]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/gt: Shrink i915_page_directory's slab bucket
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 29 Jul 2020 16:42:19 +0000 (17:42 +0100)
committerJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Mon, 7 Sep 2020 11:24:23 +0000 (14:24 +0300)
kmalloc uses power-of-two slab buckets for small allocations (up to a
few pages). Since i915_page_directory is a page of pointers, plus a
couple more, this is rounded up to 8K, and we waste nearly 50% of that
allocation. Long terms this leads to poor memory utilisation, bloating
the kernel footprint, but the problem is exacerbated by our conservative
preallocation scheme for binding VMA. As we are required to allocate all
levels for each vma just in case we need to insert them upon binding,
this leads to a large multiplication factor for a single page vma. By
halving the allocation we need for the page directory structure, we
halve the impact of that factor, bringing workloads that once fitted into
memory, hopefully back to fitting into memory.

We maintain the split between i915_page_directory and i915_page_table as
we only need half the allocation for the lowest, most populous, level.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200729164219.5737-3-chris@chris-wilson.co.uk
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
drivers/gpu/drm/i915/gt/gen6_ppgtt.c
drivers/gpu/drm/i915/gt/gen8_ppgtt.c
drivers/gpu/drm/i915/gt/intel_gtt.h
drivers/gpu/drm/i915/gt/intel_ppgtt.c

index fb702e1de73915dd9f7629f2ba34f33f3adec1c9..7e5a86b774a7a7169237b18d0ed1f63c30a6697a 100644 (file)
@@ -261,7 +261,7 @@ static void gen6_ppgtt_free_pd(struct gen6_ppgtt *ppgtt)
 
        gen6_for_all_pdes(pt, pd, pde)
                if (pt)
-                       free_px(&ppgtt->base.vm, pt);
+                       free_pt(&ppgtt->base.vm, pt);
 }
 
 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
@@ -275,7 +275,8 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
 
        mutex_destroy(&ppgtt->flush);
        mutex_destroy(&ppgtt->pin_mutex);
-       kfree(ppgtt->base.pd);
+
+       free_pd(&ppgtt->base.vm, ppgtt->base.pd);
 }
 
 static int pd_vma_set_pages(struct i915_vma *vma)
@@ -322,7 +323,7 @@ static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma)
                if (!pt || atomic_read(&pt->used))
                        continue;
 
-               free_px(&ppgtt->base.vm, pt);
+               free_pt(&ppgtt->base.vm, pt);
                pd->entry[pde] = NULL;
        }
 
@@ -447,7 +448,7 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
        ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma;
        ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
 
-       ppgtt->base.pd = __alloc_pd(sizeof(*ppgtt->base.pd));
+       ppgtt->base.pd = __alloc_pd(I915_PDES);
        if (!ppgtt->base.pd) {
                err = -ENOMEM;
                goto err_free;
@@ -468,7 +469,7 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
 err_scratch:
        free_scratch(&ppgtt->base.vm);
 err_pd:
-       kfree(ppgtt->base.pd);
+       free_pd(&ppgtt->base.vm, ppgtt->base.pd);
 err_free:
        mutex_destroy(&ppgtt->pin_mutex);
        kfree(ppgtt);
index a40cb7ea8c666cbe90c3cd1c9724b6f7412a8462..eb64f474a78c6ac529f74a3c505d73514287e4a1 100644 (file)
@@ -181,7 +181,7 @@ static void __gen8_ppgtt_cleanup(struct i915_address_space *vm,
                } while (pde++, --count);
        }
 
-       free_px(vm, pd);
+       free_px(vm, &pd->pt, lvl);
 }
 
 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
@@ -248,7 +248,7 @@ static u64 __gen8_ppgtt_clear(struct i915_address_space * const vm,
                }
 
                if (release_pd_entry(pd, idx, pt, scratch))
-                       free_px(vm, pt);
+                       free_px(vm, pt, lvl);
        } while (idx++, --len);
 
        return start;
@@ -628,7 +628,7 @@ static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt)
                err = pin_pt_dma(vm, pde->pt.base);
                if (err) {
                        i915_gem_object_put(pde->pt.base);
-                       kfree(pde);
+                       free_pd(vm, pde);
                        return err;
                }
 
@@ -648,28 +648,30 @@ gen8_alloc_top_pd(struct i915_address_space *vm)
        struct i915_page_directory *pd;
        int err;
 
-       GEM_BUG_ON(count > ARRAY_SIZE(pd->entry));
+       GEM_BUG_ON(count > I915_PDES);
 
-       pd = __alloc_pd(offsetof(typeof(*pd), entry[count]));
+       pd = __alloc_pd(count);
        if (unlikely(!pd))
                return ERR_PTR(-ENOMEM);
 
        pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
        if (IS_ERR(pd->pt.base)) {
-               kfree(pd);
-               return ERR_PTR(-ENOMEM);
+               err = PTR_ERR(pd->pt.base);
+               pd->pt.base = NULL;
+               goto err_pd;
        }
 
        err = pin_pt_dma(vm, pd->pt.base);
-       if (err) {
-               i915_gem_object_put(pd->pt.base);
-               kfree(pd);
-               return ERR_PTR(err);
-       }
+       if (err)
+               goto err_pd;
 
        fill_page_dma(px_base(pd), vm->scratch[vm->top]->encode, count);
        atomic_inc(px_used(pd)); /* mark as pinned */
        return pd;
+
+err_pd:
+       free_pd(vm, pd);
+       return ERR_PTR(err);
 }
 
 /*
index 6abab2d37b6fe0066ea7bbfa5ef9356e9b90663d..c13c650ced2242c6d70819534ce6f418be195906 100644 (file)
@@ -156,7 +156,7 @@ struct i915_page_table {
 struct i915_page_directory {
        struct i915_page_table pt;
        spinlock_t lock;
-       void *entry[512];
+       void **entry;
 };
 
 #define __px_choose_expr(x, type, expr, other) \
@@ -519,12 +519,14 @@ void free_scratch(struct i915_address_space *vm);
 struct drm_i915_gem_object *alloc_pt_dma(struct i915_address_space *vm, int sz);
 struct i915_page_table *alloc_pt(struct i915_address_space *vm);
 struct i915_page_directory *alloc_pd(struct i915_address_space *vm);
-struct i915_page_directory *__alloc_pd(size_t sz);
+struct i915_page_directory *__alloc_pd(int npde);
 
 int pin_pt_dma(struct i915_address_space *vm, struct drm_i915_gem_object *obj);
 
-void free_pt(struct i915_address_space *vm, struct i915_page_table *pt);
-#define free_px(vm, px) free_pt(vm, px_pt(px))
+void free_px(struct i915_address_space *vm,
+            struct i915_page_table *pt, int lvl);
+#define free_pt(vm, px) free_px(vm, px, 0)
+#define free_pd(vm, px) free_px(vm, px_pt(px), 1)
 
 void
 __set_pd_entry(struct i915_page_directory * const pd,
index ede6369a9092d5779592477f3837d0809d7546f9..46d9aceda64cdd260332794dec7dd974e9e58645 100644 (file)
@@ -28,14 +28,20 @@ struct i915_page_table *alloc_pt(struct i915_address_space *vm)
        return pt;
 }
 
-struct i915_page_directory *__alloc_pd(size_t sz)
+struct i915_page_directory *__alloc_pd(int count)
 {
        struct i915_page_directory *pd;
 
-       pd = kzalloc(sz, I915_GFP_ALLOW_FAIL);
+       pd = kzalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
        if (unlikely(!pd))
                return NULL;
 
+       pd->entry = kcalloc(count, sizeof(*pd->entry), I915_GFP_ALLOW_FAIL);
+       if (unlikely(!pd->entry)) {
+               kfree(pd);
+               return NULL;
+       }
+
        spin_lock_init(&pd->lock);
        return pd;
 }
@@ -44,12 +50,13 @@ struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
 {
        struct i915_page_directory *pd;
 
-       pd = __alloc_pd(sizeof(*pd));
+       pd = __alloc_pd(I915_PDES);
        if (unlikely(!pd))
                return ERR_PTR(-ENOMEM);
 
        pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
        if (IS_ERR(pd->pt.base)) {
+               kfree(pd->entry);
                kfree(pd);
                return ERR_PTR(-ENOMEM);
        }
@@ -57,9 +64,19 @@ struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
        return pd;
 }
 
-void free_pt(struct i915_address_space *vm, struct i915_page_table *pt)
+void free_px(struct i915_address_space *vm, struct i915_page_table *pt, int lvl)
 {
-       i915_gem_object_put(pt->base);
+       BUILD_BUG_ON(offsetof(struct i915_page_directory, pt));
+
+       if (lvl) {
+               struct i915_page_directory *pd =
+                       container_of(pt, typeof(*pd), pt);
+               kfree(pd->entry);
+       }
+
+       if (pt->base)
+               i915_gem_object_put(pt->base);
+
        kfree(pt);
 }
 
@@ -82,7 +99,7 @@ __set_pd_entry(struct i915_page_directory * const pd,
               u64 (*encode)(const dma_addr_t, const enum i915_cache_level))
 {
        /* Each thread pre-pins the pd, and we may have a thread per pde. */
-       GEM_BUG_ON(atomic_read(px_used(pd)) > NALLOC * ARRAY_SIZE(pd->entry));
+       GEM_BUG_ON(atomic_read(px_used(pd)) > NALLOC * I915_PDES);
 
        atomic_inc(px_used(pd));
        pd->entry[idx] = to;
@@ -263,7 +280,7 @@ void i915_vm_free_pt_stash(struct i915_address_space *vm,
        for (n = 0; n < ARRAY_SIZE(stash->pt); n++) {
                while ((pt = stash->pt[n])) {
                        stash->pt[n] = pt->stash;
-                       free_px(vm, pt);
+                       free_px(vm, pt, n);
                }
        }
 }