Do an extra check to see if in-memory tail changed while processing packets,
and if so, going back through the loop again (but only once per call to
ipath_kreceive()). In practice, this seems to be enough to guarantee that if
we crossed the clearing of an interrupt at start of ipath_intr with a
scheduled tail register update, that we'll process the "extra" packet that
lost the interrupt because we cleared it just as it was about to arrive.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The problem was that I was updating the head register multiple times in the
rcvhdrq processing loop, and setting the counter on each update. Since that
meant that the tail register was ahead of head for all but the last update, we
would get extra interrupts. The fix was to not write the counter value except
on the last update.
I also changed to update rcvhdrhead and rcvegrindexhead at most every 16
packets, if there were lots of packets in the queue (and of course, on the
last packet, regardless).
I also made some small cleanups while debugging this.
With these changes, xeon/monty typically sees two openib packets per interrupt
on sdp and ipoib, opteron/monty is about 1.25 pkts/intr.
I'm seeing about 3800 Mbit/s monty/xeon, and 5000-5100 opteron/monty with
netperf sdp. Netpipe doesn't show as good as that, peaking at about 4400 on
opteron/monty sdp. Plain ipoib xeon is about 2100+ netperf, opteron 2900+, at
128KB
Signed-off-by: olson@eng-12.pathscale.com
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Also count the number of interrupts where that works (fastrcvint). On any
interrupt where the port0 head and tail registers are not equal, just call the
ipath_kreceive code without reading the interrupt status, thus saving the
approximately 0.25usec processor stall waiting for the read to return. If any
other interrupt bits are set, or head==tail, take the normal path, but that
has been reordered to handle read ahead of pioavail. Also no longer call
ipath_kreceive() from ipath_qcheck(), because that just seems to make things
worse, and isn't really buying us anything, these days.
Also no longer loop in ipath_kreceive(); better to not hold things off too
long (I saw many cases where we would loop 4-8 times, and handle thousands (up
to 3500) in a single call).
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Made in-memory rcvhdrq tail update be in dma_alloc'ed memory, not random user
or special kernel (needed for ppc, also "just the right thing to do").
Some cleanups to make unexpected link transitions less likely to produce
complaints about packet errors, and also to not leave SMA packets stuck and
unable to go out.
A few other random debug and comment cleanups.
Always init rcvhdrq head/tail registers to 0, to avoid race conditions (should
have been that way some time ago).
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is not a DMA target, so no need to use dma_alloc_coherent on it.
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This helps us to survive better when memory is fragmented.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
These limits are somewhat artificial in that we don't actually have any
device limits. However, the verbs layer expects that such limits exist
and are enforced, so we make up arbitrary (but sensible) limits.
Signed-off-by: Robert Walsh <robert.walsh@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
There is no longer a /dev/ipath_diag file; instead, there's
/dev/ipath_diag0, 1, etc.
It's still not possible to have diags run on more than one unit at a time,
but that's easy to fix at some point.
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is needed if we wish to change the size of the resource structures.
Based on an original patch from Vivek Goyal <vgoyal@in.ibm.com>
Cc: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Extend the get_sb() filesystem operation to take an extra argument that
permits the VFS to pass in the target vfsmount that defines the mountpoint.
The filesystem is then required to manually set the superblock and root dentry
pointers. For most filesystems, this should be done with simple_set_mnt()
which will set the superblock pointer and then set the root dentry to the
superblock's s_root (as per the old default behaviour).
The get_sb() op now returns an integer as there's now no need to return the
superblock pointer.
This patch permits a superblock to be implicitly shared amongst several mount
points, such as can be done with NFS to avoid potential inode aliasing. In
such a case, simple_set_mnt() would not be called, and instead the mnt_root
and mnt_sb would be set directly.
The patch also makes the following changes:
(*) the get_sb_*() convenience functions in the core kernel now take a vfsmount
pointer argument and return an integer, so most filesystems have to change
very little.
(*) If one of the convenience function is not used, then get_sb() should
normally call simple_set_mnt() to instantiate the vfsmount. This will
always return 0, and so can be tail-called from get_sb().
(*) generic_shutdown_super() now calls shrink_dcache_sb() to clean up the
dcache upon superblock destruction rather than shrink_dcache_anon().
This is required because the superblock may now have multiple trees that
aren't actually bound to s_root, but that still need to be cleaned up. The
currently called functions assume that the whole tree is rooted at s_root,
and that anonymous dentries are not the roots of trees which results in
dentries being left unculled.
However, with the way NFS superblock sharing are currently set to be
implemented, these assumptions are violated: the root of the filesystem is
simply a dummy dentry and inode (the real inode for '/' may well be
inaccessible), and all the vfsmounts are rooted on anonymous[*] dentries
with child trees.
[*] Anonymous until discovered from another tree.
(*) The documentation has been adjusted, including the additional bit of
changing ext2_* into foo_* in the documentation.
[akpm@osdl.org: convert ipath_fs, do other stuff]
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Nathan Scott <nathans@sgi.com>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
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>
Some error paths after the mthca_alloc_mailbox() call in mthca_modify_qp()
just do a "return -EINVAL" without freeing the mailbox. Convert these
returns to "goto out" to avoid leaking the mailbox storage.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Report the true max_map_per_fmr value from mthca_query_device(),
taking into account the change in FMR remapping introduced by the
Sinai performance optimization.
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Change the mthca snoop of MADs that set PortInfo to check if the SM
has set the client reregister bit, and if it has, generate a client
reregister event. If the bit is not set, just generate a LID change
event as usual.
Signed-off-by: Leonid Arsh <leonida@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Move ipath's struct port_info into <rdma/ib_smi.h>, so that it can be
used by mthca to implement client reregister support.
Remove the __attribute__((packed)) because all the members of the struct
are naturally aligned anyway.
Signed-off-by: Leonid Arsh <leonida@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The kernel has had wait_for_completion_timeout() for a long time now.
mthca should use it to handle FW commands timing out, instead of
implementing the same thing in a much more complicated way by using
wait_for_completion() along with a timer that does complete().
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Memfree firmware is in rare cases reporting WQE index == base - 1 in
receive completion with error, instead of (rq size - 1); base is 0 in
mthca. Here is a patch to avoid kernel crash and report a correct WR
id in this case.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
mthca does not restore the following PCI-X/PCI Express registers after reset:
PCI-X device: PCI-X command register
PCI-X bridge: upstream and downstream split transaction registers
PCI Express : PCI Express device control and link control registers
This causes instability and/or bad performance on systems where one of
these registers is set to a non-default value by BIOS.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
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>
At this point, the core QP structure hasn't been initialized, so what's
in there isn't valid. Get the same information elsewhere.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The problem was that node A's sending thread, which handles sending RDMA
read response data, would write the trigger word, the last packet would
be sent, node B would send a new RDMA read request, node A's interrupt
handler would initialize s_rdma_sge, then node A's sending thread would
update s_rdma_sge. This didn't happen very often naturally but was more
frequent with 1 byte RDMA reads. Rather than adding more locking or
increasing the QP structure size and copying sge data, I modified the
copy routine to update the pointers before writing the trigger word to
avoid the update race.
Signed-off-by: Ralph Campbell <ralphc@pathscale.com>
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fixed so it works on the PE-800. It had not previously been updated to
match PE-800 receive interrupt differences from HT-400.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This is required for even semi-decent performance on OpenIB.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix NULL deref due to pcidev being clobbered before dd->ipath_f_cleanup()
was called.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix the interface version that gets exported to userspace.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Make sure modify_qp won't modify the QP if any of the changes failed.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The local loopback path for RC can lock the rkey table lock without
blocking interrupts. The receive interrupt path can then call
ipath_rkey_ok() and deadlock. Remove the redundant lock.
Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
If we post a list of length 256 exactly, 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.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Setting fw_cmd_doorbell allows FW command to be queued using posted
writes instead of requiring polling on a "go" bit, so it should be a
performance boost. However, the option causes problems with at least
some device/firmware combinations, so set the default to 0 until we
understand what's going on better.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Bryan O'Sullivan