We must not skip encoding the statistics, or the server will see an
XDR encoding error.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>
Cc: stable@vger.kernel.org # 4.0+

When we hit 22 errors, we start to overflow the memory buffers allocated
to the LAYOUTRETURN errors. The issue is that currently, RPC call reply
ordering determines how successful we are in merging errors that refer
to contiguous READ or WRITE requests.

Fix is to use an insertion sort to help detect contiguity.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

The flexfiles layout in particular, seems to want to poke around in the
O_DIRECT flags when retransmitting.
This patch sets up an interface to allow it to call back into O_DIRECT
to handle retransmission correctly. It also fixes a potential bug whereby
we could change the behaviour of O_DIRECT if an error is already pending.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Jeff reports seeing an Oops in ff_layout_alloc_lseg. Turns out
copy+paste has played cruel tricks on a nested loop.
Reported-by: NJeff Layton <jeff.layton@primarydata.com>
Cc: stable@vger.kernel.org # 4.3+
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Currently, we will only record the layoutstats correctly if the
RPC call successfully obtains a slot. If we exit before that
happens, then we may find ourselves starting the busy timer through
the call in ff_layout_(read|write)_prepare_layoutstats, but never stopping it.

The same thing happens if we're doing DA-DS.

The fix is to ensure that we catch these cases in the rpc_release()
callback.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

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

When we replay a failed read, write or commit to the dataserver, we
need to ensure that we call ff_layout_read_prepare_v3(),
ff_layout_write_prepare_v3 or ff_layout_commit_prepare_v3() so that we
reset the statistics.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Fix a bug in which flexfiles clients are falling back to I/O through the
MDS even when the FF_FLAGS_NO_IO_THRU_MDS flag is set.

The flexfiles client will always report errors through the LAYOUTRETURN
and/or LAYOUTERROR mechanisms, so it should normally be safe for it
to retry the LAYOUTGET until it fails or succeeds.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If client ever restarts IO due to some errors, we'll endup
mis-counting IO stats if we do the counting in .rpc_done
callback. Move it to .rpc_count_stats callback that is only
called when releasing RPC.
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We just need to delay and retry in these cases.
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

For ERESTARTSYS/EIO/EROFS/ENOSPC/E2BIG in layoutget, we
should just bail out instead of hiding the error and
retrying inband IO.

Change all the call sites to pop the error all the way up.
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Some servers want to be able to control the frequency with which clients
report layoutstats, for instance, in order to monitor QoS for a particular
file or set of file. In order to support this, the flexfiles layout allows
the server to pass this info as a hint in the layout payload.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

For loosely coupled pNFS/flexfiles systems, there is often no advantage
at all in going through the MDS for I/O, since the MDS is subject to
the same limitations as all other clients when talking to DSes. If a
DS is unresponsive, I/O through the MDS will fail.

For such systems, the only scalable solution is to have the pNFS clients
retry doing pNFS, and so the protocol now provides a flag that allows
the pNFS server to signal this.

If LAYOUTGET returns FF_FLAGS_NO_IO_THRU_MDS, then we should assume that
the MDS wants the client to retry using these devices, even if they were
previously marked as being unavailable. To do so, we add a helper,
ff_layout_mark_devices_valid() that will be called from layoutget.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If the pNFS/flexfiles file is mirrored, and a read to one mirror fails,
then we should bump the mirror index, so that we retry to a different
mirror. Once we've iterated through all mirrors and all failed, we can
return the layout and issue a new LAYOUTGET.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

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

When I/O cannot complete due to a fatal error on the DS, ensure that we
invalidate the corresponding layout segment and return it.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If a read-write layout has an invalid mirror, then we should
mark it as invalid, and return it.
If a read-only layout has an invalid mirror, then mark it as invalid
and check if there is still at least one valid mirror before we return
it.

Note: Also fix incorrect use of pnfs_generic_mark_devid_invalid().
We really want nfs4_mark_deviceid_unavailable().
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Unlike read layouts, the writeable layout cannot fall back to using only
one of the mirrors. It need to write to all of them.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Unlike the files layout, flexfiles does not test for the NFS_DEVICEID_INVALID
flag. Instead it relies on NFS_DEVICEID_UNAVAILABLE.
Fix is to replace with nfs4_mark_deviceid_unavailable().
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Mirrors are now shared objects, so we should not be freeing them directly
inside ff_layout_free_lseg(). We should already be doing the right thing
in _ff_layout_free_lseg(), so just let it handle things.

Also ensure that ff_layout_free_mirror() frees the RPC credential if it
is set.

Fixes: 28a0d72c ("Add refcounting to struct nfs4_ff_layout_mirror")
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

If the file was fenced and/or has been deleted on the DS, then we want
to retry pNFS after a layoutreturn with error report. If the server
cannot fix the problem, then we rely on it to tell us so in the
response to the LAYOUTGET.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

According to the flexfiles protocol, the layoutreturn should specify an
array of errors in the following format:

struct ff_ioerr4 {
	offset4        ffie_offset;
	length4        ffie_length;
	stateid4       ffie_stateid;
	device_error4  ffie_errors<>;
};

This patch fixes up the code to ensure that our ffie_errors is indeed
encoded as an array (albeit with only a single entry).
Reported-by: NTom Haynes <thomas.haynes@primarydata.com>
Cc: stable@vger.kernel.org
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

In order to ensure atomicity of updates, we merge the old layout segments
into the new ones, and then invalidate the old ones.

Also ensure that we order the list of layout segments so that
RO segments are preferred over RW.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Signed-off-by: NFengguang Wu <fengguang.wu@intel.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Allow advanced users to set the layoutstats timer in order to lengthen
or shorten the period between layoutstat transmissions to the server.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Keep the full list of mirrors in the struct nfs4_ff_layout_mirror so that
they can be shared among the layout segments that use them.
Also ensure that we send out only one copy of the layoutstats per mirror.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

When we start sharing mirrors between several lsegs, we won't be able to
keep it.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We do want to share mirrors between layout segments, so add a refcount
to enable that.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We do not want to update inode attributes with DS values.

Cc: stable@vger.kernel.org # v4.0+
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Make sure that we also handle RPC level connection and protocol
negotiation errors.
Reported-by: NTom Haynes <loghyr@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We want to ensure that the stopwatches for the busy timer and the
aggregate timer are consistent. This means that they need to use
the same start/stop times.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Otherwise we break fstest case tests/read_write/mctime.t

Does files layout need the same fix as well?

Cc: stable@vger.kernel.org # v4.0+
Cc: Anna Schumaker <anna.schumaker@netapp.com>
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

We already know that pg_lseg is NULL here.
Signed-off-by: NJeff Layton <jeff.layton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Prevent a potential deadlock.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Turned out I misinterpreted the spec...

Cc: Tom Haynes <thomas.haynes@primarydata.com>
Reported-by: NJean Spector <jean@primarydata.com>
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

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

hdr->good_bytes needs to be set to the length of the request, not
zero.

Cc: stable@vger.kernel.org # 4.0+
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

This patch ensures that we record the value of 'ffl_flags' from
the layout, and then checks for the presence of the
FF_FLAGS_NO_LAYOUTCOMMIT flag before deciding whether or not to
call pnfs_set_layoutcommit().

The effect is that servers now can decide whether or not they want
the client to call layoutcommit before returning a writeable layout.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

It looks as if xchg() and cmpxchg() are not available for 64-bit integers on sparc32:

> New breakage seen in linux-next today:
>
> ERROR: "__xchg_called_with_bad_pointer" [fs/nfs/flexfilelayout/nfs_layout_flexfiles.ko] undefined!
> ERROR: "__cmpxchg_called_with_bad_pointer" [fs/nfs/flexfilelayout/nfs_layout_flexfiles.ko] undefined!
> make[2]: *** [__modpost] Error 1
> make[1]: *** [modules] Error 2

Given that mirror ktime manipulation is already under mirror->lock, let's make use of the fact.
Reported-by: NPaul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: NPeng Tao <tao.peng@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

A ds can be associated with more than one mirror, but we currently skip
setting a mirror's credentials if we find that it's already set up with
a connected client.

The upshot is that we can end up sending DS writes with MDS credentials
instead of properly setting them up. Fix nfs4_ff_layout_prepare_ds to
always verify that the mirror's credentials are set up, even when we
have a DS that's already connected.
Reported-by: NTom Haynes <thomas.haynes@primarydata.com>
Signed-off-by: NJeff Layton <jeff.layton@primarydata.com>
Cc: stable@vger.kernel.org # 4.0+
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>