]> xenbits.xensource.com Git - qemu-xen-4.1-testing.git/commitdiff
qcow2: Avoid bounce buffers for AIO read requests
authorKevin Wolf <kwolf@redhat.com>
Mon, 13 Sep 2010 16:08:52 +0000 (18:08 +0200)
committerKevin Wolf <kwolf@redhat.com>
Tue, 21 Sep 2010 13:39:42 +0000 (15:39 +0200)
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>
block/qcow2-cluster.c
block/qcow2.c
block/qcow2.h

index cb2e33f1ac0fc3b063a3b06de73f2b79905f361e..fb4224a669ebeeff8e8b6545008aaf93fd6e2f0f 100644 (file)
@@ -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);
index a53014dbdaad00fc4e8a00419d2fdf0ea9cbbefe..4a688b48638b7fa4ab6112ea936f022f82a85a39 100644 (file)
@@ -311,8 +311,8 @@ 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)
@@ -321,7 +321,9 @@ 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;
 }
 
@@ -333,6 +335,7 @@ 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;
@@ -397,15 +400,19 @@ 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 */
@@ -415,6 +422,11 @@ 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) {
@@ -423,15 +435,17 @@ 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,
@@ -445,7 +459,7 @@ 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;
@@ -454,8 +468,11 @@ 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;
@@ -465,9 +482,23 @@ 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,
@@ -481,11 +512,8 @@ 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);
 }
 
@@ -501,13 +529,17 @@ 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;
index 3ff162efcdcd56d2a18349febcd230fcb323d292..356a34af434a73d2800a3444a00f61936240093f 100644 (file)
@@ -166,8 +166,8 @@ 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);