Skip to contentSkip to conversation

ttm/dma: Implement set_page_caching implementation in the TTM DMA pool code.

PATCH Kevin Shanahan 2 participants 2 messages Sep 29, 2011

lore
Share & tools
lore
#0 lore raw
Quoted 90 lines · Konrad Rzeszutek Wilk

. which is pretty much like the other TTM pool except it also handles moving the page to another pool list.

---
 drivers/gpu/drm/ttm/ttm_page_alloc_dma.c |   96 ++++++++++++++++++++++++++++++
 1 files changed, 96 insertions(+), 0 deletions(-)
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +73 -0
index 5909d28..cea031e 100644
@@ -1307,11 +1307,107 @@ static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
 	mutex_unlock(&_manager->lock);
 	return 0;
 }
+#ifdef CONFIG_X86
+static int ttm_dma_page_set_page_caching(struct page *p,
+					 int flags,
+					 enum ttm_caching_state c_old,
+					 enum ttm_caching_state c_new,
+					 struct device *dev)
+{
+	struct dma_pool *src, *dst;
+	enum pool_type type;
+	struct dma_page *dma_p;
+	bool found = false;
+	unsigned long irq_flags;
+	int ret = 0;
+
+	if (!p)
+		return 0;
+
+	if (PageHighMem(p))
+		return 0;
+
+	type = ttm_to_type(flags, c_old);
+	src = ttm_dma_find_pool(dev, type);
+	if (!src) {
+		WARN_ON(!src);
+		return -ENOMEM;
+	}
+	type = ttm_to_type(flags, c_new);
+	dst = ttm_dma_find_pool(dev, type);
+	if (!dst) {
+		gfp_t gfp_flags;
+		if (flags & TTM_PAGE_FLAG_DMA32)
+			gfp_flags = GFP_USER | GFP_DMA32;
+		else
+			gfp_flags = GFP_HIGHUSER;
+
+		if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
+			gfp_flags |= __GFP_ZERO;
+
+		dst = ttm_dma_pool_init(dev, gfp_flags, type);
+		if (IS_ERR_OR_NULL(dst))
+			return -ENOMEM;
+	}
+
+	dev_dbg(dev, "(%d) Caching %p (%p) from %x to %x.\n", current->pid,
+		p, page_address(p), c_old, c_new);
+
+	if (c_old != tt_cached) {
+		/* p isn't in the default caching state, set it to
+		 * writeback first to free its current memtype. */
+
+		ret = set_pages_wb(p, 1);
+		if (ret)
+			return ret;
+	}
 
+	if (c_new == tt_wc)
+		ret = set_memory_wc((unsigned long) page_address(p), 1);
+	else if (c_new == tt_uncached)
+		ret = set_pages_uc(p, 1);
+
+	if (ret)
+		return ret;
+
+	dev_dbg(src->dev, "(%s:%d) Moving %p (%p) to %s.\n", src->name,
+		current->pid, p, page_address(p), dst->name);
+
+	/* To make it faster we only take the spinlock on list
+	 * removal, and later on adding the page to the destination pool. */
+	spin_lock_irqsave(&src->lock, irq_flags);
+	list_for_each_entry(dma_p, &src->page_list, page_list) {
+		if (virt_to_page(dma_p->vaddr) != p) {
+			pr_debug("%s: (%s:%d) Skipping %p (%p) (DMA:0x%lx)\n",
+				src->dev_name, src->name, current->pid,
+				dma_p->vaddr,

drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data) · line 1393

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +10 -0
+				virt_to_page(dma_p->vaddr),
+				(unsigned long)dma_p->dma);
+			continue;
+		}
+		list_del(&dma_p->page_list);
+		src->npages_in_use -= 1;
+		found = true;
+		break;
+	}
+	spin_lock_irqsave(&src->lock, irq_flags);
        ^^^^^^^^^^^^^^^^^

Was this meant to be spin_unlock_irqrestore?

Cheers,
Kevin.

Quoted 28 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +13 -0
+	if (!found)
+		return -ENODEV;
+
+	spin_lock_irqsave(&dst->lock, irq_flags);
+	list_add(&dma_p->page_list, &dst->page_list);
+	dst->npages_in_use++;
+	spin_unlock_irqrestore(&dst->lock, irq_flags);
+	return 0;
+};
+#endif
 struct ttm_page_alloc_func ttm_page_alloc_dma = {
 	.get_pages	= ttm_dma_get_pages,
 	.put_pages	= ttm_dma_put_pages,
 	.alloc_init	= ttm_dma_page_alloc_init,
 	.alloc_fini	= ttm_dma_page_alloc_fini,
 	.debugfs	= ttm_dma_page_alloc_debugfs,
+#ifdef CONFIG_X86
+	.set_caching	= ttm_dma_page_set_page_caching,
+#endif
 };
Signature · 7 lines
1.7.4.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/etc.
#1 lore raw
Quoted 103 lines · Kevin Shanahan
Quoted 90 lines · Konrad Rzeszutek Wilk

. which is pretty much like the other TTM pool except it also handles moving the page to another pool list.

---
 drivers/gpu/drm/ttm/ttm_page_alloc_dma.c |   96 ++++++++++++++++++++++++++++++
 1 files changed, 96 insertions(+), 0 deletions(-)
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +73 -0
index 5909d28..cea031e 100644
@@ -1307,11 +1307,107 @@ static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
 	mutex_unlock(&_manager->lock);
 	return 0;
 }
+#ifdef CONFIG_X86
+static int ttm_dma_page_set_page_caching(struct page *p,
+					 int flags,
+					 enum ttm_caching_state c_old,
+					 enum ttm_caching_state c_new,
+					 struct device *dev)
+{
+	struct dma_pool *src, *dst;
+	enum pool_type type;
+	struct dma_page *dma_p;
+	bool found = false;
+	unsigned long irq_flags;
+	int ret = 0;
+
+	if (!p)
+		return 0;
+
+	if (PageHighMem(p))
+		return 0;
+
+	type = ttm_to_type(flags, c_old);
+	src = ttm_dma_find_pool(dev, type);
+	if (!src) {
+		WARN_ON(!src);
+		return -ENOMEM;
+	}
+	type = ttm_to_type(flags, c_new);
+	dst = ttm_dma_find_pool(dev, type);
+	if (!dst) {
+		gfp_t gfp_flags;
+		if (flags & TTM_PAGE_FLAG_DMA32)
+			gfp_flags = GFP_USER | GFP_DMA32;
+		else
+			gfp_flags = GFP_HIGHUSER;
+
+		if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
+			gfp_flags |= __GFP_ZERO;
+
+		dst = ttm_dma_pool_init(dev, gfp_flags, type);
+		if (IS_ERR_OR_NULL(dst))
+			return -ENOMEM;
+	}
+
+	dev_dbg(dev, "(%d) Caching %p (%p) from %x to %x.\n", current->pid,
+		p, page_address(p), c_old, c_new);
+
+	if (c_old != tt_cached) {
+		/* p isn't in the default caching state, set it to
+		 * writeback first to free its current memtype. */
+
+		ret = set_pages_wb(p, 1);
+		if (ret)
+			return ret;
+	}
 
+	if (c_new == tt_wc)
+		ret = set_memory_wc((unsigned long) page_address(p), 1);
+	else if (c_new == tt_uncached)
+		ret = set_pages_uc(p, 1);
+
+	if (ret)
+		return ret;
+
+	dev_dbg(src->dev, "(%s:%d) Moving %p (%p) to %s.\n", src->name,
+		current->pid, p, page_address(p), dst->name);
+
+	/* To make it faster we only take the spinlock on list
+	 * removal, and later on adding the page to the destination pool. */
+	spin_lock_irqsave(&src->lock, irq_flags);
+	list_for_each_entry(dma_p, &src->page_list, page_list) {
+		if (virt_to_page(dma_p->vaddr) != p) {
+			pr_debug("%s: (%s:%d) Skipping %p (%p) (DMA:0x%lx)\n",
+				src->dev_name, src->name, current->pid,
+				dma_p->vaddr,

drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data) · line 1393

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +10 -0
+				virt_to_page(dma_p->vaddr),
+				(unsigned long)dma_p->dma);
+			continue;
+		}
+		list_del(&dma_p->page_list);
+		src->npages_in_use -= 1;
+		found = true;
+		break;
+	}
+	spin_lock_irqsave(&src->lock, irq_flags);
        ^^^^^^^^^^^^^^^^^
Quoted 1 line

Was this meant to be spin_unlock_irqrestore?

Yup! Fixing it up. Hmm.... how did it run. This is called from the TTM shrinker when the memory pressure is low, unless I am reading the code wrong.