diff mbox

[v3,1/2] Add bdrv_aio_multiwrite

Message ID 1252511618-19497-2-git-send-email-kwolf@redhat.com
State Superseded
Headers show

Commit Message

Kevin Wolf Sept. 9, 2009, 3:53 p.m. UTC 
One performance problem of qcow2 during the initial image growth are
sequential writes that are not cluster aligned. In this case, when a first
requests requires to allocate a new cluster but writes only to the first
couple of sectors in that cluster, the rest of the cluster is zeroed - just
to be overwritten by the following second request that fills up the cluster.

Let's try to merge sequential write requests to the same cluster, so we can
avoid to write the zero padding to the disk in the first place.

As a nice side effect, also other formats take advantage of dealing with less
and larger requests.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 block.c       |  183 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 block.h       |   15 +++++
 block_int.h   |    6 ++
 cutils.c      |   25 ++++++++
 qemu-common.h |    1 +
 5 files changed, 230 insertions(+), 0 deletions(-)

Comments

Kevin Wolf Sept. 10, 2009, 7:07 a.m. UTC | #1 
Am 10.09.2009 01:08, schrieb Juan Quintela:
> Kevin Wolf <kwolf@redhat.com> wrote:
> 
> Hi again
> 
> 
>> +            QEMUIOVector *qiov = qemu_mallocz(sizeof(*qiov));
>> +            qemu_iovec_init(qiov,
>> +                reqs[outidx].qiov->niov + reqs[i].qiov->niov + 1);
>> +
>> +            // Add the first request to the merged one. If the requests are
>> +            // overlapping, drop the last sectors of the first request.
>> +            size = (reqs[i].sector - reqs[outidx].sector) << 9;
>> +            qemu_iovec_concat(qiov, reqs[outidx].qiov, size);
>> +
>> +            // We might need to add some zeros between the two requests
>> +            if (reqs[i].sector > oldreq_last) {
>> +                size_t zero_bytes = (reqs[i].sector - oldreq_last) << 9;
>> +                uint8_t *buf = qemu_blockalign(bs, zero_bytes);
>> +                memset(buf, 0, zero_bytes);
>> +                qemu_iovec_add(qiov, buf, zero_bytes);
>> +                mcb->callbacks[i].free_buf = buf;
>> +            }
>> +
>> +            // Add the second request
>> +            qemu_iovec_concat(qiov, reqs[i].qiov, reqs[i].qiov->size);
>> +
>> +            reqs[outidx].nb_sectors += reqs[i].nb_sectors;
>> +            reqs[outidx].qiov = qiov;
> 
> What frees reqs[outidx].qiov previous value, or new one for that matter?

Your decision where to stop quoting the code is kind of funny. The very
next line is:

mcb->callbacks[i].free_qiov = reqs[outidx].qiov;

This saves each newly allocated qiov. It is later freed in
multiwrite_user_cb. To free the original value of qiov is the task of
the caller.

> I can't see where it is done.  As far as I can see, we are losing both
> the ones that we are overwritten and the ones for the request that got merged.

Hope you see them now. :-)

Kevin
Christoph Hellwig Sept. 10, 2009, 10:44 p.m. UTC | #2 
One thing that concerns me here is that we keep adding more memory
allocations to the I/O path.  At least on fast SSDs even kernel memory
allocations are a performance problem and they're much faster than
userspace ones.
Kevin Wolf Sept. 11, 2009, 6:29 a.m. UTC | #3 
Am 11.09.2009 00:44, schrieb Christoph Hellwig:
> One thing that concerns me here is that we keep adding more memory
> allocations to the I/O path.  At least on fast SSDs even kernel memory
> allocations are a performance problem and they're much faster than
> userspace ones.  

In the non-merging case we have one additional allocation, the one for
mcb. Maybe we could do something like the AIOPool here to avoid some
mallocs.

In the merging case we need to allocate new IO vectors. I don't see a
way around this, though, and I hope that the benefit of a saved write
request outweighs the malloc costs. If not, it probably was a bad idea
to accept your suggestion to make it generic code instead of just
merging requests for qcow2.

Kevin
Christoph Hellwig Sept. 11, 2009, 6:36 p.m. UTC | #4 
On Fri, Sep 11, 2009 at 08:29:09AM +0200, Kevin Wolf wrote:
> Am 11.09.2009 00:44, schrieb Christoph Hellwig:
> > One thing that concerns me here is that we keep adding more memory
> > allocations to the I/O path.  At least on fast SSDs even kernel memory
> > allocations are a performance problem and they're much faster than
> > userspace ones.  
> 
> In the non-merging case we have one additional allocation, the one for
> mcb. Maybe we could do something like the AIOPool here to avoid some
> mallocs.
> 
> In the merging case we need to allocate new IO vectors. I don't see a
> way around this, though, and I hope that the benefit of a saved write
> request outweighs the malloc costs. If not, it probably was a bad idea
> to accept your suggestion to make it generic code instead of just
> merging requests for qcow2.

Right now I'm just thinking about the potential issues.  The problem
is that the trade offs maybe be very different for different scenarios.
For a normal disk which is by far the most common option (and will be
for a long time) the merging is a clear benefit.  For SSDs we might have
to do different optimizations.  If it's not a big issue avoiding too
many memory allocations and rather doing fewer larger ones is in general
a good thing, though.
diff mbox

Patch

diff --git a/block.c b/block.c
index 033957d..45ad6fb 100644
--- a/block.c
+++ b/block.c
@@ -1354,6 +1354,189 @@  BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
     return ret;
 }
 
+
+typedef struct MultiwriteCB {
+    int error;
+    int num_requests;
+    int num_callbacks;
+    struct {
+        BlockDriverCompletionFunc *cb;
+        void *opaque;
+        QEMUIOVector *free_qiov;
+        void *free_buf;
+    } callbacks[];
+} MultiwriteCB;
+
+static void multiwrite_user_cb(MultiwriteCB *mcb)
+{
+    int i;
+
+    for (i = 0; i < mcb->num_callbacks; i++) {
+        mcb->callbacks[i].cb(mcb->callbacks[i].opaque, mcb->error);
+        qemu_free(mcb->callbacks[i].free_qiov);
+        qemu_free(mcb->callbacks[i].free_buf);
+    }
+}
+
+static void multiwrite_cb(void *opaque, int ret)
+{
+    MultiwriteCB *mcb = opaque;
+
+    if (ret < 0) {
+        mcb->error = ret;
+        multiwrite_user_cb(mcb);
+    }
+
+    mcb->num_requests--;
+    if (mcb->num_requests == 0) {
+        if (mcb->error == 0) {
+            multiwrite_user_cb(mcb);
+        }
+        qemu_free(mcb);
+    }
+}
+
+static int multiwrite_req_compare(const void *a, const void *b)
+{
+    return (((BlockRequest*) a)->sector - ((BlockRequest*) b)->sector);
+}
+
+/*
+ * Takes a bunch of requests and tries to merge them. Returns the number of
+ * requests that remain after merging.
+ */
+static int multiwrite_merge(BlockDriverState *bs, BlockRequest *reqs,
+    int num_reqs, MultiwriteCB *mcb)
+{
+    int i, outidx;
+
+    // Sort requests by start sector
+    qsort(reqs, num_reqs, sizeof(*reqs), &multiwrite_req_compare);
+
+    // Check if adjacent requests touch the same clusters. If so, combine them,
+    // filling up gaps with zero sectors.
+    outidx = 0;
+    for (i = 1; i < num_reqs; i++) {
+        int merge = 0;
+        int64_t oldreq_last = reqs[outidx].sector + reqs[outidx].nb_sectors;
+
+        // This handles the cases that are valid for all block drivers, namely
+        // exactly sequential writes and overlapping writes.
+        if (reqs[i].sector <= oldreq_last) {
+            merge = 1;
+        }
+
+        // The block driver may decide that it makes sense to combine requests
+        // even if there is a gap of some sectors between them. In this case,
+        // the gap is filled with zeros (therefore only applicable for yet
+        // unused space in format like qcow2).
+        if (!merge && bs->drv->bdrv_merge_requests) {
+            merge = bs->drv->bdrv_merge_requests(bs, &reqs[outidx], &reqs[i]);
+        }
+
+        if (merge) {
+            size_t size;
+            QEMUIOVector *qiov = qemu_mallocz(sizeof(*qiov));
+            qemu_iovec_init(qiov,
+                reqs[outidx].qiov->niov + reqs[i].qiov->niov + 1);
+
+            // Add the first request to the merged one. If the requests are
+            // overlapping, drop the last sectors of the first request.
+            size = (reqs[i].sector - reqs[outidx].sector) << 9;
+            qemu_iovec_concat(qiov, reqs[outidx].qiov, size);
+
+            // We might need to add some zeros between the two requests
+            if (reqs[i].sector > oldreq_last) {
+                size_t zero_bytes = (reqs[i].sector - oldreq_last) << 9;
+                uint8_t *buf = qemu_blockalign(bs, zero_bytes);
+                memset(buf, 0, zero_bytes);
+                qemu_iovec_add(qiov, buf, zero_bytes);
+                mcb->callbacks[i].free_buf = buf;
+            }
+
+            // Add the second request
+            qemu_iovec_concat(qiov, reqs[i].qiov, reqs[i].qiov->size);
+
+            reqs[outidx].nb_sectors += reqs[i].nb_sectors;
+            reqs[outidx].qiov = qiov;
+
+            mcb->callbacks[i].free_qiov = reqs[outidx].qiov;
+        } else {
+            outidx++;
+            reqs[outidx].sector     = reqs[i].sector;
+            reqs[outidx].nb_sectors = reqs[i].nb_sectors;
+            reqs[outidx].qiov       = reqs[i].qiov;
+        }
+    }
+
+    return outidx + 1;
+}
+
+/*
+ * Submit multiple AIO write requests at once.
+ *
+ * On success, the function returns 0 and all requests in the reqs array have
+ * been submitted. In error case this function returns -1, and any of the
+ * requests may or may not be submitted yet. In particular, this means that the
+ * callback will be called for some of the requests, for others it won't. The
+ * caller must check the error field of the BlockRequest to wait for the right
+ * callbacks (if error != 0, no callback will be called).
+ *
+ * The implementation may modify the contents of the reqs array, e.g. to merge
+ * requests. However, the fields opaque and error are left unmodified as they
+ * are used to signal failure for a single request to the caller.
+ */
+int bdrv_aio_multiwrite(BlockDriverState *bs, BlockRequest *reqs, int num_reqs)
+{
+    BlockDriverAIOCB *acb;
+    MultiwriteCB *mcb;
+    int i;
+
+    if (num_reqs == 0) {
+        return 0;
+    }
+
+    // Create MultiwriteCB structure
+    mcb = qemu_mallocz(sizeof(*mcb) + num_reqs * sizeof(*mcb->callbacks));
+    mcb->num_requests = 0;
+    mcb->num_callbacks = num_reqs;
+
+    for (i = 0; i < num_reqs; i++) {
+        mcb->callbacks[i].cb = reqs[i].cb;
+        mcb->callbacks[i].opaque = reqs[i].opaque;
+    }
+
+    // Check for mergable requests
+    num_reqs = multiwrite_merge(bs, reqs, num_reqs, mcb);
+
+    // Run the aio requests
+    for (i = 0; i < num_reqs; i++) {
+        acb = bdrv_aio_writev(bs, reqs[i].sector, reqs[i].qiov,
+            reqs[i].nb_sectors, multiwrite_cb, mcb);
+
+        if (acb == NULL) {
+            // We can only fail the whole thing if no request has been
+            // submitted yet. Otherwise we'll wait for the submitted AIOs to
+            // complete and report the error in the callback.
+            if (mcb->num_requests == 0) {
+                reqs[i].error = EIO;
+                goto fail;
+            } else {
+                mcb->error = EIO;
+                break;
+            }
+        } else {
+            mcb->num_requests++;
+        }
+    }
+
+    return 0;
+
+fail:
+    free(mcb);
+    return -1;
+}
+
 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
 {
     acb->pool->cancel(acb);
diff --git a/block.h b/block.h
index 28bf357..ea69052 100644
--- a/block.h
+++ b/block.h
@@ -87,6 +87,21 @@  BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
                                   BlockDriverCompletionFunc *cb, void *opaque);
 void bdrv_aio_cancel(BlockDriverAIOCB *acb);
 
+typedef struct BlockRequest {
+    /* Fields to be filled by multiwrite caller */
+    int64_t sector;
+    int nb_sectors;
+    QEMUIOVector *qiov;
+    BlockDriverCompletionFunc *cb;
+    void *opaque;
+
+    /* Filled by multiwrite implementation */
+    int error;
+} BlockRequest;
+
+int bdrv_aio_multiwrite(BlockDriverState *bs, BlockRequest *reqs,
+    int num_reqs);
+
 /* sg packet commands */
 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf);
 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
diff --git a/block_int.h b/block_int.h
index 0902fd4..3e4b4a3 100644
--- a/block_int.h
+++ b/block_int.h
@@ -70,6 +70,12 @@  struct BlockDriver {
         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
         BlockDriverCompletionFunc *cb, void *opaque);
 
+    int (*bdrv_aio_multiwrite)(BlockDriverState *bs, BlockRequest *reqs,
+        int num_reqs);
+    int (*bdrv_merge_requests)(BlockDriverState *bs, BlockRequest* a,
+        BlockRequest *b);
+
+
     const char *protocol_name;
     int (*bdrv_truncate)(BlockDriverState *bs, int64_t offset);
     int64_t (*bdrv_getlength)(BlockDriverState *bs);
diff --git a/cutils.c b/cutils.c
index bd9a019..ffe5c71 100644
--- a/cutils.c
+++ b/cutils.c
@@ -151,6 +151,31 @@  void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len)
     ++qiov->niov;
 }
 
+/*
+ * Copies iovecs from src to the end dst until src is completely copied or the
+ * total size of the copied iovec reaches size. The size of the last copied
+ * iovec is changed in order to fit the specified total size if it isn't a
+ * perfect fit already.
+ */
+void qemu_iovec_concat(QEMUIOVector *dst, QEMUIOVector *src, size_t size)
+{
+    int i;
+    size_t done;
+
+    assert(dst->nalloc != -1);
+
+    done = 0;
+    for (i = 0; (i < src->niov) && (done != size); i++) {
+        if (done + src->iov[i].iov_len > size) {
+            qemu_iovec_add(dst, src->iov[i].iov_base, size - done);
+            break;
+        } else {
+            qemu_iovec_add(dst, src->iov[i].iov_base, src->iov[i].iov_len);
+        }
+        done += src->iov[i].iov_len;
+    }
+}
+
 void qemu_iovec_destroy(QEMUIOVector *qiov)
 {
     assert(qiov->nalloc != -1);
diff --git a/qemu-common.h b/qemu-common.h
index 74ac88f..f3cfb68 100644
--- a/qemu-common.h
+++ b/qemu-common.h
@@ -223,6 +223,7 @@  typedef struct QEMUIOVector {
 void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint);
 void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov);
 void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len);
+void qemu_iovec_concat(QEMUIOVector *dst, QEMUIOVector *src, size_t size);
 void qemu_iovec_destroy(QEMUIOVector *qiov);
 void qemu_iovec_reset(QEMUIOVector *qiov);
 void qemu_iovec_to_buffer(QEMUIOVector *qiov, void *buf);