1/ remove 'start' and 'end' args from nfs_file_fsync_commit().
   They aren't used.

2/ Make nfs_context_set_write_error() a "static inline" in internal.h
   so we can...

3/ Use nfs_context_set_write_error() instead of mapping_set_error()
   if nfs_pageio_add_request() fails before sending any request.
   NFS generally keeps errors in the open_context, not the mapping,
   so this is more consistent.

4/ If filemap_write_and_write_range() reports any error, still
   check ctx->error.  The value in ctx->error is likely to be
   more useful.  As part of this, NFS_CONTEXT_ERROR_WRITE is
   cleared slightly earlier, before nfs_file_fsync_commit() is called,
   rather than at the start of that function.
Signed-off-by: NNeilBrown <neilb@suse.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Tools like tcpdump and rpcdebug can be very useful. But there are
plenty of environments where they are difficult or impossible to
use. For example, we've had customers report I/O failures during
workloads so heavy that collecting network traffic or enabling
RPC debugging are themselves onerous.

The kernel's static tracepoints are lightweight (less likely to
introduce timing changes) and efficient (the trace data is compact).
They also work in scenarios where capturing network traffic is not
possible due to lack of hardware support (some InfiniBand HCAs) or
where data or network privacy is a concern.

Introduce tracepoints that show when an NFS READ, WRITE, or COMMIT
is initiated, and when it completes. Record the arguments and
results of each operation, which are not shown by existing sunrpc
module's tracepoints.

For instance, the recorded offset and count can be used to match an
"initiate" event to a "done" event. If an NFS READ result returns
fewer bytes than requested or zero, seeing the EOF flag can be
probative. Seeing an NFS4ERR_BAD_STATEID result is also indication
of a particular class of problems. The timing information attached
to each event record can often be useful as well.

Usage example:

[root@manet tmp]# trace-cmd record -e nfs:*initiate* -e nfs:*done
/sys/kernel/debug/tracing/events/nfs/*initiate*/filter
/sys/kernel/debug/tracing/events/nfs/*done/filter
Hit Ctrl^C to stop recording
^CKernel buffer statistics:
  Note: "entries" are the entries left in the kernel ring buffer and are not
        recorded in the trace data. They should all be zero.

CPU: 0
entries: 0
overrun: 0
commit overrun: 0
bytes: 3680
oldest event ts:    78.367422
now ts:   100.124419
dropped events: 0
read events: 74

... and so on.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If we skip a subrequest due to a zero refcount, we should still count
the byte range that it covered so that we accurately reconstruct the
original request size.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

That can deadlock if this is the last reference since
nfs_page_group_destroy() calls nfs_page_group_sync_on_bit().
Note that even if the page was removed from the subpage list,
the req->wb_head could still be pointing to the old head.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

It's pretty much a duplicate of nfs_scan_commit_list() that also
clears the PG_COMMIT_TO_DS flag.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Since the commit list is not ordered, it is possible for nfs_scan_commit_list
to hold a request that nfs_lock_and_join_requests() is waiting for, while
at the same time trying to grab a request that nfs_lock_and_join_requests
already holds.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Commit fbe77c30 ("NFS: move rw_mode to nfs_pageio_header")
reintroduced some pointless code that commit 518662e0 ("NFS: fix
usage of mempools.") had recently removed.

Remove it again.

Cc: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: NNeilBrown <neilb@suse.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Now that the mirror allocation has been moved, the parameter can go.
Also remove the redundant symbol export.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If a request is on the commit list, but is locked, we will currently skip
it, which can lead to livelocking when the commit count doesn't reduce
to zero.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Switch from using the inode->i_lock for this to avoid contention with
other metadata manipulation.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

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

Rather than forcing us to take the inode->i_lock just in order to bump
the number.
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>

Split out the 2 cases so that we can treat the locking differently.
The issue is that the locking in the pageswapcache cache is highly
linked to the commit list locking.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Hide the locking from nfs_lock_and_join_requests() so that we can
separate out the requirements for swapcache pages.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Fix up the test in nfs_page_group_covers_page(). The simplest implementation
is to check that we have a set of intersecting or contiguous subrequests
that connect page offset 0 to nfs_page_length(req->wb_page).
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

nfs_page_group_lock() is now always called with the 'nonblock'
parameter set to 'false'.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

At this point, we only expect ever to potentially see PG_REMOVE and
PG_TEARDOWN being set on the subrequests.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Since nfs_page_group_destroy() does not take any locks on the requests
to be freed, we need to ensure that we don't inadvertently free the
request in nfs_destroy_unlinked_subrequests() while the last reference
is being released elsewhere.

Do this by:

1) Taking a reference to the request unless it is already being freed
2) Checking (under the page group lock) if PG_TEARDOWN is already set before
   freeing an unreferenced request in nfs_destroy_unlinked_subrequests()
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

When locking the entire group in order to remove subrequests,
the locks are always taken in order, and with the page group
lock being taken after the page head is locked. The intention
is that:

1) The lock on the group head guarantees that requests may not
   be removed from the group (although new entries could be appended
   if we're not holding the group lock).
2) It is safe to drop and retake the page group lock while iterating
   through the list, in particular when waiting for a subrequest lock.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We should no longer need the inode->i_lock, now that we've
straightened out the request locking. The locking schema is now:

1) Lock page head request
2) Lock the page group
3) Lock the subrequests one by one

Note that there is a subtle race with nfs_inode_remove_request() due
to the fact that the latter does not lock the page head, when removing
it from the struct page. Only the last subrequest is locked, hence
we need to re-check that the PagePrivate(page) is still set after
we've locked all the subrequests.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

nfs_try_to_update_request() should be able to cope now.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Simplify the code, and avoid some flushes to disk.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Both nfs_destroy_unlinked_subrequests() and nfs_lock_and_join_requests()
manipulate the inode flags adjusting the NFS_I(inode)->nrequests.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We don't want nfs_lock_and_join_requests() to start fiddling with
the request before the call to nfs_page_group_sync_on_bit().
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Request offsets and sizes are not guaranteed to be stable unless you
are holding the request locked.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

All other callers of nfs_page_group_lock() appear to already hold the
page lock on the head page, so doing it in the opposite order here
is inefficient, although not deadlock prone since we roll back all
locks on contention.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Yes, this is a situation that should never happen (hence the WARN_ON)
but we should still ensure that we free up the locks and references to
the faulty pages.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Micro-optimisation to move the lockless check into the for(;;) loop.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Add a lockless check for whether or not the page might be carrying
an existing writeback before we grab the inode->i_lock.
Reported-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We don't expect the page header lock to ever be held across I/O, so
it should always be safe to wait for it, even if we're doing nonblocking
writebacks.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Now that the writes will schedule a commit on their own, we don't
need nfs_write_inode() to schedule one if there are outstanding
writes, and we're being called in non-blocking mode.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

If the page cache is being flushed, then we want to ensure that we
do start a commit once the pages are done being flushed.
If we just wait until all I/O is done to that file, we can end up
livelocking until the balance_dirty_pages() mechanism puts its
foot down and forces I/O to stop.
So instead we do more or less the same thing that O_DIRECT does,
and set up a counter to tell us when the flush is done,
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Instead of messing with the commit path which has been causing issues,
add a COMMIT op after the COPY and ask for stable copies in the first
space.

It saves a round trip, since after the COPY, the client sends a COMMIT
anyway.
Signed-off-by: NOlga Kornievskaia <kolga@netapp.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If the client receives a fatal server error from nfs_pageio_add_request(),
then we should always truncate the page on which the error occurred.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If the server has already returned a fatal write error that the user
has not yet received on this file, then don't write back the other pages.
Instead, act as if they have been sent, and have returned with the same
error.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Anna Schumaker <anna.schumaker@netapp.com>
CC: linux-nfs@vger.kernel.org
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJan Kara <jack@suse.cz>
Acked-by: NTrond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Let's try to have it in a cacheline in nfs4_proc_pgio_rpc_prepare().
Signed-off-by: NBenjamin Coddington <bcodding@redhat.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Signed-off-by: NFred Isaman <fred.isaman@gmail.com>
Fixes: 0bcbf039 ("nfs: handle request add failure properly")
Cc: stable@vger.kernel.org # v4.5+
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

When passed GFP flags that allow sleeping (such as
GFP_NOIO), mempool_alloc() will never return NULL, it will
wait until memory is available.

This means that we don't need to handle failure, but that we
do need to ensure one thread doesn't call mempool_alloc()
twice on the one pool without queuing or freeing the first
allocation.  If multiple threads did this during times of
high memory pressure, the pool could be exhausted and a
deadlock could result.

pnfs_generic_alloc_ds_commits() attempts to allocate from
the nfs_commit_mempool while already holding an allocation
from that pool.  This is not safe.  So change
nfs_commitdata_alloc() to take a flag that indicates whether
failure is acceptable.

In pnfs_generic_alloc_ds_commits(), accept failure and
handle it as we currently do.  Else where, do not accept
failure, and do not handle it.

Even when failure is acceptable, we want to succeed if
possible.  That means both
 - using an entry from the pool if there is one
 - waiting for direct reclaim is there isn't.

We call mempool_alloc(GFP_NOWAIT) to achieve the first, then
kmem_cache_alloc(GFP_NOIO|__GFP_NORETRY) to achieve the
second.  Each of these can fail, but together they do the
best they can without blocking indefinitely.

The objects returned by kmem_cache_alloc() will still be freed
by mempool_free().  This is safe as mempool_alloc() uses
exactly the same function to allocate objects (since the mempool
was created with mempool_create_slab_pool()).  The object returned
by mempool_alloc() and kmem_cache_alloc() are indistinguishable
so mempool_free() will handle both identically, either adding to the
pool or calling kmem_cache_free().

Also, don't test for failure when allocating from
nfs_wdata_mempool.
Signed-off-by: NNeilBrown <neilb@suse.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>