block/nfs: avoid BDRV_POLL_WHILE() in nfs_co_get_allocated_file_size()
PATCH 1/1 block/nfs: avoid BDRV_POLL_WHILE() in nfs_co_get_allocated_file_size()
Commit 82618d7bc341 ("block: Convert bdrv_get_allocated_file_size() to co_wrapper") made nfs_get_allocated_file_size() a coroutine. The coroutine still uses BDRV_POLL_WHILE() to wait for the NFS RPC to complete.
Take it a step further and yield the coroutine until the RPC completes. This avoids the blocking, nested event loop and unifies nfs_co_get_allocated_file_size() with the other coroutine functions that send RPCs:
- Use nfs_co_init_task() to set up a coroutine NFSRPC task.
- Take client->mutex to protect fd handler state since we're in IO_CODE.
- Use nfs_co_generic_cb() instead of a specialized callback function.
- Yield until the task completes.
Getting rid of BDRV_POLL_WHILE() helps with the multi-queue block layer effort where we don't want to take the AioContext lock.
This commit passes qemu-iotests/check -nfs, except inactivate-failure, which also fails before this commit.
--- block/nfs.c | 45 +++++++++++++++++---------------------------- 1 file changed, 17 insertions(+), 28 deletions(-)
block/nfs.c +17 -28
@@ -248,14 +248,18 @@ nfs_co_generic_cb(int ret, struct nfs_context *nfs, void *data, {NFSRPC *task = private_data; task->ret = ret; -assert (!task->st);if (task->ret >0 && task->iov) { +assert (!task->st);if (task->ret <= task->iov->size) {qemu_iovec_from_buf (task->iov,0 , data, task->ret); }else { task->ret = -EIO ; } } +if (task->ret ==0 && task->st) { +assert (!task->iov); +memcpy (task->st, data,sizeof (struct stat )); + }if (task->ret <0 ) {error_report ("NFS Error: %s" ,nfs_get_error (nfs)); }
@@ -713,29 +717,10 @@ static int nfs_has_zero_init(BlockDriverState *bs) }#if !defined(_WIN32) -/* Called (via nfs_service) with QemuMutex held. */ -static void -nfs_get_allocated_file_size_cb (int ret,struct nfs_context *nfs,void *data, -void *private_data) -{ -NFSRPC *task = private_data; - task->ret = ret; -if (task->ret ==0 ) { -memcpy (task->st, data,sizeof (struct stat )); - } -if (task->ret <0 ) { -error_report ("NFS Error: %s" ,nfs_get_error (nfs)); - } - -/* Set task->complete before reading bs->wakeup. */ -qatomic_mb_set (&task->complete,1 ); -bdrv_wakeup (task->bs); -} - static int64_t coroutine_fn nfs_co_get_allocated_file_size(BlockDriverState *bs) {NFSClient *client = bs->opaque; -NFSRPC task = {0 }; +NFSRPC task;struct stat st; if (bdrv_is_read_only(bs) &&
@@ -743,15 +728,19 @@ static int64_t coroutine_fn nfs_co_get_allocated_file_size(BlockDriverState *bs)return client->st_blocks *512 ; } - task.bs = bs; +nfs_co_init_task (bs, &task); task.st = &st; -if (nfs_fstat_async (client->context, client->fh, nfs_get_allocated_file_size_cb, - &task) !=0 ) { -return -ENOMEM ; - } + WITH_QEMU_LOCK_GUARD(&client->mutex) { +if (nfs_fstat_async (client->context, client->fh, nfs_co_generic_cb, + &task) !=0 ) { +return -ENOMEM ; + } -nfs_set_events (client); -BDRV_POLL_WHILE (bs, !task.complete); +nfs_set_events (client); + } +while (!task.complete) { +qemu_coroutine_yield (); + }return (task.ret <0 ? task.ret : st.st_blocks *512 ); }
Signature · 1 line
2.39.2
Am 04.04.2023 um 01:30 hat Stefan Hajnoczi geschrieben:
Quoted 14 lines
Commit
82618d7bc341("block: Convertbdrv_get_allocated_file_size()toco_wrapper") madenfs_get_allocated_file_size()a coroutine. The coroutine still usesBDRV_POLL_WHILE()to wait for the NFS RPC to complete.Take it a step further and yield the coroutine until the RPC completes. This avoids the blocking, nested event loop and unifies
nfs_co_get_allocated_file_size()with the other coroutine functions that send RPCs:
- Usenfs_co_init_task()to set up a coroutine NFSRPC task.
- Takeclient->mutexto protect fd handler state since we're inIO_CODE.
- Usenfs_co_generic_cb()instead of a specialized callback function.
- Yield until the task completes.
Quoted 7 lines
Getting rid of
BDRV_POLL_WHILE()helps with the multi-queue block layer effort where we don't want to take the AioContext lock.This commit passes qemu-iotests/check -nfs, except inactivate-failure, which also fails before this commit.
Signed-off-by Stefan Hajnoczi <[email protected]>
This conflicts with Paolo's commit 3fe64abc, which seems to address the same bug. Your patch does some more cleanup like removing nfs_get_allocated_file_size_cb() in favour of nfs_co_generic_cb().
Do you want to rebase and send the remaining pieces as a cleanup patch?
Kevin