This series includes shared code updates for mlx5 core driver for both
netdev and rdma subsystems.
By Saeed,
First six patches of the series are meant to address a performance issue
and should provide a performance boost for multi core IRQ interrupt hungry
workloads. The issue is fixed in the first patch, all other patches are
meant to refactor the code in light of this fix.
The problem it comes to fix, is a shared spinlock accessed across all HCA
IRQs which protects the CQ database. To solve this we simply move the CQ
database and its spinlock to be per EQ (IRQ), thus per core.
By Yonatan,
Fragmented completion queue (CQ) for RDMA,
core driver implementation to create fragmented CQ buffers rather than
one large contiguous memory buffer, the implementation scheme already
exist and used by the netdev CQs, the patch shares that code with the
rdma CQ creation flow and makes use of the new API in mlx5_ib driver.
Thanks,
Saeed.
-----BEGIN PGP SIGNATURE-----
iQEcBAABAgAGBQJajcwxAAoJEEg/ir3gV/o+uJAIAICP2LBu5ANBgX7Z7/B0IZX4
Abkfk+IKR44v+3t0BBgEAZI2A9LcTQ8pwIOeLCDpcKd78665Y7rYQL1p31WuPKc5
1WwwBesjQLAmdiNLIjUnjW1fd3YyOmqLBcaMMYIeF0m4m4pCQue1kwtxesQrhdZV
I7FBxxnQt/8o6hoh/5nmU4DzGyZaY+uR/4FxUCEreklfhNeI5M2DZKe1xQCmhsJ4
Tosh/cJ8kSBJfclH6hk1lh5eWGDYDltWCpY86KBWsYktl10VAE3P7LPmekn487ta
e6cL1NPoEHbp+/UBThkNZMzUnA7wpoNeDbTTtVZrdcL1KivwColt1r3pmS1bqEQ=
=DlfW
-----END PGP SIGNATURE-----
Merge tag 'mlx5-updates-2018-02-21' of git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux
Saeed Mahameed says:
====================
mlx5-updates-2018-02-21
This series includes shared code updates for mlx5 core driver for both
netdev and rdma subsystems.
By Saeed,
First six patches of the series are meant to address a performance issue
and should provide a performance boost for multi core IRQ interrupt hungry
workloads. The issue is fixed in the first patch, all other patches are
meant to refactor the code in light of this fix.
The problem it comes to fix, is a shared spinlock accessed across all HCA
IRQs which protects the CQ database. To solve this we simply move the CQ
database and its spinlock to be per EQ (IRQ), thus per core.
By Yonatan,
Fragmented completion queue (CQ) for RDMA,
core driver implementation to create fragmented CQ buffers rather than
one large contiguous memory buffer, the implementation scheme already
exist and used by the netdev CQs, the patch shares that code with the
rdma CQ creation flow and makes use of the new API in mlx5_ib driver.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
BNXT_RE_FLAG_TASK_IN_PROG doesn't handle multiple work
requests posted together. Track schedule of multiple
workqueue items by maintaining a per device counter
and proceed with IB dereg only if this counter is zero.
flush_workqueue is no longer required from
NETDEV_UNREGISTER path.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
During driver unload, the driver proceeds with cleanup
without waiting for the scheduled events. So the device
pointers get freed up and driver crashes when the events
are scheduled later.
Flush the bnxt_re_task work queue before starting
device removal.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Avoid system crash when destroy_qp is invoked while
the driver is processing the poll_cq. Synchronize these
functions using the cq_lock.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Driver leaves the QP memory pinned if QP create command
fails from the FW. Avoids this scenario by adding a proper
exit path if the FW command fails.
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
More testing needs to be done before enabling this feature.
Disabling the feature temporarily
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
uaccess_kernel() isn't sufficient to determine if an rdma resource is
user-mode or not. For example, resources allocated in the add_one()
function of an ib_client get falsely labeled as user mode, when they
are kernel mode allocations. EG: mad qps.
The result is that these qps are skipped over during a nldev query
because of an erroneous namespace mismatch.
So now we determine if the resource is user-mode by looking at the object
struct's uobject or similar pointer to know if it was allocated for user
mode applications.
Fixes: 02d8883f52 ("RDMA/restrack: Add general infrastructure to track RDMA resources")
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Update all the flows to ensure that function pointer exists prior
to accessing it.
This is much safer than checking the uverbs_ex_mask variable, especially
since we know that test isn't working properly and will be removed
in -next.
This prevents a user triggereable oops.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This ensures that we return the right structures back to userspace.
Otherwise, it looks like the reserved fields in the response structures
in userspace might have uninitialized data in them.
Fixes: 8b10ba783c ("RDMA/vmw_pvrdma: Add shared receive queue support")
Fixes: 29c8d9eba5 ("IB: Add vmw_pvrdma driver")
Suggested-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Bryan Tan <bryantan@vmware.com>
Reviewed-by: Aditya Sarwade <asarwade@vmware.com>
Reviewed-by: Jorgen Hansen <jhansen@vmware.com>
Signed-off-by: Adit Ranadive <aditr@vmware.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
There is no matching lock for this mutex. Git history suggests this is
just a missed remnant from an earlier version of the function before
this locking was moved into uverbs_free_xrcd.
Originally this lock was protecting the xrcd_table_delete()
=====================================
WARNING: bad unlock balance detected!
4.15.0+ #87 Not tainted
-------------------------------------
syzkaller223405/269 is trying to release lock (&uverbs_dev->xrcd_tree_mutex) at:
[<00000000b8703372>] ib_uverbs_close_xrcd+0x195/0x1f0
but there are no more locks to release!
other info that might help us debug this:
1 lock held by syzkaller223405/269:
#0: (&uverbs_dev->disassociate_srcu){....}, at: [<000000005af3b960>] ib_uverbs_write+0x265/0xef0
stack backtrace:
CPU: 0 PID: 269 Comm: syzkaller223405 Not tainted 4.15.0+ #87
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
Call Trace:
dump_stack+0xde/0x164
? dma_virt_map_sg+0x22c/0x22c
? ib_uverbs_write+0x265/0xef0
? console_unlock+0x502/0xbd0
? ib_uverbs_close_xrcd+0x195/0x1f0
print_unlock_imbalance_bug+0x131/0x160
lock_release+0x59d/0x1100
? ib_uverbs_close_xrcd+0x195/0x1f0
? lock_acquire+0x440/0x440
? lock_acquire+0x440/0x440
__mutex_unlock_slowpath+0x88/0x670
? wait_for_completion+0x4c0/0x4c0
? rdma_lookup_get_uobject+0x145/0x2f0
ib_uverbs_close_xrcd+0x195/0x1f0
? ib_uverbs_open_xrcd+0xdd0/0xdd0
ib_uverbs_write+0x7f9/0xef0
? cyc2ns_read_end+0x10/0x10
? ib_uverbs_open_xrcd+0xdd0/0xdd0
? uverbs_devnode+0x110/0x110
? cyc2ns_read_end+0x10/0x10
? cyc2ns_read_end+0x10/0x10
? sched_clock_cpu+0x18/0x200
__vfs_write+0x10d/0x700
? uverbs_devnode+0x110/0x110
? kernel_read+0x170/0x170
? __fget+0x358/0x5d0
? security_file_permission+0x93/0x260
vfs_write+0x1b0/0x550
SyS_write+0xc7/0x1a0
? SyS_read+0x1a0/0x1a0
? trace_hardirqs_on_thunk+0x1a/0x1c
entry_SYSCALL_64_fastpath+0x1e/0x8b
RIP: 0033:0x4335c9
Cc: syzkaller <syzkaller@googlegroups.com>
Cc: <stable@vger.kernel.org> # 4.11
Fixes: fd3c7904db ("IB/core: Change idr objects to use the new schema")
Reported-by: Noa Osherovich <noaos@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Once the uobj is committed it is immediately possible another thread
could destroy it, which worst case, can result in a use-after-free
of the restrack objects.
Cc: syzkaller <syzkaller@googlegroups.com>
Fixes: 08f294a152 ("RDMA/core: Add resource tracking for create and destroy CQs")
Reported-by: Noa Osherovich <noaos@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The command number is not bounds checked against the command mask before it
is shifted, resulting in an ubsan hit. This does not cause malfunction since
the command number is eventually bounds checked, but we can make this ubsan
clean by moving the bounds check to before the mask check.
================================================================================
UBSAN: Undefined behaviour in
drivers/infiniband/core/uverbs_main.c:647:21
shift exponent 207 is too large for 64-bit type 'long long unsigned int'
CPU: 0 PID: 446 Comm: syz-executor3 Not tainted 4.15.0-rc2+ #61
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
Call Trace:
dump_stack+0xde/0x164
? dma_virt_map_sg+0x22c/0x22c
ubsan_epilogue+0xe/0x81
__ubsan_handle_shift_out_of_bounds+0x293/0x2f7
? debug_check_no_locks_freed+0x340/0x340
? __ubsan_handle_load_invalid_value+0x19b/0x19b
? lock_acquire+0x440/0x440
? lock_acquire+0x19d/0x440
? __might_fault+0xf4/0x240
? ib_uverbs_write+0x68d/0xe20
ib_uverbs_write+0x68d/0xe20
? __lock_acquire+0xcf7/0x3940
? uverbs_devnode+0x110/0x110
? cyc2ns_read_end+0x10/0x10
? sched_clock_cpu+0x18/0x200
? sched_clock_cpu+0x18/0x200
__vfs_write+0x10d/0x700
? uverbs_devnode+0x110/0x110
? kernel_read+0x170/0x170
? __fget+0x35b/0x5d0
? security_file_permission+0x93/0x260
vfs_write+0x1b0/0x550
SyS_write+0xc7/0x1a0
? SyS_read+0x1a0/0x1a0
? trace_hardirqs_on_thunk+0x1a/0x1c
entry_SYSCALL_64_fastpath+0x18/0x85
RIP: 0033:0x448e29
RSP: 002b:00007f033f567c58 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX: 00007f033f5686bc RCX: 0000000000448e29
RDX: 0000000000000060 RSI: 0000000020001000 RDI: 0000000000000012
RBP: 000000000070bea0 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 00000000ffffffff
R13: 00000000000056a0 R14: 00000000006e8740 R15: 0000000000000000
================================================================================
Cc: syzkaller <syzkaller@googlegroups.com>
Cc: <stable@vger.kernel.org> # 4.5
Fixes: 2dbd5186a3 ("IB/core: IB/core: Allow legacy verbs through extended interfaces")
Reported-by: Noa Osherovich <noaos@mellanox.com>
Reviewed-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
If remove_commit fails then the lock is left locked while the uobj still
exists. Eventually the kernel will deadlock.
lockdep detects this and says:
test/4221 is leaving the kernel with locks still held!
1 lock held by test/4221:
#0: (&ucontext->cleanup_rwsem){.+.+}, at: [<000000001e5c7523>] rdma_explicit_destroy+0x37/0x120 [ib_uverbs]
Fixes: 4da70da23e ("IB/core: Explicitly destroy an object while keeping uobject")
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This is really being used as an assert that the expected usecnt
is being held and implicitly that the usecnt is valid. Rename it to
assert_uverbs_usecnt and tighten the checks to only accept valid
values of usecnt (eg 0 and < -1 are invalid).
The tigher checkes make the assertion cover more cases and is more
likely to find bugs via syzkaller/etc.
Fixes: 3832125624 ("IB/core: Add support for idr types")
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The race is between lookup_get_idr_uobject and
uverbs_idr_remove_uobj -> uverbs_uobject_put.
We deliberately do not call sychronize_rcu after the idr_remove in
uverbs_idr_remove_uobj for performance reasons, instead we call
kfree_rcu() during uverbs_uobject_put.
However, this means we can obtain pointers to uobj's that have
already been released and must protect against krefing them
using kref_get_unless_zero.
==================================================================
BUG: KASAN: use-after-free in copy_ah_attr_from_uverbs.isra.2+0x860/0xa00
Read of size 4 at addr ffff88005fda1ac8 by task syz-executor2/441
CPU: 1 PID: 441 Comm: syz-executor2 Not tainted 4.15.0-rc2+ #56
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
Call Trace:
dump_stack+0x8d/0xd4
print_address_description+0x73/0x290
kasan_report+0x25c/0x370
? copy_ah_attr_from_uverbs.isra.2+0x860/0xa00
copy_ah_attr_from_uverbs.isra.2+0x860/0xa00
? uverbs_try_lock_object+0x68/0xc0
? modify_qp.isra.7+0xdc4/0x10e0
modify_qp.isra.7+0xdc4/0x10e0
ib_uverbs_modify_qp+0xfe/0x170
? ib_uverbs_query_qp+0x970/0x970
? __lock_acquire+0xa11/0x1da0
ib_uverbs_write+0x55a/0xad0
? ib_uverbs_query_qp+0x970/0x970
? ib_uverbs_query_qp+0x970/0x970
? ib_uverbs_open+0x760/0x760
? futex_wake+0x147/0x410
? sched_clock_cpu+0x18/0x180
? check_prev_add+0x1680/0x1680
? do_futex+0x3b6/0xa30
? sched_clock_cpu+0x18/0x180
__vfs_write+0xf7/0x5c0
? ib_uverbs_open+0x760/0x760
? kernel_read+0x110/0x110
? lock_acquire+0x370/0x370
? __fget+0x264/0x3b0
vfs_write+0x18a/0x460
SyS_write+0xc7/0x1a0
? SyS_read+0x1a0/0x1a0
? trace_hardirqs_on_thunk+0x1a/0x1c
entry_SYSCALL_64_fastpath+0x18/0x85
RIP: 0033:0x448e29
RSP: 002b:00007f443fee0c58 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX: 00007f443fee16bc RCX: 0000000000448e29
RDX: 0000000000000078 RSI: 00000000209f8000 RDI: 0000000000000012
RBP: 000000000070bea0 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 00000000ffffffff
R13: 0000000000008e98 R14: 00000000006ebf38 R15: 0000000000000000
Allocated by task 1:
kmem_cache_alloc_trace+0x16c/0x2f0
mlx5_alloc_cmd_msg+0x12e/0x670
cmd_exec+0x419/0x1810
mlx5_cmd_exec+0x40/0x70
mlx5_core_mad_ifc+0x187/0x220
mlx5_MAD_IFC+0xd7/0x1b0
mlx5_query_mad_ifc_gids+0x1f3/0x650
mlx5_ib_query_gid+0xa4/0xc0
ib_query_gid+0x152/0x1a0
ib_query_port+0x21e/0x290
mlx5_port_immutable+0x30f/0x490
ib_register_device+0x5dd/0x1130
mlx5_ib_add+0x3e7/0x700
mlx5_add_device+0x124/0x510
mlx5_register_interface+0x11f/0x1c0
mlx5_ib_init+0x56/0x61
do_one_initcall+0xa3/0x250
kernel_init_freeable+0x309/0x3b8
kernel_init+0x14/0x180
ret_from_fork+0x24/0x30
Freed by task 1:
kfree+0xeb/0x2f0
mlx5_free_cmd_msg+0xcd/0x140
cmd_exec+0xeba/0x1810
mlx5_cmd_exec+0x40/0x70
mlx5_core_mad_ifc+0x187/0x220
mlx5_MAD_IFC+0xd7/0x1b0
mlx5_query_mad_ifc_gids+0x1f3/0x650
mlx5_ib_query_gid+0xa4/0xc0
ib_query_gid+0x152/0x1a0
ib_query_port+0x21e/0x290
mlx5_port_immutable+0x30f/0x490
ib_register_device+0x5dd/0x1130
mlx5_ib_add+0x3e7/0x700
mlx5_add_device+0x124/0x510
mlx5_register_interface+0x11f/0x1c0
mlx5_ib_init+0x56/0x61
do_one_initcall+0xa3/0x250
kernel_init_freeable+0x309/0x3b8
kernel_init+0x14/0x180
ret_from_fork+0x24/0x30
The buggy address belongs to the object at ffff88005fda1ab0
which belongs to the cache kmalloc-32 of size 32
The buggy address is located 24 bytes inside of
32-byte region [ffff88005fda1ab0, ffff88005fda1ad0)
The buggy address belongs to the page:
page:00000000d5655c19 count:1 mapcount:0 mapping: (null)
index:0xffff88005fda1fc0
flags: 0x4000000000000100(slab)
raw: 4000000000000100 0000000000000000 ffff88005fda1fc0 0000000180550008
raw: ffffea00017f6780 0000000400000004 ffff88006c803980 0000000000000000
page dumped because: kasan: bad access detected
Memory state around the buggy address:
ffff88005fda1980: fc fc fb fb fb fb fc fc fb fb fb fb fc fc fb fb
ffff88005fda1a00: fb fb fc fc fb fb fb fb fc fc 00 00 00 00 fc fc
ffff88005fda1a80: fb fb fb fb fc fc fb fb fb fb fc fc fb fb fb fb
ffff88005fda1b00: fc fc 00 00 00 00 fc fc fb fb fb fb fc fc fb fb
ffff88005fda1b80: fb fb fc fc fb fb fb fb fc fc fb fb fb fb fc fc
==================================================================@
Cc: syzkaller <syzkaller@googlegroups.com>
Cc: <stable@vger.kernel.org> # 4.11
Fixes: 3832125624 ("IB/core: Add support for idr types")
Reported-by: Noa Osherovich <noaos@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This clarifies the design intention that time between allocate and
commit has the uobj exclusive to the caller. We already guarantee
this by delaying publishing the uobj pointer via idr_insert,
fd_install, list_add, etc.
Additionally holding the usecnt lock during this period provides
extra clarity and more protection against future mistakes.
Fixes: 3832125624 ("IB/core: Add support for idr types")
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
If the same attribute is listed twice by the user in the ioctl attribute
list then error unwind can cause the kernel to deref garbage.
This happens when an object with WRITE access is sent twice. The second
parse properly fails but corrupts the state required for the error unwind
it triggers.
Fixing this by making duplicates in the attribute list invalid. This is
not something we need to support.
The ioctl interface is currently recommended to be disabled in kConfig.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
32 bit processes running on a 64 bit kernel call compat_ioctl so that
implementations can revise any structure layout issues. Point compat_ioctl
at our normal ioctl because:
- All our structures are designed to be the same on 32 and 64 bit, ie we
use __aligned_u64 when required and are careful to manage padding.
- Any pointers are stored in u64's and userspace is expected
to prepare them properly.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Fix a bug in uverbs_ioctl_merge that looked at the object's iterator
number instead of the method's iterator number when merging methods.
While we're at it, make the uverbs_ioctl_merge code a bit more clear
and faster.
Fixes: 118620d368 ('IB/core: Add uverbs merge trees functionality')
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The union approach will get the endianness wrong sometimes if the kernel's
pointer size is 32 bits resulting in EFAULTs when trying to copy to/from
user.
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The rule for the API is pointers less than 8 bytes are inlined into
the .data field of the attribute. Fix the creation of the driver udata
struct to follow this rule and point to the .data itself when the size
is less than 8 bytes.
Otherwise if the UHW struct is less than 8 bytes the driver will get
EFAULT during copy_from_user.
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This fixes several bugs around the copy_to/from user path:
- copy_to used the user provided size of the attribute
and could copy data beyond the end of the kernel buffer into
userspace.
- copy_from didn't know the size of the kernel buffer and
could have left kernel memory unexpectedly un-initialized.
- copy_from did not use the user length to determine if the
attribute data is inlined or not.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Resource tracking of XRCD objects is not implemented in current
version of restrack and hence can be removed.
Fixes: 02d8883f52 ("RDMA/restrack: Add general infrastructure to track RDMA resources")
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
netdev_wait_allrefs() could rebroadcast NETDEV_UNREGISTER event
multiple times until all refs are gone, which will result in calling
ipoib_delete_debug_files multiple times and printing a warning.
Remove the WARN_ONCE since checks of NULL pointers before calling
debugfs_remove are not needed.
Fixes: 771a525840 ("IB/IPoIB: ibX: failed to create mcg debug file")
Signed-off-by: Alaa Hleihel <alaa@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The current implementation of create CQ requires contiguous
memory, such requirement is problematic once the memory is
fragmented or the system is low in memory, it causes for
failures in dma_zalloc_coherent().
This patch implements new scheme of fragmented CQ to overcome
this issue by introducing new type: 'struct mlx5_frag_buf_ctrl'
to allocate fragmented buffers, rather than contiguous ones.
Base the Completion Queues (CQs) on this new fragmented buffer.
It fixes following crashes:
kworker/29:0: page allocation failure: order:6, mode:0x80d0
CPU: 29 PID: 8374 Comm: kworker/29:0 Tainted: G OE 3.10.0
Workqueue: ib_cm cm_work_handler [ib_cm]
Call Trace:
[<>] dump_stack+0x19/0x1b
[<>] warn_alloc_failed+0x110/0x180
[<>] __alloc_pages_slowpath+0x6b7/0x725
[<>] __alloc_pages_nodemask+0x405/0x420
[<>] dma_generic_alloc_coherent+0x8f/0x140
[<>] x86_swiotlb_alloc_coherent+0x21/0x50
[<>] mlx5_dma_zalloc_coherent_node+0xad/0x110 [mlx5_core]
[<>] ? mlx5_db_alloc_node+0x69/0x1b0 [mlx5_core]
[<>] mlx5_buf_alloc_node+0x3e/0xa0 [mlx5_core]
[<>] mlx5_buf_alloc+0x14/0x20 [mlx5_core]
[<>] create_cq_kernel+0x90/0x1f0 [mlx5_ib]
[<>] mlx5_ib_create_cq+0x3b0/0x4e0 [mlx5_ib]
Signed-off-by: Yonatan Cohen <yonatanc@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Changes since v1:
Added changes in these files:
drivers/infiniband/hw/usnic/usnic_transport.c
drivers/staging/lustre/lnet/lnet/lib-socket.c
drivers/target/iscsi/iscsi_target_login.c
drivers/vhost/net.c
fs/dlm/lowcomms.c
fs/ocfs2/cluster/tcp.c
security/tomoyo/network.c
Before:
All these functions either return a negative error indicator,
or store length of sockaddr into "int *socklen" parameter
and return zero on success.
"int *socklen" parameter is awkward. For example, if caller does not
care, it still needs to provide on-stack storage for the value
it does not need.
None of the many FOO_getname() functions of various protocols
ever used old value of *socklen. They always just overwrite it.
This change drops this parameter, and makes all these functions, on success,
return length of sockaddr. It's always >= 0 and can be differentiated
from an error.
Tests in callers are changed from "if (err)" to "if (err < 0)", where needed.
rpc_sockname() lost "int buflen" parameter, since its only use was
to be passed to kernel_getsockname() as &buflen and subsequently
not used in any way.
Userspace API is not changed.
text data bss dec hex filename
30108430 2633624 873672 33615726 200ef6e vmlinux.before.o
30108109 2633612 873672 33615393 200ee21 vmlinux.o
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
CC: David S. Miller <davem@davemloft.net>
CC: linux-kernel@vger.kernel.org
CC: netdev@vger.kernel.org
CC: linux-bluetooth@vger.kernel.org
CC: linux-decnet-user@lists.sourceforge.net
CC: linux-wireless@vger.kernel.org
CC: linux-rdma@vger.kernel.org
CC: linux-sctp@vger.kernel.org
CC: linux-nfs@vger.kernel.org
CC: linux-x25@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
This is the mindless scripted replacement of kernel use of POLL*
variables as described by Al, done by this script:
for V in IN OUT PRI ERR RDNORM RDBAND WRNORM WRBAND HUP RDHUP NVAL MSG; do
L=`git grep -l -w POLL$V | grep -v '^t' | grep -v /um/ | grep -v '^sa' | grep -v '/poll.h$'|grep -v '^D'`
for f in $L; do sed -i "-es/^\([^\"]*\)\(\<POLL$V\>\)/\\1E\\2/" $f; done
done
with de-mangling cleanups yet to come.
NOTE! On almost all architectures, the EPOLL* constants have the same
values as the POLL* constants do. But they keyword here is "almost".
For various bad reasons they aren't the same, and epoll() doesn't
actually work quite correctly in some cases due to this on Sparc et al.
The next patch from Al will sort out the final differences, and we
should be all done.
Scripted-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
- Clean up some function signatures in rxe for clarity
- Tidy the RDMA netlink header to remove unimplemented constants
- bnxt_re driver fixes, one is a regression this window.
- Minor hns driver fixes
- Various fixes from Dan Carpenter and his tool
- Fix IRQ cleanup race in HFI1
- HF1 performance optimizations and a fix to report counters in the right units
- Fix for an IPoIB startup sequence race with the external manager
- Oops fix for the new kabi path
- Endian cleanups for hns
- Fix for mlx5 related to the new automatic affinity support
-----BEGIN PGP SIGNATURE-----
iQIcBAABAgAGBQJaePL/AAoJELgmozMOVy/dqhsQALUhzDuuJ+/F6supjmyqZG53
Ak/PoFjTmHToGQfDq/1TRzyKwMx12aB2l6WGZc31FzhvCw4daPWkoEVKReNWUUJ+
fmESxjLgo8ZRGSqpNxn9Q8agE/I/5JZQoA8bCFCYgdZPKTPNKdtAVBphpdhmrOX4
ygjABikWf/wBsNF1A8lnX9xkfPO21cPHrFQLTnuOzOT/hc6U+PPklHSQCnS91svh
1+Pqjtssg54rxYkJqiFq3giSnfwvmAXO8WyVGmRRPFGLpB0nIjq0Sl6ZgLLClz7w
YJdiBGr7rlnNMgGCjlPU2ZO3lO6J0ytXQzFNqRqvKryXQOv+uVeJgep7WqHTcdQU
UN30FCKQMgLL/F6NF8wKaKcK4X0VgXQa7gpuH2fVSXF0c3LO3/mmWNjixbGSzT2c
Wj+EW3eOKlTddhRLhgbMOdwc32tIGhaD85z2F4+FZO+XI9ZQtJaDewWVDjYoumP/
RlDIFw+KCgSq7+UZL8CoXuh0BuS1nu9TGfkx1HW0DLMF1+yigNiswpUfksV4cISP
JqE2I3yH0A4UobD/a+f9IhIfk2MjxO0tJWNjU8IA9LXgUFlskQ6MpH/AcE9G8JNv
tlfLGR3s4PJa/7j/Iy2F84og/b/KH8v7vyj4Eknq/hLq63/BiM5wj0AUBRrGulN6
HhAMOegxGZ7IKP/y0L7I
=xwZz
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull more rdma updates from Doug Ledford:
"Items of note:
- two patches fix a regression in the 4.15 kernel. The 4.14 kernel
worked fine with NVMe over Fabrics and mlx5 adapters. That broke in
4.15. The fix is here.
- one of the patches (the endian notation patch from Lijun) looks
like a lot of lines of change, but it's mostly mechanical in
nature. It amounts to the biggest chunk of change in it (it's about
2/3rds of the overall pull request).
Summary:
- Clean up some function signatures in rxe for clarity
- Tidy the RDMA netlink header to remove unimplemented constants
- bnxt_re driver fixes, one is a regression this window.
- Minor hns driver fixes
- Various fixes from Dan Carpenter and his tool
- Fix IRQ cleanup race in HFI1
- HF1 performance optimizations and a fix to report counters in the right units
- Fix for an IPoIB startup sequence race with the external manager
- Oops fix for the new kabi path
- Endian cleanups for hns
- Fix for mlx5 related to the new automatic affinity support"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (38 commits)
net/mlx5: increase async EQ to avoid EQ overrun
mlx5: fix mlx5_get_vector_affinity to start from completion vector 0
RDMA/hns: Fix the endian problem for hns
IB/uverbs: Use the standard kConfig format for experimental
IB: Update references to libibverbs
IB/hfi1: Add 16B rcvhdr trace support
IB/hfi1: Convert kzalloc_node and kcalloc to use kcalloc_node
IB/core: Avoid a potential OOPs for an unused optional parameter
IB/core: Map iWarp AH type to undefined in rdma_ah_find_type
IB/ipoib: Fix for potential no-carrier state
IB/hfi1: Show fault stats in both TX and RX directions
IB/hfi1: Remove blind constants from 16B update
IB/hfi1: Convert PortXmitWait/PortVLXmitWait counters to flit times
IB/hfi1: Do not override given pcie_pset value
IB/hfi1: Optimize process_receive_ib()
IB/hfi1: Remove unnecessary fecn and becn fields
IB/hfi1: Look up ibport using a pointer in receive path
IB/hfi1: Optimize packet type comparison using 9B and bypass code paths
IB/hfi1: Compute BTH only for RDMA_WRITE_LAST/SEND_LAST packet
IB/hfi1: Remove dependence on qp->s_hdrwords
...
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=ttq9
-----END PGP SIGNATURE-----
Merge tag 'pci-v4.16-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci
Pull PCI updates from Bjorn Helgaas:
- skip AER driver error recovery callbacks for correctable errors
reported via ACPI APEI, as we already do for errors reported via the
native path (Tyler Baicar)
- fix DPC shared interrupt handling (Alex Williamson)
- print full DPC interrupt number (Keith Busch)
- enable DPC only if AER is available (Keith Busch)
- simplify DPC code (Bjorn Helgaas)
- calculate ASPM L1 substate parameter instead of hardcoding it (Bjorn
Helgaas)
- enable Latency Tolerance Reporting for ASPM L1 substates (Bjorn
Helgaas)
- move ASPM internal interfaces out of public header (Bjorn Helgaas)
- allow hot-removal of VGA devices (Mika Westerberg)
- speed up unplug and shutdown by assuming Thunderbolt controllers
don't support Command Completed events (Lukas Wunner)
- add AtomicOps support for GPU and Infiniband drivers (Felix Kuehling,
Jay Cornwall)
- expose "ari_enabled" in sysfs to help NIC naming (Stuart Hayes)
- clean up PCI DMA interface usage (Christoph Hellwig)
- remove PCI pool API (replaced with DMA pool) (Romain Perier)
- deprecate pci_get_bus_and_slot(), which assumed PCI domain 0 (Sinan
Kaya)
- move DT PCI code from drivers/of/ to drivers/pci/ (Rob Herring)
- add PCI-specific wrappers for dev_info(), etc (Frederick Lawler)
- remove warnings on sysfs mmap failure (Bjorn Helgaas)
- quiet ROM validation messages (Alex Deucher)
- remove redundant memory alloc failure messages (Markus Elfring)
- fill in types for compile-time VGA and other I/O port resources
(Bjorn Helgaas)
- make "pci=pcie_scan_all" work for Root Ports as well as Downstream
Ports to help AmigaOne X1000 (Bjorn Helgaas)
- add SPDX tags to all PCI files (Bjorn Helgaas)
- quirk Marvell 9128 DMA aliases (Alex Williamson)
- quirk broken INTx disable on Ceton InfiniTV4 (Bjorn Helgaas)
- fix CONFIG_PCI=n build by adding dummy pci_irqd_intx_xlate() (Niklas
Cassel)
- use DMA API to get MSI address for DesignWare IP (Niklas Cassel)
- fix endpoint-mode DMA mask configuration (Kishon Vijay Abraham I)
- fix ARTPEC-6 incorrect IS_ERR() usage (Wei Yongjun)
- add support for ARTPEC-7 SoC (Niklas Cassel)
- add endpoint-mode support for ARTPEC (Niklas Cassel)
- add Cadence PCIe host and endpoint controller driver (Cyrille
Pitchen)
- handle multiple INTx status bits being set in dra7xx (Vignesh R)
- translate dra7xx hwirq range to fix INTD handling (Vignesh R)
- remove deprecated Exynos PHY initialization code (Jaehoon Chung)
- fix MSI erratum workaround for HiSilicon Hip06/Hip07 (Dongdong Liu)
- fix NULL pointer dereference in iProc BCMA driver (Ray Jui)
- fix Keystone interrupt-controller-node lookup (Johan Hovold)
- constify qcom driver structures (Julia Lawall)
- rework Tegra config space mapping to increase space available for
endpoints (Vidya Sagar)
- simplify Tegra driver by using bus->sysdata (Manikanta Maddireddy)
- remove PCI_REASSIGN_ALL_BUS usage on Tegra (Manikanta Maddireddy)
- add support for Global Fabric Manager Server (GFMS) event to
Microsemi Switchtec switch driver (Logan Gunthorpe)
- add IDs for Switchtec PSX 24xG3 and PSX 48xG3 (Kelvin Cao)
* tag 'pci-v4.16-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci: (140 commits)
PCI: cadence: Add EndPoint Controller driver for Cadence PCIe controller
dt-bindings: PCI: cadence: Add DT bindings for Cadence PCIe endpoint controller
PCI: endpoint: Fix EPF device name to support multi-function devices
PCI: endpoint: Add the function number as argument to EPC ops
PCI: cadence: Add host driver for Cadence PCIe controller
dt-bindings: PCI: cadence: Add DT bindings for Cadence PCIe host controller
PCI: Add vendor ID for Cadence
PCI: Add generic function to probe PCI host controllers
PCI: generic: fix missing call of pci_free_resource_list()
PCI: OF: Add generic function to parse and allocate PCI resources
PCI: Regroup all PCI related entries into drivers/pci/Makefile
PCI/DPC: Reformat DPC register definitions
PCI/DPC: Add and use DPC Status register field definitions
PCI/DPC: Squash dpc_rp_pio_get_info() into dpc_process_rp_pio_error()
PCI/DPC: Remove unnecessary RP PIO register structs
PCI/DPC: Push dpc->rp_pio_status assignment into dpc_rp_pio_get_info()
PCI/DPC: Squash dpc_rp_pio_print_error() into dpc_rp_pio_get_info()
PCI/DPC: Make RP PIO log size check more generic
PCI/DPC: Rename local "status" to "dpc_status"
PCI/DPC: Squash dpc_rp_pio_print_tlp_header() into dpc_rp_pio_print_error()
...
The hip06 and hip08 run on a little endian ARM, it needs to
revise the annotations to indicate that the HW uses little
endian data in the various DMA buffers, and flow the necessary
swaps throughout.
The imm_data use big endian mode. The cpu_to_le32/le32_to_cpu
swaps are no-op for this, which makes the only substantive
change the handling of imm_data which is now mandatory swapped.
This also keep match with the userspace hns driver and resolve
the warning by sparse.
Signed-off-by: Lijun Ou <oulijun@huawei.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
We really don't want people turning this on just yet, make it very
clear with capital letters.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
These days the userspace comes from rdma-core, revise references
in the kernel to point to the current repository.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Add trace_hfi1_rcvhdr support for bypass packets.
While here, remove the etype argument as it is available
in struct hfi1_packet.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Kzalloc_node API doesn't check for overflows in size multiplication.
While kcalloc API check for overflows in size multiplication
but these implementations are not NUMA-aware.
This conversion allowed for correcting an allocation used in the hot
path to be on the local NUMA and ensure us overflow free multiplication
for the size of a memory allocation.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Kamenee Arumugam <kamenee.arumugam@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The ev_file is an optional parameter for CQ creation. If the parameter
is not passed, the ev_file pointer will be NULL. Using that pointer
to set the cq_context will result in an OOPs.
Verify that ev_file is not NULL before using.
Cc: <stable@vger.kernel.org> # 4.14.x
Fixes: 9ee79fce36 ("IB/core: Add completion queue (cq) object actions")
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
On reboot SM can program port pkey table before ipoib registered its
event handler, which could result in missing pkey event and leave root
interface with initial pkey value from index 0.
Since OPA port starts with invalid pkey in index 0, root interface will
fail to initialize and stay down with no-carrier flag.
For IB ipoib interface may end up with pkey different from value
opensm put in pkey table idx 0, resulting in connectivity issues
(different mcast groups, for example).
Close the window by calling event handler after registration
to make sure ipoib pkey is in sync with port pkey table.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The routine which shows the fault stats checks the counters
to determine whether to show any stats based on the number of
transmitted pkts/bytes for a particular opcode.
Unfortunately, it only checked the receive counters. As a result,
if any packet faults have happened for packets egressing the HFI,
those stats would not be shown.
In order to fix this, the routine is amended to also check the
TX counters. With this change the pkt/byte counts are the sum of
both TX and RX counts for the opcode.
Fixes: 1b311f8931 ("IB/hfi1: Add tx_opcode_stats like the opcode_stats")
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
These values were introduced as part of the 16B code to
account for the varying size of the LRH between the differing
packet formats.
Replace the blind constants with defines based on FIELD_SIZEOF()
calls.
Fixes: 5b6cabb0db ("IB/hfi1: Add 16B RC/UC support")
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
HFI's counters SendWaitCnt and SendWaitVlCnt are in units
of TXE cycle time (at 805MHz). OPA counters PortXmitWait and
PortVLXmtWait are in units of flit times.
Convert the counter values to flit units using following
conversion formula:
PortXmitWait =
SendWaitCnt * 2 * (4 /link_width) * (25 Gbps /link_speed)
PortVLXmitWait =
SendWaitVLCnt * 2 * (4 /link_width) * (25 Gbps /link_speed)
At link up or downgrade events, the link width can change. To ensure
accurate counter calculations, sample the counters after the events,
during counter requests, and then aggregate the OPA counters.
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Kamenee Arumugam <kamenee.arumugam@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
During PCIe Gen 3 transistion, pcie_pset is read and might be overridden
to a default value(i.e. 255) in do_pcie_gen3_transition() routine.
If the pcie_pset value is overridden then this new value will be used
during initialization of next adapter on a different card.
Introducing a new local variable to avoid modification of pcie_pset
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Bartlomiej Dudek <bartlomiej.dudek@intel.com>
Signed-off-by: Patel Jay P <jay.p.patel@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The arguments for trace_hfi1_rcvhdr() get computed every
time in the hot path regardless of the whether the trace
is on or off. This is seen to be costly with a profile.
The handling of fault inject isolates the verbs device for
all packets regardless of the presence of a RHF_DC_ERR error.
Fix the first by computing trace_hfi1_rcvhdr() arguments within
the trace itself, so that when the trace is off, the argument
data isn't computed. Fix the second by moving the error check to
handle_eflags() when an RHF error occurs and by testing for
RHF_DC_ERR before executing the reset of handle_eflags().
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
packet->fecn and packet->becn are calculated in the hot path
and are never used. Remove these fields as they show to be
costly in a profile. Also, remove initialization for
becn and fecn in process_ecn() as they're unconditionally
assigned in the function and ensure fecn and becn variables
use a boolean type.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
In the receive path, hfi1_ibport is looked up by indexing into an
array. A profile shows this to be expensive. The receive context
data has a pointer to the ibport data, use that pointer instead.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The packet type comparison used to find out if a packet is a bypass
packet in the hot path is an expensive operation as seen in a profile.
Determine packet's pkey and migration bit through the bypass and 9B
code paths instead.
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
In hfi1_rc_rcv(), BTH is computed for all packets received.
However, it's only used for packets received with opcodes
RDMA_WRITE_LAST and SEND_LAST, and it is a costly operation.
Compute BTH only in the RDMA_WRITE_LAST/SEND_LAST code path
and let the compiler handle endianness conversion for bitwise
operations.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The s_hdrwords variable was used to indicate whether a
packet was already built on a previous iteration of the
send engine. This variable assumed the protection of the
QP's RVT_S_BUSY flag, which was required since the the
QP's s_lock was dropped just prior to the packet being
queued on the one of the egress mechanisms.
Support for multiple send engine instantiations require
that the field not be used due to concurency issues.
The ps.txreq signals the "already built" without the
potential concurency issues.
Fix by getting rid of all s_hdrword usage. A wrapper
is added to test for the already built case that used to
use s_hdrwords.
What used to be stored in s_hdrwords is now in the txreq.
The PBC is not counted, but is added in the pio/sdma code
paths prior to posting the packet.
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
David S. Miller