qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [RFC PATCH 3/3] virtio-blk: introduce multiread


From: Peter Lieven
Subject: Re: [Qemu-devel] [RFC PATCH 3/3] virtio-blk: introduce multiread
Date: Thu, 04 Dec 2014 08:16:55 +0100
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.2.0

On 03.12.2014 18:25, Max Reitz wrote:
On 2014-12-02 at 15:33, Peter Lieven wrote:
this patch finally introduce multiread support to virtio-blk while
multiwrite support was there for a long time read support was missing.

To achieve this the patch does serveral things which might need futher
explaination:

  - the whole merge and multireq logic is moved from block.c into
    virtio-blk. This is move is a preparation for directly creating a
    coroutine out of virtio-blk.

  - requests are only merged if they are strictly sequential and no
    longer sorted. This simplification decreases overhead and reduces
    latency. It will also merge some requests which were unmergable before.

    The old algorithm took up to 32 requests sorted them and tried to merge
    them. The outcome was anything between 1 and 32 requests. In case of
    32 requests there were 31 requests unnecessarily delayed.

    On the other hand lets imagine e.g. 16 unmergeable requests followed
    by 32 mergable requests. The latter 32 requests would have been split
    into two 16 byte requests.

    Last the simplified logic allows for a fast path if we have only a
    single request in the multirequest. In this case the request is sent as
    ordinary request without mulltireq callbacks.

As a first benchmark I installed Ubuntu 14.04.1 on a ramdisk. The number of
merged requests is in the same order while the latency is slightly decreased.
One should not stick to much to the numbers because the number of wr_requests
are highly fluctuant. I hope the numbers show that this patch is at least
not causing to big harm:

Cmdline:
qemu-system-x86_64 -m 1024 -smp 2 -enable-kvm -cdrom 
ubuntu-14.04.1-server-amd64.iso \
  -drive if=virtio,file=/tmp/ubuntu.raw -monitor stdio

Before:
virtio0: rd_bytes=150979072 wr_bytes=2591138816 rd_operations=18475 
wr_operations=69216
          flush_operations=15343 wr_total_time_ns=26969283701 
rd_total_time_ns=1018449432
          flush_total_time_ns=562596819 rd_merged=0 wr_merged=10453

After:
virtio0: rd_bytes=148112896 wr_bytes=2845834240 rd_operations=17535 
wr_operations=72197
          flush_operations=15760 wr_total_time_ns=26104971623 
rd_total_time_ns=1012926283
          flush_total_time_ns=564414752 rd_merged=517 wr_merged=9859

Some first numbers of improved read performance while booting:

The Ubuntu 14.04.1 vServer from above:
virtio0: rd_bytes=86158336 wr_bytes=688128 rd_operations=5851 wr_operations=70
          flush_operations=4 wr_total_time_ns=10886705 
rd_total_time_ns=416688595
          flush_total_time_ns=288776 rd_merged=1297 wr_merged=2

Windows 2012R2:
virtio0: rd_bytes=176559104 wr_bytes=61859840 rd_operations=7200 
wr_operations=360
          flush_operations=68 wr_total_time_ns=34344992718 
rd_total_time_ns=134386844669
          flush_total_time_ns=18115517 rd_merged=641 wr_merged=216

Signed-off-by: Peter Lieven <address@hidden>
---
  hw/block/dataplane/virtio-blk.c |   10 +-
  hw/block/virtio-blk.c           |  222 ++++++++++++++++++++++++---------------
  include/hw/virtio/virtio-blk.h  |   23 ++--
  3 files changed, 156 insertions(+), 99 deletions(-)

So, since you CC-ed me, I guess you're expecting a reply from me. I feel like I'm saying this more often recently (about anything, not just about virtio), but I don't know anything about virtio, so you should take anything from me with a not just a grain but more of a pack of salt.

Thats fine with me, I appreciate your feedback nevertheless.


diff --git a/hw/block/dataplane/virtio-blk.c b/hw/block/dataplane/virtio-blk.c
index 1222a37..aa4ad91 100644
--- a/hw/block/dataplane/virtio-blk.c
+++ b/hw/block/dataplane/virtio-blk.c
@@ -96,9 +96,8 @@ static void handle_notify(EventNotifier *e)
      event_notifier_test_and_clear(&s->host_notifier);
      blk_io_plug(s->conf->conf.blk);
      for (;;) {
-        MultiReqBuffer mrb = {
-            .num_writes = 0,
-        };
+        MultiReqBuffer mrb_rd = {};
+        MultiReqBuffer mrb_wr = {.is_write = 1};
          int ret;
            /* Disable guest->host notifies to avoid unnecessary vmexits */
@@ -117,10 +116,11 @@ static void handle_notify(EventNotifier *e)
req->elem.in_num,
req->elem.index);
  -            virtio_blk_handle_request(req, &mrb);
+            virtio_blk_handle_request(req, &mrb_wr, &mrb_rd);
          }
  -        virtio_submit_multiwrite(s->conf->conf.blk, &mrb);
+        virtio_submit_multireq(s->conf->conf.blk, &mrb_wr);
+        virtio_submit_multireq(s->conf->conf.blk, &mrb_rd);
            if (likely(ret == -EAGAIN)) { /* vring emptied */
              /* Re-enable guest->host notifies and stop processing the vring.
diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
index 490f961..f522bfc 100644
--- a/hw/block/virtio-blk.c
+++ b/hw/block/virtio-blk.c
@@ -22,12 +22,15 @@
  #include "dataplane/virtio-blk.h"
  #include "migration/migration.h"
  #include "block/scsi.h"
+#include "block/block_int.h"
  #ifdef __linux__
  # include <scsi/sg.h>
  #endif
  #include "hw/virtio/virtio-bus.h"
  #include "hw/virtio/virtio-access.h"
  +/* #define DEBUG_MULTIREQ */
+
  VirtIOBlockReq *virtio_blk_alloc_request(VirtIOBlock *s)
  {
      VirtIOBlockReq *req = g_slice_new(VirtIOBlockReq);
@@ -88,6 +91,11 @@ static void virtio_blk_rw_complete(void *opaque, int ret)
        trace_virtio_blk_rw_complete(req, ret);
  +#ifdef DEBUG_MULTIREQ
+    printf("virtio_blk_rw_complete p %p ret %d\n",
+           req, ret);
+#endif
+
      if (ret) {
          int p = virtio_ldl_p(VIRTIO_DEVICE(req->dev), &req->out.type);
          bool is_read = !(p & VIRTIO_BLK_T_OUT);
@@ -257,24 +265,63 @@ static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
      virtio_blk_free_request(req);
  }
  -void virtio_submit_multiwrite(BlockBackend *blk, MultiReqBuffer *mrb)
+static void virtio_multireq_cb(void *opaque, int ret)
+{
+    MultiReqBuffer *mrb = opaque;
+    int i;
+#ifdef DEBUG_MULTIREQ
+    printf("virtio_multireq_cb: p %p sector_num %ld nb_sectors %d is_write %d 
num_reqs %d\n",
+           mrb, mrb->sector_num, mrb->nb_sectors, mrb->is_write, 
mrb->num_reqs);
+#endif
+    for (i = 0; i < mrb->num_reqs; i++) {
+        virtio_blk_rw_complete(mrb->reqs[i], ret);
+    }
+
+    qemu_iovec_destroy(&mrb->qiov);
+    g_free(mrb);
+}
+
+void virtio_submit_multireq(BlockBackend *blk, MultiReqBuffer *mrb0)
  {
-    int i, ret;
+    MultiReqBuffer *mrb = NULL;
  -    if (!mrb->num_writes) {
+    if (!mrb0->num_reqs) {
          return;
      }
  -    ret = blk_aio_multiwrite(blk, mrb->blkreq, mrb->num_writes);
-    if (ret != 0) {
-        for (i = 0; i < mrb->num_writes; i++) {
-            if (mrb->blkreq[i].error) {
- virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
-            }
+    if (mrb0->num_reqs == 1) {
+        if (mrb0->is_write) {
+            blk_aio_writev(blk, mrb0->sector_num, &mrb0->reqs[0]->qiov, 
mrb0->nb_sectors,
+                           virtio_blk_rw_complete, mrb0->reqs[0]);
+        } else {
+            blk_aio_readv(blk, mrb0->sector_num, &mrb0->reqs[0]->qiov, 
mrb0->nb_sectors,
+                          virtio_blk_rw_complete, mrb0->reqs[0]);
          }
+        qemu_iovec_destroy(&mrb0->qiov);
+        mrb0->num_reqs = 0;
+        return;
+    }
+
+    mrb = g_malloc(sizeof(MultiReqBuffer));
+    memcpy(mrb, mrb0, sizeof(MultiReqBuffer));
+    mrb0->num_reqs = 0;

I guess you're making mrb0->num_reqs == 1 a special case so you don't have to duplicate mrb0 in that case. But why are you duplicating it at all? And you're just moving over the IO vector without duplicating mrb->qiov.iov, which seems risky to me considering the qemu_iovec_destroy() in virtio_multireq_cb().

I could imagine you're duplicating it so the request can run in the background while you're again filling up the MultiReqBuffer, but that doesn't fit together with not duplicating it in the case of mrb0->num_reqs == 1 which can run in the background just the same.

I need to duplicate it in the > 1 case because MultiReqBuffer might be recycled 
before the virtio_multireq_cb callback is fired. In the == 1 case this can't 
happen because I pass the qiov from req[0] and use
the normal callback.

However, your concern reminds me of an optimization I had in mind, but that I 
forgot in this RFC patch.
In case of == 1 there is a useless init and destroy of mrb->qiov and 1 
qemu_iovec_concat. I will adjust this
the next version by creating mrb->qiov from the req[]->qiov array only I enter 
the part of virtio_submit_multireq
for mrb->num_reqs > 1.


+
+#ifdef DEBUG_MULTIREQ
+    printf("virtio_submit_multireq: p %p sector_num %ld nb_sectors %d is_write %d 
num_reqs %d\n",
+           mrb, mrb->sector_num, mrb->nb_sectors, mrb->is_write, 
mrb->num_reqs);
+#endif
+
+    if (mrb->is_write) {
+        blk_aio_writev(blk, mrb->sector_num, &mrb->qiov, mrb->nb_sectors,
+                       virtio_multireq_cb, mrb);
+    } else {
+        blk_aio_readv(blk, mrb->sector_num, &mrb->qiov, mrb->nb_sectors,
+                      virtio_multireq_cb, mrb);
      }
  -    mrb->num_writes = 0;
+    block_acct_merge_done(blk_get_stats(blk),
+                          mrb->is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ,
+                          mrb->num_reqs - 1);
  }
    static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer 
*mrb)
@@ -283,9 +330,9 @@ static void virtio_blk_handle_flush(VirtIOBlockReq *req, 
MultiReqBuffer *mrb)
                       BLOCK_ACCT_FLUSH);
        /*
-     * Make sure all outstanding writes are posted to the backing device.
+     * Make sure all outstanding requests are posted to the backing device.
       */
-    virtio_submit_multiwrite(req->dev->blk, mrb);
+    virtio_submit_multireq(req->dev->blk, mrb);
      blk_aio_flush(req->dev->blk, virtio_blk_flush_complete, req);
  }
  @@ -308,61 +355,8 @@ static bool virtio_blk_sect_range_ok(VirtIOBlock *dev,
      return true;
  }
  -static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
-{
-    BlockRequest *blkreq;
-    uint64_t sector;
-
-    sector = virtio_ldq_p(VIRTIO_DEVICE(req->dev), &req->out.sector);
-
-    trace_virtio_blk_handle_write(req, sector, req->qiov.size / 512);
-
-    if (!virtio_blk_sect_range_ok(req->dev, sector, req->qiov.size)) {
-        virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
-        virtio_blk_free_request(req);
-        return;
-    }
-
-    block_acct_start(blk_get_stats(req->dev->blk), &req->acct, req->qiov.size,
-                     BLOCK_ACCT_WRITE);
-
-    if (mrb->num_writes == VIRTIO_BLK_MAX_MERGE_REQS) {
-        virtio_submit_multiwrite(req->dev->blk, mrb);
-    }
-
-    blkreq = &mrb->blkreq[mrb->num_writes];
-    blkreq->sector = sector;
-    blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
-    blkreq->qiov = &req->qiov;
-    blkreq->cb = virtio_blk_rw_complete;
-    blkreq->opaque = req;
-    blkreq->error = 0;
-
-    mrb->num_writes++;
-}
-
-static void virtio_blk_handle_read(VirtIOBlockReq *req)
-{
-    uint64_t sector;
-
-    sector = virtio_ldq_p(VIRTIO_DEVICE(req->dev), &req->out.sector);
-
-    trace_virtio_blk_handle_read(req, sector, req->qiov.size / 512);
-
-    if (!virtio_blk_sect_range_ok(req->dev, sector, req->qiov.size)) {
-        virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
-        virtio_blk_free_request(req);
-        return;
-    }
-
-    block_acct_start(blk_get_stats(req->dev->blk), &req->acct, req->qiov.size,
-                     BLOCK_ACCT_READ);
-    blk_aio_readv(req->dev->blk, sector, &req->qiov,
-                  req->qiov.size / BDRV_SECTOR_SIZE,
-                  virtio_blk_rw_complete, req);
-}
-
-void virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb)
+void virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb_wr,
+                               MultiReqBuffer *mrb_rd)
  {
      uint32_t type;
      struct iovec *in_iov = req->elem.in_sg;
@@ -397,7 +391,7 @@ void virtio_blk_handle_request(VirtIOBlockReq *req, 
MultiReqBuffer *mrb)
      type = virtio_ldl_p(VIRTIO_DEVICE(req->dev), &req->out.type);
        if (type & VIRTIO_BLK_T_FLUSH) {
-        virtio_blk_handle_flush(req, mrb);
+        virtio_blk_handle_flush(req, mrb_wr);
      } else if (type & VIRTIO_BLK_T_SCSI_CMD) {
          virtio_blk_handle_scsi(req);
      } else if (type & VIRTIO_BLK_T_GET_ID) {
@@ -414,13 +408,71 @@ void virtio_blk_handle_request(VirtIOBlockReq *req, 
MultiReqBuffer *mrb)
          iov_from_buf(in_iov, in_num, 0, serial, size);
          virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
          virtio_blk_free_request(req);
-    } else if (type & VIRTIO_BLK_T_OUT) {
-        qemu_iovec_init_external(&req->qiov, iov, out_num);
-        virtio_blk_handle_write(req, mrb);
-    } else if (type == VIRTIO_BLK_T_IN || type == VIRTIO_BLK_T_BARRIER) {
-        /* VIRTIO_BLK_T_IN is 0, so we can't just & it. */
-        qemu_iovec_init_external(&req->qiov, in_iov, in_num);
-        virtio_blk_handle_read(req);
+    } else if (type & VIRTIO_BLK_T_OUT || type == VIRTIO_BLK_T_IN || type == 
VIRTIO_BLK_T_BARRIER) {

Please don't omit the comment about VIRTIO_BLK_T_IN being 0, I spend quite a while thinking about why VIRTIO_BLK_T_BARRIER is seemingly handled just the same as VIRTIO_BLK_T_IN. I still don't know what VIRTIO_BLK_T_BARRIER is supposed to do (and I don't really want to grab the virtio spec right now, although I should do it some time...), but now it at least seems to me like it doesn't really matter.

Actually I don't know as well. I just copied the behaviour. But I will try to 
find out what VIRTIO_BLK_BARRIER is supposed to do.


+        bool is_write = type & VIRTIO_BLK_T_OUT;
+        MultiReqBuffer *mrb = is_write ? mrb_wr : mrb_rd;
+        int64_t sector_num = virtio_ldq_p(VIRTIO_DEVICE(req->dev), 
&req->out.sector);
+        BlockDriverState *bs = blk_bs(req->dev->blk);
+        int nb_sectors = 0;
+        int merge = 1;

I shouldn't be criticizing such small issues in an RFC, but it's the most I can 
do for a virtio-blk series: I'd like this to be a bool instead of an int, 
please.

Judging from the not-so-quick-in-quantity-but-pretty-brief-in-quality scan through this RFC, it looks OK to me. There are some minor things (e.g. "too" instead of "to" in the commit message) which should be fixed eventually, but logic-wise, I didn't spot anything major.

Perhaps we could put the merge logic into an own function, though, just like 
bdrv_aio_multiwrite() did with multiwrite_merge().

I also though about that, but the merge logic itself is pretty compact now and 
deals with Virtio specific structs. I think there will be much more code
and more complicated code if I add a generic version. I don't know if an driver 
will adopt the multiread stuff. No driver has done since 2009 actually.


I trust you that overlapping requests are not an issue, at least when not sorting the requests. Other than that, nothing bad from me. The patch looks good, but I don't want to be too optimistic, as I don't want to definitely judge anything related to virtio-blk. :-)

(Apart from the first two patches of this series, they were completely fine, 
but I seem to remember having them reviewed once already...)

Mostly, but I had to change a few things. So I couldn't copy the Reviewed-by.

Thank you,
Peter



reply via email to

[Prev in Thread] Current Thread [Next in Thread]