sdp: device_removal_lock should not be a spinlock because module removal takes a long time.
In order to prevent a situation that many CPUs are stack with read-spinlock
waiting for remove-device write-lock to be freed, I replaced the rw-spinlock
with rw-semaphore. That way the blocked processes goes to sleep and are not
blocking other processes.
Eldad Zinger [Thu, 17 Jun 2010 06:57:57 +0000 (09:57 +0300)]
sdp: Fix for warning message when receiving with MSG_PEEK flag, and free skb that is not needed any more after all data was read from it.
'rx_sa->used', unlike 'offset', was not updated when MSG_PEEK flag was up and
that lead to the behavior that bytes considered to be available to copy while
the sequence offset showed that the bytes already consumed.
The solution is to discard any use of 'rx_sa->used' and use 'offset' instead.
An skb of SDP_MID_SRCAVAIL is not needed anymore when all its inline data was
consumed and the RDMA operation was canceled (due to MSG_PEEK flag).
This fix eliminates the warning message: "Trying to read beyond SKB".
Amir Vadai [Wed, 16 Jun 2010 08:54:11 +0000 (11:54 +0300)]
sdp: Fix iperf multistream hanging
When trying to send and no credits available. Sometime coulnd't get
credit update, because interrupts are turned off by default.
Added poll RX CQ to fix in this case.
Amir Vadai [Thu, 10 Jun 2010 08:32:50 +0000 (11:32 +0300)]
sdp: Fix bad handling of small rcvbuf size in zcopy
- Do not send RdmaRdComp when no buffers
- Same for SendSm
- post at least 3 buffers in RX to have the minimal number of credits
- make purge_tx_ring ignore WR used by RDMA
- fixed a typo, to reschedule tx_cq_poll timer according to tx queue and
not to rx queue (!)
- Allow credit updates when less than half RX Q is filled
Amir Vadai [Wed, 9 Jun 2010 09:40:23 +0000 (12:40 +0300)]
sdp: cleanup skb allocations
- Bad sizing of inline data on send sockets had implications on the
performance.
- All sent data is placed on the skb itself (unless accumulated by nagle)
- Do not count sdp header twice when allocating skb
added some likely/unlikely
Eldad Zinger [Sun, 30 May 2010 11:03:43 +0000 (14:03 +0300)]
sdp: device removal rewritten for a stability improvement.
main changes:
1. device_removal_lock is better used.
2. sdp_dev is marked NULL in order to prevent new sockets born to the removed
device.
3. new timeout functionality used when a reference count was taken for the CMA
to return, but the CMA won't be invoked because rdma_id was destroyed.
Eldad Zinger [Tue, 4 May 2010 08:43:29 +0000 (11:43 +0300)]
sdp: BUG2031 - sdp_cma_handler() won't be invoked after last ref count removed
rdma_id will be destructed in cases where the sdp_cma_handler() is not required any more.
This will disable asynchronous-events handling after the last reference count removed.
Kernel always crashed in the following test case:
user program: socket+bind+listen+accept. socket accpted.
shell: rmmod mlx4_ib
user program: close <<< CRASH
The fix closes any socket that its ib_device is the device being removed.
Amir Vadai [Thu, 15 Apr 2010 08:57:11 +0000 (11:57 +0300)]
sdp: Don't try to allocate FMR larger than RLIMIT_MEMLOCK
During ZCopy. If don't have CAP_IPC_LOCK capability and current
max number of locked pages is smaller than the buffer size, split
the send into small fragments.
Amir Vadai [Thu, 15 Apr 2010 08:52:19 +0000 (11:52 +0300)]
sdp: Don't count sdp header twice when calculating size_goal
sizeof(struct sdp_bsdh) is included inside skb->len. Ignore it
when calculating maximum payload of the skb.
This mistake caused every BCopy send of 32K (and its multiples) to
be split and thus got bad performance.
sdp: added differentiation between bind failures of sdp.
When bind()ing in mode 'BOTH', bind(sdp_sock) might fail if:
1. the IP&port is already bounded.
2. the IP is not part of IB network.
previous implementation returned errno=EADDRINUSE either way.
Only the first case should fail the bind(), the second is legitimate
because the TCP socket will hanle the connection.
This fix corresponds to a fix in libsdp.
Amir Vadai [Thu, 25 Feb 2010 09:43:03 +0000 (11:43 +0200)]
sdp: SendSM wasn't sent sometimes after getting SrcAvailCancel
* skb was freed if rx_sa is aborted - preventing SendSM
to be sent.
* Didn't update rx_sa->used in case of SrcAvailCancel
and therefore not sending RdmaRdCompl.
This also caused the next read to fail because offset
wasn't updated
Amir Vadai [Wed, 24 Feb 2010 08:59:31 +0000 (10:59 +0200)]
sdp: Fix bug in crossing SrcAvail
* Handle RdmaRdCompl in interrupt, before splitted to two Q's
This way the handling is sequencial, and no race could occure
between RdmaRdCompl and SrcAvailCancel
* Fixed an error when checking that RdmaRdCompl is not for
old SrcAvail
If sdp_add_device() fails, there is no client data stored in the IB device,
leading to a kernel crash when a connection is being established. Fix this
by rejecting connections when the device is not initialized.
Also, fix a bad goto target in an error case early in sdp_init_qp().
Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com> Signed-off-by: Amir Vadai <amirv@mellanox.co.il>
Amir Vadai [Tue, 16 Feb 2010 15:36:00 +0000 (17:36 +0200)]
sdp: Fix bugs in huge paged HW's
* Protect some constants that are based on PAGE_SIZE:
- FMR size
- xmit_goal
* renamed SDP_HEAD_SIZE => SDP_SKB_HEAD_SIZE
* removed unneeded special IA64 code due to changes here
Amir Vadai [Sun, 24 Jan 2010 15:12:34 +0000 (17:12 +0200)]
sdp: must use ib_sg_dma_*, not sg_dma_* for mapping
This fixes OFED bug 1895, althoug some warnings are still generated,
when running qperf sdp_bw with large sizes (using zcopy), on the
truescale adapters.
Signed-off-by: Dave Olson <dave.olson@qlogic.com> Signed-off-by: Amir Vadai <amirv@mellanox.co.il>
Amir Vadai [Tue, 24 Nov 2009 07:32:39 +0000 (09:32 +0200)]
sdp: fixed BUG1796 - running out of memory on rx
rcv queue could grow endlessly because minimal RX buffers in QP
was set to SDP_MIN_TX_CREDITS + 1 - so there always were credits
available for the sender.
Jack Morgenstein [Sun, 30 Aug 2009 14:16:13 +0000 (17:16 +0300)]
sdp: incorrect SDP_FMR_SIZE on 32-bit machines
On 32-bit machines, sizeof (u64 *) is 4 bytes (size of a ***pointer***).
However, the max SDP FMR pool size should be PAGE_SIZE / sizeof(an mtt entry) --
and mtt entries are u64's (or __be64's).
This resulted in SDP requesting twice as many entries per pool on 32-bit machines
as could fit on a single page -- with the result that the fmr pool allocation failed
at driver startup.
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
projects / users / jedix / linux-maple.git / log