On rmmod of irdma, the PBLE object memory is not being freed. PBLE object
memory are not statically pre-allocated at function initialization time
unlike other HMC objects. PBLEs objects and the Segment Descriptors (SD)
for it can be dynamically allocated during scale up and SD's remain
allocated till function deinitialization.
Fix this leak by adding IRDMA_HMC_IW_PBLE to the iw_hmc_obj_types[] table
and skip pbles in irdma_create_hmc_obj but not in irdma_del_hmc_objects().
Fixes: 44d9e52977 ("RDMA/irdma: Implement device initialization definitions")
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Link: https://lore.kernel.org/r/20230315145231.931-3-shiraz.saleem@intel.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Refactor HW statistics which,
- Unifies HW statistics support for all HW generations.
- Unifies support of 32- and 64-bit counters.
- Removes duplicated code and simplifies implementation.
- Fixes roll-over handling.
- Removes unneeded last_hw_stats.
With new implementation, there is no separate handling and no separate
arrays for 32- and 64-bit counters (offsets, regs, values). Instead,
there is a HW stats map array for each HW revision, which defines
HW-specific width and location of each counter in the statistics buffer.
Once the statistics are gathered (either via CQP op, or by reading HW
registers), counter values are extracted from the statistics buffer using
the stats map and the delta between the last and new values is computed.
Finally, the counter values in rdma_hw_stats are incremented by those
deltas.
From the OS perspective, all the counters are 64-bit and their order in
rdma_hw_stats->value[] array, as well as in irdma_hw_stat_names[], is the
same for all HW gens. New statistics should always be added at the end.
Signed-off-by: Krzysztof Czurylo <krzysztof.czurylo@intel.com>
Signed-off-by: Youvaraj Sagar <youvaraj.sagar@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Link: https://lore.kernel.org/r/20230315145305.955-2-shiraz.saleem@intel.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Quite a small cycle this time, even with the rc8. I suppose everyone
went to sleep over xmas.
- Minor driver updates for hfi1, cxgb4, erdma, hns, irdma, mlx5, siw,
mana
- inline CQE support for hns
- Have mlx5 display device error codes
- Pinned DMABUF support for irdma
- Continued rxe cleanups, particularly converting the MRs to use
xarray
- Improvements to what can be cached in the mlx5 mkey cache"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (61 commits)
IB/mlx5: Extend debug control for CC parameters
IB/hfi1: Fix sdma.h tx->num_descs off-by-one errors
IB/hfi1: Fix math bugs in hfi1_can_pin_pages()
RDMA/irdma: Add support for dmabuf pin memory regions
RDMA/mlx5: Use query_special_contexts for mkeys
net/mlx5e: Use query_special_contexts for mkeys
net/mlx5: Change define name for 0x100 lkey value
net/mlx5: Expose bits for querying special mkeys
RDMA/rxe: Fix missing memory barriers in rxe_queue.h
RDMA/mana_ib: Fix a bug when the PF indicates more entries for registering memory on first packet
RDMA/rxe: Remove rxe_alloc()
RDMA/cma: Distinguish between sockaddr_in and sockaddr_in6 by size
Subject: RDMA/rxe: Handle zero length rdma
iw_cxgb4: Fix potential NULL dereference in c4iw_fill_res_cm_id_entry()
RDMA/mlx5: Use rdma_umem_for_each_dma_block()
RDMA/umem: Remove unused 'work' member from struct ib_umem
RDMA/irdma: Cap MSIX used to online CPUs + 1
RDMA/mlx5: Check reg_create() create for errors
RDMA/restrack: Correct spelling
RDMA/cxgb4: Fix potential null-ptr-deref in pass_establish()
...
This is a followup to the EFA dmabuf[1]. Irdma driver currently does
not support on-demand-paging(ODP). So it uses habanalabs as the
dmabuf exporter, and irdma as the importer to allow for peer2peer
access through libibverbs.
In this commit, the function ib_umem_dmabuf_get_pinned() is used.
This function is introduced in EFA dmabuf[1] which allows the driver
to get a dmabuf umem which is pinned and does not require move_notify
callback implementation. The returned umem is pinned and DMA mapped
like standard cpu umems, and is released through ib_umem_release().
[1]https://lore.kernel.org/lkml/20211007114018.GD2688930@ziepe.ca/t/
Link: https://lore.kernel.org/r/20230217011425.498847-1-yanjun.zhu@intel.com
Reviewed-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The opcode written by HW, in the RQ CQE, is the
RoCEv2/iWARP protocol opcode from the received
packet and not the SW opcode as currently assumed.
Fix this by returning the raw operation type and
queue type in the CQE to irdma_process_cqe and add
2 helpers set_ib_wc_op_sq set_ib_wc_op_rq to map
IRDMA HW op types to IB op types.
Note that for iWARP, only Write with Immediate is
supported so the opcode can only be IB_WC_RECV_RDMA_WITH_IMM
when there is immediate data present.
Fixes: b48c24c2d7 ("RDMA/irdma: Implement device supported verb APIs")
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Link: https://lore.kernel.org/r/20221115011701.1379-3-shiraz.saleem@intel.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
A number of asynchronous event (AE) ids were not aligned to the
correct flush_code and event_type. Fix these up so that the
correct IBV error and event codes are returned to application.
Also, add handling for new AE ids like IRDMA_AE_INVALID_REQUEST to
return the correct WC error code.
Fixes: 44d9e52977 ("RDMA/irdma: Implement device initialization definitions")
Signed-off-by: Sindhu-Devale <sindhu.devale@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Link: https://lore.kernel.org/r/20220907191324.1173-2-shiraz.saleem@intel.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When a QP and a MR on a local host are in different PDs, the HW generates
an asynchronous event (AE). The same AE is generated when a QP and a MW
are in different PDs during a bind operation. Return the more appropriate
IBV_WC_MW_BIND_ERR for the latter case by checking the OP type from the
CQE in error.
Fixes: 551c46edc7 ("RDMA/irdma: Add user/kernel shared libraries")
Signed-off-by: Sindhu-Devale <sindhu.devale@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Link: https://lore.kernel.org/r/20220906223244.1119-4-shiraz.saleem@intel.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"This cycle we got a new RDMA driver "ERDMA" for the Alibaba cloud
environment. Otherwise the changes are dominated by rxe fixes.
There is another RDMA driver on the list that might get merged next
cycle, 'MANA' for the Azure cloud environment.
Summary:
- Bug fixes and small features for irdma, hns, siw, qedr, hfi1, mlx5
- General spelling/grammer fixes
- rdma cm can follow changes in neighbours for control packets
- Significant amounts of rxe fixes and spec compliance changes
- Use the modern NAPI API
- Use the bitmap API instead of open coding
- Performance improvements for rtrs
- Add the ERDMA driver for Alibaba cloud
- Fix a use after free bug in SRP"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (99 commits)
RDMA/ib_srpt: Unify checking rdma_cm_id condition in srpt_cm_req_recv()
RDMA/rxe: Fix error unwind in rxe_create_qp()
RDMA/mlx5: Add missing check for return value in get namespace flow
RDMA/rxe: Split qp state for requester and completer
RDMA/rxe: Generate error completion for error requester QP state
RDMA/rxe: Update wqe_index for each wqe error completion
RDMA/srpt: Fix a use-after-free
RDMA/srpt: Introduce a reference count in struct srpt_device
RDMA/srpt: Duplicate port name members
IB/qib: Fix repeated "in" within comments
RDMA/erdma: Add driver to kernel build environment
RDMA/erdma: Add the ABI definitions
RDMA/erdma: Add the erdma module
RDMA/erdma: Add connection management (CM) support
RDMA/erdma: Add verbs implementation
RDMA/erdma: Add verbs header file
RDMA/erdma: Add event queue implementation
RDMA/erdma: Add cmdq implementation
RDMA/erdma: Add main include file
RDMA/erdma: Add the hardware related definitions
...
Setting err_rq_idx_valid field in QP context when the AE source of the
AEQE is not associated with an RQ causes the firmware flush to fail.
Set err_rq_idx_valid field in QP context only if it is associated with an
RQ. Additionally, cleanup the redundant setting of this field in
irdma_process_aeq.
Fixes: 44d9e52977 ("RDMA/irdma: Implement device initialization definitions")
Link: https://lore.kernel.org/r/20220705230815.265-8-shiraz.saleem@intel.com
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Adapt the resource distribution algorithm in irdma_cfg_fpm_val to be more
QP oriented. If the configuration is too big for the available memory,
trim the MR and PBLE's first before trimming the QPs. This also avoids
having to double QPs requested as input to algorithm for GEN1 devices.
Link: https://lore.kernel.org/r/20220705230815.265-5-shiraz.saleem@intel.com
Signed-off-by: Nayan Kumar <nayan.kumar@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Taking the qos_mutex to process RoCEv2 QP's on netdev events causes a
kernel splat.
Fix this by removing the handling for RoCEv2 in
irdma_cm_teardown_connections that uses the mutex. This handling is only
needed for iWARP to avoid having connections established while the link is
down or having connections remain functional after the IP address is
removed.
BUG: sleeping function called from invalid context at kernel/locking/mutex.
Call Trace:
kernel: dump_stack+0x66/0x90
kernel: ___might_sleep.cold.92+0x8d/0x9a
kernel: mutex_lock+0x1c/0x40
kernel: irdma_cm_teardown_connections+0x28e/0x4d0 [irdma]
kernel: ? check_preempt_curr+0x7a/0x90
kernel: ? select_idle_sibling+0x22/0x3c0
kernel: ? select_task_rq_fair+0x94c/0xc90
kernel: ? irdma_exec_cqp_cmd+0xc27/0x17c0 [irdma]
kernel: ? __wake_up_common+0x7a/0x190
kernel: irdma_if_notify+0x3cc/0x450 [irdma]
kernel: ? sched_clock_cpu+0xc/0xb0
kernel: irdma_inet6addr_event+0xc6/0x150 [irdma]
Fixes: 146b9756f1 ("RDMA/irdma: Add connection manager")
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
x722 does not support 1GB page size but the irdma driver incorrectly
advertises 1GB page size support for x722 device to ib_core to compute the
best page size to use on this MR. This could lead to incorrect start
offsets computed by hardware on the MR.
Fixes: b48c24c2d7 ("RDMA/irdma: Implement device supported verb APIs")
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
HW flushes after QP in error state is not reliable. This can lead to
application hang waiting on a completion for outstanding WRs. Implement a
SW mechanism to generate completions for any outstanding WR's after the QP
is modified to error.
This is accomplished by starting a delayed worker after the QP is modified
to error and the HW flush is performed. The worker will generate
completions that will be returned to the application when it polls the
CQ. This mechanism only applies to Kernel applications.
Link: https://lore.kernel.org/r/20220425181624.1617-1-shiraz.saleem@intel.com
Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
QP destroy is synchronous and waits for its refcnt to be decremented in
irdma_cm_node_free_cb (for iWARP) which fires after the RCU grace period
elapses.
Applications running a large number of connections are exposed to high
wait times on destroy QP for events like SIGABORT.
The long pole for this wait time is the firing of the call_rcu callback
during a CM node destroy which can be slow. It holds the QP reference
count and blocks the destroy QP from completing.
call_rcu only needs to make sure that list walkers have a reference to the
cm_node object before freeing it and thus need to wait for grace period
elapse. The rest of the connection teardown in irdma_cm_node_free_cb is
moved out of the grace period wait in irdma_destroy_connection. Also,
replace call_rcu with a simple kfree_rcu as it just needs to do a kfree on
the cm_node
Fixes: 146b9756f1 ("RDMA/irdma: Add connection manager")
Link: https://lore.kernel.org/r/20220425181703.1634-3-shiraz.saleem@intel.com
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When connection establishment fails in iWARP mode, an app can drain the
QPs and hang because flush isn't issued when the QP is modified from RTR
state to error. Issue a flush in this case using function
irdma_cm_disconn().
Update irdma_cm_disconn() to do flush when cm_id is NULL, which is the
case when the QP is in RTR state and there is an error in the connection
establishment.
Fixes: b48c24c2d7 ("RDMA/irdma: Implement device supported verb APIs")
Link: https://lore.kernel.org/r/20220425181703.1634-2-shiraz.saleem@intel.com
Signed-off-by: Tatyana Nikolova <tatyana.e.nikolova@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
There is a deadlock in irdma_cleanup_cm_core(), which is shown below:
(Thread 1) | (Thread 2)
| irdma_schedule_cm_timer()
irdma_cleanup_cm_core() | add_timer()
spin_lock_irqsave() //(1) | (wait a time)
... | irdma_cm_timer_tick()
del_timer_sync() | spin_lock_irqsave() //(2)
(wait timer to stop) | ...
We hold cm_core->ht_lock in position (1) of thread 1 and use
del_timer_sync() to wait timer to stop, but timer handler also need
cm_core->ht_lock in position (2) of thread 2. As a result,
irdma_cleanup_cm_core() will block forever.
This patch removes the check of timer_pending() in
irdma_cleanup_cm_core(), because the del_timer_sync() function will just
return directly if there isn't a pending timer. As a result, the lock is
redundant, because there is no resource it could protect.
Link: https://lore.kernel.org/r/20220418153322.42524-1-duoming@zju.edu.cn
Signed-off-by: Duoming Zhou <duoming@zju.edu.cn>
Reviewed-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Split out flags from ib_device::device_cap_flags that are only used
internally to the kernel into kernel_cap_flags that is not part of the
uapi. This limits the device_cap_flags to being the same bitmap that will
be copied to userspace.
This cleanly splits out the uverbs flags from the kernel flags to avoid
confusion in the flags bitmap.
Add some short comments describing which each of the kernel flags is
connected to. Remove unused kernel flags.
Link: https://lore.kernel.org/r/0-v2-22c19e565eef+139a-kern_caps_jgg@nvidia.com
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <mgurtovoy@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
