Figuring out the root case for the REMOVE/CLOSE race and
suggesting the solution was done by Neil Brown.

Currently what happens is that direct IO calls hold a reference
on the open context which is decremented as an asynchronous task
in the nfs_direct_complete(). Before reference is decremented,
control is returned to the application which is free to close the
file. When close is being processed, it decrements its reference
on the open_context but since directIO still holds one, it doesn't
sent a close on the wire. It returns control to the application
which is free to do other operations. For instance, it can delete a
file. Direct IO is finally releasing its reference and triggering
an asynchronous close. Which races with the REMOVE. On the server,
REMOVE can be processed before the CLOSE, failing the REMOVE with
EACCES as the file is still opened.
Signed-off-by: NOlga Kornievskaia <kolga@netapp.com>
Suggested-by: NNeil Brown <neilb@suse.com>
CC: stable@vger.kernel.org
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

I measured a 50% throughput regression for large direct writes.

The observed on-the-wire behavior is that the client sends every
NFS WRITE twice: once as an UNSTABLE WRITE plus a COMMIT, and once
as a FILE_SYNC WRITE.

This is because the nfs_write_match_verf() check in
nfs_direct_commit_complete() fails for every WRITE.

Buffered writes use nfs_write_completion(), which sets req->wb_verf
correctly. Direct writes use nfs_direct_write_completion(), which
does not set req->wb_verf at all. This leaves req->wb_verf set to
all zeroes for every direct WRITE, and thus
nfs_direct_commit_completion() always sets NFS_ODIRECT_RESCHED_WRITES.

This fix appears to restore nearly all of the lost performance.

Fixes: 1f28476d ("NFS: Fix O_DIRECT commit verifier handling")
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The variable result is being initialized with a value that is never read
and it is being updated later with a new value.  The initialization is
redundant and can be removed.

Addresses-Coverity: ("Unused value")
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

If we have to retransmit requests, try to join their page groups
first.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

nfs_direct_write_scan_commit_list() will lock the request and bump
the reference count, but we also need to account for the reference
that was taken when we initially added the request to the commit list.

Fixes: fb5f7f20 ("NFS: commit errors should be fatal")
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Move the pNFS commit related operations into a separate structure
that can be carried by the pnfs_ds_commit_info.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Remove the unused bucket array in struct pnfs_ds_commit_info.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Instead of trying to save the commit verifiers and checking them against
previous writes, adopt the same strategy as for buffered writes, of
just checking the verifiers at commit time.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Fix the O_DIRECT code to avoid retries if the COMMIT fails with a fatal
error.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

Add a pNFS callback to allow the O_DIRECT code to release the DS
commitinfo when freeing the dreq.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

When we have multiple layout segments with different lists of mirrored
data, we need to track the commits on a per layout segment basis.
This patch adds a list to support this tracking in struct
pnfs_ds_commit_info.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

When dreq is allocated by nfs_direct_req_alloc(), dreq->kref is
initialized to 2. Therefore we need to call nfs_direct_req_release()
twice to release the allocated dreq. Usually it is called in
nfs_file_direct_{read, write}() and nfs_direct_complete().

However, current code only calls nfs_direct_req_relese() once if
nfs_get_lock_context() fails in nfs_file_direct_{read, write}().
So, that case would result in memory leak.

Fix this by adding the missing call.
Signed-off-by: NMisono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

The 'hdr->good_bytes' is defined as the number of bytes we expect to
read or write starting at offset hdr->io_start. In the case of a partial
read/write we may end up adjusting hdr->args.offset and hdr->args.count
to skip I/O for data that was already read/written, and so we must ensure
the calculation takes that into account.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of making assumptions about the commit verifier contents, change
the commit code to ensure we always check that the verifier was set
by the XDR code.

Fixes: f54bcf2e ("pnfs: Prepare for flexfiles by pulling out common code")
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

We no longer need the extra mirror length tracking in the O_DIRECT code,
as we are able to track the maximum contiguous length in dreq->max_count.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

When a series of O_DIRECT reads or writes are truncated, either due to
eof or due to an error, then we should return the number of contiguous
bytes that were received/sent starting at the offset specified by the
application.

Currently, we are failing to correctly check contiguity, and so we're
failing the generic/465 in xfstests when the race between the read
and write RPCs causes the file to get extended while the 2 reads are
outstanding. If the first read RPC call wins the race and returns with
eof set, we should treat the second read RPC as being truncated.
Reported-by: NSu Yanjun <suyj.fnst@cn.fujitsu.com>
Fixes: 1ccbad9f ("nfs: fix DIO good bytes calculation")
Cc: stable@vger.kernel.org # 4.1+
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

If the attempt to resend the I/O results in no bytes being read/written,
we must ensure that we report the error.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Fixes: 0a00b77b ("nfs: mirroring support for direct io")
Cc: stable@vger.kernel.org # v3.20+

Add SPDX license identifiers to all files which:

 - Have no license information of any form

 - Have EXPORT_.*_SYMBOL_GPL inside which was used in the
   initial scan/conversion to ignore the file

These files fall under the project license, GPL v2 only. The resulting SPDX
license identifier is:

  GPL-2.0-only
Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Signed-off-by: NGreg Kroah-Hartman <gregkh@linuxfoundation.org>

When the client is reading or writing using pNFS, and hits an error
on the DS, then it typically sends a LAYOUTERROR and/or LAYOUTRETURN
to the MDS, before redirtying the failed pages, and going for a new
round of reads/writebacks. The problem is that if the server has no
way to fix the DS, then we may need a way to interrupt this loop
after a set number of attempts have been made.
This patch adds an optional module parameter that allows the admin
to specify how many times to retry the read/writeback process before
failing with a fatal error.
The default behaviour is to retry forever.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

All the callers of nfs_create_request() are now creating page group
heads, so we can remove the redundant 'last' page argument.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Allow the caller to pass error information when cleaning up a failed
I/O request so that we can conditionally take action to cancel the
request altogether if the error turned out to be fatal.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

In several places we're just moving the struct nfs_page from one list to
another by first removing from the existing list, then adding to the new
one.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

When we use direct_IO with an NFS backing store, we can trigger a
WARNING in __set_page_dirty(), as below, since we're dirtying the page
unnecessarily in nfs_direct_read_completion().

To fix, replicate the logic in commit 53cbf3b1 ("fs: direct-io:
don't dirtying pages for ITER_BVEC/ITER_KVEC direct read").

Other filesystems that implement direct_IO handle this; most use
blockdev_direct_IO(). ceph and cifs have similar logic.

mount 127.0.0.1:/export /nfs
dd if=/dev/zero of=/nfs/image bs=1M count=200
losetup --direct-io=on -f /nfs/image
mkfs.btrfs /dev/loop0
mount -t btrfs /dev/loop0 /mnt/

kernel: WARNING: CPU: 0 PID: 8067 at fs/buffer.c:580 __set_page_dirty+0xaf/0xd0
kernel: Modules linked in: loop(E) nfsv3(E) rpcsec_gss_krb5(E) nfsv4(E) dns_resolver(E) nfs(E) fscache(E) nfsd(E) auth_rpcgss(E) nfs_acl(E) lockd(E) grace(E) fuse(E) tun(E) ip6t_rpfilter(E) ipt_REJECT(E) nf_
kernel:  snd_seq(E) snd_seq_device(E) snd_pcm(E) video(E) snd_timer(E) snd(E) soundcore(E) ip_tables(E) xfs(E) libcrc32c(E) sd_mod(E) sr_mod(E) cdrom(E) ata_generic(E) pata_acpi(E) crc32c_intel(E) ahci(E) li
kernel: CPU: 0 PID: 8067 Comm: kworker/0:2 Tainted: G            E     4.20.0-rc1.master.20181111.ol7.x86_64 #1
kernel: Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006
kernel: Workqueue: nfsiod rpc_async_release [sunrpc]
kernel: RIP: 0010:__set_page_dirty+0xaf/0xd0
kernel: Code: c3 48 8b 02 f6 c4 04 74 d4 48 89 df e8 ba 05 f7 ff 48 89 c6 eb cb 48 8b 43 08 a8 01 75 1f 48 89 d8 48 8b 00 a8 04 74 02 eb 87 <0f> 0b eb 83 48 83 e8 01 eb 9f 48 83 ea 01 0f 1f 00 eb 8b 48 83 e8
kernel: RSP: 0000:ffffc1c8825b7d78 EFLAGS: 00013046
kernel: RAX: 000fffffc0020089 RBX: fffff2b603308b80 RCX: 0000000000000001
kernel: RDX: 0000000000000001 RSI: ffff9d11478115c8 RDI: ffff9d11478115d0
kernel: RBP: ffffc1c8825b7da0 R08: 0000646f6973666e R09: 8080808080808080
kernel: R10: 0000000000000001 R11: 0000000000000000 R12: ffff9d11478115d0
kernel: R13: ffff9d11478115c8 R14: 0000000000003246 R15: 0000000000000001
kernel: FS:  0000000000000000(0000) GS:ffff9d115ba00000(0000) knlGS:0000000000000000
kernel: CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
kernel: CR2: 00007f408686f640 CR3: 0000000104d8e004 CR4: 00000000000606f0
kernel: DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
kernel: DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
kernel: Call Trace:
kernel:  __set_page_dirty_buffers+0xb6/0x110
kernel:  set_page_dirty+0x52/0xb0
kernel:  nfs_direct_read_completion+0xc4/0x120 [nfs]
kernel:  nfs_pgio_release+0x10/0x20 [nfs]
kernel:  rpc_free_task+0x30/0x70 [sunrpc]
kernel:  rpc_async_release+0x12/0x20 [sunrpc]
kernel:  process_one_work+0x174/0x390
kernel:  worker_thread+0x4f/0x3e0
kernel:  kthread+0x102/0x140
kernel:  ? drain_workqueue+0x130/0x130
kernel:  ? kthread_stop+0x110/0x110
kernel:  ret_from_fork+0x35/0x40
kernel: ---[ end trace 01341980905412c9 ]---
Signed-off-by: NDave Kleikamp <dave.kleikamp@oracle.com>
Signed-off-by: NSantosh Shilimkar <santosh.shilimkar@oracle.com>

[forward-ported to v4.20]
Signed-off-by: NCalum Mackay <calum.mackay@oracle.com>
Reviewed-by: NDave Kleikamp <dave.kleikamp@oracle.com>
Reviewed-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

When a direct-write completes, a work_struct is schedule to handle
the completion.
When NFS is being used for swap, the direct write might be a swap-out,
so memory allocation can block until the write completes.
The work queue currently used is not WQ_MEM_RECLAIM, so tasks
can block waiting for memory - this leads to deadlock.

So use nfsiod_workqueue instead.  This will always have a running
thread, and work items should never block waiting for memory.
Signed-off-by: NNeil Brown <neilb@suse.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The start offset needs to be of type loff_t.

Fixed: 5fadeb47 ("nfs: count DIO good bytes correctly with mirroring")
Cc: stable@vger.kernel.org # v4.0+
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If some of the WRITE calls making up an O_DIRECT write syscall fail,
we neglect to commit, even if some of the WRITEs succeed.

We also depend on the commit code to free the reference count on the
nfs_page taken in the "if (request_commit)" case at the end of
nfs_direct_write_completion().  The problem was originally noticed
because ENOSPC's encountered partway through a write would result in a
closed file being sillyrenamed when it should have been unlinked.
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

The commit lists can get very large, so using the inode->i_lock can
end up affecting general metadata performance.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Just remove the function and have the caller use nfs_release_request()
instead.
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

It leaves the iterator advanced by the amount of IO it has requested
instead of the amount actually transferred.  Among other things,
that confuses the hell out of generic_file_splice_read().
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

This was entirely automated, using the script by Al:

  PATT='^[[:blank:]]*#[[:blank:]]*include[[:blank:]]*<asm/uaccess.h>'
  sed -i -e "s!$PATT!#include <linux/uaccess.h>!" \
        $(git grep -l "$PATT"|grep -v ^include/linux/uaccess.h)

to do the replacement at the end of the merge window.
Requested-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This parameter hasn't been used since 2a009ec9 (Linux 3.13-rc3), so
let's remove it from this function.
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

There is only one waiter for the completion, therefore there
is no need to use complete_all(). Let's make that clear by
using complete() instead of complete_all().

nfs_file_direct_write() or nfs_file_direct_read() allocated a request
object via nfs_direct_req_alloc(), which initializes the
completion. The request object then is freed later in the exit path.
Between the initialization and the release either
nfs_direct_write_schedule_iovec() resp
nfs_direct_read_schedule_iovec() are called which will asynchronously
process the request. The calling function waits via nfs_direct_wait()
till the async work has been done. Thus there is only one waiter on
the completion.

nfs_direct_pgio_init() and nfs_direct_read_completion() are passed via
function pointers to nfs pageio. The first function does a ref
counting (get_dreq() and put_dreq()) which ensures that
nfs_direct_read_completion() and nfs_direct_read_schedule_iovec() only
call the completion path once.

The usage pattern of the completion is:

waiter context                          waker context

nfs_file_direct_write()
  dreq = nfs_direct_req_alloc()
    init_completion()
  nfs_direct_write_schedule_iovec()
  nfs_direct_wait()
    wait_for_completion_killable()

                                        nfs_direct_write_schedule_work()
                                          nfs_direct_complete()
                                            complete()

nfs_file_direct_read()
  dreq = nfs_direct_req_all()
    init_completion()
  nfs_direct_read_schedule_iovec()
  nfs_direct_wait()
    wait_for_completion_killable()
                                        nfs_direct_read_schedule_iovec()
                                          nfs_direct_complete()
                                            complete()

                                        nfs_direct_read_completion()
                                          nfs_direct_complete()
                                            complete()
Signed-off-by: NDaniel Wagner <daniel.wagner@bmw-carit.de>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

There is only one caller that sets the "write" argument to true,
so just move the call to nfs_zap_mapping() and get rid of the
now redundant argument.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Allow dio requests to be scheduled in parallel, but ensuring that they
do not conflict with buffered I/O.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

On success, the RPC callbacks will ensure that we make the appropriate calls
to nfs_writeback_update_inode()
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We should not be interested in looking at the value of the stable field,
since that could take any value.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

if we read or wrote something, we must report it
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: NJeff Layton <jlayton@poochiereds.net>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>