- 08 Jan, 2022 20 commits
-
-
Chuck Lever authored
On the wire, I observed NFSv4 OPEN(CREATE) operations sometimes returning a reasonable-looking value in the cinfo.before field and zero in the cinfo.after field. RFC 8881 Section 10.8.1 says: > When a client is making changes to a given directory, it needs to > determine whether there have been changes made to the directory by > other clients. It does this by using the change attribute as > reported before and after the directory operation in the associated > change_info4 value returned for the operation. and > ... The post-operation change > value needs to be saved as the basis for future change_info4 > comparisons. A good quality client implementation therefore saves the zero cinfo.after value. During a subsequent OPEN operation, it will receive a different non-zero value in the cinfo.before field for that directory, and it will incorrectly believe the directory has changed, triggering an undesirable directory cache invalidation. There are filesystem types where fs_supports_change_attribute() returns false, tmpfs being one. On NFSv4 mounts, this means the fh_getattr() call site in fill_pre_wcc() and fill_post_wcc() is never invoked. Subsequently, nfsd4_change_attribute() is invoked with an uninitialized @stat argument. In fill_pre_wcc(), @stat contains stale stack garbage, which is then placed on the wire. In fill_post_wcc(), ->fh_post_wc is all zeroes, so zero is placed on the wire. Both of these values are meaningless. This fix can be applied immediately to stable kernels. Once there are more regression tests in this area, this optimization can be attempted again. Fixes: 428a23d2 ("nfsd: skip some unnecessary stats in the v4 case") Signed-off-by:
Chuck Lever <chuck.lever@oracle.com>
-
Chuck Lever authored
According to commit bbf2f098 ("nfsd: Reset the boot verifier on all write I/O errors"), the Linux NFS server forces all clients to resend pending unstable writes if any server-side write or commit operation encounters an error (say, ENOSPC). This is a rare and quite exceptional event that could require administrative recovery action, so it should be made trace-able. Example trace event: nfsd-938 [002] 7174.945558: nfsd_writeverf_reset: boot_time= 61cc920d xid=0xdcd62036 error=-28 new verifier=0x08aecc6142515904 Signed-off-by:
-
Chuck Lever authored
Clean up: These functions handle what the specs call a write verifier, which in the Linux NFS server implementation is now divorced from the server's boot instance Signed-off-by: -
Chuck Lever authored
There are two boot-time fields in struct nfsd_net: one called boot_time and one called nfssvc_boot. The latter is used only to form write verifiers, but its documenting comment declares: /* Time of server startup */ Since commit 27c438f5 ("nfsd: Support the server resetting the boot verifier"), this field can be reset at any time; it's no longer tied to server restart. So that comment is stale. Also, according to pahole, struct timespec64 is 16 bytes long on x86_64. The nfssvc_boot field is used only to form a write verifier, which is 8 bytes long. Let's clarify this situation by manufacturing an 8-byte verifier in nfs_reset_boot_verifier() and storing only that in struct nfsd_net. We're grabbing 128 bits of time, so compress all of those into a 64-bit verifier instead of throwing out the high-order bits. In the future, the siphash_key can be re-used for other hashed objects per-nfsd_net. Signed-off-by: -
Chuck Lever authored
When vfs_iter_write() starts to fail because a file system is full, a bunch of writes can fail at once with ENOSPC. These writes repeatedly invoke nfsd_reset_boot_verifier() in quick succession. Ensure that the time it grabs doesn't go backwards due to an ntp adjustment going on at the same time. Signed-off-by: -
Trond Myklebust authored
Since a clone error commit can cause the boot verifier to change, we should trace those errors. Signed-off-by:
Trond Myklebust <trond.myklebust@hammerspace.com> Signed-off-by:
-
Chuck Lever authored
Since this pointer is used repeatedly, move it to a stack variable. Signed-off-by: -
Chuck Lever authored
Since this pointer is used repeatedly, move it to a stack variable. Signed-off-by: -
Chuck Lever authored
The RWF_SYNC and !RWF_SYNC arms are now exactly alike except that the RWF_SYNC arm resets the boot verifier twice in a row. Fix that redundancy and de-duplicate the code. Signed-off-by: -
Trond Myklebust authored
The nfsd_file nf_rwsem is currently being used to separate file write and commit instances to ensure that we catch errors and apply them to the correct write/commit. We can improve scalability at the expense of a little accuracy (some extra false positives) by replacing the nf_rwsem with more careful use of the errseq_t mechanism to track errors across the different operations. Signed-off-by:
-
Chuck Lever authored
RFC 8881 explains the purpose of the write verifier this way: > The final portion of the result is the field writeverf. This field > is the write verifier and is a cookie that the client can use to > determine whether a server has changed instance state (e.g., server > restart) between a call to WRITE and a subsequent call to either > WRITE or COMMIT. But then it says: > This cookie MUST be unchanged during a single instance of the > NFSv4.1 server and MUST be unique between instances of the NFSv4.1 > server. If the cookie changes, then the client MUST assume that > any data written with an UNSTABLE4 value for committed and an old > writeverf in the reply has been lost and will need to be > recovered. RFC 1813 has similar language for NFSv3. NFSv2 does not have a write verifier since it doesn't implement the COMMIT procedure. Since commit 19e0663f ("nfsd: Ensure sampling of the write verifier is atomic with the write"), the Linux NFS server has returned a boot-time-based verifier for UNSTABLE WRITEs, but a zero verifier for FILE_SYNC and DATA_SYNC WRITEs. FILE_SYNC and DATA_SYNC WRITEs are not followed up with a COMMIT, so there's no need for clients to compare verifiers for stable writes. However, by returning a different verifier for stable and unstable writes, the above commit puts the Linux NFS server a step farther out of compliance with the first MUST above. At least one NFS client (FreeBSD) noticed the difference, making this a potential regression. Reported-by:
Rick Macklem <rmacklem@uoguelph.ca> Link: https://lore.kernel.org/linux-nfs/YQXPR0101MB096857EEACF04A6DF1FC6D9BDD749@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM/T/ Fixes: 19e0663f ("nfsd: Ensure sampling of the write verifier is atomic with the write") Signed-off-by:
-
Jeff Layton authored
If we get back -EOPENSTALE from an NFSv4 open, then we either got some unhandled error or the inode we got back was not the same as the one associated with the dentry. We really have no recourse in that situation other than to retry the open, and if it fails to just return nfserr_stale back to the client. Signed-off-by:
Jeff Layton <jeff.layton@primarydata.com> Signed-off-by:
Lance Shelton <lance.shelton@hammerspace.com> Signed-off-by:
-
Jeff Layton authored
The NFS client can occasionally return EREMOTEIO when signalling issues with the server. ...map to NFSERR_IO. Signed-off-by:
-
Peng Tao authored
Now that we have open file cache, it is possible that another client deletes the file and DP will not know about it. Then IO to MDS would fail with BADSTATEID and knfsd would start state recovery, which should fail as well and then nfs read/write will fail with EBADF. And it triggers a WARN() in nfserrno(). -----------[ cut here ]------------ WARNING: CPU: 0 PID: 13529 at fs/nfsd/nfsproc.c:758 nfserrno+0x58/0x70 [nfsd]() nfsd: non-standard errno: -9 modules linked in: nfsv3 nfs_layout_flexfiles rpcsec_gss_krb5 nfsv4 dns_resolver nfs fscache ip6t_rpfilter ip6t_REJECT nf_reject_ipv6 xt_connt pata_acpi floppy CPU: 0 PID: 13529 Comm: nfsd Tainted: G W 4.1.5-00307-g6e6579b #7 Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 09/30/2014 0000000000000000 00000000464e6c9c ffff88079085fba8 ffffffff81789936 0000000000000000 ffff88079085fc00 ffff88079085fbe8 ffffffff810a08ea ffff88079085fbe8 ffff88080f45c900 ffff88080f627d50 ffff880790c46a48 all Trace: [<ffffffff81789936>] dump_stack+0x45/0x57 [<ffffffff810a08ea>] warn_slowpath_common+0x8a/0xc0 [<ffffffff810a0975>] warn_slowpath_fmt+0x55/0x70 [<ffffffff81252908>] ? splice_direct_to_actor+0x148/0x230 [<ffffffffa02fb8c0>] ? fsid_source+0x60/0x60 [nfsd] [<ffffffffa02f9918>] nfserrno+0x58/0x70 [nfsd] [<ffffffffa02fba57>] nfsd_finish_read+0x97/0xb0 [nfsd] [<ffffffffa02fc7a6>] nfsd_splice_read+0x76/0xa0 [nfsd] [<ffffffffa02fcca1>] nfsd_read+0xc1/0xd0 [nfsd] [<ffffffffa0233af2>] ? svc_tcp_adjust_wspace+0x12/0x30 [sunrpc] [<ffffffffa03073da>] nfsd3_proc_read+0xba/0x150 [nfsd] [<ffffffffa02f7a03>] nfsd_dispatch+0xc3/0x210 [nfsd] [<ffffffffa0233af2>] ? svc_tcp_adjust_wspace+0x12/0x30 [sunrpc] [<ffffffffa0232913>] svc_process_common+0x453/0x6f0 [sunrpc] [<ffffffffa0232cc3>] svc_process+0x113/0x1b0 [sunrpc] [<ffffffffa02f740f>] nfsd+0xff/0x170 [nfsd] [<ffffffffa02f7310>] ? nfsd_destroy+0x80/0x80 [nfsd] [<ffffffff810bf3a8>] kthread+0xd8/0xf0 [<ffffffff810bf2d0>] ? kthread_create_on_node+0x1b0/0x1b0 [<ffffffff817912a2>] ret_from_fork+0x42/0x70 [<ffffffff810bf2d0>] ? kthread_create_on_node+0x1b0/0x1b0 Signed-off-by:
Peng Tao <tao.peng@primarydata.com> Signed-off-by:
-
Chuck Lever authored
The Linux NFS server currently responds to a zero-length NFSv3 WRITE request with NFS3ERR_IO. It responds to a zero-length NFSv4 WRITE with NFS4_OK and count of zero. RFC 1813 says of the WRITE procedure's @count argument: count The number of bytes of data to be written. If count is 0, the WRITE will succeed and return a count of 0, barring errors due to permissions checking. RFC 8881 has similar language for NFSv4, though NFSv4 removed the explicit @count argument because that value is already contained in the opaque payload array. The synthetic client pynfs's WRT4 and WRT15 tests do emit zero- length WRITEs to exercise this spec requirement. Commit fdec6114 ("nfsd4: zero-length WRITE should succeed") addressed the same problem there with the same fix. But interestingly the Linux NFS client does not appear to emit zero- length WRITEs, instead squelching them. I'm not aware of a test that can generate such WRITEs for NFSv3, so I wrote a naive C program to generate a zero-length WRITE and test this fix. Fixes: 8154ef27 ("NFSD: Clean up legacy NFS WRITE argument XDR decoders") Reported-by: -
Vasily Averin authored
nbl allocated in nfsd4_lock can be released by a several ways: directly in nfsd4_lock(), via nfs4_laundromat(), via another nfs command RELEASE_LOCKOWNER or via nfsd4_callback. This structure should be refcounted to be used and released correctly in all these cases. Refcount is initialized to 1 during allocation and is incremented when nbl is added into nbl_list/nbl_lru lists. Usually nbl is linked into both lists together, so only one refcount is used for both lists. However nfsd4_lock() should keep in mind that nbl can be present in one of lists only. This can happen if nbl was handled already by nfs4_laundromat/nfsd4_callback/etc. Refcount is decremented if vfs_lock_file() returns FILE_LOCK_DEFERRED, because nbl can be handled already by nfs4_laundromat/nfsd4_callback/etc. Refcount is not changed in find_blocked_lock() because of it reuses counter released after removing nbl from lists. Signed-off-by:
Vasily Averin <vvs@virtuozzo.com> Reviewed-by:
Jeff Layton <jlayton@kernel.org> Signed-off-by:
-
J. Bruce Fields authored
NFSv4.1 supports an optional lock notification feature which notifies the client when a lock comes available. (Normally NFSv4 clients just poll for locks if necessary.) To make that work, we need to request a blocking lock from the filesystem. We turned that off for NFS in commit f657f8ee ("nfs: don't atempt blocking locks on nfs reexports") [sic] because it actually blocks the nfsd thread while waiting for the lock. Thanks to Vasily Averin for pointing out that NFS isn't the only filesystem with that problem. Any filesystem that leaves ->lock NULL will use posix_lock_file(), which does the right thing. Simplest is just to assume that any filesystem that defines its own ->lock is not safe to request a blocking lock from. So, this patch mostly reverts commit f657f8ee ("nfs: don't atempt blocking locks on nfs reexports") [sic] and commit b840be2f ("lockd: don't attempt blocking locks on nfs reexports"), and instead uses a check of ->lock (Vasily's suggestion) to decide whether to support blocking lock notifications on a given filesystem. Also add a little documentation. Perhaps someday we could add back an export flag later to allow filesystems with "good" ->lock methods to support blocking lock notifications. Reported-by:
J. Bruce Fields <bfields@redhat.com> [ cel: Description rewritten to address checkpatch nits ] [ cel: Fixed warning when SUNRPC debugging is disabled ] [ cel: Fixed NULL check ] Signed-off-by:
-
Chuck Lever authored
Clean up. Trond points out that xdr_stream_decode_uint32_array() does the same thing as nfsd4_decode_bitmap4(). Suggested-by:
Trond Myklebust <trondmy@hammerspace.com> Signed-off-by:
-
J. Bruce Fields authored
The use of the bitmaps is confusing. Add a cross-reference to make it easier to find the existing comment. Add an updated reference with URL to make it quicker to look up. And a bit more editorializing about the value of this. Signed-off-by:
-
Chuck Lever authored
Clean up: The garbage_args and cant_encode tracepoints report the same information as each other, so combine them into a single tracepoint class to reduce code duplication and slightly reduce the size of trace.o. Signed-off-by:
-
- 13 Dec, 2021 20 commits
-
-
Chuck Lever authored
I'm about to add more information to the server-side SUNRPC tracepoints, so I'm going to offset the increased trace log consumption by getting rid of some tracepoints that fire frequently but don't offer much value. trace_svc_xprt_received() was useful for debugging, perhaps, but is not generally informative. trace_svc_handle_xprt() reports largely the same information as trace_svc_xdr_recvfrom(). As a clean-up, rename trace_svc_xprt_do_enqueue() to match svc_xprt_dequeue(). Signed-off-by: -
NeilBrown authored
We currently have a 'laundrette' for closing cached files - a different work-item for each network-namespace. These 'laundrettes' (aka struct nfsd_fcache_disposal) are currently on a list, and are freed using rcu. The list is not necessary as we have a per-namespace structure (struct nfsd_net) which can hold a link to the nfsd_fcache_disposal. The use of kfree_rcu is also unnecessary as the cache is cleaned of all files associated with a given namespace, and no new files can be added, before the nfsd_fcache_disposal is freed. So add a '->fcache_disposal' link to nfsd_net, and discard the list management and rcu usage. Signed-off-by:
NeilBrown <neilb@suse.de> Signed-off-by:
-
Jiapeng Chong authored
Eliminate the follow smatch warning: fs/nfsd/nfs4xdr.c:4766 nfsd4_encode_read_plus_hole() warn: inconsistent indenting. Reported-by:
Abaci Robot <abaci@linux.alibaba.com> Signed-off-by:
Jiapeng Chong <jiapeng.chong@linux.alibaba.com> Signed-off-by:
-
Chuck Lever authored
Commit 7142b98d ("nfsd: Clean up drc cache in preparation for global spinlock elimination"), billed as a clean-up, added be32_to_cpu() to the DRC hash function without explanation. That commit removed two comments that state that byte-swapping in the hash function is unnecessary without explaining whether there was a need for that change. On some Intel CPUs, the swab32 instruction is known to cause a CPU pipeline stall. be32_to_cpu() does not add extra randomness, since the hash multiplication is done /before/ shifting to the high-order bits of the result. As a micro-optimization, remove the unnecessary transform from the DRC hash function. Signed-off-by:
-
NeilBrown authored
Now that thread management is consistent there is no need for nfs-callback to use svc_create_pooled() as introduced in Commit df807fff ("NFSv4.x/callback: Create the callback service through svc_create_pooled"). So switch back to svc_create(). If service pools were configured, but the number of threads were left at '1', nfs callback may not work reliably when svc_create_pooled() is used. Signed-off-by:
-
NeilBrown authored
svc_set_num_threads() does everything that lockd_start_svc() does, except set sv_maxconn. It also (when passed 0) finds the threads and stops them with kthread_stop(). So move the setting for sv_maxconn, and use svc_set_num_thread() We now don't need nlmsvc_task. Now that we use svc_set_num_threads() it makes sense to set svo_module. This request that the thread exists with module_put_and_exit(). Also fix the documentation for svo_module to make this explicit. svc_prepare_thread is now only used where it is defined, so it can be made static. Signed-off-by:
-
NeilBrown authored
Currently 'pooled' services hold a reference on the pool_map, and 'unpooled' services do not. svc_destroy() uses the presence of ->svo_function (via svc_serv_is_pooled()) to determine if the reference should be dropped. There is no direct correlation between being pooled and the use of svo_function, though in practice, lockd is the only non-pooled service, and the only one not to use svo_function. This is untidy and would cause problems if we changed lockd to use svc_set_num_threads(), which requires the use of ->svo_function. So change the test for "is the service pooled" to "is sv_nrpools > 1". This means that when svc_pool_map_get() returns 1, it must NOT take a reference to the pool. We discard svc_serv_is_pooled(), and test sv_nrpools directly. Signed-off-by:
-
NeilBrown authored
These definitions are not used outside of svc.c, and there is no evidence that they ever have been. So move them into svc.c and make the declarations 'static'. Signed-off-by:
-
NeilBrown authored
lockd_create_svc() already does an svc_get() if the service already exists, so it is more like a "get" than a "create". So: - Move the increment of nlmsvc_users into the function as well - rename to lockd_get(). It is now the inverse of lockd_put(). Signed-off-by:
-
NeilBrown authored
There is some cleanup that is duplicated in lockd_down() and the failure path of lockd_up(). Factor these out into a new lockd_put() and call it from both places. lockd_put() does *not* take the mutex - that must be held by the caller. It decrements nlmsvc_users and if that reaches zero, it cleans up. Signed-off-by:
-
NeilBrown authored
The normal place to call svc_exit_thread() is from the thread itself just before it exists. Do this for lockd. This means that nlmsvc_rqst is not used out side of lockd_start_svc(), so it can be made local to that function, and renamed to 'rqst'. Signed-off-by:
-
NeilBrown authored
lockd_start_svc() only needs to be called once, just after the svc is created. If the start fails, the svc is discarded too. It thus makes sense to call lockd_start_svc() from lockd_create_svc(). This allows us to remove the test against nlmsvc_rqst at the start of lockd_start_svc() - it must always be NULL. lockd_up() only held an extra reference on the svc until a thread was created - then it dropped it. The thread - and thus the extra reference - will remain until kthread_stop() is called. Now that the thread is created in lockd_create_svc(), the extra reference can be dropped there. So the 'serv' variable is no longer needed in lockd_up(). Signed-off-by:
-
NeilBrown authored
Now that the network status notifiers use nlmsvc_serv rather then nlmsvc_rqst the management can be simplified. Notifier unregistration synchronises with any pending notifications so providing we unregister before nlm_serv is freed no further interlock is required. So we move the unregister call to just before the thread is killed (which destroys the service) and just before the service is destroyed in the failure-path of lockd_up(). Then nlm_ntf_refcnt and nlm_ntf_wq can be removed. Signed-off-by:
-
NeilBrown authored
lockd has two globals - nlmsvc_task and nlmsvc_rqst - but mostly it wants the 'struct svc_serv', and when it doesn't want it exactly it can get to what it wants from the serv. This patch is a first step to removing nlmsvc_task and nlmsvc_rqst. It introduces nlmsvc_serv to store the 'struct svc_serv*'. This is set as soon as the serv is created, and cleared only when it is destroyed. Signed-off-by:
-
NeilBrown authored
nfsd currently maintains an open-coded read/write semaphore (refcount and wait queue) for each network namespace to ensure the nfs service isn't shut down while the notifier is running. This is excessive. As there is unlikely to be contention between notifiers and they run without sleeping, a single spinlock is sufficient to avoid problems. Signed-off-by:
-
NeilBrown authored
The ->svo_setup callback serves no purpose. It is always called from within the same module that chooses which callback is needed. So discard it and call the relevant function directly. Now that svc_set_num_threads() is no longer used remove it and rename svc_set_num_threads_sync() to remove the "_sync" suffix. Signed-off-by:
-
NeilBrown authored
nfsd cannot currently use svc_set_num_threads_sync. It instead uses svc_set_num_threads which does *not* wait for threads to all exit, and has a separate mechanism (nfsd_shutdown_complete) to wait for completion. The reason that nfsd is unlike other services is that nfsd threads can exit separately from svc_set_num_threads being called - they die on receipt of SIGKILL. Also, when the last thread exits, the service must be shut down (sockets closed). For this, the nfsd_mutex needs to be taken, and as that mutex needs to be held while svc_set_num_threads is called, the one cannot wait for the other. This patch changes the nfsd thread so that it can drop the ref on the service without blocking on nfsd_mutex, so that svc_set_num_threads_sync can be used: - if it can drop a non-last reference, it does that. This does not trigger shutdown and does not require a mutex. This will likely happen for all but the last thread signalled, and for all threads being shut down by nfsd_shutdown_threads() - if it can get the mutex without blocking (trylock), it does that and then drops the reference. This will likely happen for the last thread killed by SIGKILL - Otherwise there might be an unrelated task holding the mutex, possibly in another network namespace, or nfsd_shutdown_threads() might be just about to get a reference on the service, after which we can drop ours safely. We cannot conveniently get wakeup notifications on these events, and we are unlikely to need to, so we sleep briefly and check again. With this we can discard nfsd_shutdown_complete and nfsd_complete_shutdown(), and switch to svc_set_num_threads_sync. Signed-off-by:
-
NeilBrown authored
There is nothing happening in the start of nfsd() that requires protection by the mutex, so don't take it until shutting down the thread - which does still require protection - but only for nfsd_put(). Signed-off-by:
-
NeilBrown authored
Using sv_lock means we don't need to hold the service mutex over these updates. In particular, svc_exit_thread() no longer requires synchronisation, so threads can exit asynchronously. Note that we could use an atomic_t, but as there are many more read sites than writes, that would add unnecessary noise to the code. Some reads are already racy, and there is no need for them to not be. Signed-off-by:
-
NeilBrown authored
This allows us to move the updates for th_cnt out of the mutex. This is a step towards reducing mutex coverage in nfsd(). Signed-off-by:
-
