fuse: fix accounting background requests (v2)
Get series with b4
b4 am [email protected]Versions newest first
- v1 Latest4 patches 1 replyIn review
Changed in v2 from the cover letter
- rebased on for-next branch of the fuse tree
- fixed race when processing request begins before init-reply came
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/4 | fuse: make request allocations for background processing explicitfs/ fs/ fs/ +2 more | +31 | 0 replies | |
| 2/4 | fuse: add flag fc->uninitializedfs/ fs/ fs/ +1 more | +9 | 0 replies | |
| 3/4 | fuse: skip blocking on allocations of synchronous requestsfs/ | +18 | 0 replies | |
| 4/4 | fuse: implement exclusive wakeup for blocked_waitqfs/ | +16 | 0 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [PATCH 0/4] fuse: fix accounting background requests (v2)
Hi,
The feature was added long time ago (commit 08a53cdc...) with the comment:
Quoted 10 lines
A task may have at most one synchronous request allocated. So these requests need not be otherwise limited.
However the number of background requests (release, forget, asynchronous reads, interrupted requests) can grow indefinitely. This can be used by a malicous user to cause FUSE to allocate arbitrary amounts of unswappable kernel memory, denying service.
For this reason add a limit for the number of background requests, and block allocations of new requests until the number goes bellow the limit.
However, the implementation suffers from the following problems:
1. Latency of synchronous requests. As soon as fc->num_background hits the limit, all allocations are blocked: both for synchronous and background requests. This is unnecessary - as the comment cited above states, synchronous requests need not be limited (by fuse). Moreover, sometimes it's very inconvenient. For example, a dozen of tasks aggressively writing to mmap()-ed area may block 'ls' for long while (>1min in my experiments).
2. Thundering herd problem. When fc->num_background falls below the limit, request_end() calls wake_up_all(&fc->blocked_waitq). This wakes up all waiters while it's not impossible that the first waiter getting new request will immediately put it to background increasing fc->num_background again. (experimenting with mmap()-ed writes I observed 2x slowdown as compared with fuse after applying this patch-set)
The patch-set re-works fuse_get_req (and its callers) to throttle only requests intended for background processing. Having this done, it becomes possible to use exclusive wakeups in chained manner: request_end() wakes up a waiter, the waiter allocates new request and submits it for background processing, the processing ends in request_end() where another wakeup happens an so on.
Changed in v2:
- rebased on for-next branch of the fuse tree
- fixed race when processing request begins before init-reply came
Thanks,
Maxim
---
Maxim V. Patlasov (4):
fuse: make request allocations for background processing explicit
fuse: add flag fc->uninitialized
fuse: skip blocking on allocations of synchronous requests
fuse: implement exclusive wakeup for blocked_waitq
fs/fuse/cuse.c | 3 ++ fs/fuse/dev.c | 69 +++++++++++++++++++++++++++++++++++++++++++----------- fs/fuse/file.c | 6 +++-- fs/fuse/fuse_i.h | 8 ++++++ fs/fuse/inode.c | 4 +++ 5 files changed, 73 insertions(+), 17 deletions(-)
Signature · 1 line
Signature