From 1486e716880d4a569ba4afa6d19660b3ccc458a4 Mon Sep 17 00:00:00 2001 From: Kevin Wolf Date: Tue, 23 Nov 2010 10:31:45 -0200 Subject: [RHEL6 qemu-kvm PATCH 17/21] qcow2: Avoid bounce buffers for AIO read requests RH-Author: Kevin Wolf Message-id: <1290508306-11607-10-git-send-email-kwolf@redhat.com> Patchwork-id: 13833 O-Subject: [RHEL-6.1 qemu-kvm PATCH v2 09/10] qcow2: Avoid bounce buffers for AIO read requests Bugzilla: 653972 RH-Acked-by: Avi Kivity RH-Acked-by: Christoph Hellwig RH-Acked-by: Jes Sorensen Bugzilla: 653972 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 (cherry picked from commit bd28f835652e396841bb73080a07bcc08fe21bf0) --- block/qcow2-cluster.c | 8 ++++- block/qcow2.c | 86 +++++++++++++++++++++++++++++++++--------------- block/qcow2.h | 4 +- 3 files changed, 68 insertions(+), 30 deletions(-) Signed-off-by: Eduardo Habkost --- block/qcow2-cluster.c | 8 ++++- block/qcow2.c | 86 +++++++++++++++++++++++++++++++++--------------- block/qcow2.h | 4 +- 3 files changed, 68 insertions(+), 30 deletions(-) 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 @@ static int qcow_read(BlockDriverState *bs, int64_t sector_num, 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 @@ static int qcow_read(BlockDriverState *bs, int64_t sector_num, 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 9c02067..780e282 100644 --- a/block/qcow2.c +++ b/block/qcow2.c @@ -313,8 +313,8 @@ static int qcow_is_allocated(BlockDriverState *bs, int64_t sector_num, } /* 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) @@ -323,7 +323,9 @@ int qcow2_backing_read1(BlockDriverState *bs, 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; } @@ -335,6 +337,7 @@ typedef struct QCowAIOCB { 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; @@ -399,15 +402,19 @@ static void qcow_aio_read_cb(void *opaque, int ret) /* 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 */ @@ -417,6 +424,11 @@ static void qcow_aio_read_cb(void *opaque, int ret) /* 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) { @@ -425,15 +437,17 @@ static void qcow_aio_read_cb(void *opaque, int ret) 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, @@ -447,7 +461,7 @@ static void qcow_aio_read_cb(void *opaque, int ret) } } 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; @@ -456,8 +470,11 @@ static void qcow_aio_read_cb(void *opaque, int ret) /* 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; @@ -467,9 +484,23 @@ static void qcow_aio_read_cb(void *opaque, int ret) 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, @@ -483,11 +514,8 @@ static void qcow_aio_read_cb(void *opaque, int ret) 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); } @@ -503,13 +531,17 @@ static QCowAIOCB *qcow_aio_setup(BlockDriverState *bs, 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 d976e2b..023fa67 100644 --- a/block/qcow2.h +++ b/block/qcow2.h @@ -159,8 +159,8 @@ static inline int64_t align_offset(int64_t offset, int n) // 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); -- 1.7.3.2