drm/i915/gtt: Pull global wc page stash under its own locking
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 4 Jul 2018 18:55:18 +0000 (19:55 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 4 Jul 2018 20:23:11 +0000 (21:23 +0100)
Currently, the wc-stash used for providing flushed WC pages ready for
constructing the page directories is assumed to be protected by the
struct_mutex. However, we want to remove this global lock and so must
install a replacement global lock for accessing the global wc-stash (the
per-vm stash continues to be guarded by the vm).

We need to push ahead on this patch due to an oversight in hastily
removing the struct_mutex guard around the igt_ppgtt_alloc selftest. No
matter, it will prove very useful (i.e. will be required) in the near
future.

v2: Restore the onstack stash so that we can drop the vm->mutex in
future across the allocation.
v3: Restore the lost pagevec_init of the onstack allocation, and repaint
function names.
v4: Reorder init so that we don't try and use i915_address_space before
it is ininitialised.

Fixes: 1f6f00238abf ("drm/i915/selftests: Drop struct_mutex around lowlevel pggtt allocation")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180704185518.4193-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem_context.c
drivers/gpu/drm/i915/i915_gem_gtt.c
drivers/gpu/drm/i915/i915_gem_gtt.h
drivers/gpu/drm/i915/selftests/huge_pages.c
drivers/gpu/drm/i915/selftests/i915_gem_gtt.c
drivers/gpu/drm/i915/selftests/mock_gtt.c

index ce7d0633288427f0a81c85a84298f8fb861f4b3c..4f9191ab3e126143044888a839d1c344a26b2521 100644 (file)
@@ -952,7 +952,7 @@ struct i915_gem_mm {
        /**
         * Small stash of WC pages
         */
-       struct pagevec wc_stash;
+       struct pagestash wc_stash;
 
        /**
         * tmpfs instance used for shmem backed objects
index ccf463ab6562faed7762cdd29ffd4f9a5292c178..985ef70d9416d2ce5fcbadc217db8bff7601a397 100644 (file)
@@ -374,7 +374,7 @@ i915_gem_create_context(struct drm_i915_private *dev_priv,
        if (USES_FULL_PPGTT(dev_priv)) {
                struct i915_hw_ppgtt *ppgtt;
 
-               ppgtt = i915_ppgtt_create(dev_priv, file_priv, ctx->name);
+               ppgtt = i915_ppgtt_create(dev_priv, file_priv);
                if (IS_ERR(ppgtt)) {
                        DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
                                         PTR_ERR(ppgtt));
index ba1aed07d6dce9ae07e39cf3dd78292027adafc4..5fb29959057922cb81ef21faa25828ec65b504c6 100644 (file)
@@ -375,37 +375,70 @@ static gen6_pte_t iris_pte_encode(dma_addr_t addr,
        return pte;
 }
 
+static void stash_init(struct pagestash *stash)
+{
+       pagevec_init(&stash->pvec);
+       spin_lock_init(&stash->lock);
+}
+
+static struct page *stash_pop_page(struct pagestash *stash)
+{
+       struct page *page = NULL;
+
+       spin_lock(&stash->lock);
+       if (likely(stash->pvec.nr))
+               page = stash->pvec.pages[--stash->pvec.nr];
+       spin_unlock(&stash->lock);
+
+       return page;
+}
+
+static void stash_push_pagevec(struct pagestash *stash, struct pagevec *pvec)
+{
+       int nr;
+
+       spin_lock_nested(&stash->lock, SINGLE_DEPTH_NESTING);
+
+       nr = min_t(int, pvec->nr, pagevec_space(&stash->pvec));
+       memcpy(stash->pvec.pages + stash->pvec.nr,
+              pvec->pages + pvec->nr - nr,
+              sizeof(pvec->pages[0]) * nr);
+       stash->pvec.nr += nr;
+
+       spin_unlock(&stash->lock);
+
+       pvec->nr -= nr;
+}
+
 static struct page *vm_alloc_page(struct i915_address_space *vm, gfp_t gfp)
 {
-       struct pagevec *pvec = &vm->free_pages;
-       struct pagevec stash;
+       struct pagevec stack;
+       struct page *page;
 
        if (I915_SELFTEST_ONLY(should_fail(&vm->fault_attr, 1)))
                i915_gem_shrink_all(vm->i915);
 
-       if (likely(pvec->nr))
-               return pvec->pages[--pvec->nr];
+       page = stash_pop_page(&vm->free_pages);
+       if (page)
+               return page;
 
        if (!vm->pt_kmap_wc)
                return alloc_page(gfp);
 
-       /* A placeholder for a specific mutex to guard the WC stash */
-       lockdep_assert_held(&vm->i915->drm.struct_mutex);
-
        /* Look in our global stash of WC pages... */
-       pvec = &vm->i915->mm.wc_stash;
-       if (likely(pvec->nr))
-               return pvec->pages[--pvec->nr];
+       page = stash_pop_page(&vm->i915->mm.wc_stash);
+       if (page)
+               return page;
 
        /*
-        * Otherwise batch allocate pages to amoritize cost of set_pages_wc.
+        * Otherwise batch allocate pages to amortize cost of set_pages_wc.
         *
         * We have to be careful as page allocation may trigger the shrinker
         * (via direct reclaim) which will fill up the WC stash underneath us.
         * So we add our WB pages into a temporary pvec on the stack and merge
         * them into the WC stash after all the allocations are complete.
         */
-       pagevec_init(&stash);
+       pagevec_init(&stack);
        do {
                struct page *page;
 
@@ -413,59 +446,67 @@ static struct page *vm_alloc_page(struct i915_address_space *vm, gfp_t gfp)
                if (unlikely(!page))
                        break;
 
-               stash.pages[stash.nr++] = page;
-       } while (stash.nr < pagevec_space(pvec));
+               stack.pages[stack.nr++] = page;
+       } while (pagevec_space(&stack));
 
-       if (stash.nr) {
-               int nr = min_t(int, stash.nr, pagevec_space(pvec));
-               struct page **pages = stash.pages + stash.nr - nr;
+       if (stack.nr && !set_pages_array_wc(stack.pages, stack.nr)) {
+               page = stack.pages[--stack.nr];
 
-               if (nr && !set_pages_array_wc(pages, nr)) {
-                       memcpy(pvec->pages + pvec->nr,
-                              pages, sizeof(pages[0]) * nr);
-                       pvec->nr += nr;
-                       stash.nr -= nr;
-               }
+               /* Merge spare WC pages to the global stash */
+               stash_push_pagevec(&vm->i915->mm.wc_stash, &stack);
+
+               /* Push any surplus WC pages onto the local VM stash */
+               if (stack.nr)
+                       stash_push_pagevec(&vm->free_pages, &stack);
+       }
 
-               pagevec_release(&stash);
+       /* Return unwanted leftovers */
+       if (unlikely(stack.nr)) {
+               WARN_ON_ONCE(set_pages_array_wb(stack.pages, stack.nr));
+               __pagevec_release(&stack);
        }
 
-       return likely(pvec->nr) ? pvec->pages[--pvec->nr] : NULL;
+       return page;
 }
 
 static void vm_free_pages_release(struct i915_address_space *vm,
                                  bool immediate)
 {
-       struct pagevec *pvec = &vm->free_pages;
+       struct pagevec *pvec = &vm->free_pages.pvec;
+       struct pagevec stack;
 
+       lockdep_assert_held(&vm->free_pages.lock);
        GEM_BUG_ON(!pagevec_count(pvec));
 
        if (vm->pt_kmap_wc) {
-               struct pagevec *stash = &vm->i915->mm.wc_stash;
-
-               /* When we use WC, first fill up the global stash and then
+               /*
+                * When we use WC, first fill up the global stash and then
                 * only if full immediately free the overflow.
                 */
+               stash_push_pagevec(&vm->i915->mm.wc_stash, pvec);
 
-               lockdep_assert_held(&vm->i915->drm.struct_mutex);
-               if (pagevec_space(stash)) {
-                       do {
-                               stash->pages[stash->nr++] =
-                                       pvec->pages[--pvec->nr];
-                               if (!pvec->nr)
-                                       return;
-                       } while (pagevec_space(stash));
-
-                       /* As we have made some room in the VM's free_pages,
-                        * we can wait for it to fill again. Unless we are
-                        * inside i915_address_space_fini() and must
-                        * immediately release the pages!
-                        */
-                       if (!immediate)
-                               return;
-               }
+               /*
+                * As we have made some room in the VM's free_pages,
+                * we can wait for it to fill again. Unless we are
+                * inside i915_address_space_fini() and must
+                * immediately release the pages!
+                */
+               if (pvec->nr <= (immediate ? 0 : PAGEVEC_SIZE - 1))
+                       return;
 
+               /*
+                * We have to drop the lock to allow ourselves to sleep,
+                * so take a copy of the pvec and clear the stash for
+                * others to use it as we sleep.
+                */
+               stack = *pvec;
+               pagevec_reinit(pvec);
+               spin_unlock(&vm->free_pages.lock);
+
+               pvec = &stack;
                set_pages_array_wb(pvec->pages, pvec->nr);
+
+               spin_lock(&vm->free_pages.lock);
        }
 
        __pagevec_release(pvec);
@@ -481,8 +522,38 @@ static void vm_free_page(struct i915_address_space *vm, struct page *page)
         * unconditional might_sleep() for everybody.
         */
        might_sleep();
-       if (!pagevec_add(&vm->free_pages, page))
+       spin_lock(&vm->free_pages.lock);
+       if (!pagevec_add(&vm->free_pages.pvec, page))
                vm_free_pages_release(vm, false);
+       spin_unlock(&vm->free_pages.lock);
+}
+
+static void i915_address_space_init(struct i915_address_space *vm,
+                                   struct drm_i915_private *dev_priv)
+{
+       GEM_BUG_ON(!vm->total);
+       drm_mm_init(&vm->mm, 0, vm->total);
+       vm->mm.head_node.color = I915_COLOR_UNEVICTABLE;
+
+       stash_init(&vm->free_pages);
+
+       INIT_LIST_HEAD(&vm->active_list);
+       INIT_LIST_HEAD(&vm->inactive_list);
+       INIT_LIST_HEAD(&vm->unbound_list);
+
+       list_add_tail(&vm->global_link, &dev_priv->vm_list);
+}
+
+static void i915_address_space_fini(struct i915_address_space *vm)
+{
+       spin_lock(&vm->free_pages.lock);
+       if (pagevec_count(&vm->free_pages.pvec))
+               vm_free_pages_release(vm, true);
+       GEM_BUG_ON(pagevec_count(&vm->free_pages.pvec));
+       spin_unlock(&vm->free_pages.lock);
+
+       drm_mm_takedown(&vm->mm);
+       list_del(&vm->global_link);
 }
 
 static int __setup_page_dma(struct i915_address_space *vm,
@@ -1562,6 +1633,8 @@ static struct i915_hw_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915)
        if (!ppgtt)
                return ERR_PTR(-ENOMEM);
 
+       kref_init(&ppgtt->ref);
+
        ppgtt->vm.i915 = i915;
        ppgtt->vm.dma = &i915->drm.pdev->dev;
 
@@ -1569,6 +1642,8 @@ static struct i915_hw_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915)
                1ULL << 48 :
                1ULL << 32;
 
+       i915_address_space_init(&ppgtt->vm, i915);
+
        /* There are only few exceptions for gen >=6. chv and bxt.
         * And we are not sure about the latter so play safe for now.
         */
@@ -2068,11 +2143,15 @@ static struct i915_hw_ppgtt *gen6_ppgtt_create(struct drm_i915_private *i915)
        if (!ppgtt)
                return ERR_PTR(-ENOMEM);
 
+       kref_init(&ppgtt->base.ref);
+
        ppgtt->base.vm.i915 = i915;
        ppgtt->base.vm.dma = &i915->drm.pdev->dev;
 
        ppgtt->base.vm.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
 
+       i915_address_space_init(&ppgtt->base.vm, i915);
+
        ppgtt->base.vm.allocate_va_range = gen6_alloc_va_range;
        ppgtt->base.vm.clear_range = gen6_ppgtt_clear_range;
        ppgtt->base.vm.insert_entries = gen6_ppgtt_insert_entries;
@@ -2105,30 +2184,6 @@ err_free:
        return ERR_PTR(err);
 }
 
-static void i915_address_space_init(struct i915_address_space *vm,
-                                   struct drm_i915_private *dev_priv,
-                                   const char *name)
-{
-       drm_mm_init(&vm->mm, 0, vm->total);
-       vm->mm.head_node.color = I915_COLOR_UNEVICTABLE;
-
-       INIT_LIST_HEAD(&vm->active_list);
-       INIT_LIST_HEAD(&vm->inactive_list);
-       INIT_LIST_HEAD(&vm->unbound_list);
-
-       list_add_tail(&vm->global_link, &dev_priv->vm_list);
-       pagevec_init(&vm->free_pages);
-}
-
-static void i915_address_space_fini(struct i915_address_space *vm)
-{
-       if (pagevec_count(&vm->free_pages))
-               vm_free_pages_release(vm, true);
-
-       drm_mm_takedown(&vm->mm);
-       list_del(&vm->global_link);
-}
-
 static void gtt_write_workarounds(struct drm_i915_private *dev_priv)
 {
        /* This function is for gtt related workarounds. This function is
@@ -2199,8 +2254,7 @@ __hw_ppgtt_create(struct drm_i915_private *i915)
 
 struct i915_hw_ppgtt *
 i915_ppgtt_create(struct drm_i915_private *i915,
-                 struct drm_i915_file_private *fpriv,
-                 const char *name)
+                 struct drm_i915_file_private *fpriv)
 {
        struct i915_hw_ppgtt *ppgtt;
 
@@ -2208,8 +2262,6 @@ i915_ppgtt_create(struct drm_i915_private *i915,
        if (IS_ERR(ppgtt))
                return ppgtt;
 
-       kref_init(&ppgtt->ref);
-       i915_address_space_init(&ppgtt->vm, i915, name);
        ppgtt->vm.file = fpriv;
 
        trace_i915_ppgtt_create(&ppgtt->vm);
@@ -2788,7 +2840,7 @@ int i915_gem_init_aliasing_ppgtt(struct drm_i915_private *i915)
        struct i915_hw_ppgtt *ppgtt;
        int err;
 
-       ppgtt = i915_ppgtt_create(i915, ERR_PTR(-EPERM), "[alias]");
+       ppgtt = i915_ppgtt_create(i915, ERR_PTR(-EPERM));
        if (IS_ERR(ppgtt))
                return PTR_ERR(ppgtt);
 
@@ -2918,7 +2970,7 @@ void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
 
        ggtt->vm.cleanup(&ggtt->vm);
 
-       pvec = &dev_priv->mm.wc_stash;
+       pvec = &dev_priv->mm.wc_stash.pvec;
        if (pvec->nr) {
                set_pages_array_wb(pvec->pages, pvec->nr);
                __pagevec_release(pvec);
@@ -3518,6 +3570,8 @@ int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
        struct i915_ggtt *ggtt = &dev_priv->ggtt;
        int ret;
 
+       stash_init(&dev_priv->mm.wc_stash);
+
        INIT_LIST_HEAD(&dev_priv->vm_list);
 
        /* Note that we use page colouring to enforce a guard page at the
@@ -3526,7 +3580,7 @@ int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
         * and beyond the end of the GTT if we do not provide a guard.
         */
        mutex_lock(&dev_priv->drm.struct_mutex);
-       i915_address_space_init(&ggtt->vm, dev_priv, "[global]");
+       i915_address_space_init(&ggtt->vm, dev_priv);
        if (!HAS_LLC(dev_priv) && !USES_PPGTT(dev_priv))
                ggtt->vm.mm.color_adjust = i915_gtt_color_adjust;
        mutex_unlock(&dev_priv->drm.struct_mutex);
index 9a4824cae68d2dcc6a34a311de841e5ca0ca3974..f298e72b79ca231c05ed496777fba0247c5b7e37 100644 (file)
@@ -270,6 +270,11 @@ struct i915_vma_ops {
        void (*clear_pages)(struct i915_vma *vma);
 };
 
+struct pagestash {
+       spinlock_t lock;
+       struct pagevec pvec;
+};
+
 struct i915_address_space {
        struct drm_mm mm;
        struct drm_i915_private *i915;
@@ -324,7 +329,7 @@ struct i915_address_space {
         */
        struct list_head unbound_list;
 
-       struct pagevec free_pages;
+       struct pagestash free_pages;
        bool pt_kmap_wc;
 
        /* FIXME: Need a more generic return type */
@@ -615,8 +620,7 @@ void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv);
 int i915_ppgtt_init_hw(struct drm_i915_private *dev_priv);
 void i915_ppgtt_release(struct kref *kref);
 struct i915_hw_ppgtt *i915_ppgtt_create(struct drm_i915_private *dev_priv,
-                                       struct drm_i915_file_private *fpriv,
-                                       const char *name);
+                                       struct drm_i915_file_private *fpriv);
 void i915_ppgtt_close(struct i915_address_space *vm);
 static inline void i915_ppgtt_get(struct i915_hw_ppgtt *ppgtt)
 {
index b5e87fcdcdae422e7726fbbca6a7c77edcbc43f1..39699939587dac858051b099fd796fbf1688a7a8 100644 (file)
@@ -1694,7 +1694,7 @@ int i915_gem_huge_page_mock_selftests(void)
        dma_coerce_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(39));
 
        mutex_lock(&dev_priv->drm.struct_mutex);
-       ppgtt = i915_ppgtt_create(dev_priv, ERR_PTR(-ENODEV), "mock");
+       ppgtt = i915_ppgtt_create(dev_priv, ERR_PTR(-ENODEV));
        if (IS_ERR(ppgtt)) {
                err = PTR_ERR(ppgtt);
                goto out_unlock;
index a7956ecc3e1fe4706a2d04e74a8c3402ec693121..4bfb0537f9be7e66016885e763ccaa0a98fb84fb 100644 (file)
@@ -1004,7 +1004,7 @@ static int exercise_ppgtt(struct drm_i915_private *dev_priv,
                return PTR_ERR(file);
 
        mutex_lock(&dev_priv->drm.struct_mutex);
-       ppgtt = i915_ppgtt_create(dev_priv, file->driver_priv, "mock");
+       ppgtt = i915_ppgtt_create(dev_priv, file->driver_priv);
        if (IS_ERR(ppgtt)) {
                err = PTR_ERR(ppgtt);
                goto out_unlock;
index 6a7f4da7b523e98ebce684e3f91a9673473c2cc6..0da5b8c6d91262b3242f2e5454414cf79e0c164c 100644 (file)
@@ -124,7 +124,7 @@ void mock_init_ggtt(struct drm_i915_private *i915)
        ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
        ggtt->vm.vma_ops.clear_pages = clear_pages;
 
-       i915_address_space_init(&ggtt->vm, i915, "global");
+       i915_address_space_init(&ggtt->vm, i915);
 }
 
 void mock_fini_ggtt(struct drm_i915_private *i915)