| Message ID | 87pn3zlk8u.fsf@notabene.neil.brown.name (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | NFS: switch nfsiod to be an UNBOUND workqueue. | expand |
On Fri, Nov 27, 2020 at 11:24:33AM +1100, NeilBrown wrote: > > nfsiod is currently a concurrency-managed workqueue (CMWQ). > This means that workitems scheduled to nfsiod on a given CPU are queued > behind all other work items queued on any CMWQ on the same CPU. This > can introduce unexpected latency. > > Occaionally nfsiod can even cause excessive latency. If the work item > to complete a CLOSE request calls the final iput() on an inode, the > address_space of that inode will be dismantled. This takes time > proportional to the number of in-memory pages, which on a large host > working on large files (e.g.. 5TB), can be a large number of pages > resulting in a noticable number of seconds. > > We can avoid these latency problems by switching nfsiod to WQ_UNBOUND. > This causes each concurrent work item to gets a dedicated thread which > can be scheduled to an idle CPU. > > There is precedent for this as several other filesystems use WQ_UNBOUND > workqueue for handling various async events. > > Signed-off-by: NeilBrown <neilb@suse.de> Acked-by: Tejun Heo <tj@kernel.org> Thanks.
diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c index aa6493905bbe..43af053f467a 100644 --- a/fs/nfs/inode.c +++ b/fs/nfs/inode.c @@ -2180,7 +2180,7 @@ static int nfsiod_start(void) { struct workqueue_struct *wq; dprintk("RPC: creating workqueue nfsiod\n"); - wq = alloc_workqueue("nfsiod", WQ_MEM_RECLAIM, 0); + wq = alloc_workqueue("nfsiod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0); if (wq == NULL) return -ENOMEM; nfsiod_workqueue = wq;
nfsiod is currently a concurrency-managed workqueue (CMWQ). This means that workitems scheduled to nfsiod on a given CPU are queued behind all other work items queued on any CMWQ on the same CPU. This can introduce unexpected latency. Occaionally nfsiod can even cause excessive latency. If the work item to complete a CLOSE request calls the final iput() on an inode, the address_space of that inode will be dismantled. This takes time proportional to the number of in-memory pages, which on a large host working on large files (e.g.. 5TB), can be a large number of pages resulting in a noticable number of seconds. We can avoid these latency problems by switching nfsiod to WQ_UNBOUND. This causes each concurrent work item to gets a dedicated thread which can be scheduled to an idle CPU. There is precedent for this as several other filesystems use WQ_UNBOUND workqueue for handling various async events. Signed-off-by: NeilBrown <neilb@suse.de> --- fs/nfs/inode.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)