Clean up. There are a few places where we want to probe the server, but
don't actually care about the fsinfo result. Change these to use
nfs_probe_server(), which handles the fattr allocation for us.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

And rename it to nfs_probe_server(). I also change it to take the nfs_fh
as an argument so callers can choose what filehandle to probe.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

And call it before doing an FSINFO probe to reset to the baseline
capabilities before probing.
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

dprintk call sites that display no other information than the
function name can be replaced with use of the trace "function" or
"function_graph" plug-ins.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Introduce a single tracepoint that can replace simple dprintk call
sites in upper layer "rpc_call_done" callbacks. Example:

   kworker/u24:2-1254  [001]   771.026677: rpc_stats_latency:    task:00000001@00000002 xid=0x16a6f3c0 rpcbindv2 GETPORT backlog=446 rtt=101 execute=555
   kworker/u24:2-1254  [001]   771.026677: rpc_task_call_done:   task:00000001@00000002 flags=ASYNC|DYNAMIC|SOFT|SOFTCONN|SENT runstate=RUNNING|ACTIVE status=0 action=rpcb_getport_done
   kworker/u24:2-1254  [001]   771.026678: rpcb_setport:         task:00000001@00000002 status=0 port=20048
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

These new events report slightly different information for readpage
and readpages/readahead.

For readpage:
             fsx-1387  [006]   380.761896: nfs_aop_readpage:    fileid=00:28:2 fhandle=0x36fbbe51 version=1752899355910932437 offset=131072
             fsx-1387  [006]   380.761900: nfs_aop_readpage_done: fileid=00:28:2 fhandle=0x36fbbe51 version=1752899355910932437 offset=131072 ret=0

The index of a synchronous single-page read is reported.

For readpages:

             fsx-1387  [006]   380.760847: nfs_aop_readahead:   fileid=00:28:2 fhandle=0x36fbbe51 version=1752899355909932456 nr_pages=3
             fsx-1387  [006]   380.760853: nfs_aop_readahead_done: fileid=00:28:2 fhandle=0x36fbbe51 version=1752899355909932456 nr_pages=3 ret=0

The count of pages requested is reported. nfs_readpages does not
wait for the READ requests to complete.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Clean up: BIT() is preferred over open-coding the shift.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

For certain special cases, RPC-related tracepoints record a -1 as
the task ID or the client ID. It's ugly for a trace event to display
4 billion in these cases.

To help keep SUNRPC tracepoints consistent, create a macro that
defines the print format specifiers for tk_pid and cl_clid. At some
point in the future we might try tk_pid with a wider range of values
than 0..64K so this makes it easier to make that change.

RPC tracepoints now look like this:

<...>-1276  [009]   149.720358: rpc_clnt_new:         client=00000005 peer=[192.168.2.55]:20049 program=nfs server=klimt.ib

<...>-1342  [004]   149.921234: rpc_xdr_recvfrom:     task:0000001a@00000005 head=[0xff1242d9ab6dc01c,144] page=0 tail=[(nil),0] len=144
<...>-1342  [004]   149.921235: xprt_release_cong:    task:0000001a@00000005 snd_task:ffffffff cong=256 cwnd=16384
<...>-1342  [004]   149.921235: xprt_put_cong:        task:0000001a@00000005 snd_task:ffffffff cong=0 cwnd=16384
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Clean up: this field is no longer used.

xprt_rdma_pad_optimize is also no longer used, but is left in place
because it is part of the kernel/userspace API.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

This is a buffer to be left persistently registered while a
connection is up. Connection tear-down will automatically DMA-unmap,
invalidate, and dereg the MR. A persistently registered buffer is
lower in cost to provide, and it can never be coalesced into the
RDMA segment that carries the data payload.

An RPC that provisions a Write chunk with a non-aligned length now
uses this MR rather than the tail buffer of the RPC's rq_rcv_buf.
Reviewed-By: Tom Talpey <tom@talpey.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Fixes: 61ca2c4a ("NFS: Only reference user namespace from nfs4idmap struct instead of cred")
Signed-off-by: Alexey Gladkov <legion@kernel.org>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Save some space in the nfs_inode by setting up an anonymous union with
the fields that are peculiar to a specific type of filesystem object.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Fixes the following WARN_ON
WARNING: CPU: 2 PID: 18678 at fs/nfs/inode.c:123 nfs_clear_inode+0x3b/0x50 [nfs]
...
Call Trace:
  nfs4_evict_inode+0x57/0x70 [nfsv4]
  evict+0xd1/0x180
  dispose_list+0x48/0x60
  evict_inodes+0x156/0x190
  generic_shutdown_super+0x37/0x110
  nfs_kill_super+0x1d/0x40 [nfs]
  deactivate_locked_super+0x36/0xa0
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

We mustn't call nfs_wb_all() on anything other than a regular file.
Furthermore, we can exit early when we don't hold a delegation.
Reported-by: David Wysochanski <dwysocha@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Olga reports seeing the following Oops when doing O_DIRECT writes to a
pNFS flexfiles server:

Oops: 0000 [#1] SMP PTI
CPU: 1 PID: 234186 Comm: kworker/u8:1 Not tainted 5.15.0-rc4+ #4
Hardware name: Red Hat KVM/RHEL-AV, BIOS 1.13.0-2.module+el8.3.0+7353+9de0a3cc 04/01/2014
Workqueue: nfsiod rpc_async_release [sunrpc]
RIP: 0010:nfs_mark_request_commit+0x12/0x30 [nfs]
Code: ff ff be 03 00 00 00 e8 ac 34 83 eb e9 29 ff ff
ff e8 22 bc d7 eb 66 90 0f 1f 44 00 00 48 85 f6 74 16 48 8b 42 10 48
8b 40 18 <48> 8b 40 18 48 85 c0 74 05 e9 70 fc 15 ec 48 89 d6 e9 68 ed
ff ff
RSP: 0018:ffffa82f0159fe00 EFLAGS: 00010286
RAX: 0000000000000000 RBX: ffff8f3393141880 RCX: 0000000000000000
RDX: ffffa82f0159fe08 RSI: ffff8f3381252500 RDI: ffff8f3393141880
RBP: ffff8f33ac317c00 R08: 0000000000000000 R09: ffff8f3487724cb0
R10: 0000000000000008 R11: 0000000000000001 R12: 0000000000000001
R13: ffff8f3485bccee0 R14: ffff8f33ac317c10 R15: ffff8f33ac317cd8
FS:  0000000000000000(0000) GS:ffff8f34fbc80000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000018 CR3: 0000000122120006 CR4: 0000000000770ee0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
PKRU: 55555554
Call Trace:
 nfs_direct_write_completion+0x13b/0x250 [nfs]
 rpc_free_task+0x39/0x60 [sunrpc]
 rpc_async_release+0x29/0x40 [sunrpc]
 process_one_work+0x1ce/0x370
 worker_thread+0x30/0x380
 ? process_one_work+0x370/0x370
 kthread+0x11a/0x140
 ? set_kthread_struct+0x40/0x40
 ret_from_fork+0x22/0x30
Reported-by: Olga Kornievskaia <aglo@umich.edu>
Fixes: 9c455a8c ("NFS/pNFS: Clean up pNFS commit operations")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

If O_DIRECT bumps the commit_info rpcs_out field, then that could lead
to fsync() hangs. The fix is to ensure that O_DIRECT calls
nfs_commit_end().

Fixes: 723c921e ("sched/wait, fs/nfs: Convert wait_on_atomic_t() usage to the new wait_var_event() API")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Partially revert commit 2ce209c4 ("NFS: Wait for requests that are
locked on the commit list"), since it can lead to deadlocks between
commit requests and nfs_join_page_group().
For now we should assume that any locked requests on the commit list are
either about to be removed and committed by another task, or the writes
they describe are about to be retransmitted. In either case, we should
not need to worry.

Fixes: 2ce209c4 ("NFS: Wait for requests that are locked on the commit list")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Generate a trace event whenever the NFS client modifies the size of
a file. These new events aid troubleshooting workloads that trigger
races around size updates.

There are four new trace points, all named nfs_size_something so
they are easy to grep for or enable as a group with a single glob.

Size updated on the server:

  kworker/u24:10-194   [010]   369.939174: nfs_size_update:      fileid=00:28:2 fhandle=0x36fbbe51 version=1752899344277980615 cursize=250471 newsize=172083

Server-side size update reported via NFSv3 WCC attributes:

             fsx-1387  [006]   380.760686: nfs_size_wcc:         fileid=00:28:2 fhandle=0x36fbbe51 version=1752899355909932456 cursize=146792 newsize=171216

File has been truncated locally:

             fsx-1387  [007]   369.437421: nfs_size_truncate:    fileid=00:28:2 fhandle=0x36fbbe51 version=1752899231200117272 cursize=215244 newsize=0

File has been extended locally:

             fsx-1387  [007]   369.439213: nfs_size_grow:        fileid=00:28:2 fhandle=0x36fbbe51 version=1752899343704248410 cursize=258048 newsize=262144
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

The current range of RPC task PIDs is 0..65535. That's not adequate
for distinguishing tasks across multiple rpc_clnts running high
throughput workloads.

To help relieve this situation and to reduce the bottleneck of
having a single atomic for assigning all RPC task PIDs, assign task
PIDs per rpc_clnt.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Clean up: TRACE_DEFINE_ENUM is unnecessary because the target
symbols are all C macros, not enums.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

_nfs4_pnfs_v3/v4_ds_connect do
   some work
   smp_wmb
   ds->ds_clp = clp;

And nfs4_ff_layout_prepare_ds currently does
   smp_rmb
   if(ds->ds_clp)
      ...

This patch places the smp_rmb after the if. This ensures that following
reads only happen once nfs4_ff_layout_prepare_ds has checked that data
has been properly initialized.

Fixes: d67ae825 ("pnfs/flexfiles: Add the FlexFile Layout Driver")
Signed-off-by: Baptiste Lepers <baptiste.lepers@gmail.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Remove cache invalidations that are already covered by change attribute
updates.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

The original premise in commit 83672d39 ("NFS: Fix directory caching
problem - with test case and patch.") was that readdirplus was caching
attribute information and replaying it later. This is no longer the
case.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

If the directory changed while we were revalidating the dentry, then
don't update the dentry verifier. There is no value in setting the
verifier to an older value, and we could end up overwriting a more up to
date verifier from a parallel revalidation.

Fixes: efeda80d ("NFSv4: Fix revalidation of dentries with delegations")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>

If a user is doing 'ls -l', we have a heuristic in GETATTR that tells
the readdir code to try to use READDIRPLUS in order to refresh the inode
attributes. In certain cirumstances, we also try to invalidate the
remaining directory entries in order to ensure this refresh.

If there are multiple readers of the directory, we probably should avoid
invalidating the page cache, since the heuristic breaks down in that
situation anyway.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>

The check for duplicate readdir cookies should only care if the change
attribute is invalid or the data cache is invalid.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>

If we want to revalidate the directory, then just mark the change
attribute as invalid.

Fixes: 13c0b082 ("NFS: Replace use of NFS_INO_REVAL_PAGECACHE when checking cache validity")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>

NFS_INO_DATA_INVAL_DEFER and NFS_INO_INVALID_DATA should be considered
mutually exclusive.

Fixes: 1c341b77 ("NFS: Add deferred cache invalidation for close-to-open consistency violations")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Reviewed-by: Benjamin Coddington <bcodding@redhat.com>

Both NFSv3 and NFSv2 generate their change attribute from the ctime
value that was supplied by the server. However the problem is that there
are plenty of servers out there with ctime resolutions of 1ms or worse.
In a modern performance system, this is insufficient when trying to
decide which is the most recent set of attributes when, for instance, a
READ or GETATTR call races with a WRITE or SETATTR.

For this reason, let's revert to labelling the NFSv2/v3 change
attributes as NFS4_CHANGE_TYPE_IS_UNDEFINED. This will ensure we protect
against such races.

Fixes: 7b24dacf ("NFS: Another inode revalidation improvement")
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Tested-by: Chuck Lever <chuck.lever@oracle.com>

NFS4_CREATE_EXCLUSIVE does not allow the caller to set an access mode,
so for most Linux filesystems, the access call ends up returning no
permissions. However both NFS4_CREATE_EXCLUSIVE4_1 and
NFS4_CREATE_GUARDED allow the client to set the access mode.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

All these bits are being used as bit locks.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

The clearing of the XPRT_LOCKED bit has to happen after we clear
xprt->snd_task, but we don't require any extra memory barriers after
that.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

The only check for RPC_TASK_RUNNING is the one in rpc_make_runnable(),
which happens under the same spin lock held when we call
rpc_clear_running().

Ditto, the last check for RPC_TASK_QUEUED in rpc_execute() is performed
under the same lock as the one held when we call rpc_clear_queued().
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Don't let xprtiod pre-empt softirq.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

Allow tasks that need to pre-empt rpciod/xprtiod to do so when it is
safe.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

The premise of commit 6f9f1728 ("SUNRPC: Mitigate cond_resched() in
xprt_transmit()") was that cond_resched() is expensive and unnecessary
when there has been just a single send.
The point of cond_resched() is to ensure that tasks that should pre-empt
this one get a chance to do so when it is safe to do so. The code prior
to commit 6f9f1728 failed to take into account that it was keeping a
rpc_task pinned for longer than it needed to, and so rather than doing a
full revert, let's just move the cond_resched.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

If the cached credential exists but doesn't have any expiration callback
then exit early.
Fix up atomicity issues when replacing the credential with a new one
since the existing code could lead to refcount leaks.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

After the success of an operation such as rmdir() or unlink(), we expect
to add the dentry back to the dcache as an ordinary negative dentry.
However in NFS, unless it is labelled with the appropriate verifier for
the parent directory state, then nfs_lookup_revalidate will end up
discarding that dentry and forcing a new lookup.

The fix is to ensure that we relabel the dentry appropriately on
success.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

After the success of an operation such as link(), or symlink(), we
expect to add the dentry back to the dcache as an ordinary positive
dentry.
However in NFS, unless it is labelled with the appropriate verifier for
the parent directory state, then nfs_lookup_revalidate will end up
discarding that dentry and forcing a new lookup.

The fix is to ensure that we relabel the dentry appropriately on
success.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>