A netfs read request can run in one of two modes: for synchronous reads
writes, the app thread does the collection of results and for asynchronous
reads, this is offloaded to a worker thread. This is controlled by the
NETFS_RREQ_OFFLOAD_COLLECTION flag.
Now, if a subrequest incurs an error, the NETFS_RREQ_PAUSE flag is set to
stop the issuing loop temporarily from issuing more subrequests until a
retry is successful or the request is abandoned.
When the issuing loop sees NETFS_RREQ_PAUSE, it jumps to
netfs_wait_for_pause() which will wait for the PAUSE flag to be cleared -
and whilst it is waiting, it will call out to the collector as more results
acrue... But this is the wrong thing to do if OFFLOAD_COLLECTION is set as
we can then end up with both the app thread and the work item collecting
results simultaneously.
This manifests itself occasionally when running the generic/323 xfstest
against multichannel cifs as an oops that's a bit random but frequently
involving io_submit() (the test does lots of simultaneous async DIO reads).
Fix this by only doing the collection in netfs_wait_for_pause() if the
NETFS_RREQ_OFFLOAD_COLLECTION is not set.
Fixes: e2d46f2ec3 ("netfs: Change the read result collector to only use one work item")
Reported-by: Steve French <stfrench@microsoft.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20250314164201.1993231-2-dhowells@redhat.com
Acked-by: "Paulo Alcantara (Red Hat)" <pc@manguebit.com>
cc: Paulo Alcantara <pc@manguebit.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: linux-cifs@vger.kernel.org
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Fix a number of hangs in the netfslib read-retry code, including:
(1) netfs_reissue_read() doubles up the getting of references on
subrequests, thereby leaking the subrequest and causing inode eviction
to wait indefinitely. This can lead to the kernel reporting a hang in
the filesystem's evict_inode().
Fix this by removing the get from netfs_reissue_read() and adding one
to netfs_retry_read_subrequests() to deal with the one place that
didn't double up.
(2) The loop in netfs_retry_read_subrequests() that retries a sequence of
failed subrequests doesn't record whether or not it retried the one
that the "subreq" pointer points to when it leaves the loop. It may
not if renegotiation/repreparation of the subrequests means that fewer
subrequests are needed to span the cumulative range of the sequence.
Because it doesn't record this, the piece of code that discards
now-superfluous subrequests doesn't know whether it should discard the
one "subreq" points to - and so it doesn't.
Fix this by noting whether the last subreq it examines is superfluous
and if it is, then getting rid of it and all subsequent subrequests.
If that one one wasn't superfluous, then we would have tried to go
round the previous loop again and so there can be no further unretried
subrequests in the sequence.
(3) netfs_retry_read_subrequests() gets yet an extra ref on any additional
subrequests it has to get because it ran out of ones it could reuse to
to renegotiation/repreparation shrinking the subrequests.
Fix this by removing that extra ref.
(4) In netfs_retry_reads(), it was using wait_on_bit() to wait for
NETFS_SREQ_IN_PROGRESS to be cleared on all subrequests in the
sequence - but netfs_read_subreq_terminated() is now using a wait
queue on the request instead and so this wait will never finish.
Fix this by waiting on the wait queue instead. To make this work, a
new flag, NETFS_RREQ_RETRYING, is now set around the wait loop to tell
the wake-up code to wake up the wait queue rather than requeuing the
request's work item.
Note that this flag replaces the NETFS_RREQ_NEED_RETRY flag which is
no longer used.
(5) Whilst not strictly anything to do with the hang,
netfs_retry_read_subrequests() was also doubly incrementing the
subreq_counter and re-setting the debug index, leaving a gap in the
trace. This is also fixed.
One of these hangs was observed with 9p and with cifs. Others were forced
by manual code injection into fs/afs/file.c. Firstly, afs_prepare_read()
was created to provide an changing pattern of maximum subrequest sizes:
static int afs_prepare_read(struct netfs_io_subrequest *subreq)
{
struct netfs_io_request *rreq = subreq->rreq;
if (!S_ISREG(subreq->rreq->inode->i_mode))
return 0;
if (subreq->retry_count < 20)
rreq->io_streams[0].sreq_max_len =
umax(200, 2222 - subreq->retry_count * 40);
else
rreq->io_streams[0].sreq_max_len = 3333;
return 0;
}
and pointed to by afs_req_ops. Then the following:
struct netfs_io_subrequest *subreq = op->fetch.subreq;
if (subreq->error == 0 &&
S_ISREG(subreq->rreq->inode->i_mode) &&
subreq->retry_count < 20) {
subreq->transferred = subreq->already_done;
__clear_bit(NETFS_SREQ_HIT_EOF, &subreq->flags);
__set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
afs_fetch_data_notify(op);
return;
}
was inserted into afs_fetch_data_success() at the beginning and struct
netfs_io_subrequest given an extra field, "already_done" that was set to
the value in "subreq->transferred" by netfs_reissue_read().
When reading a 4K file, the subrequests would get gradually smaller, a new
subrequest would be allocated around the 3rd retry and then eventually be
rendered superfluous when the 20th retry was hit and the limit on the first
subrequest was eased.
Fixes: e2d46f2ec3 ("netfs: Change the read result collector to only use one work item")
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20250212222402.3618494-2-dhowells@redhat.com
Tested-by: Marc Dionne <marc.dionne@auristor.com>
Tested-by: Steve French <stfrench@microsoft.com>
cc: Ihor Solodrai <ihor.solodrai@pm.me>
cc: Eric Van Hensbergen <ericvh@kernel.org>
cc: Latchesar Ionkov <lucho@ionkov.net>
cc: Dominique Martinet <asmadeus@codewreck.org>
cc: Christian Schoenebeck <linux_oss@crudebyte.com>
cc: Paulo Alcantara <pc@manguebit.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: v9fs@lists.linux.dev
cc: linux-cifs@vger.kernel.org
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Pull vfs netfs updates from Christian Brauner:
"This contains read performance improvements and support for monolithic
single-blob objects that have to be read/written as such (e.g. AFS
directory contents). The implementation of the two parts is interwoven
as each makes the other possible.
- Read performance improvements
The read performance improvements are intended to speed up some
loss of performance detected in cifs and to a lesser extend in afs.
The problem is that we queue too many work items during the
collection of read results: each individual subrequest is collected
by its own work item, and then they have to interact with each
other when a series of subrequests don't exactly align with the
pattern of folios that are being read by the overall request.
Whilst the processing of the pages covered by individual
subrequests as they complete potentially allows folios to be woken
in parallel and with minimum delay, it can shuffle wakeups for
sequential reads out of order - and that is the most common I/O
pattern.
The final assessment and cleanup of an operation is then held up
until the last I/O completes - and for a synchronous sequential
operation, this means the bouncing around of work items just adds
latency.
Two changes have been made to make this work:
(1) All collection is now done in a single "work item" that works
progressively through the subrequests as they complete (and
also dispatches retries as necessary).
(2) For readahead and AIO, this work item be done on a workqueue
and can run in parallel with the ultimate consumer of the data;
for synchronous direct or unbuffered reads, the collection is
run in the application thread and not offloaded.
Functions such as smb2_readv_callback() then just tell netfslib
that the subrequest has terminated; netfslib does a minimal bit of
processing on the spot - stat counting and tracing mostly - and
then queues/wakes up the worker. This simplifies the logic as the
collector just walks sequentially through the subrequests as they
complete and walks through the folios, if buffered, unlocking them
as it goes. It also keeps to a minimum the amount of latency
injected into the filesystem's low-level I/O handling
The way netfs supports filesystems using the deprecated
PG_private_2 flag is changed: folios are flagged and added to a
write request as they complete and that takes care of scheduling
the writes to the cache. The originating read request can then just
unlock the pages whatever happens.
- Single-blob object support
Single-blob objects are files for which the content of the file
must be read from or written to the server in a single operation
because reading them in parts may yield inconsistent results. AFS
directories are an example of this as there exists the possibility
that the contents are generated on the fly and would differ between
reads or might change due to third party interference.
Such objects will be written to and retrieved from the cache if one
is present, though we allow/may need to propose multiple
subrequests to do so. The important part is that read from/write to
the *server* is monolithic.
Single blob reading is, for the moment, fully synchronous and does
result collection in the application thread and, also for the
moment, the API is supplied the buffer in the form of a folio_queue
chain rather than using the pagecache.
- Related afs changes
This series makes a number of changes to the kafs filesystem,
primarily in the area of directory handling:
- AFS's FetchData RPC reply processing is made partially
asynchronous which allows the netfs_io_request's outstanding
operation counter to be removed as part of reducing the
collection to a single work item.
- Directory and symlink reading are plumbed through netfslib using
the single-blob object API and are now cacheable with fscache.
This also allows the afs_read struct to be eliminated and
netfs_io_subrequest to be used directly instead.
- Directory and symlink content are now stored in a folio_queue
buffer rather than in the pagecache. This means we don't require
the RCU read lock and xarray iteration to access it, and folios
won't randomly disappear under us because the VM wants them
back.
- The vnode operation lock is changed from a mutex struct to a
private lock implementation. The problem is that the lock now
needs to be dropped in a separate thread and mutexes don't
permit that.
- When a new directory or symlink is created, we now initialise it
locally and mark it valid rather than downloading it (we know
what it's likely to look like).
- We now use the in-directory hashtable to reduce the number of
entries we need to scan when doing a lookup. The edit routines
have to maintain the hash chains.
- Cancellation (e.g. by signal) of an async call after the
rxrpc_call has been set up is now offloaded to the worker thread
as there will be a notification from rxrpc upon completion. This
avoids a double cleanup.
- A "rolling buffer" implementation is created to abstract out the
two separate folio_queue chaining implementations I had (one for
read and one for write).
- Functions are provided to create/extend a buffer in a folio_queue
chain and tear it down again.
This is used to handle AFS directories, but could also be used to
create bounce buffers for content crypto and transport crypto.
- The was_async argument is dropped from netfs_read_subreq_terminated()
Instead we wake the read collection work item by either queuing it
or waking up the app thread.
- We don't need to use BH-excluding locks when communicating between
the issuing thread and the collection thread as neither of them now
run in BH context.
- Also included are a number of new tracepoints; a split of the
netfslib write collection code to put retrying into its own file
(it gets more complicated with content encryption).
- There are also some minor fixes AFS included, including fixing the
AFS directory format struct layout, reducing some directory
over-invalidation and making afs_mkdir() translate EEXIST to
ENOTEMPY (which is not available on all systems the servers
support).
- Finally, there's a patch to try and detect entry into the folio
unlock function with no folio_queue structs in the buffer (which
isn't allowed in the cases that can get there).
This is a debugging patch, but should be minimal overhead"
* tag 'vfs-6.14-rc1.netfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (31 commits)
netfs: Report on NULL folioq in netfs_writeback_unlock_folios()
afs: Add a tracepoint for afs_read_receive()
afs: Locally initialise the contents of a new symlink on creation
afs: Use the contained hashtable to search a directory
afs: Make afs_mkdir() locally initialise a new directory's content
netfs: Change the read result collector to only use one work item
afs: Make {Y,}FS.FetchData an asynchronous operation
afs: Fix cleanup of immediately failed async calls
afs: Eliminate afs_read
afs: Use netfslib for symlinks, allowing them to be cached
afs: Use netfslib for directories
afs: Make afs_init_request() get a key if not given a file
netfs: Add support for caching single monolithic objects such as AFS dirs
netfs: Add functions to build/clean a buffer in a folio_queue
afs: Add more tracepoints to do with tracking validity
cachefiles: Add auxiliary data trace
cachefiles: Add some subrequest tracepoints
netfs: Remove some extraneous directory invalidations
afs: Fix directory format encoding struct
afs: Fix EEXIST error returned from afs_rmdir() to be ENOTEMPTY
...
Netfslib needs to be able to handle kernel-initiated asynchronous DIO that
is supplied with a bio_vec[] array. Currently, because of the async flag,
this gets passed to netfs_extract_user_iter() which throws a warning and
fails because it only handles IOVEC and UBUF iterators. This can be
triggered through a combination of cifs and a loopback blockdev with
something like:
mount //my/cifs/share /foo
dd if=/dev/zero of=/foo/m0 bs=4K count=1K
losetup --sector-size 4096 --direct-io=on /dev/loop2046 /foo/m0
echo hello >/dev/loop2046
This causes the following to appear in syslog:
WARNING: CPU: 2 PID: 109 at fs/netfs/iterator.c:50 netfs_extract_user_iter+0x170/0x250 [netfs]
and the write to fail.
Fix this by removing the check in netfs_unbuffered_write_iter_locked() that
causes async kernel DIO writes to be handled as userspace writes. Note
that this change relies on the kernel caller maintaining the existence of
the bio_vec array (or kvec[] or folio_queue) until the op is complete.
Fixes: 153a9961b5 ("netfs: Implement unbuffered/DIO write support")
Reported-by: Nicolas Baranger <nicolas.baranger@3xo.fr>
Closes: https://lore.kernel.org/r/fedd8a40d54b2969097ffa4507979858@3xo.fr/
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/608725.1736275167@warthog.procyon.org.uk
Tested-by: Nicolas Baranger <nicolas.baranger@3xo.fr>
Acked-by: Paulo Alcantara (Red Hat) <pc@manguebit.com>
cc: Steve French <smfrench@gmail.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-cifs@vger.kernel.org
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
It seems that it's possible to get to netfs_writeback_unlock_folios() with
an empty rolling buffer during buffered writes. This should not be
possible as the rolling buffer is initialised as the write request is set
up and thereafter maintains at least one folio_queue struct therein until
it gets destroyed. This allows lockless addition and removal of
folio_queue structs in the buffer because, unlike with a ring buffer, the
producer and consumer each only need to look at and alter one pointer into
the buffer.
Now, the rolling buffer is only used for buffered I/O operations as
netfs_collect_write_results() should only call
netfs_writeback_unlock_folios() if the request is of origin type
NETFS_WRITEBACK, NETFS_WRITETHROUGH or NETFS_PGPRIV2_COPY_TO_CACHE.
So it would seem that one of the following occurred: (1) I/O started before
the request was fully initialised, (2) the origin got switched mid-flow or
(3) the request has already been freed and this is a UAF error. I think
the last is the most likely.
Make netfs_writeback_unlock_folios() report information about the request
and subrequests if folioq is seen to be NULL to try and help debug this,
throw a warning and return.
Note that this does not try to fix the problem.
Reported-by: syzbot+af5c06208fa71bf31b16@syzkaller.appspotmail.com
Link: https://syzkaller.appspot.com/bug?extid=af5c06208fa71bf31b16
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/ZxshMEW4U7MTgQYa@gmail.com/
Link: https://lore.kernel.org/r/20241216204124.3752367-33-dhowells@redhat.com
cc: Chang Yu <marcus.yu.56@gmail.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Change the way netfslib collects read results to do all the collection for
a particular read request using a single work item that walks along the
subrequest queue as subrequests make progress or complete, unlocking folios
progressively rather than doing the unlock in parallel as parallel requests
come in.
The code is remodelled to be more like the write-side code, though only
using a single stream. This makes it more directly comparable and thus
easier to duplicate fixes between the two sides.
This has a number of advantages:
(1) It's simpler. There doesn't need to be a complex donation mechanism
to handle mismatches between the size and alignment of subrequests and
folios. The collector unlocks folios as the subrequests covering each
complete.
(2) It should cause less scheduler overhead as there's a single work item
in play unlocking pages in parallel when a read gets split up into a
lot of subrequests instead of one per subrequest.
Whilst the parallellism is nice in theory, in practice, the vast
majority of loads are sequential reads of the whole file, so
committing a bunch of threads to unlocking folios out of order doesn't
help in those cases.
(3) It should make it easier to implement content decryption. A folio
cannot be decrypted until all the requests that contribute to it have
completed - and, again, most loads are sequential and so, most of the
time, we want to begin decryption sequentially (though it's great if
the decryption can happen in parallel).
There is a disadvantage in that we're losing the ability to decrypt and
unlock things on an as-things-arrive basis which may affect some
applications.
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20241216204124.3752367-28-dhowells@redhat.com
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Now that directory and symlink reads go through netfslib, the afs_read
struct is mostly redundant with almost all data duplicated in the
netfs_io_request and netfs_io_subrequest structs that are also available
any time we're doing a fetch.
Eliminate afs_read by moving the one field we still need there to the
afs_call struct (we may be given a different amount of data than what we
asked for and have to track what remains of that) and using the
netfs_io_subrequest directly instead.
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20241216204124.3752367-24-dhowells@redhat.com
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Add two netfslib functions to build up or clean up a buffer in a
folio_queue. The first, netfs_alloc_folioq_buffer() will add folios to a
buffer, extending up at least to the given size. If it can, it will add
multipage folios. The folios are optionally have the mapping set and will
have the index set according to the distance from the front of the folio
queue.
The second function will free up a folio queue and put any folios in the
queue that have the first mark set.
The netfs_folio tracepoint is also altered to cope with folios that have a
NULL mapping, and the folios being added/put will have trace lines emitted
and will be accounted in the stats.
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20241216204124.3752367-19-dhowells@redhat.com
cc: Jeff Layton <jlayton@kernel.org>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: netfs@lists.linux.dev
cc: linux-afs@lists.infradead.org
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
A rolling buffer is a series of folios held in a list of folio_queues. New
folios and folio_queue structs may be inserted at the head simultaneously
with spent ones being removed from the tail without the need for locking.
The rolling buffer includes an iov_iter and it has to be careful managing
this as the list of folio_queues is extended such that an oops doesn't
incurred because the iterator was pointing to the end of a folio_queue
segment that got appended to and then removed.
We need to use the mechanism twice, once for read and once for write, and,
in future patches, we will use a second rolling buffer to handle bounce
buffering for content encryption.
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20241216204124.3752367-6-dhowells@redhat.com
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
netfs: Fix is-caching check in read-retry
The read-retry code checks the NETFS_RREQ_COPY_TO_CACHE flag to determine
if there might be failed reads from the cache that need turning into reads
from the server, with the intention of skipping the complicated part if it
can. The code that set the flag, however, got lost during the read-side
rewrite.
Fix the check to see if the cache_resources are valid instead. The flag
can then be removed.
Fixes: ee4cdf7ba8 ("netfs: Speed up buffered reading")
Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/3752048.1734381285@warthog.procyon.org.uk
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
