]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
net/mlx4: Get rid of page operation after dma_alloc_coherent
authorStephen Warren <swarren@nvidia.com>
Thu, 3 Jan 2019 17:23:23 +0000 (10:23 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 7 Jan 2019 13:14:17 +0000 (05:14 -0800)
This patch solves a crash at the time of mlx4 driver unload or system
shutdown. The crash occurs because dma_alloc_coherent() returns one
value in mlx4_alloc_icm_coherent(), but a different value is passed to
dma_free_coherent() in mlx4_free_icm_coherent(). In turn this is because
when allocated, that pointer is passed to sg_set_buf() to record it,
then when freed it is re-calculated by calling
lowmem_page_address(sg_page()) which returns a different value. Solve
this by recording the value that dma_alloc_coherent() returns, and
passing this to dma_free_coherent().

This patch is roughly equivalent to commit 378efe798ecf ("RDMA/hns: Get
rid of page operation after dma_alloc_coherent").

Based-on-code-from: Christoph Hellwig <hch@lst.de>
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx4/icm.c
drivers/net/ethernet/mellanox/mlx4/icm.h

index 4b4351141b94c7853e5519afc5bf48b8374cc196..76b84d08a058b1e692b55140878d3a93424fb2db 100644 (file)
@@ -57,12 +57,12 @@ static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chu
        int i;
 
        if (chunk->nsg > 0)
-               pci_unmap_sg(dev->persist->pdev, chunk->mem, chunk->npages,
+               pci_unmap_sg(dev->persist->pdev, chunk->sg, chunk->npages,
                             PCI_DMA_BIDIRECTIONAL);
 
        for (i = 0; i < chunk->npages; ++i)
-               __free_pages(sg_page(&chunk->mem[i]),
-                            get_order(chunk->mem[i].length));
+               __free_pages(sg_page(&chunk->sg[i]),
+                            get_order(chunk->sg[i].length));
 }
 
 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
@@ -71,9 +71,9 @@ static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *
 
        for (i = 0; i < chunk->npages; ++i)
                dma_free_coherent(&dev->persist->pdev->dev,
-                                 chunk->mem[i].length,
-                                 lowmem_page_address(sg_page(&chunk->mem[i])),
-                                 sg_dma_address(&chunk->mem[i]));
+                                 chunk->buf[i].size,
+                                 chunk->buf[i].addr,
+                                 chunk->buf[i].dma_addr);
 }
 
 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
@@ -111,22 +111,21 @@ static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order,
        return 0;
 }
 
-static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
-                                   int order, gfp_t gfp_mask)
+static int mlx4_alloc_icm_coherent(struct device *dev, struct mlx4_icm_buf *buf,
+                                  int order, gfp_t gfp_mask)
 {
-       void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
-                                      &sg_dma_address(mem), gfp_mask);
-       if (!buf)
+       buf->addr = dma_alloc_coherent(dev, PAGE_SIZE << order,
+                                      &buf->dma_addr, gfp_mask);
+       if (!buf->addr)
                return -ENOMEM;
 
-       if (offset_in_page(buf)) {
-               dma_free_coherent(dev, PAGE_SIZE << order,
-                                 buf, sg_dma_address(mem));
+       if (offset_in_page(buf->addr)) {
+               dma_free_coherent(dev, PAGE_SIZE << order, buf->addr,
+                                 buf->dma_addr);
                return -ENOMEM;
        }
 
-       sg_set_buf(mem, buf, PAGE_SIZE << order);
-       sg_dma_len(mem) = PAGE_SIZE << order;
+       buf->size = PAGE_SIZE << order;
        return 0;
 }
 
@@ -159,21 +158,21 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 
        while (npages > 0) {
                if (!chunk) {
-                       chunk = kmalloc_node(sizeof(*chunk),
+                       chunk = kzalloc_node(sizeof(*chunk),
                                             gfp_mask & ~(__GFP_HIGHMEM |
                                                          __GFP_NOWARN),
                                             dev->numa_node);
                        if (!chunk) {
-                               chunk = kmalloc(sizeof(*chunk),
+                               chunk = kzalloc(sizeof(*chunk),
                                                gfp_mask & ~(__GFP_HIGHMEM |
                                                             __GFP_NOWARN));
                                if (!chunk)
                                        goto fail;
                        }
+                       chunk->coherent = coherent;
 
-                       sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
-                       chunk->npages = 0;
-                       chunk->nsg    = 0;
+                       if (!coherent)
+                               sg_init_table(chunk->sg, MLX4_ICM_CHUNK_LEN);
                        list_add_tail(&chunk->list, &icm->chunk_list);
                }
 
@@ -186,10 +185,10 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 
                if (coherent)
                        ret = mlx4_alloc_icm_coherent(&dev->persist->pdev->dev,
-                                                     &chunk->mem[chunk->npages],
-                                                     cur_order, mask);
+                                               &chunk->buf[chunk->npages],
+                                               cur_order, mask);
                else
-                       ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
+                       ret = mlx4_alloc_icm_pages(&chunk->sg[chunk->npages],
                                                   cur_order, mask,
                                                   dev->numa_node);
 
@@ -205,7 +204,7 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
                if (coherent)
                        ++chunk->nsg;
                else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
-                       chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
+                       chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->sg,
                                                chunk->npages,
                                                PCI_DMA_BIDIRECTIONAL);
 
@@ -220,7 +219,7 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
        }
 
        if (!coherent && chunk) {
-               chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
+               chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->sg,
                                        chunk->npages,
                                        PCI_DMA_BIDIRECTIONAL);
 
@@ -320,7 +319,7 @@ void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
        u64 idx;
        struct mlx4_icm_chunk *chunk;
        struct mlx4_icm *icm;
-       struct page *page = NULL;
+       void *addr = NULL;
 
        if (!table->lowmem)
                return NULL;
@@ -336,28 +335,49 @@ void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
 
        list_for_each_entry(chunk, &icm->chunk_list, list) {
                for (i = 0; i < chunk->npages; ++i) {
+                       dma_addr_t dma_addr;
+                       size_t len;
+
+                       if (table->coherent) {
+                               len = chunk->buf[i].size;
+                               dma_addr = chunk->buf[i].dma_addr;
+                               addr = chunk->buf[i].addr;
+                       } else {
+                               struct page *page;
+
+                               len = sg_dma_len(&chunk->sg[i]);
+                               dma_addr = sg_dma_address(&chunk->sg[i]);
+
+                               /* XXX: we should never do this for highmem
+                                * allocation.  This function either needs
+                                * to be split, or the kernel virtual address
+                                * return needs to be made optional.
+                                */
+                               page = sg_page(&chunk->sg[i]);
+                               addr = lowmem_page_address(page);
+                       }
+
                        if (dma_handle && dma_offset >= 0) {
-                               if (sg_dma_len(&chunk->mem[i]) > dma_offset)
-                                       *dma_handle = sg_dma_address(&chunk->mem[i]) +
-                                               dma_offset;
-                               dma_offset -= sg_dma_len(&chunk->mem[i]);
+                               if (len > dma_offset)
+                                       *dma_handle = dma_addr + dma_offset;
+                               dma_offset -= len;
                        }
+
                        /*
                         * DMA mapping can merge pages but not split them,
                         * so if we found the page, dma_handle has already
                         * been assigned to.
                         */
-                       if (chunk->mem[i].length > offset) {
-                               page = sg_page(&chunk->mem[i]);
+                       if (len > offset)
                                goto out;
-                       }
-                       offset -= chunk->mem[i].length;
+                       offset -= len;
                }
        }
 
+       addr = NULL;
 out:
        mutex_unlock(&table->mutex);
-       return page ? lowmem_page_address(page) + offset : NULL;
+       return addr ? addr + offset : NULL;
 }
 
 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
index c9169a490557cc28a865ef98cbe0676fd3999d33..d199874b1c074cafd7a740e6806761c279e476cd 100644 (file)
@@ -47,11 +47,21 @@ enum {
        MLX4_ICM_PAGE_SIZE      = 1 << MLX4_ICM_PAGE_SHIFT,
 };
 
+struct mlx4_icm_buf {
+       void                    *addr;
+       size_t                  size;
+       dma_addr_t              dma_addr;
+};
+
 struct mlx4_icm_chunk {
        struct list_head        list;
        int                     npages;
        int                     nsg;
-       struct scatterlist      mem[MLX4_ICM_CHUNK_LEN];
+       bool                    coherent;
+       union {
+               struct scatterlist      sg[MLX4_ICM_CHUNK_LEN];
+               struct mlx4_icm_buf     buf[MLX4_ICM_CHUNK_LEN];
+       };
 };
 
 struct mlx4_icm {
@@ -114,12 +124,18 @@ static inline void mlx4_icm_next(struct mlx4_icm_iter *iter)
 
 static inline dma_addr_t mlx4_icm_addr(struct mlx4_icm_iter *iter)
 {
-       return sg_dma_address(&iter->chunk->mem[iter->page_idx]);
+       if (iter->chunk->coherent)
+               return iter->chunk->buf[iter->page_idx].dma_addr;
+       else
+               return sg_dma_address(&iter->chunk->sg[iter->page_idx]);
 }
 
 static inline unsigned long mlx4_icm_size(struct mlx4_icm_iter *iter)
 {
-       return sg_dma_len(&iter->chunk->mem[iter->page_idx]);
+       if (iter->chunk->coherent)
+               return iter->chunk->buf[iter->page_idx].size;
+       else
+               return sg_dma_len(&iter->chunk->sg[iter->page_idx]);
 }
 
 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm);