While marshaling an RPC/RDMA request, the inline_{rsize,wsize}
settings determine whether an inline request is used, or whether
read or write chunks lists are built. The current default value of
these settings is 1024. Any RPC request smaller than 1024 bytes is
sent to the NFS server completely inline.
rpcrdma_buffer_create() allocates and pre-registers a set of RPC
buffers for each transport instance, also based on the inline rsize
and wsize settings.
RPC/RDMA requests and replies are built in these buffers. However,
if an RPC/RDMA request is expected to be larger than 1024, a buffer
has to be allocated and registered for that RPC, and deregistered
and released when the RPC is complete. This is known has a
"hardway allocation."
Since the introduction of NFSv4, the size of RPC requests has become
larger, and hardway allocations are thus more frequent. Hardway
allocations are significant overhead, and they waste the existing
RPC buffers pre-allocated by rpcrdma_buffer_create().
We'd like fewer hardway allocations.
Increasing the size of the pre-registered buffers is the most direct
way to do this. However, a blanket increase of the inline thresholds
has interoperability consequences.
On my 64-bit system, rpcrdma_buffer_create() requests roughly 7000
bytes for each RPC request buffer, using kmalloc(). Due to internal
fragmentation, this wastes nearly 1200 bytes because kmalloc()
already returns an 8192-byte piece of memory for a 7000-byte
allocation request, though the extra space remains unused.
So let's round up the size of the pre-allocated buffers, and make
use of the unused space in the kmalloc'd memory.
This change reduces the amount of hardway allocated memory for an
NFSv4 general connectathon run from 1322092 to 9472 bytes (99%).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Sagi Grimberg <sagig@dev.mellanox.co.il> points out that a steady
stream of CQ events could starve other work because of the boundless
loop pooling in rpcrdma_{send,recv}_poll().
Instead of a (potentially infinite) while loop, return after
collecting a budgeted number of completions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Acked-by: Sagi Grimberg <sagig@dev.mellanox.co.il>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Change the completion handlers to grab up to 16 items per
ib_poll_cq() call. No extra ib_poll_cq() is needed if fewer than 16
items are returned.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Skip the ib_poll_cq() after re-arming, if the provider knows there
are no additional items waiting. (Have a look at commit ed23a727 for
more details).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The current CQ handler uses the ib_wc.opcode field to distinguish
between event types. However, the contents of that field are not
reliable if the completion status is not IB_WC_SUCCESS.
When an error completion occurs on a send event, the CQ handler
schedules a tasklet with something that is not a struct rpcrdma_rep.
This is never correct behavior, and sometimes it results in a panic.
To resolve this issue, split the completion queue into a send CQ and
a receive CQ. The send CQ handler now handles only struct rpcrdma_mw
wr_id's, and the receive CQ handler now handles only struct
rpcrdma_rep wr_id's.
Fix suggested by Shirley Ma <shirley.ma@oracle.com>
Reported-by: Rafael Reiter <rafael.reiter@ims.co.at>
Fixes: 5c635e09ce
BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=73211
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Klemens Senn <klemens.senn@ims.co.at>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: rpcrdma_ep_destroy() returns a value that is used
only to print a debugging message. rpcrdma_ep_destroy() already
prints debugging messages in all error cases.
Make rpcrdma_ep_destroy() return void instead.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
If the selected memory registration mode is not supported by the
underlying provider/HCA, the NFS mount command reports that there was
an invalid mount option, and fails. This is misleading.
Reporting a problem allocating memory is a lot closer to the truth.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
An audit of in-kernel RDMA providers that do not support the FRMR
memory registration shows that several of them support MTHCAFMR.
Prefer MTHCAFMR when FRMR is not supported.
If MTHCAFMR is not supported, only then choose ALLPHYSICAL.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
All kernel RDMA providers except amso1100 support either MTHCAFMR
or FRMR, both of which are faster than REGISTER. amso1100 can
continue to use ALLPHYSICAL.
The only other ULP consumer in the kernel that uses the reg_phys_mr
verb is Lustre.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The MEMWINDOWS and MEMWINDOWS_ASYNC memory registration modes were
intended as stop-gap modes before the introduction of FRMR. They
are now considered obsolete.
MEMWINDOWS_ASYNC is also considered unsafe because it can leave
client memory registered and exposed for an indeterminant time after
each I/O.
At this point, the MEMWINDOWS modes add needless complexity, so
remove them.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Clean up: This memory registration mode is slow and was never
meant for use in production environments. Remove it to reduce
implementation complexity.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
An IB provider can invoke rpcrdma_conn_func() in an IRQ context,
thus rpcrdma_conn_func() cannot be allowed to directly invoke
generic RPC functions like xprt_wake_pending_tasks().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Two memory region leaks were found during testing:
1. rpcrdma_buffer_create: While allocating RPCRDMA_FRMR's
ib_alloc_fast_reg_mr is called and then ib_alloc_fast_reg_page_list is
called. If ib_alloc_fast_reg_page_list returns an error it bails out of
the routine dropping the last ib_alloc_fast_reg_mr frmr region creating a
memory leak. Added code to dereg the last frmr if
ib_alloc_fast_reg_page_list fails.
2. rpcrdma_buffer_destroy: While cleaning up, the routine will only free
the MR's on the rb_mws list if there are rb_send_bufs present. However, in
rpcrdma_buffer_create while the rb_mws list is being built if one of the MR
allocation requests fail after some MR's have been allocated on the rb_mws
list the routine never gets to create any rb_send_bufs but instead jumps to
the rpcrdma_buffer_destroy routine which will never free the MR's on rb_mws
list because the rb_send_bufs were never created. This leaks all the MR's
on the rb_mws list that were created prior to one of the MR allocations
failing.
Issue(2) was seen during testing. Our adapter had a finite number of MR's
available and we created enough connections to where we saw an MR
allocation failure on our Nth NFS connection request. After the kernel
cleaned up the resources it had allocated for the Nth connection we noticed
that FMR's had been leaked due to the coding error described above.
Issue(1) was seen during a code review while debugging issue(2).
Signed-off-by: Allen Andrews <allen.andrews@emulex.com>
Reviewed-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Some rdma devices don't support a fast register page list depth of
at least RPCRDMA_MAX_DATA_SEGS. So xprtrdma needs to chunk its fast
register regions according to the minimum of the device max supported
depth or RPCRDMA_MAX_DATA_SEGS.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The object and block layouts already exist in their own
subdirectories. This patch completes the set!
Note that as a layout denotes nfs4 already, I stripped
that prefix out of the file names.
Signed-off-by: Tom Haynes <Thomas.Haynes@primarydata.com>
Acked-by: Jeff Layton <jlayton@poochiereds.net>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Those flags are obsolete and checking them can incorrectly cause
remount operations to fail.
Signed-off-by: Scott Mayhew <smayhew@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Place the call to resend the failed GETATTR under the error handler so that
when appropriate, the GETATTR is retried more than once.
The server can fail the GETATTR op in the OPEN compound with a recoverable
error such as NFS4ERR_DELAY. In the case of an O_EXCL open, the server has
created the file, so a retrans of the OPEN call will fail with NFS4ERR_EXIST.
Signed-off-by: Andy Adamson <andros@netapp.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
We cannot allow nfs_page_group_lock to use TASK_KILLABLE here, since
the loop would cause a busy wait if somebody kills the task.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
nfs_read_completion relied on the fact that there was a 1:1 mapping
of page to nfs_request, but this has now changed.
Regions not covered by a request have already been zeroed elsewhere.
Signed-off-by: Weston Andros Adamson <dros@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Use the new pg_test interface to adjust requests to fit in the current
stripe / segment.
Signed-off-by: Weston Andros Adamson <dros@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Chuck Lever
Allen Andrews
Steve Wise
Tom Haynes
Trond Myklebust
Scott Mayhew
Andy Adamson
David Rientjes
Weston Andros Adamson