qcow2: Avoid bounce buffers for AIO read requests
qcow2 used to use bounce buffers for any AIO requests. This does not only imply
unnecessary copying, but also unbounded allocations which should be avoided.
This patch removes bounce buffers from the normal AIO read path, and constrains
them to a constant size for encrypted images.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index cb2e33f..fb4224a 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -350,6 +350,8 @@
BDRVQcowState *s = bs->opaque;
int ret, index_in_cluster, n, n1;
uint64_t cluster_offset;
+ struct iovec iov;
+ QEMUIOVector qiov;
while (nb_sectors > 0) {
n = nb_sectors;
@@ -364,7 +366,11 @@
if (!cluster_offset) {
if (bs->backing_hd) {
/* read from the base image */
- n1 = qcow2_backing_read1(bs->backing_hd, sector_num, buf, n);
+ iov.iov_base = buf;
+ iov.iov_len = n * 512;
+ qemu_iovec_init_external(&qiov, &iov, 1);
+
+ n1 = qcow2_backing_read1(bs->backing_hd, &qiov, sector_num, n);
if (n1 > 0) {
BLKDBG_EVENT(bs->file, BLKDBG_READ_BACKING);
ret = bdrv_read(bs->backing_hd, sector_num, buf, n1);
diff --git a/block/qcow2.c b/block/qcow2.c
index a53014d..4a688b4 100644
--- a/block/qcow2.c
+++ b/block/qcow2.c
@@ -311,8 +311,8 @@
}
/* handle reading after the end of the backing file */
-int qcow2_backing_read1(BlockDriverState *bs,
- int64_t sector_num, uint8_t *buf, int nb_sectors)
+int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
+ int64_t sector_num, int nb_sectors)
{
int n1;
if ((sector_num + nb_sectors) <= bs->total_sectors)
@@ -321,7 +321,9 @@
n1 = 0;
else
n1 = bs->total_sectors - sector_num;
- memset(buf + n1 * 512, 0, 512 * (nb_sectors - n1));
+
+ qemu_iovec_memset(qiov, 0, 512 * (nb_sectors - n1));
+
return n1;
}
@@ -333,6 +335,7 @@
void *orig_buf;
int remaining_sectors;
int cur_nr_sectors; /* number of sectors in current iteration */
+ uint64_t bytes_done;
uint64_t cluster_offset;
uint8_t *cluster_data;
BlockDriverAIOCB *hd_aiocb;
@@ -397,15 +400,19 @@
/* nothing to do */
} else {
if (s->crypt_method) {
- qcow2_encrypt_sectors(s, acb->sector_num, acb->buf, acb->buf,
- acb->cur_nr_sectors, 0,
- &s->aes_decrypt_key);
+ qcow2_encrypt_sectors(s, acb->sector_num, acb->cluster_data,
+ acb->cluster_data, acb->cur_nr_sectors, 0, &s->aes_decrypt_key);
+ qemu_iovec_reset(&acb->hd_qiov);
+ qemu_iovec_copy(&acb->hd_qiov, acb->qiov, acb->bytes_done,
+ acb->cur_nr_sectors * 512);
+ qemu_iovec_from_buffer(&acb->hd_qiov, acb->cluster_data,
+ 512 * acb->cur_nr_sectors);
}
}
acb->remaining_sectors -= acb->cur_nr_sectors;
acb->sector_num += acb->cur_nr_sectors;
- acb->buf += acb->cur_nr_sectors * 512;
+ acb->bytes_done += acb->cur_nr_sectors * 512;
if (acb->remaining_sectors == 0) {
/* request completed */
@@ -415,6 +422,11 @@
/* prepare next AIO request */
acb->cur_nr_sectors = acb->remaining_sectors;
+ if (s->crypt_method) {
+ acb->cur_nr_sectors = MIN(acb->cur_nr_sectors,
+ QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
+ }
+
ret = qcow2_get_cluster_offset(bs, acb->sector_num << 9,
&acb->cur_nr_sectors, &acb->cluster_offset);
if (ret < 0) {
@@ -423,15 +435,17 @@
index_in_cluster = acb->sector_num & (s->cluster_sectors - 1);
+ qemu_iovec_reset(&acb->hd_qiov);
+ qemu_iovec_copy(&acb->hd_qiov, acb->qiov, acb->bytes_done,
+ acb->cur_nr_sectors * 512);
+
if (!acb->cluster_offset) {
+
if (bs->backing_hd) {
/* read from the base image */
- n1 = qcow2_backing_read1(bs->backing_hd, acb->sector_num,
- acb->buf, acb->cur_nr_sectors);
+ n1 = qcow2_backing_read1(bs->backing_hd, &acb->hd_qiov,
+ acb->sector_num, acb->cur_nr_sectors);
if (n1 > 0) {
- acb->hd_iov.iov_base = (void *)acb->buf;
- acb->hd_iov.iov_len = acb->cur_nr_sectors * 512;
- qemu_iovec_init_external(&acb->hd_qiov, &acb->hd_iov, 1);
BLKDBG_EVENT(bs->file, BLKDBG_READ_BACKING_AIO);
acb->hd_aiocb = bdrv_aio_readv(bs->backing_hd, acb->sector_num,
&acb->hd_qiov, acb->cur_nr_sectors,
@@ -445,7 +459,7 @@
}
} else {
/* Note: in this case, no need to wait */
- memset(acb->buf, 0, 512 * acb->cur_nr_sectors);
+ qemu_iovec_memset(&acb->hd_qiov, 0, 512 * acb->cur_nr_sectors);
ret = qcow_schedule_bh(qcow_aio_read_bh, acb);
if (ret < 0)
goto done;
@@ -454,8 +468,11 @@
/* add AIO support for compressed blocks ? */
if (qcow2_decompress_cluster(bs, acb->cluster_offset) < 0)
goto done;
- memcpy(acb->buf, s->cluster_cache + index_in_cluster * 512,
- 512 * acb->cur_nr_sectors);
+
+ qemu_iovec_from_buffer(&acb->hd_qiov,
+ s->cluster_cache + index_in_cluster * 512,
+ 512 * acb->cur_nr_sectors);
+
ret = qcow_schedule_bh(qcow_aio_read_bh, acb);
if (ret < 0)
goto done;
@@ -465,9 +482,23 @@
goto done;
}
- acb->hd_iov.iov_base = (void *)acb->buf;
- acb->hd_iov.iov_len = acb->cur_nr_sectors * 512;
- qemu_iovec_init_external(&acb->hd_qiov, &acb->hd_iov, 1);
+ if (s->crypt_method) {
+ /*
+ * For encrypted images, read everything into a temporary
+ * contiguous buffer on which the AES functions can work.
+ */
+ if (!acb->cluster_data) {
+ acb->cluster_data =
+ qemu_mallocz(QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
+ }
+
+ assert(acb->cur_nr_sectors <=
+ QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
+ qemu_iovec_reset(&acb->hd_qiov);
+ qemu_iovec_add(&acb->hd_qiov, acb->cluster_data,
+ 512 * acb->cur_nr_sectors);
+ }
+
BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
acb->hd_aiocb = bdrv_aio_readv(bs->file,
(acb->cluster_offset >> 9) + index_in_cluster,
@@ -481,11 +512,8 @@
return;
done:
- if (acb->qiov->niov > 1) {
- qemu_iovec_from_buffer(acb->qiov, acb->orig_buf, acb->qiov->size);
- qemu_vfree(acb->orig_buf);
- }
acb->common.cb(acb->common.opaque, ret);
+ qemu_iovec_destroy(&acb->hd_qiov);
qemu_aio_release(acb);
}
@@ -501,13 +529,17 @@
acb->hd_aiocb = NULL;
acb->sector_num = sector_num;
acb->qiov = qiov;
- if (qiov->niov > 1) {
- acb->buf = acb->orig_buf = qemu_blockalign(bs, qiov->size);
- if (is_write)
- qemu_iovec_to_buffer(qiov, acb->buf);
- } else {
+
+ if (!is_write) {
+ qemu_iovec_init(&acb->hd_qiov, qiov->niov);
+ } else if (qiov->niov == 1) {
acb->buf = (uint8_t *)qiov->iov->iov_base;
+ } else {
+ acb->buf = acb->orig_buf = qemu_blockalign(bs, qiov->size);
+ qemu_iovec_to_buffer(qiov, acb->buf);
}
+
+ acb->bytes_done = 0;
acb->remaining_sectors = nb_sectors;
acb->cur_nr_sectors = 0;
acb->cluster_offset = 0;
diff --git a/block/qcow2.h b/block/qcow2.h
index 3ff162e..356a34a 100644
--- a/block/qcow2.h
+++ b/block/qcow2.h
@@ -166,8 +166,8 @@
// FIXME Need qcow2_ prefix to global functions
/* qcow2.c functions */
-int qcow2_backing_read1(BlockDriverState *bs,
- int64_t sector_num, uint8_t *buf, int nb_sectors);
+int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
+ int64_t sector_num, int nb_sectors);
/* qcow2-refcount.c functions */
int qcow2_refcount_init(BlockDriverState *bs);
