For debugging purpose it will be easier to understand if prefetch works
okay if it has its own counter. Introduce ODP prefetch counter and count
per MR the total number of prefetched pages.
In addition remove comment which is not relevant anymore and anyway not in
the correct place.
Link: https://lore.kernel.org/r/20200621104147.53795-1-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add support to get resource dump in raw format. It enable drivers to
return the entire device specific QP/CQ/MR context without a need from the
driver to set each field separately.
The raw query returns only the device specific data, general data is still
returned by using the existing queries.
Example:
$ rdma res show mr dev mlx5_1 mrn 2 -r -j
[{"ifindex":7,"ifname":"mlx5_1",
"data":[0,4,255,254,0,0,0,0,0,0,0,0,16,28,0,216,...]}]
Link: https://lore.kernel.org/r/20200623113043.1228482-9-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Allow a ULP to ask the core to provide a completion queue based on a
least-used search on a per-device CQ pools. The device CQ pools grow in a
lazy fashion when more CQs are requested.
This feature reduces the amount of interrupts when using many QPs. Using
shared CQs allows for more effcient completion handling. It also reduces
the amount of overhead needed for CQ contexts.
Test setup:
Intel(R) Xeon(R) Platinum 8176M CPU @ 2.10GHz servers.
Running NVMeoF 4KB read IOs over ConnectX-5EX across Spectrum switch.
TX-depth = 32. The patch was applied in the nvme driver on both the target
and initiator. Four controllers are accessed from each core. In the
current test case we have exposed sixteen NVMe namespaces using four
different subsystems (four namespaces per subsystem) from one NVM port.
Each controller allocated X queues (RDMA QPs) and attached to Y CQs.
Before this series we had X == Y, i.e for four controllers we've created
total of 4X QPs and 4X CQs. In the shared case, we've created 4X QPs and
only X CQs which means that we have four controllers that share a
completion queue per core. Until fourteen cores there is no significant
change in performance and the number of interrupts per second is less than
a million in the current case.
==================================================
|Cores|Current KIOPs |Shared KIOPs |improvement|
|-----|---------------|--------------|-----------|
|14 |2332 |2723 |16.7% |
|-----|---------------|--------------|-----------|
|20 |2086 |2712 |30% |
|-----|---------------|--------------|-----------|
|28 |1971 |2669 |35.4% |
|=================================================
|Cores|Current avg lat|Shared avg lat|improvement|
|-----|---------------|--------------|-----------|
|14 |767us |657us |14.3% |
|-----|---------------|--------------|-----------|
|20 |1225us |943us |23% |
|-----|---------------|--------------|-----------|
|28 |1816us |1341us |26.1% |
========================================================
|Cores|Current interrupts|Shared interrupts|improvement|
|-----|------------------|-----------------|-----------|
|14 |1.6M/sec |0.4M/sec |72% |
|-----|------------------|-----------------|-----------|
|20 |2.8M/sec |0.6M/sec |72.4% |
|-----|------------------|-----------------|-----------|
|28 |2.9M/sec |0.8M/sec |63.4% |
====================================================================
|Cores|Current 99.99th PCTL lat|Shared 99.99th PCTL lat|improvement|
|-----|------------------------|-----------------------|-----------|
|14 |67ms |6ms |90.9% |
|-----|------------------------|-----------------------|-----------|
|20 |5ms |6ms |-10% |
|-----|------------------------|-----------------------|-----------|
|28 |8.7ms |6ms |25.9% |
|===================================================================
Performance improvement with sixteen disks (sixteen CQs per core) is
comparable.
Link: https://lore.kernel.org/r/1590568495-101621-3-git-send-email-yaminf@mellanox.com
Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
While creating a uobject every create reaches a point where the uobject is
fully initialized. For ioctls that go on to copy_to_user this means they
need to open code the destruction of a fully created uobject - ie the
RDMA_REMOVE_DESTROY sort of flow.
Open coding this creates bugs, eg the CQ does not properly flush the
events list when it does its error unwind.
Provide a uverbs_finalize_uobj_create() function which indicates that the
uobject is fully initialized and that abort should call to destroy_hw to
destroy the uobj->object and related.
Methods can call this function if they go on to have error cases after
setting uobj->object. Once done those error cases can simply do return,
without an error unwind.
Link: https://lore.kernel.org/r/20200519072711.257271-2-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Adds capability to create a qpn to be recognized as an accelerated
UD QP for ipoib.
This is accomplished by reserving 0x81 in byte[0] of the qpn as the
prefix for these qp types and reserving qpns between 0x810000 and
0x81ffff.
The hfi1 capability mask already contained a flag for the VNIC netdev.
This has been renamed and extended to include both VNIC and ipoib.
The rvt code to allocate qps now recognizes this flag and sets 0x81
into byte[0] of the qpn.
The code to allocate qpns is modified to reset the qpn numbering when it
is detected that a value is located in byte[0] for a UD QP and it is a
qpn being requested for net dev use. If it is a regular UD QP then it is
allowable to have bits set in byte[0] of the qpn and provide the
previously normal behavior.
The code to free the qpn now checks for the AIP prefix value of 0x81 and
removes it from the qpn before being freed so that the lower 16 bit
number can be reused.
This patch requires minor changes in the IB core and ipoib to facilitate
the creation of accelerated UP QPs.
Link: https://lore.kernel.org/r/20200511160607.173205.11757.stgit@awfm-01.aw.intel.com
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Gary Leshner <Gary.S.Leshner@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This patch implements the mechanism to accelerate the transmit side of
a multiple transmit queue RDMA netdev by submitting the packets to
the SDMA engine directly instead of sending through the verbs layer.
This patch also changes the UD/SEND_ONLY op to output the entropy value
in byte 0 of deth[1]. UD/SEND_ONLY_WITH_IMMEDIATE uses the previous
behavior with no entropy value being output.
The code in the ipoib rdma netdev which submits tx requests upon
successful submission will call trace_sdma_output_ibhdr to output
the ibhdr to the trace buffer.
Link: https://lore.kernel.org/r/20200511160548.173205.45616.stgit@awfm-01.aw.intel.com
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Gary Leshner <Gary.S.Leshner@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When a client is added it isn't allowed to fail, but all the client's have
various failure paths within their add routines.
This creates the very fringe condition where the client was added, failed
during add and didn't set the client_data. The core code will then still
call other client_data centric ops like remove(), rename(), get_nl_info(),
and get_net_dev_by_params() with NULL client_data - which is confusing and
unexpected.
If the add() callback fails, then do not call any more client ops for the
device, even remove.
Remove all the now redundant checks for NULL client_data in ops callbacks.
Update all the add() callbacks to return error codes
appropriately. EOPNOTSUPP is used for cases where the ULP does not support
the ib_device - eg because it only works with IB.
Link: https://lore.kernel.org/r/20200421172440.387069-1-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The current codebase makes use of the zero-length array language
extension to the C90 standard, but the preferred mechanism to declare
variable-length types such as these ones is a flexible array member[1][2],
introduced in C99:
struct foo {
int stuff;
struct boo array[];
};
By making use of the mechanism above, we will get a compiler warning
in case the flexible array does not occur last in the structure, which
will help us prevent some kind of undefined behavior bugs from being
inadvertently introduced[3] to the codebase from now on.
Also, notice that, dynamic memory allocations won't be affected by
this change:
"Flexible array members have incomplete type, and so the sizeof operator
may not be applied. As a quirk of the original implementation of
zero-length arrays, sizeof evaluates to zero."[1]
This issue was found with the help of Coccinelle.
[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html
[2] https://github.com/KSPP/linux/issues/21
[3] commit 7649773293 ("cxgb3/l2t: Fix undefined behaviour")
Link: https://lore.kernel.org/r/20200213010425.GA13068@embeddedor.com
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com> # added a few more
From https://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma
Leon Romanovsky says:
====================
Use ODP MRs for kernel ULPs
The following series extends MR creation routines to allow creation of
user MRs through kernel ULPs as a proxy. The immediate use case is to
allow RDS to work over FS-DAX, which requires ODP (on-demand-paging)
MRs to be created and such MRs were not possible to create prior this
series.
The first part of this patchset extends RDMA to have special verb
ib_reg_user_mr(). The common use case that uses this function is a
userspace application that allocates memory for HCA access but the
responsibility to register the memory at the HCA is on an kernel ULP.
This ULP acts as an agent for the userspace application.
The second part provides advise MR functionality for ULPs. This is
integral part of ODP flows and used to trigger pagefaults in advance
to prepare memory before running working set.
The third part is actual user of those in-kernel APIs.
====================
* tag 'rds-odp-for-5.5':
net/rds: Use prefetch for On-Demand-Paging MR
net/rds: Handle ODP mr registration/unregistration
net/rds: Detect need of On-Demand-Paging memory registration
RDMA/mlx5: Fix handling of IOVA != user_va in ODP paths
IB/mlx5: Mask out unsupported ODP capabilities for kernel QPs
RDMA/mlx5: Don't fake udata for kernel path
IB/mlx5: Add ODP WQE handlers for kernel QPs
IB/core: Add interface to advise_mr for kernel users
IB/core: Introduce ib_reg_user_mr
IB: Allow calls to ib_umem_get from kernel ULPs
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Allow ULPs to call advise_mr, so they can control ODP regions
in the same way as user space applications.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Add ib_reg_user_mr() for kernel ULPs to register user MRs.
The common use case that uses this function is a userspace application
that allocates memory for HCA access but the responsibility to register
the memory at the HCA is on an kernel ULP. This ULP that acts as an agent
for the userspace application.
This function is intended to be used without a user context so vendor
drivers need to be aware of calling reg_user_mr() device operation with
udata equal to NULL.
Among all drivers, i40iw is the only driver which relies on presence
of udata, so check udata existence for that driver.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Guy Levi <guyle@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Currently when the low level driver notifies Pkey, GID, and port change
events they are notified to the registered handlers in the order they are
registered.
IB core and other ULPs such as IPoIB are interested in GID, LID, Pkey
change events.
Since all GID queries done by ULPs are serviced by IB core, and the IB
core deferes cache updates to a work queue, it is possible for other
clients to see stale cache data when they handle their own events.
For example, the below call tree shows how ipoib will call
rdma_query_gid() concurrently with the update to the cache sitting in the
WQ.
mlx5_ib_handle_event()
ib_dispatch_event()
ib_cache_event()
queue_work() -> slow cache update
[..]
ipoib_event()
queue_work()
[..]
work handler
ipoib_ib_dev_flush_light()
__ipoib_ib_dev_flush()
ipoib_dev_addr_changed_valid()
rdma_query_gid() <- Returns old GID, cache not updated.
Move all the event dispatch to a work queue so that the cache update is
always done before any clients are notified.
Fixes: f35faa4ba9 ("IB/core: Simplify ib_query_gid to always refer to cache")
Link: https://lore.kernel.org/r/20191212113024.336702-3-leon@kernel.org
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Sample trace events:
kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE
<idle>-0 [002] 120.056292: cq_schedule: cq.id=4
kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt
kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1
<idle>-0 [002] 120.067503: cq_schedule: cq.id=4
kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt
kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1
<idle>-0 [002] 120.067657: cq_schedule: cq.id=4
kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt
kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1
...
systemd-1 [002] 122.392653: cq_schedule: cq.id=4
kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt
kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16
kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16
kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16
kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16
kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3
Several features to note in this output:
- The WCE count and context type are reported at allocation time
- The CPU and kworker for each CQ is evident
- The CQ's restracker ID is tagged on each trace event
- CQ poll scheduling latency is measured
- Details about how often single completions occur versus multiple
completions are evident
- The cost of the ULP's completion handler is recorded
Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Pull hmm updates from Jason Gunthorpe:
"This is another round of bug fixing and cleanup. This time the focus
is on the driver pattern to use mmu notifiers to monitor a VA range.
This code is lifted out of many drivers and hmm_mirror directly into
the mmu_notifier core and written using the best ideas from all the
driver implementations.
This removes many bugs from the drivers and has a very pleasing
diffstat. More drivers can still be converted, but that is for another
cycle.
- A shared branch with RDMA reworking the RDMA ODP implementation
- New mmu_interval_notifier API. This is focused on the use case of
monitoring a VA and simplifies the process for drivers
- A common seq-count locking scheme built into the
mmu_interval_notifier API usable by drivers that call
get_user_pages() or hmm_range_fault() with the VA range
- Conversion of mlx5 ODP, hfi1, radeon, nouveau, AMD GPU, and Xen
GntDev drivers to the new API. This deletes a lot of wonky driver
code.
- Two improvements for hmm_range_fault(), from testing done by Ralph"
* tag 'for-linus-hmm' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
mm/hmm: remove hmm_range_dma_map and hmm_range_dma_unmap
mm/hmm: make full use of walk_page_range()
xen/gntdev: use mmu_interval_notifier_insert
mm/hmm: remove hmm_mirror and related
drm/amdgpu: Use mmu_interval_notifier instead of hmm_mirror
drm/amdgpu: Use mmu_interval_insert instead of hmm_mirror
drm/amdgpu: Call find_vma under mmap_sem
nouveau: use mmu_interval_notifier instead of hmm_mirror
nouveau: use mmu_notifier directly for invalidate_range_start
drm/radeon: use mmu_interval_notifier_insert
RDMA/hfi1: Use mmu_interval_notifier_insert for user_exp_rcv
RDMA/odp: Use mmu_interval_notifier_insert()
mm/hmm: define the pre-processor related parts of hmm.h even if disabled
mm/hmm: allow hmm_range to be used with a mmu_interval_notifier or hmm_mirror
mm/mmu_notifier: add an interval tree notifier
mm/mmu_notifier: define the header pre-processor parts even if disabled
mm/hmm: allow snapshot of the special zero page
