Most NFS RPCs place their large payload argument at the end of the
RPC header (eg, NFSv3 WRITE). For NFSv3 WRITE and SYMLINK, RPC/RDMA
sends the complete RPC header inline, and the payload argument in
the read list. Data in the read list is the last part of the XDR
stream.

One important case is not like this, however. NFSv4 COMPOUND is a
counted array of operations. A WRITE operation, with its large data
payload, can appear in the middle of the compound's operations
array. Thus NFSv4 WRITE compounds can have header content after the
WRITE payload.

The Linux client, for example, performs an NFSv4 WRITE like this:

  { PUTFH, WRITE, GETATTR }

Though RFC 5667 is not precise about this, the proper way to convey
this compound is to place the GETATTR inline, _after_ the front of
the RPC header. The receiver inserts the read list payload into the
XDR stream after the initial WRITE arguments, and before the GETATTR
operation, thanks to the value of the read list "position" field.

The Linux client currently sends the GETATTR at the end of the
RPC/RDMA read list, which is incorrect. It will be corrected in the
future.

The Linux server currently rejects NFSv4 compounds with inline
content after the read list. For the above NFSv4 WRITE compound, the
NFS compound header indicates there are three operations, but the
server finds nonsense when it looks in the XDR stream for the third
operation, and the compound fails with OP_ILLEGAL.

Move trailing inline content to the end of the XDR buffer's page
list. This presents incoming NFSv4 WRITE compounds to NFSD in the
same way the socket transport does.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

This is a pre-requisite for a subsequent patch.

Read list XDR round-up needs to be done _before_ additional inline
content is copied to the end of the XDR buffer's page list. Move
the logic added by commit e560e3b5 ("svcrdma: Add zero padding
if the client doesn't send it").
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Currently the Linux server can not decode RDMA_NOMSG type requests.
Operations whose length exceeds the fixed size of RDMA SEND buffers,
like large NFSv4 CREATE(NF4LNK) operations, must be conveyed via
RDMA_NOMSG.

For an RDMA_MSG type request, the client sends the RPC/RDMA, RPC
headers, and some or all of the NFS arguments via RDMA SEND.

For an RDMA_NOMSG type request, the client sends just the RPC/RDMA
header via RDMA SEND. The request's read list contains elements for
the entire RPC message, including the RPC header.

NFSD expects the RPC/RMDA header and RPC header to be contiguous in
page zero of the XDR buffer. Add logic in the RDMA READ path to make
the read list contents land where the server prefers, when the
incoming message is a type RDMA_NOMSG message.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

An RPC/RDMA client may send large RPC arguments via a read
list. This is a list of scatter/gather elements which convey
RPC call arguments too large to fit in a small RDMA SEND.

Each entry in the read list has a "position" field, whose value is
the byte offset in the XDR stream where the data in that entry is to
be inserted. Entries which share the same "position" value make up
the same RPC argument. The receiver inserts entries with the same
position field value in list order into the XDR stream.

Currently the Linux NFS/RDMA server cannot handle receiving read
chunks in more than one position, mostly because no current client
sends read lists with elements in more than one position. As a
sanity check, ensure that all received chunks have the same
"rc_position."
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

The RDMA reader function doesn't change once an svcxprt_rdma is
instantiated. Instead of checking sc_devcap during every incoming
RPC, set the reader function once when the connection is accepted.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

xdr_start() can return the wrong rmsgp address if an assumption
about how the xdr_buf was constructed changes.  When it gets it
wrong, the client receives a reply that has gibberish in the
RPC/RDMA header, preventing it from matching a waiting RPC request.

Instead, make (and document) just one assumption: that the RDMA
header for the client's RPC call is at the start of the first page
in rq_pages.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Current convention is to avoid using BUG_ON() in places where an
oops could cause complete system failure.

Replace BUG_ON() call sites in svcrdma with an assertion error
message and allow execution to continue safely.

Some BUG_ON() calls are removed because they have never fired in
production (that we are aware of).

Some WARN_ON() calls are also replaced where a back trace is not
helpful; e.g., in a workqueue task.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

The byte_count argument is not used, and the function is called
only from one place.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Nit: remove an unused variable to squelch a compiler warning.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Nit: Fix inconsistent white space in dprintk messages.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

A struct xdr_stream at a page boundary might point to the end of one
page or the beginning of the next, but xdr_truncate_encode isn't
prepared to handle the former.

This can cause corruption of NFSv4 READDIR replies in the case that a
readdir entry that would have exceeded the client's dircount/maxcount
limit would have ended exactly on a 4k page boundary.  You're more
likely to hit this case on large directories.

Other xdr_truncate_encode callers are probably also affected.
Reported-by: NHolger Hoffstätte <holger.hoffstaette@googlemail.com>
Tested-by: NHolger Hoffstätte <holger.hoffstaette@googlemail.com>
Fixes: 3e19ce76 "rpc: xdr_truncate_encode"
Cc: stable@vger.kernel.org
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

There is nice kernel helper to escape a given strings by provided rules. Let's
use it instead of custom approach.
Signed-off-by: NAndy Shevchenko <andriy.shevchenko@linux.intel.com>
[bfields@redhat.com: fix length calculation]
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

...move the WARN_ON_ONCE inside the following if block since they use
the same condition.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

These were useful when I was tracking down a race condition between
svc_xprt_do_enqueue and svc_get_next_xprt.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Testing has shown that the pool->sp_lock can be a bottleneck on a busy
server. Every time data is received on a socket, the server must take
that lock in order to dequeue a thread from the sp_threads list.

Address this problem by eliminating the sp_threads list (which contains
threads that are currently idle) and replacing it with a RQ_BUSY flag in
svc_rqst. This allows us to walk the sp_all_threads list under the
rcu_read_lock and find a suitable thread for the xprt by doing a
test_and_set_bit.

Note that we do still have a potential atomicity problem however with
this approach.  We don't want svc_xprt_do_enqueue to set the
rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned
zero (which indicates that the thread was idle). But, by the time we
check that, the bit could be flipped by a waking thread.

To address this, we acquire a new per-rqst spinlock (rq_lock) and take
that before doing the test_and_set_bit. If that returns false, then we
can set rq_xprt and drop the spinlock. Then, when the thread wakes up,
it must set the bit under the same spinlock and can trust that if it was
already set then the rq_xprt is also properly set.

With this scheme, the case where we have an idle thread no longer needs
to take the highly contended pool->sp_lock at all, and that removes the
bottleneck.

That still leaves one issue: What of the case where we walk the whole
sp_all_threads list and don't find an idle thread? Because the search is
lockess, it's possible for the queueing to race with a thread that is
going to sleep. To address that, we queue the xprt and then search again.

If we find an idle thread at that point, we can't attach the xprt to it
directly since that might race with a different thread waking up and
finding it.  All we can do is wake the idle thread back up and let it
attempt to find the now-queued xprt.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Tested-by: NChris Worley <chris.worley@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

In a later patch, we'll be removing some spinlocking around the socket
and thread queueing code in order to fix some contention problems. At
that point, the stats counters will no longer be protected by the
sp_lock.

Change the counters to atomic_long_t fields, except for the
"sockets_queued" counter which will still be manipulated under a
spinlock.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Tested-by: NChris Worley <chris.worley@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

...also make the manipulation of sp_all_threads list use RCU-friendly
functions.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Tested-by: NChris Worley <chris.worley@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Currently all svc_create callers pass in NULL for the shutdown parm,
which then gets fixed up to be svc_rpcb_cleanup if the service uses
rpcbind.

Simplify this by instead having the the only caller that requires it
(lockd) pass in svc_rpcb_cleanup and get rid of the special casing.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

The way that svc_wake_up works is a bit inefficient. It walks all of the
available pools for a service and either wakes up a task in each one or
sets the SP_TASK_PENDING flag in each one.

When svc_wake_up is called, there is no need to wake up more than one
thread to do this work. In practice, only lockd currently uses this
function and it's single threaded anyway. Thus, this just boils down to
doing a wake up of a thread in pool 0 or setting a single flag.

Eliminate the for loop in this function and change it to just operate on
pool 0. Also update the comments that sit above it and get rid of some
code that has been commented out for years now.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

In a later patch, we'll want to be able to handle this flag without
holding the sp_lock. Change this field to an unsigned long flags
field, and declare a new flag in it that can be managed with atomic
bitops.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

In a later patch, we're going to need some atomic bit flags. Since that
field will need to be an unsigned long, we mitigate that space
consumption by migrating some other bitflags to the new field. Start
with the rq_secure flag.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Currently, it leaks when the allocation fails.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

All it does is indicate whether a xprt has already been deleted from
a list or not, which is unnecessary since we use list_del_init and it's
always set and checked under the sv_lock anyway.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Add a new directory heirarchy under the debugfs sunrpc/ directory:

    sunrpc/
        rpc_xprt/
            <xprt id>/

Within that directory, we can put files that give info about the
xprts. We do have the (minor) problem that there is no succinct,
unique identifier for rpc_xprts. So we generate them synthetically
with a static atomic_t counter.

For now, this directory just holds an "info" file, but we may add
other files to it in the future.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

It's possible to get a dump of the RPC task queue by writing a value to
/proc/sys/sunrpc/rpc_debug. If you write any value to that file, you get
a dump of the RPC client task list into the log buffer. This is a rather
inconvenient interface however, and makes it hard to get immediate info
about the task queue.

Add a new directory hierarchy under debugfs:

    sunrpc/
        rpc_clnt/
            <clientid>/

Within each clientid directory we create a new "tasks" file that will
dump info similar to what shows up in the log buffer, but with a few
small differences -- we avoid printing raw kernel addresses in favor of
symbolic names and the XID is also displayed.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Occasionally mountstats reports a negative retransmission rate.
Ensure that two RPCs completing concurrently don't confuse the sums
in the transport's op_metrics array.

Since pNFS filelayout can invoke rpc_count_iostats() on another
transport from xprt_release(), we can't rely on simply holding the
transport_lock in xprt_release(). There's nothing for it but hard
serialization. One spin lock per RPC operation should make this as
painless as it can be.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

An async error upcall is a hard error, and should be reported in
the system log.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The Linux NFS/RDMA server used to reject NFSv3 WRITE requests when
pad optimization was enabled. That bug was fixed by commit
e560e3b5 ("svcrdma: Add zero padding if the client doesn't send
it").

We can now enable pad optimization on the client, which helps
performance and is supported now by both Linux and Solaris servers.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Currently rpcrdma_flush_cqs() attempts to avoid code duplication,
and simply invokes rpcrdma_recvcq_upcall and rpcrdma_sendcq_upcall.

1. rpcrdma_flush_cqs() can run concurrently with provider upcalls.
   Both flush_cqs() and the upcalls were invoking ib_poll_cq() in
   different threads using the same wc buffers (ep->rep_recv_wcs
   and ep->rep_send_wcs), added by commit 1c00dd07 ("xprtrmda:
   Reduce calls to ib_poll_cq() in completion handlers").

   During transport disconnect processing, this sometimes resulted
   in the same reply getting added to the rpcrdma_tasklets_g list
   more than once, which corrupted the list.

2. The upcall functions drain only a limited number of CQEs,
   thanks to the poll budget added by commit 8301a2c0
   ("xprtrdma: Limit work done by completion handler").

Fixes: a7bc211a ("xprtrdma: On disconnect, don't ignore ... ")
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=276Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Restore the separate function that schedules the reply handling
tasklet. I need to call it from two different paths.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

When using RPCRDMA_MTHCAFMR memory registration, after a few
transport disconnect / reconnect cycles, ib_map_phys_fmr() starts to
return EINVAL because the provider has exhausted its map pool.

Make sure that all FMRs are unmapped during transport disconnect,
and that ->send_request remarshals them during an RPC retransmit.
This resets the transport's MRs to ensure that none are leaked
during a disconnect.
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

Recent work made FRMR registration and invalidation completions
unsignaled. This greatly reduces the adapter interrupt rate.

Every so often, however, a posted send Work Request is allowed to
signal. Otherwise, the provider's Work Queue will wrap and the
workload will hang.

The number of Work Requests that are allowed to remain unsignaled is
determined by the value of req_cqinit. Currently, this is set to the
size of the send Work Queue divided by two, minus 1.

For FRMR, the send Work Queue is the maximum number of concurrent
RPCs (currently 32) times the maximum number of Work Requests an
RPC might use (currently 7, though some adapters may need more).

For mlx4, this is 224 entries. This leaves completion signaling
disabled for 111 send Work Requests.

Some providers hold back dispatching Work Requests until a CQE is
generated.  If completions are disabled, then no CQEs are generated
for quite some time, and that can stall the Work Queue.

I've seen this occur running xfstests generic/113 over NFSv4, where
eventually, posting a FAST_REG_MR Work Request fails with -ENOMEM
because the Work Queue has overflowed. The connection is dropped
and re-established.

Cap the rep_cqinit setting so completions are not left turned off
for too long.

BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=269Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

The RPC/RDMA send_request method and the chunk registration code
expects an errno from the registration function. This allows
the upper layers to distinguish between a recoverable failure
(for example, temporary memory exhaustion) and a hard failure
(for example, a bug in the registration logic).
Signed-off-by: NChuck Lever <chuck.lever@oracle.com>
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

It's always set to the same value as CONFIG_TRACEPOINTS, so we can just
use that instead.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

It's always set to whatever CONFIG_SUNRPC_DEBUG is, so just use that.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Add tracepoints inside the main loop on xs_tcp_data_recv that allow
us to keep an eye on what's happening during each phase of it.
Signed-off-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>