dma-contiguous: provide the ability to reserve per-numa CMA

Right now, drivers like ARM SMMU are using dma_alloc_coherent() to get
coherent DMA buffers to save their command queues and page tables. As
there is only one default CMA in the whole system, SMMUs on nodes other
than node0 will get remote memory. This leads to significant latency.

This patch provides per-numa CMA so that drivers like SMMU can get local
memory. Tests show localizing CMA can decrease dma_unmap latency much.
For instance, before this patch, SMMU on node2  has to wait for more than
560ns for the completion of CMD_SYNC in an empty command queue; with this
patch, it needs 240ns only.

A positive side effect of this patch would be improving performance even
further for those users who are worried about performance more than DMA
security and use iommu.passthrough=1 to skip IOMMU. With local CMA, all
drivers can get local coherent DMA buffers.

Also, this patch changes the default CONFIG_CMA_AREAS to 19 in NUMA. As
1+CONFIG_CMA_AREAS should be quite enough for most servers on the market
even they enable both hugetlb_cma and pernuma_cma.
2 numa nodes: 2(hugetlb) + 2(pernuma) + 1(default global cma) = 5
4 numa nodes: 4(hugetlb) + 4(pernuma) + 1(default global cma) = 9
8 numa nodes: 8(hugetlb) + 8(pernuma) + 1(default global cma) = 17

Signed-off-by: Barry Song <song.bao.hua@hisilicon.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

1 files changed, 90 insertions, 10 deletions

diff --git a/kernel/dma/contiguous.c b/kernel/dma/contiguous.c
index cff7e60968b9..aa53384fd7dc 100644
--- a/kernel/dma/contiguous.c
+++ b/kernel/dma/contiguous.c
@@ -69,6 +69,19 @@ static int __init early_cma(char *p)
 }
 early_param("cma", early_cma);
 
+#ifdef CONFIG_DMA_PERNUMA_CMA
+
+static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES];
+static phys_addr_t pernuma_size_bytes __initdata;
+
+static int __init early_cma_pernuma(char *p)
+{
+	pernuma_size_bytes = memparse(p, &p);
+	return 0;
+}
+early_param("cma_pernuma", early_cma_pernuma);
+#endif
+
 #ifdef CONFIG_CMA_SIZE_PERCENTAGE
 
 static phys_addr_t __init __maybe_unused cma_early_percent_memory(void)
@@ -96,6 +109,34 @@ static inline __maybe_unused phys_addr_t cma_early_percent_memory(void)
 
 #endif
 
+#ifdef CONFIG_DMA_PERNUMA_CMA
+void __init dma_pernuma_cma_reserve(void)
+{
+	int nid;
+
+	if (!pernuma_size_bytes)
+		return;
+
+	for_each_online_node(nid) {
+		int ret;
+		char name[20];
+		struct cma **cma = &dma_contiguous_pernuma_area[nid];
+
+		snprintf(name, sizeof(name), "pernuma%d", nid);
+		ret = cma_declare_contiguous_nid(0, pernuma_size_bytes, 0, 0,
+						 0, false, name, cma, nid);
+		if (ret) {
+			pr_warn("%s: reservation failed: err %d, node %d", __func__,
+				ret, nid);
+			continue;
+		}
+
+		pr_debug("%s: reserved %llu MiB on node %d\n", __func__,
+			(unsigned long long)pernuma_size_bytes / SZ_1M, nid);
+	}
+}
+#endif
+
 /**
  * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling
  * @limit: End address of the reserved memory (optional, 0 for any).
@@ -228,23 +269,44 @@ static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp)
  * @size:  Requested allocation size.
  * @gfp:   Allocation flags.
  *
- * This function allocates contiguous memory buffer for specified device. It
- * tries to use device specific contiguous memory area if available, or the
- * default global one.
+ * tries to use device specific contiguous memory area if available, or it
+ * tries to use per-numa cma, if the allocation fails, it will fallback to
+ * try default global one.
  *
- * Note that it byapss one-page size of allocations from the global area as
- * the addresses within one page are always contiguous, so there is no need
- * to waste CMA pages for that kind; it also helps reduce fragmentations.
+ * Note that it bypass one-page size of allocations from the per-numa and
+ * global area as the addresses within one page are always contiguous, so
+ * there is no need to waste CMA pages for that kind; it also helps reduce
+ * fragmentations.
  */
 struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
 {
+#ifdef CONFIG_DMA_PERNUMA_CMA
+	int nid = dev_to_node(dev);
+#endif
+
 	/* CMA can be used only in the context which permits sleeping */
 	if (!gfpflags_allow_blocking(gfp))
 		return NULL;
 	if (dev->cma_area)
 		return cma_alloc_aligned(dev->cma_area, size, gfp);
-	if (size <= PAGE_SIZE || !dma_contiguous_default_area)
+	if (size <= PAGE_SIZE)
+		return NULL;
+
+#ifdef CONFIG_DMA_PERNUMA_CMA
+	if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) {
+		struct cma *cma = dma_contiguous_pernuma_area[nid];
+		struct page *page;
+
+		if (cma) {
+			page = cma_alloc_aligned(cma, size, gfp);
+			if (page)
+				return page;
+		}
+	}
+#endif
+	if (!dma_contiguous_default_area)
 		return NULL;
+
 	return cma_alloc_aligned(dma_contiguous_default_area, size, gfp);
 }
 
@@ -261,9 +323,27 @@ struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
  */
 void dma_free_contiguous(struct device *dev, struct page *page, size_t size)
 {
-	if (!cma_release(dev_get_cma_area(dev), page,
-			 PAGE_ALIGN(size) >> PAGE_SHIFT))
-		__free_pages(page, get_order(size));
+	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
+
+	/* if dev has its own cma, free page from there */
+	if (dev->cma_area) {
+		if (cma_release(dev->cma_area, page, count))
+			return;
+	} else {
+		/*
+		 * otherwise, page is from either per-numa cma or default cma
+		 */
+#ifdef CONFIG_DMA_PERNUMA_CMA
+		if (cma_release(dma_contiguous_pernuma_area[page_to_nid(page)],
+					page, count))
+			return;
+#endif
+		if (cma_release(dma_contiguous_default_area, page, count))
+			return;
+	}
+
+	/* not in any cma, free from buddy */
+	__free_pages(page, get_order(size));
 }
 
 /*