drm/i915: Keep a small stash of preallocated WC pages

We use WC pages for coherent writes into the ppGTT on !llc
architectures. However, to create a WC page requires a stop_machine(),
i.e. is very slow. To compensate we currently keep a per-vm cache of
recently freed pages, but we still see the slow startup of new contexts.
We can amoritize that cost slightly by allocating WC pages in small
batches (PAGEVEC_SIZE == 14) and since creating a WC page implies a
stop_machine() there is no penalty for keeping that stash global.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170822173828.5932-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_drv.h

@@ -1465,6 +1465,11 @@ struct i915_gem_mm {
 	struct llist_head free_list;
 	struct work_struct free_work;
 
+	/**
+	 * Small stash of WC pages
+	 */
+	struct pagevec wc_stash;
+
 	/** Usable portion of the GTT for GEM */
 	dma_addr_t stolen_base; /* limited to low memory (32-bit) */
 

drivers/gpu/drm/i915/i915_gem_gtt.c

@@ -356,39 +356,86 @@ static gen6_pte_t iris_pte_encode(dma_addr_t addr,
 
 static struct page *vm_alloc_page(struct i915_address_space *vm, gfp_t gfp)
 {
-	struct page *page;
+	struct pagevec *pvec = &vm->free_pages;
 
 	if (I915_SELFTEST_ONLY(should_fail(&vm->fault_attr, 1)))
 		i915_gem_shrink_all(vm->i915);
 
-	if (vm->free_pages.nr)
-		return vm->free_pages.pages[--vm->free_pages.nr];
+	if (likely(pvec->nr))
+		return pvec->pages[--pvec->nr];
+
+	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 = alloc_page(gfp);
-	if (!page)
+	/* Otherwise batch allocate pages to amoritize cost of set_pages_wc. */
+	do {
+		struct page *page;
+
+		page = alloc_page(gfp);
+		if (unlikely(!page))
+			break;
+
+		pvec->pages[pvec->nr++] = page;
+	} while (pagevec_space(pvec));
+
+	if (unlikely(!pvec->nr))
 		return NULL;
 
-	if (vm->pt_kmap_wc)
-		set_pages_array_wc(&page, 1);
+	set_pages_array_wc(pvec->pages, pvec->nr);
 
-	return page;
+	return pvec->pages[--pvec->nr];
 }
 
-static void vm_free_pages_release(struct i915_address_space *vm)
+static void vm_free_pages_release(struct i915_address_space *vm,
+				  bool immediate)
 {
-	GEM_BUG_ON(!pagevec_count(&vm->free_pages));
+	struct pagevec *pvec = &vm->free_pages;
+
+	GEM_BUG_ON(!pagevec_count(pvec));
+
+	if (vm->pt_kmap_wc) {
+		struct pagevec *stash = &vm->i915->mm.wc_stash;
 
-	if (vm->pt_kmap_wc)
-		set_pages_array_wb(vm->free_pages.pages,
-				   pagevec_count(&vm->free_pages));
+		/* When we use WC, first fill up the global stash and then
+		 * only if full immediately free the overflow.
+		 */
+
+		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;
+		}
+
+		set_pages_array_wb(pvec->pages, pvec->nr);
+	}
 
-	__pagevec_release(&vm->free_pages);
+	__pagevec_release(pvec);
 }
 
 static void vm_free_page(struct i915_address_space *vm, struct page *page)
 {
 	if (!pagevec_add(&vm->free_pages, page))
-		vm_free_pages_release(vm);
+		vm_free_pages_release(vm, false);
 }
 
 static int __setup_page_dma(struct i915_address_space *vm,
@@ -452,12 +499,31 @@ static void fill_page_dma_32(struct i915_address_space *vm,
 static int
 setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
 {
-	return __setup_page_dma(vm, &vm->scratch_page, gfp | __GFP_ZERO);
+	struct page *page;
+	dma_addr_t addr;
+
+	page = alloc_page(gfp | __GFP_ZERO);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	addr = dma_map_page(vm->dma, page, 0, PAGE_SIZE,
+			    PCI_DMA_BIDIRECTIONAL);
+	if (unlikely(dma_mapping_error(vm->dma, addr))) {
+		__free_page(page);
+		return -ENOMEM;
+	}
+
+	vm->scratch_page.page = page;
+	vm->scratch_page.daddr = addr;
+	return 0;
 }
 
 static void cleanup_scratch_page(struct i915_address_space *vm)
 {
-	cleanup_page_dma(vm, &vm->scratch_page);
+	struct i915_page_dma *p = &vm->scratch_page;
+
+	dma_unmap_page(vm->dma, p->daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+	__free_page(p->page);
 }
 
 static struct i915_page_table *alloc_pt(struct i915_address_space *vm)
@@ -1337,18 +1403,18 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
 		1ULL << 48 :
 		1ULL << 32;
 
-	ret = gen8_init_scratch(&ppgtt->base);
-	if (ret) {
-		ppgtt->base.total = 0;
-		return ret;
-	}
-
 	/* There are only few exceptions for gen >=6. chv and bxt.
 	 * And we are not sure about the latter so play safe for now.
 	 */
 	if (IS_CHERRYVIEW(dev_priv) || IS_BROXTON(dev_priv))
 		ppgtt->base.pt_kmap_wc = true;
 
+	ret = gen8_init_scratch(&ppgtt->base);
+	if (ret) {
+		ppgtt->base.total = 0;
+		return ret;
+	}
+
 	if (use_4lvl(vm)) {
 		ret = setup_px(&ppgtt->base, &ppgtt->pml4);
 		if (ret)
@@ -1872,7 +1938,7 @@ static void i915_address_space_init(struct i915_address_space *vm,
 static void i915_address_space_fini(struct i915_address_space *vm)
 {
 	if (pagevec_count(&vm->free_pages))
-		vm_free_pages_release(vm);
+		vm_free_pages_release(vm, true);
 
 	i915_gem_timeline_fini(&vm->timeline);
 	drm_mm_takedown(&vm->mm);
@@ -2598,6 +2664,7 @@ void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
 {
 	struct i915_ggtt *ggtt = &dev_priv->ggtt;
 	struct i915_vma *vma, *vn;
+	struct pagevec *pvec;
 
 	ggtt->base.closed = true;
 
@@ -2621,6 +2688,13 @@ void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
 	}
 
 	ggtt->base.cleanup(&ggtt->base);
+
+	pvec = &dev_priv->mm.wc_stash;
+	if (pvec->nr) {
+		set_pages_array_wb(pvec->pages, pvec->nr);
+		__pagevec_release(pvec);
+	}
+
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 
 	arch_phys_wc_del(ggtt->mtrr);