A timing issue can occur where qib_mr_dereg can return -EBUSY if the
MR use count is not zero.
This can occur if the MR is de-registered while RDMA read response
packets are being progressed from the SDMA ring. The suspicion is
that the peer sent an RDMA read request, which has already been copied
across to the peer. The peer sees the completion of his request and
then communicates to the responder that the MR is not needed any
longer. The responder tries to de-register the MR, catching some
responses remaining in the SDMA ring holding the MR use count.
The code now uses a get/put paradigm to track MR use counts and
coordinates with the MR de-registration process using a completion
when the count has reached zero. A timeout on the delay is in place
to catch other EBUSY issues.
The reference count protocol is as follows:
- The return to the user counts as 1
- A reference from the lk_table or the qib_ibdev counts as 1.
- Transient I/O operations increase/decrease as necessary
A lot of code duplication has been folded into the new routines
init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit
initialization of fields to zero is now handled by kzalloc().
Also, duplicated code 'while.*num_sge' that decrements reference
counts have been consolidated in qib_put_ss().
Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This patch reorganizes the QP and devdata files to be more cache line aware.
qib_qp fields in particular are split into read-mostly, send, and receive fields.
qib_devdata fields are split into read-mostly and read/write fields
Testing has show that bidirectional tests improve by as much as 100%
with this patch.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Use a bit in wc_flags rather then a whole integer to hold the
"checksum OK" flag. By itself, this change doesn't reduce the size of
struct ib_wc on 64bit machines -- it stays on 56 bytes because of
padding. However, it will allow to add more fields in the future
without enlarging the struct. Also, it will let us have a unified
approach with future libibverbs checksum offload reporting, because a
bit flag doesn't break the library ABI.
This patch was suggested during conversation with Liran Liss
<liranl@mellanox.com>.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The following panic can occur when flushing a QP:
RIP: 0010:[<ffffffffa0168e8b>] [<ffffffffa0168e8b>] qib_send_complete+0x3b/0x190 [ib_qib]
RSP: 0018:ffff8803cdc6fc90 EFLAGS: 00010046
RAX: 0000000000000000 RBX: ffff8803d84ba000 RCX: 0000000000000000
RDX: 0000000000000005 RSI: ffffc90015a53430 RDI: ffff8803d84ba000
RBP: ffff8803cdc6fce0 R08: ffff8803cdc6fc90 R09: 0000000000000001
R10: 00000000ffffffff R11: 0000000000000000 R12: ffff8803d84ba0c0
R13: ffff8803d84ba5cc R14: 0000000000000800 R15: 0000000000000246
FS: 0000000000000000(0000) GS:ffff880036600000(0000) knlGS:0000000000000000
CS: 0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: 0000000000000034 CR3: 00000003e44f9000 CR4: 00000000000406f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process qib/0 (pid: 1350, threadinfo ffff8803cdc6e000, task ffff88042728a100)
Stack:
53544c5553455201 0000000100000005 0000000000000000 ffff8803d84ba000
0000000000000000 0000000000000000 0000000000000000 0000000000000000
0000000000000000 0000000000000001 ffff8803cdc6fd30 ffffffffa0165d7a
Call Trace:
[<ffffffffa0165d7a>] qib_make_rc_req+0x36a/0xe80 [ib_qib]
[<ffffffffa0165a10>] ? qib_make_rc_req+0x0/0xe80 [ib_qib]
[<ffffffffa01698b3>] qib_do_send+0xf3/0xb60 [ib_qib]
[<ffffffff814db757>] ? thread_return+0x4e/0x777
[<ffffffffa01697c0>] ? qib_do_send+0x0/0xb60 [ib_qib]
[<ffffffff81088bf0>] worker_thread+0x170/0x2a0
[<ffffffff8108e530>] ? autoremove_wake_function+0x0/0x40
[<ffffffff81088a80>] ? worker_thread+0x0/0x2a0
[<ffffffff8108e1c6>] kthread+0x96/0xa0
[<ffffffff8100c1ca>] child_rip+0xa/0x20
[<ffffffff8108e130>] ? kthread+0x0/0xa0
[<ffffffff8100c1c0>] ? child_rip+0x0/0x20
RIP [<ffffffffa0168e8b>] qib_send_complete+0x3b/0x190 [ib_qib]
The RC error state flush logic in qib_make_rc_req() could return all
of the acked wqes and potentially have emptied the queue. It would
then unconditionally try return a flush completion via
qib_send_complete() for an invalid wqe, or worse a valid one that is
not queued. The panic results when the completion code tries to
maintain an MR reference count for a NULL MR.
This fix modifies logic to only send one completion per
qib_make_rc_req() call and changing the completion status from
IB_WC_SUCCESS to IB_WC_WR_FLUSH_ERR as the completions progress.
The outer loop will call as many times as necessary to flush the queue.
Reviewed-by: Ram Vepa <ram.vepa@qlogic.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Review of qib_ruc_check_hdr() shows that the s_lock is not required in
the normal case. The r_lock is held in all cases, and protects the qp
fields that are read.
The s_lock will be needed to around the call to qib_migrate_qp() to
insure that the send engine sees a consistent set of fields.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
A new field is added to qib_qp called timeout_jiffies. It is
initialized upon create and modify.
The field is now used instead of a computation based on qp->timeout.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Store both the encoded and decoded MTU in the QP structure as a minor
optimization for UC/RC receive routines.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The memset for zeroing work completions had been unconditional.
This patch removes the memset and moves the zeroing into the work
completion with a more explicit field by field set. With this patch,
non-ONLY/non-LAST packets will avoid the overhead since they will not
generate a completion.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
There is a double completion associated with error handling for RC QPs.
The sequence is:
- The do_rc_ack() routine fields an RNR nack and there are 0
rnr_retries configured on the QP.
- qib_error_qp() stops the pending timer
- qib_rc_send_complete() is called from sdma_complete()
- qib_rc_send_complete() starts the timer because the msb of the psn
just completed says an ack is needed.
- a bunch of flushes occur as ipoib posts WQEs to an error'ed QP
- rc_timeout() calls qib_restart_rc()
- qib_restart_rc() calls qib_send_complete() with a
IB_WC_RETRY_EXC_ERR on a wqe that has already been completed in the
past
The fix avoids starting the timer since another packet will never
arrive.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The following panic BUG_ON occurs during qib testing:
Kernel BUG at include/linux/timer.h:82
RIP [<ffffffff881f7109>] :ib_qib:start_timer+0x73/0x89
RSP <ffffffff80425bd0>
<0>Kernel panic - not syncing: Fatal exception
<0>Dumping qib trace buffer from panic
qib_set_lid INFO: IB0:1 got a lid: 0xf8
Done dumping qib trace buffer
BUG: warning at kernel/panic.c:137/panic() (Tainted: G
The flaw is due to a missing state test when processing responses that
results in an add_timer() call when the same timer is already queued.
This code was executing in parallel with a QP destroy on another CPU
that had changed the state to reset, but the missing test caused to
response handling code to run on into the panic.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Currently on receipt of a response message (ACKs, RDMA Response,
Atomic Responses etc.) if the SDMA completion counter is not advanced
the driver delays the completion of the WQE. In most cases this is
overly pessimistic as the response (ACK) to a previously transmitted
send implies that the send is complete. Ensure that SDMA queue is
progressed appropriately before determining if a send has delayed
completions.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
See table 35 in IBA - the header order for RDMA_WRITE_ONLY_WITH_IMMEDIATE
and SEND_LAST_WITH_IMMEDIATE is different: the RDMA_WRITE_ONLY has
a RETH header before the immediate data, so we need a different code path
to extract the immediate data.
I tested this with a userspace app that does RDMA_WRITE with immediate
on a QLE7140.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When transitioning a QP to the error state, in progress RWQEs need to
be marked complete. This also involves releasing the reference count
to the memory regions referenced in the SGEs. The locking in the
receive packet processing wasn't sufficient to prevent qib_error_qp()
from modifying the r_sge state at the same time, thus leading to
kernel panics.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add a low-level IB driver for QLogic PCIe adapters.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Mike Marciniszyn