The commit noted below fixed a case where a pq is left on the sdma wait
list.

It however missed another case.

user_sdma_send_pkts() has two calls from hfi1_user_sdma_process_request().

If the first one fails as indicated by -EBUSY, the pq will be placed on
the waitlist as by design.

If the second call then succeeds, the pq is still on the waitlist setting
up a race with the interrupt handler if a subsequent request uses a
different SDMA engine

Fix by deleting the first call.

The use of pcount and the intent to send a short burst of packets followed
by the larger balance of packets was never correctly implemented, because
the two calls always send pcount packets no matter what.  A subsequent
patch will correct that issue.

Fixes: 9a293d1e ("IB/hfi1: Ensure pq is not left on waitlist")
Link: https://lore.kernel.org/r/20200504130917.175613.43231.stgit@awfm-01.aw.intel.com
Cc: <stable@vger.kernel.org>
Reviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The following warning can occur when a pq is left on the dmawait list and
the pq is then freed:

  WARNING: CPU: 47 PID: 3546 at lib/list_debug.c:29 __list_add+0x65/0xc0
  list_add corruption. next->prev should be prev (ffff939228da1880), but was ffff939cabb52230. (next=ffff939cabb52230).
  Modules linked in: mmfs26(OE) mmfslinux(OE) tracedev(OE) 8021q garp mrp ib_isert iscsi_target_mod target_core_mod crc_t10dif crct10dif_generic opa_vnic rpcrdma ib_iser libiscsi scsi_transport_iscsi ib_ipoib(OE) bridge stp llc iTCO_wdt iTCO_vendor_support intel_powerclamp coretemp intel_rapl iosf_mbi kvm_intel kvm irqbypass crct10dif_pclmul crct10dif_common crc32_pclmul ghash_clmulni_intel aesni_intel lrw gf128mul glue_helper ablk_helper cryptd ast ttm drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops drm pcspkr joydev drm_panel_orientation_quirks i2c_i801 mei_me lpc_ich mei wmi ipmi_si ipmi_devintf ipmi_msghandler nfit libnvdimm acpi_power_meter acpi_pad hfi1(OE) rdmavt(OE) rdma_ucm ib_ucm ib_uverbs ib_umad rdma_cm ib_cm iw_cm ib_core binfmt_misc numatools(OE) xpmem(OE) ip_tables
  nfsv3 nfs_acl nfs lockd grace sunrpc fscache igb ahci libahci i2c_algo_bit dca libata ptp pps_core crc32c_intel [last unloaded: i2c_algo_bit]
  CPU: 47 PID: 3546 Comm: wrf.exe Kdump: loaded Tainted: G W OE ------------ 3.10.0-957.41.1.el7.x86_64 #1
  Hardware name: HPE.COM HPE SGI 8600-XA730i Gen10/X11DPT-SB-SG007, BIOS SBED1229 01/22/2019
  Call Trace:
  [<ffffffff91f65ac0>] dump_stack+0x19/0x1b
  [<ffffffff91898b78>] __warn+0xd8/0x100
  [<ffffffff91898bff>] warn_slowpath_fmt+0x5f/0x80
  [<ffffffff91a1dabe>] ? ___slab_alloc+0x24e/0x4f0
  [<ffffffff91b97025>] __list_add+0x65/0xc0
  [<ffffffffc03926a5>] defer_packet_queue+0x145/0x1a0 [hfi1]
  [<ffffffffc0372987>] sdma_check_progress+0x67/0xa0 [hfi1]
  [<ffffffffc03779d2>] sdma_send_txlist+0x432/0x550 [hfi1]
  [<ffffffff91a20009>] ? kmem_cache_alloc+0x179/0x1f0
  [<ffffffffc0392973>] ? user_sdma_send_pkts+0xc3/0x1990 [hfi1]
  [<ffffffffc0393e3a>] user_sdma_send_pkts+0x158a/0x1990 [hfi1]
  [<ffffffff918ab65e>] ? try_to_del_timer_sync+0x5e/0x90
  [<ffffffff91a3fe1a>] ? __check_object_size+0x1ca/0x250
  [<ffffffffc0395546>] hfi1_user_sdma_process_request+0xd66/0x1280 [hfi1]
  [<ffffffffc034e0da>] hfi1_aio_write+0xca/0x120 [hfi1]
  [<ffffffff91a4245b>] do_sync_readv_writev+0x7b/0xd0
  [<ffffffff91a4409e>] do_readv_writev+0xce/0x260
  [<ffffffff918df69f>] ? pick_next_task_fair+0x5f/0x1b0
  [<ffffffff918db535>] ? sched_clock_cpu+0x85/0xc0
  [<ffffffff91f6b16a>] ? __schedule+0x13a/0x860
  [<ffffffff91a442c5>] vfs_writev+0x35/0x60
  [<ffffffff91a4447f>] SyS_writev+0x7f/0x110
  [<ffffffff91f78ddb>] system_call_fastpath+0x22/0x27

The issue happens when wait_event_interruptible_timeout() returns a value
<= 0.

In that case, the pq is left on the list. The code continues sending
packets and potentially can complete the current request with the pq still
on the dmawait list provided no descriptor shortage is seen.

If the pq is torn down in that state, the sdma interrupt handler could
find the now freed pq on the list with list corruption or memory
corruption resulting.

Fix by adding a flush routine to ensure that the pq is never on a list
after processing a request.

A follow-up patch series will address issues with seqlock surfaced in:
https://lore.kernel.org/r/20200320003129.GP20941@ziepe.ca

The seqlock use for sdma will then be converted to a spin lock since the
list_empty() doesn't need the protection afforded by the sequence lock
currently in use.

Fixes: a0d40693 ("staging/rdma/hfi1: Add page lock limit check for SDMA requests")
Link: https://lore.kernel.org/r/20200320200200.23203.37777.stgit@awfm-01.aw.intel.comReviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Cleaning up a pq can result in the following warning and panic:

  WARNING: CPU: 52 PID: 77418 at lib/list_debug.c:53 __list_del_entry+0x63/0xd0
  list_del corruption, ffff88cb2c6ac068->next is LIST_POISON1 (dead000000000100)
  Modules linked in: mmfs26(OE) mmfslinux(OE) tracedev(OE) 8021q garp mrp ib_isert iscsi_target_mod target_core_mod crc_t10dif crct10dif_generic opa_vnic rpcrdma ib_iser libiscsi scsi_transport_iscsi ib_ipoib(OE) bridge stp llc iTCO_wdt iTCO_vendor_support intel_powerclamp coretemp intel_rapl iosf_mbi kvm_intel kvm irqbypass crct10dif_pclmul crct10dif_common crc32_pclmul ghash_clmulni_intel ast aesni_intel ttm lrw gf128mul glue_helper ablk_helper drm_kms_helper cryptd syscopyarea sysfillrect sysimgblt fb_sys_fops drm pcspkr joydev lpc_ich mei_me drm_panel_orientation_quirks i2c_i801 mei wmi ipmi_si ipmi_devintf ipmi_msghandler nfit libnvdimm acpi_power_meter acpi_pad hfi1(OE) rdmavt(OE) rdma_ucm ib_ucm ib_uverbs ib_umad rdma_cm ib_cm iw_cm ib_core binfmt_misc numatools(OE) xpmem(OE) ip_tables
   nfsv3 nfs_acl nfs lockd grace sunrpc fscache igb ahci i2c_algo_bit libahci dca ptp libata pps_core crc32c_intel [last unloaded: i2c_algo_bit]
  CPU: 52 PID: 77418 Comm: pvbatch Kdump: loaded Tainted: G           OE  ------------   3.10.0-957.38.3.el7.x86_64 #1
  Hardware name: HPE.COM HPE SGI 8600-XA730i Gen10/X11DPT-SB-SG007, BIOS SBED1229 01/22/2019
  Call Trace:
   [<ffffffff90365ac0>] dump_stack+0x19/0x1b
   [<ffffffff8fc98b78>] __warn+0xd8/0x100
   [<ffffffff8fc98bff>] warn_slowpath_fmt+0x5f/0x80
   [<ffffffff8ff970c3>] __list_del_entry+0x63/0xd0
   [<ffffffff8ff9713d>] list_del+0xd/0x30
   [<ffffffff8fddda70>] kmem_cache_destroy+0x50/0x110
   [<ffffffffc0328130>] hfi1_user_sdma_free_queues+0xf0/0x200 [hfi1]
   [<ffffffffc02e2350>] hfi1_file_close+0x70/0x1e0 [hfi1]
   [<ffffffff8fe4519c>] __fput+0xec/0x260
   [<ffffffff8fe453fe>] ____fput+0xe/0x10
   [<ffffffff8fcbfd1b>] task_work_run+0xbb/0xe0
   [<ffffffff8fc2bc65>] do_notify_resume+0xa5/0xc0
   [<ffffffff90379134>] int_signal+0x12/0x17
  BUG: unable to handle kernel NULL pointer dereference at 0000000000000010
  IP: [<ffffffff8fe1f93e>] kmem_cache_close+0x7e/0x300
  PGD 2cdab19067 PUD 2f7bfdb067 PMD 0
  Oops: 0000 [#1] SMP
  Modules linked in: mmfs26(OE) mmfslinux(OE) tracedev(OE) 8021q garp mrp ib_isert iscsi_target_mod target_core_mod crc_t10dif crct10dif_generic opa_vnic rpcrdma ib_iser libiscsi scsi_transport_iscsi ib_ipoib(OE) bridge stp llc iTCO_wdt iTCO_vendor_support intel_powerclamp coretemp intel_rapl iosf_mbi kvm_intel kvm irqbypass crct10dif_pclmul crct10dif_common crc32_pclmul ghash_clmulni_intel ast aesni_intel ttm lrw gf128mul glue_helper ablk_helper drm_kms_helper cryptd syscopyarea sysfillrect sysimgblt fb_sys_fops drm pcspkr joydev lpc_ich mei_me drm_panel_orientation_quirks i2c_i801 mei wmi ipmi_si ipmi_devintf ipmi_msghandler nfit libnvdimm acpi_power_meter acpi_pad hfi1(OE) rdmavt(OE) rdma_ucm ib_ucm ib_uverbs ib_umad rdma_cm ib_cm iw_cm ib_core binfmt_misc numatools(OE) xpmem(OE) ip_tables
   nfsv3 nfs_acl nfs lockd grace sunrpc fscache igb ahci i2c_algo_bit libahci dca ptp libata pps_core crc32c_intel [last unloaded: i2c_algo_bit]
  CPU: 52 PID: 77418 Comm: pvbatch Kdump: loaded Tainted: G        W  OE  ------------   3.10.0-957.38.3.el7.x86_64 #1
  Hardware name: HPE.COM HPE SGI 8600-XA730i Gen10/X11DPT-SB-SG007, BIOS SBED1229 01/22/2019
  task: ffff88cc26db9040 ti: ffff88b5393a8000 task.ti: ffff88b5393a8000
  RIP: 0010:[<ffffffff8fe1f93e>]  [<ffffffff8fe1f93e>] kmem_cache_close+0x7e/0x300
  RSP: 0018:ffff88b5393abd60  EFLAGS: 00010287
  RAX: 0000000000000000 RBX: ffff88cb2c6ac000 RCX: 0000000000000003
  RDX: 0000000000000400 RSI: 0000000000000400 RDI: ffffffff9095b800
  RBP: ffff88b5393abdb0 R08: ffffffff9095b808 R09: ffffffff8ff77c19
  R10: ffff88b73ce1f160 R11: ffffddecddde9800 R12: ffff88cb2c6ac000
  R13: 000000000000000c R14: ffff88cf3fdca780 R15: 0000000000000000
  FS:  00002aaaaab52500(0000) GS:ffff88b73ce00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000010 CR3: 0000002d27664000 CR4: 00000000007607e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  PKRU: 55555554
  Call Trace:
   [<ffffffff8fe20d44>] __kmem_cache_shutdown+0x14/0x80
   [<ffffffff8fddda78>] kmem_cache_destroy+0x58/0x110
   [<ffffffffc0328130>] hfi1_user_sdma_free_queues+0xf0/0x200 [hfi1]
   [<ffffffffc02e2350>] hfi1_file_close+0x70/0x1e0 [hfi1]
   [<ffffffff8fe4519c>] __fput+0xec/0x260
   [<ffffffff8fe453fe>] ____fput+0xe/0x10
   [<ffffffff8fcbfd1b>] task_work_run+0xbb/0xe0
   [<ffffffff8fc2bc65>] do_notify_resume+0xa5/0xc0
   [<ffffffff90379134>] int_signal+0x12/0x17
  Code: 00 00 ba 00 04 00 00 0f 4f c2 3d 00 04 00 00 89 45 bc 0f 84 e7 01 00 00 48 63 45 bc 49 8d 04 c4 48 89 45 b0 48 8b 80 c8 00 00 00 <48> 8b 78 10 48 89 45 c0 48 83 c0 10 48 89 45 d0 48 8b 17 48 39
  RIP  [<ffffffff8fe1f93e>] kmem_cache_close+0x7e/0x300
   RSP <ffff88b5393abd60>
  CR2: 0000000000000010

The panic is the result of slab entries being freed during the destruction
of the pq slab.

The code attempts to quiesce the pq, but looking for n_req == 0 doesn't
account for new requests.

Fix the issue by using SRCU to get a pq pointer and adjust the pq free
logic to NULL the fd pq pointer prior to the quiesce.

Fixes: e87473bc ("IB/hfi1: Only set fd pointer when base context is completely initialized")
Link: https://lore.kernel.org/r/20200210131033.87408.81174.stgit@awfm-01.aw.intel.comReviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The call to sdma_progress() is called outside the wait lock.

In this case, there is a race condition where sdma_progress() can return
false and the sdma_engine can idle.  If that happens, there will be no
more sdma interrupts to cause the wakeup and the user_sdma xmit will hang.

Fix by moving the lock to enclose the sdma_progress() call.

Also, delete busycount. The need for this was removed by:
commit bcad2913 ("IB/hfi1: Serve the most starved iowait entry first")

Cc: <stable@vger.kernel.org>
Fixes: 77241056 ("IB/hfi1: add driver files")
Reviewed-by: NGary Leshner <Gary.S.Leshner@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

ACK packets are generally associated with request completion and resource
release and therefore should be sent first. This patch optimizes the
send engine by using the following policies:
(1) QPs with RVT_S_ACK_PENDING bit set in qp->s_flags or qpriv->s_flags
should have their priority incremented;
(2) QPs with ACK or TID-ACK packet queued should have their priority
incremented;
(3) When a QP is queued to the wait list due to resource constraints, it
will be queued to the head if it has ACK packet to send;
(4) When selecting qps to run from the wait list, the one with the highest
priority and starve_cnt will be selected; each priority will be equivalent
to a fixed number of starve_cnt (16).
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This patch integrates TID RDMA WRITE protocol into normal RDMA verbs
framework. The TID RDMA WRITE protocol is an end-to-end protocol
between the hfi1 drivers on two OPA nodes that converts a qualified
RDMA WRITE request into a TID RDMA WRITE request to avoid data copying
on the responder side.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Commit 4e045572 ("IB/hfi1: Add unique txwait_lock for txreq events")
laid the ground work to support per resource waiting locking.

This patch adds that with a lock unique to each sdma engine and pio
sendcontext and makes necessary changes for verbs, PSM, and vnic to use
the new locks.

This is particularly beneficial for smaller messages that will exhaust
resources at a faster rate.

Fixes: 77241056 ("IB/hfi1: add driver files")
Reviewed-by: NGary Leshner <Gary.S.Leshner@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Current implementation allows each qp to have only one send engine.  As
such, each qp has only one list to queue prebuilt packets when send engine
resources are not available. To improve performance, it is desired to
support multiple send engines for each qp.

This patch creates the framework to support two send engines
(two legs) for each qp for the TID RDMA protocol, which can be easily
extended to support more send engines. It achieves the goal by creating a
leg specific struct, iowait_work in the iowait struct, to hold the
work_struct and the tx_list as well as a pointer to the parent iowait
struct.

The hfi1_pkt_state now has an additional field to record the current legs
work structure and that is now passed to all egress waiters to determine
the leg that needs to wait via a new iowait helper.  The APIs are adjusted
to use the new leg specific struct as required.

Many new and modified helpers are added to support this change.
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

If the number of packets in a user sdma request does not match
the actual iovectors being sent, sdma_cleanup can be called on
an uninitialized request structure, resulting in a crash similar
to this:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000008
IP: [<ffffffffc0ae8bb7>] __sdma_txclean+0x57/0x1e0 [hfi1]
PGD 8000001044f61067 PUD 1052706067 PMD 0
Oops: 0000 [#1] SMP
CPU: 30 PID: 69912 Comm: upsm Kdump: loaded Tainted: G           OE
------------   3.10.0-862.el7.x86_64 #1
Hardware name: Intel Corporation S2600KPR/S2600KPR, BIOS
SE5C610.86B.01.01.0019.101220160604 10/12/2016
task: ffff8b331c890000 ti: ffff8b2ed1f98000 task.ti: ffff8b2ed1f98000
RIP: 0010:[<ffffffffc0ae8bb7>]  [<ffffffffc0ae8bb7>] __sdma_txclean+0x57/0x1e0
[hfi1]
RSP: 0018:ffff8b2ed1f9bab0  EFLAGS: 00010286
RAX: 0000000000008b2b RBX: ffff8b2adf6e0000 RCX: 0000000000000000
RDX: 00000000000000a0 RSI: ffff8b2e9eedc540 RDI: ffff8b2adf6e0000
RBP: ffff8b2ed1f9bad8 R08: 0000000000000000 R09: ffffffffc0b04a06
R10: ffff8b331c890190 R11: ffffe6ed00bf1840 R12: ffff8b3315480000
R13: ffff8b33154800f0 R14: 00000000fffffff2 R15: ffff8b2e9eedc540
FS:  00007f035ac47740(0000) GS:ffff8b331e100000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000008 CR3: 0000000c03fe6000 CR4: 00000000001607e0
Call Trace:
 [<ffffffffc0b0570d>] user_sdma_send_pkts+0xdcd/0x1990 [hfi1]
 [<ffffffff9fe75fb0>] ? gup_pud_range+0x140/0x290
 [<ffffffffc0ad3105>] ? hfi1_mmu_rb_insert+0x155/0x1b0 [hfi1]
 [<ffffffffc0b0777b>] hfi1_user_sdma_process_request+0xc5b/0x11b0 [hfi1]
 [<ffffffffc0ac193a>] hfi1_aio_write+0xba/0x110 [hfi1]
 [<ffffffffa001a2bb>] do_sync_readv_writev+0x7b/0xd0
 [<ffffffffa001bede>] do_readv_writev+0xce/0x260
 [<ffffffffa022b089>] ? tty_ldisc_deref+0x19/0x20
 [<ffffffffa02268c0>] ? n_tty_ioctl+0xe0/0xe0
 [<ffffffffa001c105>] vfs_writev+0x35/0x60
 [<ffffffffa001c2bf>] SyS_writev+0x7f/0x110
 [<ffffffffa051f7d5>] system_call_fastpath+0x1c/0x21
Code: 06 49 c7 47 18 00 00 00 00 0f 87 89 01 00 00 5b 41 5c 41 5d 41 5e 41 5f
5d c3 66 2e 0f 1f 84 00 00 00 00 00 48 8b 4e 10 48 89 fb <48> 8b 51 08 49 89 d4
83 e2 0c 41 81 e4 00 e0 00 00 48 c1 ea 02
RIP  [<ffffffffc0ae8bb7>] __sdma_txclean+0x57/0x1e0 [hfi1]
 RSP <ffff8b2ed1f9bab0>
CR2: 0000000000000008

There are two exit points from user_sdma_send_pkts().  One (free_tx)
merely frees the slab entry and one (free_txreq) cleans the sdma_txreq
prior to freeing the slab entry.   The free_txreq variation can only be
called after one of the sdma_init*() variations has been called.

In the panic case, the slab entry had been allocated but not inited.

Fix the issue by exiting through free_tx thus avoiding sdma_clean().

Cc: <stable@vger.kernel.org> # 4.9.x+
Fixes: 77241056 ("IB/hfi1: add driver files")
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NLukasz Odzioba <lukasz.odzioba@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

If the set_txreq_header_agh() function returns an error, the exit path
is chosen.

In this path, the code fails to set the return value.  This will cause
the caller to not realize an error has occurred.

Set the return value correctly in the error path.
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Hardware limits the maximum number of packets to u16 packets.

Match that size for all relevant sequence numbers in the user_sdma
engine.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Packet queue state is over used to determine SDMA descriptor
availablitity and packet queue request state.

cpu 0  ret = user_sdma_send_pkts(req, pcount);
cpu 0  if (atomic_read(&pq->n_reqs))
cpu 1  IRQ user_sdma_txreq_cb calls pq_update() (state to _INACTIVE)
cpu 0        xchg(&pq->state, SDMA_PKT_Q_ACTIVE);

At this point pq->n_reqs == 0 and pq->state is incorrectly
SDMA_PKT_Q_ACTIVE.  The close path will hang waiting for the state
to return to _INACTIVE.

This can also change the state from _DEFERRED to _ACTIVE.  However,
this is a mostly benign race.

Remove the racy code path.

Use n_reqs to determine if a packet queue is active or not.
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

pq_update() can only be called in two places: from the completion
function when the complete (npkts) sequence of packets has been
submitted and processed, or from setup function if a subset of the
packets were submitted (i.e. the error path).

Currently both paths can call pq_update() if an error occurrs.  This
race will cause the n_req value to go negative, hanging file_close(),
or cause a crash by freeing the txlist more than once.

Several variables are used to determine SDMA send state.  Most of
these are unnecessary, and have code inspectible races between the
setup function and the completion function, in both the send path and
the error path.

The request 'status' value can be set by the setup or by the
completion function.  This is code inspectibly racy.  Since the status
is not needed in the completion code or by the caller it has been
removed.

The request 'done' value races between usage by the setup and the
completion function.  The completion function does not need this.
When the number of processed packets matches npkts, it is done.

The 'has_error' value races between usage of the setup and the
completion function.  This can cause incorrect error handling and leave
the n_req in an incorrect value (i.e. negative).

Simplify the code by removing all of the unneeded state checks and
variables.

Clean up iovs node when it is freed.

Eliminate race conditions in the error path:

If all packets are submitted, the completion handler will set the
completion status correctly (ok or aborted).

If all packets are not submitted, the caller must wait until the
submitted packets have completed, and then set the completion status.

These two change eliminate the race condition in the error path.
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

A for loop condition of data_iovs in user_sdma_free_request
is unnecessarily repeated before the loop as an if check.

Remove the if enveloping the loop.
Reviewed-by: NJakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: NJan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE()

Please do not apply this to mainline directly, instead please re-run the
coccinelle script shown below and apply its output.

For several reasons, it is desirable to use {READ,WRITE}_ONCE() in
preference to ACCESS_ONCE(), and new code is expected to use one of the
former. So far, there's been no reason to change most existing uses of
ACCESS_ONCE(), as these aren't harmful, and changing them results in
churn.

However, for some features, the read/write distinction is critical to
correct operation. To distinguish these cases, separate read/write
accessors must be used. This patch migrates (most) remaining
ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following
coccinelle script:

----
// Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and
// WRITE_ONCE()

// $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch

virtual patch

@ depends on patch @
expression E1, E2;
@@

- ACCESS_ONCE(E1) = E2
+ WRITE_ONCE(E1, E2)

@ depends on patch @
expression E;
@@

- ACCESS_ONCE(E)
+ READ_ONCE(E)
----
Signed-off-by: NMark Rutland <mark.rutland@arm.com>
Signed-off-by: NPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: davem@davemloft.net
Cc: linux-arch@vger.kernel.org
Cc: mpe@ellerman.id.au
Cc: shuah@kernel.org
Cc: snitzer@redhat.com
Cc: thor.thayer@linux.intel.com
Cc: tj@kernel.org
Cc: viro@zeniv.linux.org.uk
Cc: will.deacon@arm.com
Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.comSigned-off-by: NIngo Molnar <mingo@kernel.org>

Remove the debug trace statement in pin_sdma_pages() that
gets executed when there is a memory allocation failure as
the trace message doesn't help with debugging the memory
allocation failure.

Cc: Leon Romanovsky <leon@kernel.org>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

AHG_HEADER_SET macro doesn't conform to the coding standards as it can
affect the control flow. Convert the macro AHG_HEADER_SET into an inline
function ahg_header_set().

Cc: Leon Romanovsky <leon@kernel.org>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The hfi1_cdbg() macro can be instantiated in the hot path even when it
is not in use.  This shows up on perf profiles.

Rework the macros (for SDMA and MMU), to use the trace interface directly
to eliminate this performance hit.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Clean up user_sdma.c by moving the structure and MACRO definitions into
the header file user_sdma.h
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Clean up user_exp_rcv.c file by moving structure definitions into header
file user_exp_rcv.h. Since these structure definitions depend on the
structure definitions in mmu_rb.h, move #include "mmu_rb.h" above
the include "user_exp_rcv.h" or include of header files that include
user_exp_rcv.h
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

num_user_pages() function has been defined in both user_exp_rcv.c file
and user_sdma.c file. Move the function definition to a header file so
there is only one definition in the source repo.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

In pin_vector_pages() function, if there is any error while pinning the
pages or while adding a pinned buffer to the cache, the bail out code
needs to unpin any pinned pages that are not in the cache and adjust the
n_locked counter that counts the total pages pinned. The current bail
out code doesn't seem to be doing it right in two cases:

1. Before pinning required pages for a buffer, the SDMA pinned buffer
cache is searched to see if the virtual address range that needs to be
pinned is already pinned. If there isn't a hit in the cache, a new node
is created for the buffer and is added to the cache after the buffer is
pinned. If adding the new node to the cache fails, the n_locked count is
decremented properly but the pinned pages are not freed. This commit
fixes this issue.

2. If there is a hit in the SDMA cache, but the cached buffer doesn't
have enough pages to cover the entire address range that needs to be
pinned, the node for the cached buffer is extracted from the cache,
remaining pages needed are pinned and added to the node. The node is
finally added back into the cache. If there is an error pinning the
extra pages, the bail out code frees all the pages in the node but the
n_locked count is not being decremented by the no of pages in the node
that are freed. This commit fixes this issue.

This commit fixes the above two issues by creating a new function that
frees the pages in a node and decrements the n_locked count by the
number of pages freed.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Clean up pin_vector_pages() function by moving page pinning related code
to a separate function since it really stands on its own.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

user_sdma_send_pkts() function is unnecessarily long. Clean it up by
moving some of its code into separate functions.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Performance analysis shows that the cache callback function
sdma_kmem_cache_ctor contributes to 1/2 of the kmem_cache_allocs
time.

Since all of the fields in the allocated data structure are initialized
in the code path, remove the _ctor function.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

PIO/SDMA send logic now uses the hdr_type field to determine
the type of packet that has been constructed. Based on the hdr_type,
certain things such as PBC flags, padding count and the LT extra
trailing bytes are determined.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The allocate_ctxt() function adds the context to the fd data structure.
Since the context is not completely initialized, this can cause confusion
as to whether the context is valid or not.

Move the fd reference from allocate_ctxt() to setup_base_ctxt().
Update the necessary functions to be aware of this move.
Reviewed-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Several data members of the user context have become unused over time.
Cleaning them up.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: NIra Weiny <ira.weiny@intel.com>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

When an egress resource(SDMA descriptors, pio credits) is not available,
a sending thread will be put on the resource's wait queue. When the
resource becomes available again, up to a fixed number of sending threads
can be awakened sequentially and removed from the wait queue, depending
on the number of waiting threads and the number of free resources. Since
each awakened sending thread will send as many packets as possible, it
is highly likely that the first sending thread will consume all the
egress resources. Subsequently, it will be put back to the end of the wait
queue. Depending on the timing when the later sending threads wake up,
they may not be able to send any packet and be again put back to the end
of the wait queue sequentially, right behind the first sending thread.
This starvation cycle continues until some sending threads exceed their
retry limit and consequently fail.

This patch fixes the issue by two simple approaches:
(1) Any starved sending thread will be put to the head of the wait queue
while a served sending thread will be put to the tail;
(2) The most starved sending thread will be served first.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Declarations and code in common between verbs and PSM are now moved
to exp_rcv.[ch].
Reviewed-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Atomic bit tests are used to single errors and the
completion of request submissions. These operations
don't need to be atomic and show to be expensive on
the profile.

Replace each atomic bit operation with a bool type
and a READ_ONCE/WRITE_ONCE pairing.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The atomic SDMA_REQ_SEND_DONE bit is set by the
process-level code, and then the same process-level
code uses the bit to test that all packets have been
submitted incurring a costly atomic read.

Use a bool type with a READ_ONCE/WRITE_ONCE
pairing for this bit, and use the same condition that
is used to set the bit to test that all packets have
been submitted.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The current user SDMA request structure layout has holes.
The cachelines can be reduced to improve cacheline trading.
Separate fields in the following categories: mostly read,
writable and shared with interrupt.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

An RB tree is used for the SDMA pinning cache. Cache
entries are extracted and reinserted from the tree
in case the address range for it changes. However,
if the address range for the entry doesn't change,
deleting the entry from the RB tree is not necessary.

This affects performance since the tree needs to be
rebalanced for each insertion, and this happens in
the hot path. Optimize RB search by not removing
entries when it's not needed.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The tx request is unnecessarily initialized in the hot
code path with memset(), however, there's no need to do
this as most fields are initialized later on. this
initialization shows to be costly in the profile.
Remove unnecessary initialization from tx request and make
sure all variables are initialized properly.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The error path for context initialization is not consistent. Cleanup all
resources on failure.

Removed unused variable user_event_mask.

Add the _BASE_FAILED bit to the event flags so that a base context can
notify waiting sub contexts that they cannot continue.

Running out of sub contexts is an EBUSY result, not EINVAL.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The AHG index is only accessed in the request call
from user space, so there's no need for atomic semantics.
Replace atomic operations for SDMA_REQ_HAVE_AHG bit
with a test of the AHG index.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Since almost all functions that use the hfi1_filedata get the pointer
from the file pointer, simplify by only passing the hfi1_filedata pointer.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: NIra Weiny <ira.weiny@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

To improve the readability of function prototypes, give the parameters
names.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Div instructions show costly in profiles when
the tx request header is set. Using right shift
instead of a divide operation reduces the cycles
spent in the function that sets the tx request
header as shown in the profile. Use right shift
operation instead.

Profile before change:
43.24%  009
|
|--23.41%-- user_sdma_send_pkts
|          |
|          |--99.90%-- hfi1_user_sdma_process_requestAfter:

Profile after change:
45.75%  009
|
|--14.81%-- user_sdma_send_pkts
|          |
|          |--99.95%-- hfi1_user_sdma_process_request
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>