Using CX-3 virtual functions, either from a bare-metal machine or
pass-through from a VM, MAD packets are proxied through the PF driver.
Since the VF drivers have separate name spaces for MAD Transaction Ids
(TIDs), the PF driver has to re-map the TIDs and keep the book keeping
in a cache.
Following the RDMA Connection Manager (CM) protocol, it is clear when
an entry has to evicted form the cache. But life is not perfect,
remote peers may die or be rebooted. Hence, it's a timeout to wipe out
a cache entry, when the PF driver assumes the remote peer has gone.
During workloads where a high number of QPs are destroyed concurrently,
excessive amount of CM DREQ retries has been observed
The problem can be demonstrated in a bare-metal environment, where two
nodes have instantiated 8 VFs each. This using dual ported HCAs, so we
have 16 vPorts per physical server.
64 processes are associated with each vPort and creates and destroys
one QP for each of the remote 64 processes. That is, 1024 QPs per
vPort, all in all 16K QPs. The QPs are created/destroyed using the
CM.
When tearing down these 16K QPs, excessive CM DREQ retries (and
duplicates) are observed. With some cat/paste/awk wizardry on the
infiniband_cm sysfs, we observe as sum of the 16 vPorts on one of the
nodes:
cm_rx_duplicates:
dreq 2102
cm_rx_msgs:
drep 1989
dreq 6195
rep 3968
req 4224
rtu 4224
cm_tx_msgs:
drep 4093
dreq 27568
rep 4224
req 3968
rtu 3968
cm_tx_retries:
dreq 23469
Note that the active/passive side is equally distributed between the
two nodes.
Enabling pr_debug in cm.c gives tons of:
[171778.814239] <mlx4_ib> mlx4_ib_multiplex_cm_handler: id{slave:
1,sl_cm_id: 0xd393089f} is NULL!
By increasing the CM_CLEANUP_CACHE_TIMEOUT from 5 to 30 seconds, the
tear-down phase of the application is reduced from approximately 90 to
50 seconds. Retries/duplicates are also significantly reduced:
cm_rx_duplicates:
dreq 2460
[]
cm_tx_retries:
dreq 3010
req 47
Increasing the timeout further didn't help, as these duplicates and
retries stems from a too short CMA timeout, which was 20 (~4 seconds)
on the systems. By increasing the CMA timeout to 22 (~17 seconds), the
numbers fell down to about 10 for both of them.
Adjustment of the CMA timeout is not part of this commit.
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Acked-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
CM REQs cannot be successfully retried, because a new pv_cm_id is
created for each request, without checking if one already exists.
By checking if an id exists before creating one, the bug is fixed.
This bug can be provoked by running an RDMA CM user-land application,
but inserting a five seconds delay before the rdma_accept() call on
the passive side. This delay is larger than the default CMA timeout,
and triggers a retry from the active side. The retried REQ will use
another pv_cm_id (the cm_id on the wire). This confuses the CM
protocol and two REJs are sent from the passive side.
Here is an excerpt from ibdump running without the patch:
3.285092 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.382711 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.382861 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject
7.387644 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject
and here is the same with bug fix applied:
3.251010 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.349387 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
8.258443 LID: 4 -> LID: 4 SDP 290 CM: ConnectReply(SDP Hello)
8.259890 LID: 4 -> LID: 4 InfiniBand 290 CM: ReadyToUse
Suggested-by: Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com>
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Reported-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Tested-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Acked-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The last two actual parameters when calling id_map_find_by_sl_id()
from id_map_get() are swapped. However, the same formal parameters to
id_map_get() have them swapped as well, inverting the effect of the
first error.
This commit improves readability, but makes no functional change to
the code.
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Change and simplify the code to match the variable name. This commit
improves readability but makes no functional change to the code.
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Suggested-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The prints after [k|v][m|z|c]alloc() functions are not needed,
because in case of failure, allocator will print their internal
error prints anyway.
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
If a GUID is not found, the 64-bit GUID printed in the message log
warning should converted to host-endian order for printing.
Found by Doug Ledford and Hal Rosenstock. Fix suggested by Hal.
Signed-off-by: Hal Rosenstock <hal@dev.mellanox.co.il>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
* Handle CM_SIDR_REQ_ATTR_ID and CM_SIDR_REP_ATTR_ID
in multiplex_cm_handler and demux_cm_handler.
* Handle Service ID Resolution messages and REQ messages
separately, for their formats are different.
Signed-off-by: Shani Michaeli <shanim@mellanox.com>
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This requires the following modifications:
1. Fix build_mlx4_header to properly fill in the ETH fields
2. Adjust mux and demux QP1 flow to support RoCE.
This commit still assumes only one GID per slave for RoCE.
The commit enabling multiple GIDs is a subsequent commit, and
is done separately because of its complexity.
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 6a9200603d ("IB/mlx4: convert to idr_alloc()") forgot to remove
idr_pre_get() call in mlx4_ib_cm_paravirt_init(). It's unnecessary and
idr_pre_get() will soon be deprecated. Remove it.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jack Morgenstein <jackm@dev.mellanox.co.il>
Cc: Or Gerlitz <ogerlitz@mellanox.com>
Cc: Roland Dreier <roland@purestorage.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
MAX_IDR_MASK is another weirdness in the idr interface. As idr covers
whole positive integer range, it's defined as 0x7fffffff or INT_MAX.
Its usage in idr_find(), idr_replace() and idr_remove() is bizarre.
They basically mask off the sign bit and operate on the rest, so if
the caller, by accident, passes in a negative number, the sign bit
will be masked off and the remaining part will be used as if that was
the input, which is worse than crashing.
The constant is visible in idr.h and there are several users in the
kernel.
* drivers/i2c/i2c-core.c:i2c_add_numbered_adapter()
Basically used to test if adap->nr is a negative number which isn't
-1 and returns -EINVAL if so. idr_alloc() already has negative
@start checking (w/ WARN_ON_ONCE), so this can go away.
* drivers/infiniband/core/cm.c:cm_alloc_id()
drivers/infiniband/hw/mlx4/cm.c:id_map_alloc()
Used to wrap cyclic @start. Can be replaced with max(next, 0).
Note that this type of cyclic allocation using idr is buggy. These
are prone to spurious -ENOSPC failure after the first wraparound.
* fs/super.c:get_anon_bdev()
The ID allocated from ida is masked off before being tested whether
it's inside valid range. ida allocated ID can never be a negative
number and the masking is unnecessary.
Update idr_*() functions to fail with -EINVAL when negative @id is
specified and update other MAX_IDR_MASK users as described above.
This leaves MAX_IDR_MASK without any user, remove it and relocate
other MAX_IDR_* constants to lib/idr.c.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jean Delvare <khali@linux-fr.org>
Cc: Roland Dreier <roland@kernel.org>
Cc: Sean Hefty <sean.hefty@intel.com>
Cc: Hal Rosenstock <hal.rosenstock@gmail.com>
Cc: "Marciniszyn, Mike" <mike.marciniszyn@intel.com>
Cc: Jack Morgenstein <jackm@dev.mellanox.co.il>
Cc: Or Gerlitz <ogerlitz@mellanox.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Acked-by: Wolfram Sang <wolfram@the-dreams.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert to the much saner new idr interface.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jack Morgenstein <jackm@dev.mellanox.co.il>
Cc: Or Gerlitz <ogerlitz@mellanox.com>
Cc: Roland Dreier <roland@purestorage.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
lockdep warns about taking a hard-irq-unsafe lock (sriov->id_map_lock)
inside a hard-irq-safe lock (sriov->going_down_lock).
Since id_map_lock is never taken in the interrupt context, we can
simply reverse the order of taking the two spinlocks, thus avoiding
the warning and the depencency.
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Roland Dreier <roland@purestorage.com>
To avoid name conflicts:
drivers/video/riva/fbdev.c:281:9: sparse: preprocessor token MAX_LEVEL redefined
While at it, also make the other names more consistent and add
parentheses.
[akpm@linux-foundation.org: repair fallout]
[sfr@canb.auug.org.au: IB/mlx4: fix for MAX_ID_MASK to MAX_IDR_MASK name change]
Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Cc: Bernd Petrovitsch <bernd@petrovitsch.priv.at>
Cc: walter harms <wharms@bfs.de>
Cc: Glauber Costa <glommer@parallels.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Roland Dreier <roland@purestorage.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In CM para-virtualization:
1. Incoming requests are steered to the correct vHCA according to the
embedded GID.
2. Communication IDs on outgoing requests are replaced by a globally
unique ID, generated by the PPF, since there is no synchronization
of ID generation between guests (and so these IDs are not
guaranteed to be globally unique). The guest's comm ID is stored,
and is returned to the response MAD when it arrives.
Signed-off-by: Amir Vadai <amirv@mellanox.co.il>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Matthew Wilcox