block: Don't use guest sector size for qemu_blockalign()

bs->buffer_alignment is set by the device emulation and contains the
logical block size of the guest device. This isn't something that the
block layer should know, and even less something to use for determining
the right alignment of buffers to be used for the host.

The new BlockLimits field opt_mem_alignment tells the qemu block layer
the optimal alignment to be used so that no bounce buffer must be used
in the driver.

This patch may change the buffer alignment from 4k to 512 for all
callers that used qemu_blockalign() with the top-level image format
BlockDriverState. The value was never propagated to other levels in the
tree, so in particular raw-posix never required anything else than 512.

While on disks with 4k sectors direct I/O requires a 4k alignment,
memory may still be okay when aligned to 512 byte boundaries. This is
what must have happened in practice, because otherwise this would
already have failed earlier. Therefore I don't expect regressions even
with this intermediate state. Later, raw-posix can implement the hook
and expose a different memory alignment requirement.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Wenchao Xia <xiawenc@linux.vnet.ibm.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>

3 files changed, 26 insertions, 3 deletions

diff --git a/block.c b/block.c
index 275d387354..e5f214d135 100644
--- a/block.c
+++ b/block.c
@@ -218,6 +218,16 @@ static void bdrv_io_limits_intercept(BlockDriverState *bs,
     qemu_co_queue_next(&bs->throttled_reqs[is_write]);
 }
 
+size_t bdrv_opt_mem_align(BlockDriverState *bs)
+{
+    if (!bs || !bs->drv) {
+        /* 4k should be on the safe side */
+        return 4096;
+    }
+
+    return bs->bl.opt_mem_alignment;
+}
+
 /* check if the path starts with "<protocol>:" */
 static int path_has_protocol(const char *path)
 {
@@ -497,6 +507,9 @@ int bdrv_refresh_limits(BlockDriverState *bs)
     if (bs->file) {
         bdrv_refresh_limits(bs->file);
         bs->bl.opt_transfer_length = bs->file->bl.opt_transfer_length;
+        bs->bl.opt_mem_alignment = bs->file->bl.opt_mem_alignment;
+    } else {
+        bs->bl.opt_mem_alignment = 512;
     }
 
     if (bs->backing_hd) {
@@ -504,6 +517,9 @@ int bdrv_refresh_limits(BlockDriverState *bs)
         bs->bl.opt_transfer_length =
             MAX(bs->bl.opt_transfer_length,
                 bs->backing_hd->bl.opt_transfer_length);
+        bs->bl.opt_mem_alignment =
+            MAX(bs->bl.opt_mem_alignment,
+                bs->backing_hd->bl.opt_mem_alignment);
     }
 
     /* Then let the driver override it */
@@ -4797,7 +4813,7 @@ void bdrv_set_buffer_alignment(BlockDriverState *bs, int align)
 
 void *qemu_blockalign(BlockDriverState *bs, size_t size)
 {
-    return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
+    return qemu_memalign(bdrv_opt_mem_align(bs), size);
 }
 
 /*
@@ -4806,12 +4822,13 @@ void *qemu_blockalign(BlockDriverState *bs, size_t size)
 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
 {
     int i;
+    size_t alignment = bdrv_opt_mem_align(bs);
 
     for (i = 0; i < qiov->niov; i++) {
-        if ((uintptr_t) qiov->iov[i].iov_base % bs->buffer_alignment) {
+        if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
             return false;
         }
-        if (qiov->iov[i].iov_len % bs->buffer_alignment) {
+        if (qiov->iov[i].iov_len % alignment) {
             return false;
         }
     }
diff --git a/include/block/block.h b/include/block/block.h
index 00a8790660..332ebb94fb 100644
--- a/include/block/block.h
+++ b/include/block/block.h
@@ -428,6 +428,9 @@ void bdrv_img_create(const char *filename, const char *fmt,
                      char *options, uint64_t img_size, int flags,
                      Error **errp, bool quiet);
 
+/* Returns the alignment in bytes that is required so that no bounce buffer
+ * is required throughout the stack */
+size_t bdrv_opt_mem_align(BlockDriverState *bs);
 void bdrv_set_buffer_alignment(BlockDriverState *bs, int align);
 void *qemu_blockalign(BlockDriverState *bs, size_t size);
 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov);
diff --git a/include/block/block_int.h b/include/block/block_int.h
index f6fa1f6f36..74a78a6831 100644
--- a/include/block/block_int.h
+++ b/include/block/block_int.h
@@ -258,6 +258,9 @@ typedef struct BlockLimits {
 
     /* optimal transfer length in sectors */
     int opt_transfer_length;
+
+    /* memory alignment so that no bounce buffer is needed */
+    size_t opt_mem_alignment;
 } BlockLimits;
 
 /*