We discovered a problem when running IPoIB applications on multiple
CPUs on an Altix system. Many messages such as:
ib_mthca 0002:01:00.0: SQ 000014 full (19941644 head, 19941707 tail, 64 max, 0 nreq)
appear in syslog, and the driver wedges up.
Apparently this is because writes to the doorbells from different CPUs
reach the device out of order. The following patch adds mmiowb() calls
after doorbell rings to ensure the doorbell writes are ordered.
Signed-off-by: Arthur Kepner <akepner@sgi.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
All HCAs (not just mem-free) need a spare SRQ entry, so bump srq->max
by 1 in all cases.
Noted by Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Pass a struct ib_udata to the low-level driver's ->modify_srq() and
->modify_qp() methods, so that it can get to the device-specific data
passed in by the userspace driver.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Mem-free HCAs always keep one spare SRQ WQE, so the SRQ limit cannot
be set beyond srq->max - 1.
Signed-off-by: Dotan Barak <dotanb@mellanox.co.il>
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Documentation/infiniband/core_locking.txt says:
All of the methods in struct ib_device exported by a low-level
driver must be fully reentrant. The low-level driver is required to
perform all synchronization necessary to maintain consistency, even
if multiple function calls using the same object are run
simultaneously.
However, mthca's modify_qp, modify_srq and resize_cq methods are
currently not reentrant. Add a mutex to the QP, SRQ and CQ structures
so that these calls can be properly serialized.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
If we post a list of length exactly a multiple of 256, nreq in
doorbell gets set to 256 which is wrong: it should be encoded by 0.
This is because we only zero it out on the next WR, which may not be
there. The solution is to ring the doorbell after posting a WQE, not
before posting the next one.
This is the same bug that we just fixed for QPs with non-shared RQ.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix races in in destroying various objects. If a destroy routine
waits for an object to become free by doing
wait_event(&obj->wait, !atomic_read(&obj->refcount));
/* now clean up and destroy the object */
and another place drops a reference to the object by doing
if (atomic_dec_and_test(&obj->refcount))
wake_up(&obj->wait);
then this is susceptible to a race where the wait_event() and final
freeing of the object occur between the atomic_dec_and_test() and the
wake_up(). And this is a use-after-free, since wake_up() will be
called on part of the already-freed object.
Fix this in mthca by replacing the atomic_t refcounts with plain old
integers protected by a spinlock. This makes it possible to do the
decrement of the reference count and the wake_up() so that it appears
as a single atomic operation to the code waiting on the wait queue.
While touching this code, also simplify mthca_cq_clean(): the CQ being
cleaned cannot go away, because it still has a QP attached to it. So
there's no reason to be paranoid and look up the CQ by number; it's
perfectly safe to use the pointer that the callers already have.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The driver allocates SRQ WQEs size with a power of 2 size both for
Tavor and for memfree. For Tavor, however, the hardware only requires
the WQE size to be a multiple of 16, not a power of 2, and the max
number of scatter-gather allowed is reported accordingly by the
firmware (and this is the value currently returned by
ib_query_device() and ibv_query_device()).
If the max number of scatter/gather entries reported by the FW is used
when creating an SRQ, the creation will fail for Tavor, since the
required WQE size will be increased to the next power of 2, which
turns out to be larger than the device permitted max WQE size (which
is not a power of 2).
This patch reduces the reported SRQ max wqe size so that it can be used
successfully in creating an SRQ on Tavor HCAs.
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Quite a few cleanup functions in mthca were marked as __devexit.
However, they could also be called from error paths during
initialization, so they cannot be marked that way. Just delete all of
the incorrect annotations.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The previous patch for Tavor broke MemFree logic.
The driver should perform limit check only for Tavor. For MemFree,
the check is incorrect, since ds (WQE stride) is always a power-of-2
(although the max_desc_size may not be).
In Tavor, however, WQE stride and desc_size are the same, and are not
necessarily power-of-2. The check was really for the WQE stride (and
it Tavor, we use max_desc_size for the stride).
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When setting the shared receive queue (SRQ) watermark in a modify SRQ
operation, make sure that the supplied value is not larger than the
full size of the SRQ.
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Guarantee the calculated work queue entry size does not exceed the max
allowable WQE size when creating an SRQ. This is a problem with Arbel
in Tavor-compatibility mode because the current WQE size computation
method rounds up to next power of 2.
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix endianness handling of srq_limit: it is big-endian in the context
structure, so we need to swab it before returning it.
Also add support for srq_limit query for Tavor (non-MemFree) HCAs.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
MemFree devices need to reserve one shared receive queue (SRQ) work
request for internal use, so the capacity returned from the create_srq
and query_srq methods should be srq->max - 1.
Signed-off-by: Dotan Barak <dotanb@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Have mthca's create_srq method return the actual capacity of the SRQ
that gets created. Also update comments in <rdma/ib_verbs.h> to
clarify that this is what is expected from ib_create_srq().
Signed-off-by: Dotan Barak <dotanb@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Thinko: 64 bytes is the minimum SRQ WQE size (not the maximum).
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
In Tavor mode, when posting a long list of receive work requests, a
doorbell must be rung every 256 requests. Add code to do this when
required.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix more include file problems that surfaced since I submitted the previous
fix-missing-includes.patch. This should now allow not to include sched.h
from module.h, which is done by a followup patch.
Signed-off-by: Tim Schmielau <tim@physik3.uni-rostock.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Fix wqe_to_link() to use a structure field that we know is definitely
always unused for receive work requests, so that it really avoids the
free list corruption bug that the comment claims it does.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Check the sizes of CQs, QPs and SRQs when creating objects, and fail
instead of creating too-big queues. Also return real limits instead
of just plausible-sounding values from mthca_query_device().
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The hardware relies on us keeping one extra work request that never
gets used in SRQs. Add checks to the SRQ work request posting
functions so that they fail when someone is about to use up that extra
work request, rather than when someone uses the very last work request.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Our hardware supports generating an event when the number of receives
posted to a shared receive queue (SRQ) falls below a user-specified
limit. Implement mthca_modify_srq() to arm the limit, and add code to
handle dispatching SRQ events when they occur.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Userspace SRQs don't have a buffer allocated for them in the kernel, so
it doesn't make sense to set srq->last during initialization. In fact,
this can crash trying to follow a nonexistent buffer pointer.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The error handling paths in mthca_tavor_post_srq_recv() and
mthca_arbel_post_srq_recv() are quite bogus, the result of a
screwed up merge. Fix them so they work as intended.
Pointed out by Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix posting first WQE for mem-free HCAs: we need to link to previous
WQE even in that case. While we're at it, simplify code for
Tavor-mode HCAs. We don't really need the conditional test there
either; we can similarly always link to the previous WQE.
Based on Michael S. Tsirkin's analogous fix for userspace libmthca.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
