compared to normal x86 3 level page table is the PDPEs are
> > > + * programmed via register.
> > > + */
> > > +#define GEN8_PDPE_SHIFT 30
> > > +#define GEN8_PDPE_MASK 0x3
> > > +#define GEN8_PDE_SHIFT 21
> > > +#define GEN8_PDE_MASK 0x1ff
> > > +#define GEN8_PTE_SHIFT 12
> > > +#define GEN8_PTE_MASK 0x1ff
> > >
> > > #define PPAT_UNCACHED_INDEX (_PAGE_PWT | _PAGE_PCD)
> > > #define PPAT_CACHED_PDE_INDEX 0 /* WB LLC */
> > > @@ -261,32 +273,36 @@ static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
> > > struct i915_hw_ppgtt *ppgtt =
> > > container_of(vm, struct i915_hw_ppgtt, base);
> > > gen8_gtt_pte_t *pt_vaddr, scratch_pte;
> > > - unsigned first_entry = start >> PAGE_SHIFT;
> > > + unsigned which_pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
> > > + unsigned which_pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
> > > + unsigned which_pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
> > > unsigned num_entries = length >> PAGE_SHIFT;
> > > - unsigned act_pt = first_entry / GEN8_PTES_PER_PAGE;
> > > - unsigned first_pte = first_entry % GEN8_PTES_PER_PAGE;
> > > unsigned last_pte, i;
> > >
> > > scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
> > > I915_CACHE_LLC, use_scratch);
> > >
> > > while (num_entries) {
> > > - struct page *page_table = &ppgtt->gen8_pt_pages[act_pt];
> > > + struct page *page_table = ppgtt->gen8_pt_pages[which_pdpe][which_pde];
> > >
> > > - last_pte = first_pte + num_entries;
> > > + last_pte = which_pte + num_entries;
> > > if (last_pte > GEN8_PTES_PER_PAGE)
> > > last_pte = GEN8_PTES_PER_PAGE;
> > >
> > > pt_vaddr = kmap_atomic(page_table);
> > >
> > > - for (i = first_pte; i < last_pte; i++)
> > > + for (i = which_pte; i < last_pte; i++) {
> > > pt_vaddr[i] = scratch_pte;
> > > + num_entries--;
> > > + BUG_ON(num_entries < 0);
> >
> > num_entries is unsigned.
>
> This was already changed per Chris' request.
>
> >
> > > + }
> > >
> > > kunmap_atomic(pt_vaddr);
> > >
> > > - num_entries -= last_pte - first_pte;
> > > - first_pte = 0;
> > > - act_pt++;
> > > + which_pte = 0;
> > > + if (which_pde + 1 == GEN8_PDES_PER_PAGE)
> > > + which_pdpe++;
> > > + which_pde = (which_pde + 1) & GEN8_PDE_MASK;
> > > }
> > > }
> > >
> > > @@ -298,39 +314,57 @@ static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
> > > struct i915_hw_ppgtt *ppgtt =
> > > container_of(vm, struct i915_hw_ppgtt, base);
> > > gen8_gtt_pte_t *pt_vaddr;
> > > - unsigned first_entry = start >> PAGE_SHIFT;
> > > - unsigned act_pt = first_entry / GEN8_PTES_PER_PAGE;
> > > - unsigned act_pte = first_entry % GEN8_PTES_PER_PAGE;
> > > + unsigned which_pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
> > > + unsigned which_pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
> > > + unsigned which_pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
> > > struct sg_page_iter sg_iter;
> > >
> > > pt_vaddr = NULL;
> > > +
> > > for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
> > > + if (WARN_ON(which_pdpe >= GEN8_LEGACY_PDPS))
> > > + break;
> > > +
> > > if (pt_vaddr == NULL)
> > > - pt_vaddr = kmap_atomic(&ppgtt->gen8_pt_pages[act_pt]);
> > > + pt_vaddr = kmap_atomic(ppgtt->gen8_pt_pages[which_pdpe][which_pde]);
> > >
> > > - pt_vaddr[act_pte] =
> > > + pt_vaddr[which_pte] =
> > > gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
> > > cache_level, true);
> > > - if (++act_pte == GEN8_PTES_PER_PAGE) {
> > > + if (++which_pte == GEN8_PTES_PER_PAGE) {
> > > kunmap_atomic(pt_vaddr);
> > > pt_vaddr = NULL;
> > > - act_pt++;
> > > - act_pte = 0;
> > > + if (which_pde + 1 == GEN8_PDES_PER_PAGE)
> > > + which_pdpe++;
> >
> > Afaics which_pde = (which_pde + 1) & GEN8_PDE_MASK; is missing here.
> >
>
> This was already changed per Chris' request.
>
> > > + which_pte = 0;
> >
> >
> >
> > > }
> > > }
> > > if (pt_vaddr)
> > > kunmap_atomic(pt_vaddr);
> > > }
> > >
> > > -static void gen8_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
> > > +static void gen8_free_page_tables(struct page **pt_pages)
> > > +{
> > > + int i;
> > > +
> > > + if (pt_pages == NULL)
> > > + return;
> > > +
> > > + for (i = 0; i < GEN8_PDES_PER_PAGE; i++)
> > > + if (pt_pages[i])
> > > + __free_pages(pt_pages[i], 0);
> > > +}
> > > +
> > > +static void gen8_ppgtt_free(const struct i915_hw_ppgtt *ppgtt)
> > > {
> > > int i;
> > >
> > > - for (i = 0; i < ppgtt->num_pd_pages ; i++)
> > > + for (i = 0; i < ppgtt->num_pd_pages; i++) {
> > > + gen8_free_page_tables(ppgtt->gen8_pt_pages[i]);
> > > kfree(ppgtt->gen8_pt_dma_addr[i]);
> > > + }
> > >
> > > kfree(ppgtt->gen8_pt_dma_addr);
> > > - __free_pages(ppgtt->gen8_pt_pages, get_order(ppgtt->num_pt_pages << PAGE_SHIFT));
> > > __free_pages(ppgtt->pd_pages, get_order(ppgtt->num_pd_pages << PAGE_SHIFT));
> > > }
> > >
> > > @@ -369,20 +403,59 @@ static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
> > > gen8_ppgtt_free(ppgtt);
> > > }
> > >
> > > +static struct page **__gen8_alloc_page_tables(void)
> > > +{
> > > + struct page **pt_pages;
> > > + int i;
> > > +
> > > + pt_pages = kcalloc(GEN8_PDES_PER_PAGE, sizeof(struct page *), GFP_KERNEL);
> > > + if (!pt_pages)
> > > + return ERR_PTR(-ENOMEM);
> > > +
> > > + for (i = 0; i < GEN8_PDES_PER_PAGE; i++) {
> > > + pt_pages[i] = alloc_page(GFP_KERNEL);
> > > + if (!pt_pages[i])
> > > + goto bail;
> > > + }
> > > +
> > > + return pt_pages;
> > > +
> > > +bail:
> > > + gen8_free_page_tables(pt_pages);
> > > + kfree(pt_pages);
> > > + return ERR_PTR(-ENOMEM);
> > > +}
> > > +
> > > static int gen8_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt,
> > > const int max_pdp)
> > > {
> > > - struct page *pt_pages;
> > > + struct page **pt_pages[GEN8_LEGACY_PDPS];
> > > const int num_pt_pages = GEN8_PDES_PER_PAGE * max_pdp;
> > > + int i, ret;
> > >
> > > - pt_pages = alloc_pages(GFP_KERNEL, get_order(num_pt_pages << PAGE_SHIFT));
> > > - if (!pt_pages)
> > > - return -ENOMEM;
> > > + for (i = 0; i < max_pdp; i++) {
> > > + pt_pages[i] = __gen8_alloc_page_tables();
> > > + if (IS_ERR(pt_pages[i])) {
> > > + ret = PTR_ERR(pt_pages[i]);
> > > + goto unwind_out;
> > > + }
> > > + }
> > > +
> > > + /* NB: Avoid touching gen8_pt_pages until last to keep the allocation,
> > > + * "atomic" - for cleanup purposes.
> > > + */
> > > + for (i = 0; i < max_pdp; i++)
> > > + ppgtt->gen8_pt_pages[i] = pt_pages[i];
> > >
> > > - ppgtt->gen8_pt_pages = pt_pages;
> > > ppgtt->num_pt_pages = 1 << get_order(num_pt_pages << PAGE_SHIFT);
> > >
> > > return 0;
> > > +
> > > +unwind_out:
> > > + while(i--)
> > > + gen8_free_page_tables(pt_pages[i]);
> >
> > I guess Ville commented on this issue, but pt_pages would be leaked
> > here.
>
> I think Ville was referring to a different issue. The PPGTT struct
> itself isn't freed (if I understood his point correctly, which was
> indeed an issue). Forgive my ignorance here but I don't see where the
> leak is. __gen8_alloc_page_tables() appears to always free if it failed,
Yes that cleanup inside __gen8_alloc_page_tables is ok.
> and unwind out should work backwards. Can you show me what I've missed?
As I understand after
pt_pages[i] = __gen8_alloc_page_tables();
we have a kcalloc()'d buffer in pt_pages[i]. Each entry of this buffer
points to an alloc_page()'d page.
Then on error at unwind_out for 0..i-1 we call gen8_free_page_tables()
which will do only a __free_pages() on each of the above alloc_page'd
entry. But then we still have the kcalloc'd buffer, which I can't see
being freed anywhere.
--Imre
>
> >
> > > +
> > > + return ret;
> > > }
> > >
> > > static int gen8_ppgtt_allocate_dma(struct i915_hw_ppgtt *ppgtt)
> > > @@ -475,7 +548,7 @@ static int gen8_ppgtt_setup_page_tables(struct i915_hw_ppgtt *ppgtt,
> > > struct page *p;
> > > int ret;
> > >
> > > - p = &ppgtt->gen8_pt_pages[pd * GEN8_PDES_PER_PAGE + pt];
> > > + p = ppgtt->gen8_pt_pages[pd][pt];
> > > pt_addr = pci_map_page(ppgtt->base.dev->pdev,
> > > p, 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
> > > ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pt_addr);
> >
>
>
>
More information about the Intel-gfx
mailing list