vhost: fix mmap not aligned with hugepage size
authorJianfeng Tan <jianfeng.tan@intel.com>
Thu, 12 Nov 2015 06:06:01 +0000 (14:06 +0800)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Tue, 24 Nov 2015 20:34:11 +0000 (21:34 +0100)
This patch fixes a bug under lower version linux kernel, mmap()
fails when length is not aligned with hugepage size. mmap()
without flag of MAP_ANONYMOUS, should be called with length
argument aligned with hugepagesz at older longterm version
Linux, like 2.6.32 and 3.2.72, or mmap() will fail with EINVAL.
This bug was fixed in Linux kernel by commit:
dab2d3dc45ae7343216635d981d43637e1cb7d45
To avoid failure, make sure in caller to keep length aligned.

Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
lib/librte_vhost/vhost_user/virtio-net-user.c

index d07452a..99da029 100644 (file)
@@ -74,7 +74,6 @@ free_mem_region(struct virtio_net *dev)
 {
        struct orig_region_map *region;
        unsigned int idx;
-       uint64_t alignment;
 
        if (!dev || !dev->mem)
                return;
@@ -82,12 +81,8 @@ free_mem_region(struct virtio_net *dev)
        region = orig_region(dev->mem, dev->mem->nregions);
        for (idx = 0; idx < dev->mem->nregions; idx++) {
                if (region[idx].mapped_address) {
-                       alignment = region[idx].blksz;
-                       munmap((void *)(uintptr_t)
-                               RTE_ALIGN_FLOOR(
-                                       region[idx].mapped_address, alignment),
-                               RTE_ALIGN_CEIL(
-                                       region[idx].mapped_size, alignment));
+                       munmap((void *)(uintptr_t)region[idx].mapped_address,
+                                       region[idx].mapped_size);
                        close(region[idx].fd);
                }
        }
@@ -147,6 +142,18 @@ user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
                /* This is ugly */
                mapped_size = memory.regions[idx].memory_size +
                        memory.regions[idx].mmap_offset;
+
+               /* mmap() without flag of MAP_ANONYMOUS, should be called
+                * with length argument aligned with hugepagesz at older
+                * longterm version Linux, like 2.6.32 and 3.2.72, or
+                * mmap() will fail with EINVAL.
+                *
+                * to avoid failure, make sure in caller to keep length
+                * aligned.
+                */
+               alignment = get_blk_size(pmsg->fds[idx]);
+               mapped_size = RTE_ALIGN_CEIL(mapped_size, alignment);
+
                mapped_address = (uint64_t)(uintptr_t)mmap(NULL,
                        mapped_size,
                        PROT_READ | PROT_WRITE, MAP_SHARED,
@@ -154,9 +161,11 @@ user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
                        0);
 
                RTE_LOG(INFO, VHOST_CONFIG,
-                       "mapped region %d fd:%d to %p sz:0x%"PRIx64" off:0x%"PRIx64"\n",
+                       "mapped region %d fd:%d to:%p sz:0x%"PRIx64" "
+                       "off:0x%"PRIx64" align:0x%"PRIx64"\n",
                        idx, pmsg->fds[idx], (void *)(uintptr_t)mapped_address,
-                       mapped_size, memory.regions[idx].mmap_offset);
+                       mapped_size, memory.regions[idx].mmap_offset,
+                       alignment);
 
                if (mapped_address == (uint64_t)(uintptr_t)MAP_FAILED) {
                        RTE_LOG(ERR, VHOST_CONFIG,
@@ -166,7 +175,7 @@ user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
 
                pregion_orig[idx].mapped_address = mapped_address;
                pregion_orig[idx].mapped_size = mapped_size;
-               pregion_orig[idx].blksz = get_blk_size(pmsg->fds[idx]);
+               pregion_orig[idx].blksz = alignment;
                pregion_orig[idx].fd = pmsg->fds[idx];
 
                mapped_address +=  memory.regions[idx].mmap_offset;
@@ -193,11 +202,8 @@ user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
 
 err_mmap:
        while (idx--) {
-               alignment = pregion_orig[idx].blksz;
-               munmap((void *)(uintptr_t)RTE_ALIGN_FLOOR(
-                       pregion_orig[idx].mapped_address, alignment),
-                       RTE_ALIGN_CEIL(pregion_orig[idx].mapped_size,
-                                       alignment));
+               munmap((void *)(uintptr_t)pregion_orig[idx].mapped_address,
+                               pregion_orig[idx].mapped_size);
                close(pregion_orig[idx].fd);
        }
        free(dev->mem);