[PATCH 08/16] nvme: Fix coroutine waking

Hanna Czenczek posted 16 patches 2 weeks, 3 days ago
Maintainers: Paolo Bonzini <pbonzini@redhat.com>, "Alex Bennée" <alex.bennee@linaro.org>, Kevin Wolf <kwolf@redhat.com>, Hanna Reitz <hreitz@redhat.com>, Stefan Hajnoczi <stefanha@redhat.com>, Fam Zheng <fam@euphon.net>, Ronnie Sahlberg <ronniesahlberg@gmail.com>, Peter Lieven <pl@dlhnet.de>, "Philippe Mathieu-Daudé" <philmd@linaro.org>, Ilya Dryomov <idryomov@gmail.com>, "Richard W.M. Jones" <rjones@redhat.com>, Stefan Weil <sw@weilnetz.de>
There is a newer version of this series
[PATCH 08/16] nvme: Fix coroutine waking
Posted by Hanna Czenczek 2 weeks, 3 days ago
nvme wakes the request coroutine via qemu_coroutine_enter() from a BH
scheduled in the BDS AioContext.  This may not be the same context as
the one in which the request originally ran, which would be wrong:
- It could mean we enter the coroutine before it yields,
- We would move the coroutine in to a different context.

(Can be reproduced with multiqueue by adding a usleep(100000) before the
`while (data.ret == -EINPROGRESS)` loop.)

To fix that, schedule nvme_rw_cb_bh() in the coroutine AioContext.
(Just like in the preceding iscsi and nfs patches, we could drop the
trivial nvme_rw_cb_bh() and just use aio_co_wake() directly, but don’t
do that to keep replay_bh_schedule_oneshot_event().)

With this, we can remove NVMeCoData.ctx.

Note the check of data->co == NULL to bypass the BH/yield combination in
case nvme_rw_cb() is called from nvme_submit_command(): We probably want
to keep this fast path for performance reasons, but we have to be quite
careful about it:
- We cannot overload .ret for this, but have to use a dedicated
  .skip_yield field.  Otherwise, if nvme_rw_cb() runs in a different
  thread than the coroutine, it may see .ret set and skip the yield,
  while nvme_rw_cb() will still schedule a BH for waking.   Therefore,
  the signal to skip the yield can only be set in nvme_rw_cb() if waking
  too is skipped, which is independent from communicating the return
  value.
- We can only skip the yield if nvme_rw_cb() actually runs in the
  request coroutine.  Otherwise (specifically if they run in different
  AioContexts), the order between this function’s execution and the
  coroutine yielding (or not yielding) is not reliable.
- There is no point to yielding in a loop; there are no spurious wakes,
  so once we yield, we will only be re-entered once the command is done.
  Replace `while` by `if`.

Cc: qemu-stable@nongnu.org
Signed-off-by: Hanna Czenczek <hreitz@redhat.com>
---
 block/nvme.c | 56 +++++++++++++++++++++-------------------------------
 1 file changed, 23 insertions(+), 33 deletions(-)

diff --git a/block/nvme.c b/block/nvme.c
index 7ed5f570bc..4b1f623e7d 100644
--- a/block/nvme.c
+++ b/block/nvme.c
@@ -1176,8 +1176,8 @@ fail:
 
 typedef struct {
     Coroutine *co;
+    bool skip_yield;
     int ret;
-    AioContext *ctx;
 } NVMeCoData;
 
 static void nvme_rw_cb_bh(void *opaque)
@@ -1189,12 +1189,22 @@ static void nvme_rw_cb_bh(void *opaque)
 static void nvme_rw_cb(void *opaque, int ret)
 {
     NVMeCoData *data = opaque;
+    AioContext *ctx;
+
     data->ret = ret;
-    if (!data->co) {
-        /* The rw coroutine hasn't yielded, don't try to enter. */
-        return;
+
+    if (data->co == qemu_coroutine_self()) {
+        /*
+         * Fast path: We are inside of the request coroutine (through
+         * nvme_submit_command, nvme_deferred_fn, nvme_process_completion).
+         * We can set data->skip_yield here to keep the coroutine from
+         * yielding, and then we don't need to schedule a BH to wake it.
+         */
+        data->skip_yield = true;
+    } else {
+        ctx = qemu_coroutine_get_aio_context(data->co);
+        replay_bh_schedule_oneshot_event(ctx, nvme_rw_cb_bh, data);
     }
-    replay_bh_schedule_oneshot_event(data->ctx, nvme_rw_cb_bh, data);
 }
 
 static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
@@ -1217,10 +1227,7 @@ static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
         .cdw12 = cpu_to_le32(cdw12),
     };
-    NVMeCoData data = {
-        .ctx = bdrv_get_aio_context(bs),
-        .ret = -EINPROGRESS,
-    };
+    NVMeCoData data = { .co = qemu_coroutine_self() };
 
     trace_nvme_prw_aligned(s, is_write, offset, bytes, flags, qiov->niov);
     assert(s->queue_count > 1);
@@ -1235,9 +1242,7 @@ static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
         return r;
     }
     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
-
-    data.co = qemu_coroutine_self();
-    while (data.ret == -EINPROGRESS) {
+    if (!data.skip_yield) {
         qemu_coroutine_yield();
     }
 
@@ -1332,18 +1337,13 @@ static coroutine_fn int nvme_co_flush(BlockDriverState *bs)
         .opcode = NVME_CMD_FLUSH,
         .nsid = cpu_to_le32(s->nsid),
     };
-    NVMeCoData data = {
-        .ctx = bdrv_get_aio_context(bs),
-        .ret = -EINPROGRESS,
-    };
+    NVMeCoData data = { .co = qemu_coroutine_self() };
 
     assert(s->queue_count > 1);
     req = nvme_get_free_req(ioq);
     assert(req);
     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
-
-    data.co = qemu_coroutine_self();
-    if (data.ret == -EINPROGRESS) {
+    if (!data.skip_yield) {
         qemu_coroutine_yield();
     }
 
@@ -1383,10 +1383,7 @@ static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
     };
 
-    NVMeCoData data = {
-        .ctx = bdrv_get_aio_context(bs),
-        .ret = -EINPROGRESS,
-    };
+    NVMeCoData data = { .co = qemu_coroutine_self() };
 
     if (flags & BDRV_REQ_MAY_UNMAP) {
         cdw12 |= (1 << 25);
@@ -1404,9 +1401,7 @@ static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
     assert(req);
 
     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
-
-    data.co = qemu_coroutine_self();
-    while (data.ret == -EINPROGRESS) {
+    if (!data.skip_yield) {
         qemu_coroutine_yield();
     }
 
@@ -1433,10 +1428,7 @@ static int coroutine_fn nvme_co_pdiscard(BlockDriverState *bs,
         .cdw11 = cpu_to_le32(1 << 2), /*deallocate bit*/
     };
 
-    NVMeCoData data = {
-        .ctx = bdrv_get_aio_context(bs),
-        .ret = -EINPROGRESS,
-    };
+    NVMeCoData data = { .co = qemu_coroutine_self() };
 
     if (!s->supports_discard) {
         return -ENOTSUP;
@@ -1479,9 +1471,7 @@ static int coroutine_fn nvme_co_pdiscard(BlockDriverState *bs,
     trace_nvme_dsm(s, offset, bytes);
 
     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
-
-    data.co = qemu_coroutine_self();
-    while (data.ret == -EINPROGRESS) {
+    if (!data.skip_yield) {
         qemu_coroutine_yield();
     }
 
-- 
2.51.0


Re: [PATCH 08/16] nvme: Fix coroutine waking
Posted by Kevin Wolf 2 weeks, 2 days ago
Am 28.10.2025 um 17:33 hat Hanna Czenczek geschrieben:
> nvme wakes the request coroutine via qemu_coroutine_enter() from a BH
> scheduled in the BDS AioContext.  This may not be the same context as
> the one in which the request originally ran, which would be wrong:
> - It could mean we enter the coroutine before it yields,
> - We would move the coroutine in to a different context.
> 
> (Can be reproduced with multiqueue by adding a usleep(100000) before the
> `while (data.ret == -EINPROGRESS)` loop.)
> 
> To fix that, schedule nvme_rw_cb_bh() in the coroutine AioContext.
> (Just like in the preceding iscsi and nfs patches, we could drop the
> trivial nvme_rw_cb_bh() and just use aio_co_wake() directly, but don’t
> do that to keep replay_bh_schedule_oneshot_event().)
> 
> With this, we can remove NVMeCoData.ctx.
> 
> Note the check of data->co == NULL to bypass the BH/yield combination in
> case nvme_rw_cb() is called from nvme_submit_command(): We probably want
> to keep this fast path for performance reasons, but we have to be quite
> careful about it:
> - We cannot overload .ret for this, but have to use a dedicated
>   .skip_yield field.  Otherwise, if nvme_rw_cb() runs in a different
>   thread than the coroutine, it may see .ret set and skip the yield,
>   while nvme_rw_cb() will still schedule a BH for waking.   Therefore,
>   the signal to skip the yield can only be set in nvme_rw_cb() if waking
>   too is skipped, which is independent from communicating the return
>   value.
> - We can only skip the yield if nvme_rw_cb() actually runs in the
>   request coroutine.  Otherwise (specifically if they run in different
>   AioContexts), the order between this function’s execution and the
>   coroutine yielding (or not yielding) is not reliable.
> - There is no point to yielding in a loop; there are no spurious wakes,
>   so once we yield, we will only be re-entered once the command is done.
>   Replace `while` by `if`.
> 
> Cc: qemu-stable@nongnu.org
> Signed-off-by: Hanna Czenczek <hreitz@redhat.com>

As suggested in an earlier patch, I prefer to keep setting -EINPROGRESS,
even if we don't check for it elsewhere, for better debugability.

Kevin