Currently there isn't a clear layer separation between the counters and
the steering code, whereas the steering code is doing redundant access
to the counter struct.
Separate the fs.c and counters.c, where fs code won't access or be
aware of counter structs but only the objects it needs.
As a result, move mlx5_rdma_counter struct from the header file to be
an internal struct for the counters file only.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Edward Srouji <edwards@nvidia.com>
Link: https://patch.msgid.link/9854d1fdb140e4a6552b7a2fd1a028cfe488a935.1753004208.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
As part of this enhancement, allow the creation of an MKEY associated
with a DMA handle.
Additional notes:
MKEYs with TPH (i.e. TLP Processing Hints) attributes are currently not
UMR-capable unless explicitly enabled by firmware or hardware.
Therefore, to maintain such MKEYs in the MR cache, the TPH fields have
been added to the rb_key structure, with a dedicated hash bucket.
The ability to bypass the kernel verbs flow and create an MKEY with TPH
attributes using DEVX has been restricted. TPH must follow the standard
InfiniBand flow, where a DMAH is created with the appropriate security
checks and management mechanisms in place.
DMA handles are currently not supported in conjunction with On-Demand
Paging (ODP).
Re-registration of memory regions originally created with TPH attributes
is currently not supported.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Edward Srouji <edwards@nvidia.com>
Link: https://patch.msgid.link/1c485651cf8417694ddebb80446c5093d5a791a9.1752752567.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Extend UVERBS_METHOD_REG_MR to get DMAH and pass it to all drivers.
It will be used in mlx5 driver as part of the next patch from the
series.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Edward Srouji <edwards@nvidia.com>
Link: https://patch.msgid.link/2ae1e628c0675db81f092cc00d3ad6fbf6139405.1752752567.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The current implementation of DMABUF memory registration uses a fixed
page size for the memory key (mkey), which can lead to suboptimal
performance when the underlying memory layout may offer better page
size.
The optimization improves performance by reducing the number of page
table entries required for the mkey, leading to less MTT/KSM descriptors
that the HCA must go through to find translations, fewer cache-lines,
and shorter UMR work requests on mkey updates such as when
re-registering or reusing a cacheable mkey.
To ensure safe page size updates, the implementation uses a 5-step
process:
1. Make the first X entries non-present, while X is calculated to be
minimal according to a large page shift that can be used to cover the
MR length.
2. Update the page size to the large supported page size
3. Load the remaining N-X entries according to the (optimized)
page shift
4. Update the page size according to the (optimized) page shift
5. Load the first X entries with the correct translations
This ensures that at no point is the MR accessible with a partially
updated translation table, maintaining correctness and preventing
access to stale or inconsistent mappings, such as having an mkey
advertising the new page size while some of the underlying page table
entries still contain the old page size translations.
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/bc05a6b2142c02f96a90635f9a4458ee4bbbf39f.1751979184.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Align the capabilities checked when using the log_page_size 6th bit in the
mkey context to the PRM definition. The upper and lower bounds are set by
max/min caps, and modification of the 6th bit by UMR is allowed only when
a specific UMR cap is set.
Current implementation falsely assumes all page sizes up-to 2^63 are
supported when the UMR cap is set. In case the upper bound cap is lower
than 63, this might result a FW syndrome on mkey creation, e.g:
mlx5_core 0000:c1:00.0: mlx5_cmd_out_err:832:(pid 0): CREATE_MKEY(0×200) op_mod(0×0) failed, status bad parameter(0×3), syndrome (0×38a711), err(-22)
Previous cap enforcement is still correct for all current HW, FW and
driver combinations. However, this patch aligns the code to be PRM
compliant in the general case.
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/eab4eeb4785105a4bb5eb362dc0b3662cd840412.1751979184.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The MLX5_IB_STAGE_UAR stage in the RDMA driver is redundant and should
be removed.
Responsibility for initializing the device's UAR pointer
(mdev->priv.uar) lies with mlx5_core, which already sets it during the
mlx5_load() process.
At present, the RDMA UAR stage overwrites this pointer, which was
correctly initialized by mlx5_core, creating the risk of inconsistency.
Ownership and management of the UAR pointer should remain exclusively
within mlx5_core.
In the current upstream code, we luckily receive the same pointer, since
mlx5_get_uars_page() still finds available BF registers for that UAR,
allowing it to be shared.
However, future changes in mlx5_core may expose this flaw.
For instance, if mlx5_alloc_bfreg() is invoked twice before the RDMA UAR
stage runs, the RDMA driver may overwrite the UAR allocated by
mlx5_core.
This could lead to real bugs. For example, if mlx5_ib is unloaded
(rmmod), it might free the UAR, leaving mlx5_core with a dangling
reference to an invalid UAR.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Fan Li <fanl@nvidia.com>
Link: https://patch.msgid.link/feaa84ec6f20468b4935c439923e9266122a93d0.1747134130.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reuse newly added DMA API to cache IOVA and only link/unlink pages
in fast path for UMEM ODP flow.
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
As a preparation to remove dma_list, store access mask in PFN pointer
and not in dma_addr_t.
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Add support to allow optional-counters binding to a QP, whereas when
a bind operation is requested depending on the counter optional-counter
binding state the driver will determine if to also add optional-counters
to this QP binding.
The optional-counter binding is done by simply adding a steering
rule for the specific optional-counter condition with the additional
match over that QP number.
Note that optional-counters per QP rules are handled on an earlier prio
than per device counters, and per device counter correctness is maintained
by core whereas it is responsible to sum active counters when checking device
counter and to add them to history count when they are deallocated.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/2cad1b891a6641ae61fe8d92f867e1059121813a.1741875070.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add the following optional counters:
rdma_tx_packets,rdma_rx_bytes,rdma_rx_packets,rdma_tx_bytes.
Which counts all RDMA packets/bytes sent and received per link.
Note that since each direction packet and byte counter are shared,
the counter is only reset when both counters of that direction
are removed. But from user-perspective each can be enabled/disabled separately.
The counters can be enabled using:
sudo rdma stat set link rocep8s0f0/1 optional-counters rdma_tx_packets
And can be seen using:
rdma stat -j show link rocep8s0f0/1
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/9f2753ad636f21704416df64b47395c8991d1123.1741875070.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This patch adds RDMA_TRANSPORT_RX and RDMA_TRANSPORT_TX as a new flow
table type for matcher creation.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Link: https://patch.msgid.link/2287d8c50483e880450c7e8e08d9de34cdec1b14.1741261611.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The current ODP counters represent the total number of pages
handled, but it is not enough to understand the effectiveness
of these operations.
Extend the ODP counters to include the number of times page fault
and invalidation events were handled.
Example for a single page fault handling 512 pages:
- page_fault: incremented by 512 (total pages)
- page_fault_handled: incremented by 1 (operation count)
The same example is applicable for page invalidation too.
Previous output:
$ rdma stat mr
dev rocep8s0f0 mrn 8 page_faults 27 page_invalidations 0 page_prefetch 29
New output:
$ rdma stat mr
dev rocep8s0f0 mrn 21 page_faults 512 page_faults_handled 1
page_invalidations 0 page_invalidations_handled 0 page_prefetch 51200
Signed-off-by: Chiara Meiohas <cmeiohas@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/b18f29ed1392996ade66e9e6c45f018925253f6a.1733234165.git.leonro@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Support QP with out-of-order (OOO) capabilities enabled.
This allows WRs on the receiver side of the QP to be consumed OOO,
permitting the sender side to transmit messages without guaranteeing
arrival order on the receiver side.
When enabled, the completion ordering of WRs remains in-order,
regardless of the Receive WRs consumption order.
RDMA Read and RDMA Atomic operations on the responder side continue to
be executed in-order, while the ordering of data placement for RDMA
Write and Send operations is not guaranteed.
Atomic operations larger than 8 bytes are currently not supported.
Therefore, when this feature is enabled, the created QP restricts its
atomic support to 8 bytes at most.
In addition, when querying the device, a new flag is returned in
response to indicate that the Kernel supports OOO QP.
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Reviewed-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://patch.msgid.link/06ac609a5f358c8fb0a090d22c61a2f9329d82e6.1725362773.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The IB layer provides a common interface to store and get net
devices associated to an IB device port (ib_device_set_netdev()
and ib_device_get_netdev()).
Previously, mlx5_ib stored and managed the associated net devices
internally.
Replace internal net device management in mlx5_ib with
ib_device_set_netdev() when attaching/detaching a net device and
ib_device_get_netdev() when retrieving the net device.
Export ib_device_get_netdev().
For mlx5 representors/PFs/VFs and lag creation we replace the netdev
assignments with the IB set/get netdev functions.
In active-backup mode lag the active slave net device is stored in the
lag itself. To assure the net device stored in a lag bond IB device is
the active slave we implement the following:
- mlx5_core: when modifying the slave of a bond we send the internal driver event
MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE.
- mlx5_ib: when catching the event call ib_device_set_netdev()
This patch also ensures the correct IB events are sent in switchdev lag.
While at it, when in multiport eswitch mode, only a single IB device is
created for all ports. The said IB device will receive all netdev events
of its VFs once loaded, thus to avoid overwriting the mapping of PF IB
device to PF netdev, ignore NETDEV_REGISTER events if the ib device has
already been mapped to a netdev.
Signed-off-by: Chiara Meiohas <cmeiohas@nvidia.com>
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/20240909173025.30422-6-michaelgur@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Implicit MRs in ODP memory scheme require allocating a private null mkey
and assigning the mkey and va differently in the KSM mkey.
The page faults are received on the null mkey so we also add storing the
null mkey in the odp_mkey xarray.
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/20240909100504.29797-8-michaelgur@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Protect the usage of the 6th bit with the relevant capability to ensure
we are using the new page sizes with FW that supports the bit extension.
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/20240909100504.29797-2-michaelgur@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Fix the cleanup of the temp cache entries that are dynamically created
in the MR cache.
The cleanup of the temp cache entries is currently scheduled only when a
new entry is created. Since in the cleanup of the entries only the mkeys
are destroyed and the cache entry stays in the cache, subsequent
registrations might reuse the entry and it will eventually be filled with
new mkeys without cleanup ever getting scheduled again.
On workloads that register and deregister MRs with a wide range of
properties we see the cache ends up holding many cache entries, each
holding the max number of mkeys that were ever used through it.
Additionally, as the cleanup work is scheduled to run over the whole
cache, any mkey that is returned to the cache after the cleanup was
scheduled will be held for less than the intended 30 seconds timeout.
Solve both issues by dropping the existing remove_ent_work and reusing
the existing per-entry work to also handle the temp entries cleanup.
Schedule the work to run with a 30 seconds delay every time we push an
mkey to a clean temp entry.
This ensures the cleanup runs on each entry only 30 seconds after the
first mkey was pushed to an empty entry.
As we have already been distinguishing between persistent and temp entries
when scheduling the cache_work_func, it is not being scheduled in any
other flows for the temp entries.
Another benefit from moving to a per-entry cleanup is we now not
required to hold the rb_tree mutex, thus enabling other flow to run
concurrently.
Fixes: dd1b913fb0 ("RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow")
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://patch.msgid.link/e4fa4bb03bebf20dceae320f26816cd2dde23a26.1725362530.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add support for DMABUF MR registrations with Data-direct device.
Upon userspace calling to register a DMABUF MR with the data direct bit
set, the below algorithm will be followed.
1) Obtain a pinned DMABUF umem from the IB core using the user input
parameters (FD, offset, length) and the DMA PF device. The DMA PF
device is needed to allow the IOMMU to enable the DMA PF to access the
user buffer over PCI.
2) Create a KSM MKEY by setting its entries according to the user buffer
VA to IOVA mapping, with the MKEY being the data direct device-crossed
MKEY. This KSM MKEY is umrable and will be used as part of the MR cache.
The PD for creating it is the internal device 'data direct' kernel one.
3) Create a crossing MKEY that points to the KSM MKEY using the crossing
access mode.
4) Manage the KSM MKEY by adding it to a list of 'data direct' MKEYs
managed on the mlx5_ib device.
5) Return the crossing MKEY to the user, created with its supplied PD.
Upon DMA PF unbind flow, the driver will revoke the KSM entries.
The final deregistration will occur under the hood once the application
deregisters its MKEY.
Notes:
- This version supports only the PINNED UMEM mode, so there is no
dependency on ODP.
- The IOVA supplied by the application must be system page aligned due to
HW translations of KSM.
- The crossing MKEY will not be umrable or part of the MR cache, as we
cannot change its crossed (i.e. KSM) MKEY over UMR.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://patch.msgid.link/1f99d8020ed540d9702b9e2252a145a439609ba6.1722512548.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pass uverbs_attr_bundle as part of '.reg_user_mr_dmabuf' API instead of
udata.
This enables passing some new ioctl attributes to the drivers, as will
be introduced in the next patches for mlx5 driver.
Change the involved drivers accordingly.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://patch.msgid.link/9a25b2fc02443f7c36c2d93499ae25252b6afd40.1722512548.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add the NET device initialization flow to utilize the 'data
direct' device.
When a NET mlx5_ib device is capable of 'data direct', the following
sequence of actions will occur:
- Find its affiliated 'data direct' VUID via a firmware command.
- Create its own private PD and 'data direct' mkey.
- Register to be notified when its 'data direct' driver is probed or removed.
The DMA device of the affiliated 'data direct' device, including the
private PD and the 'data direct' mkey, will be used later during MR
registrations that request the data direct functionality.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://patch.msgid.link/b11fa87b2a65bce4db8d40341bb6cee490fa4d06.1722512548.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Introduce the 'data direct' driver for a ConnectX-8 Data Direct device.
The 'data direct' driver functions as the affiliated DMA device for one
or more capable mlx5_ib devices. This DMA device, as the name suggests,
is used exclusively for DMA operations. It can be considered a DMA engine
managed by a PF/VF, lacking network capabilities and having minimal overall
capabilities.
Consequently, the DMA NIC PF will not be exposed to or directly used by
software applications. The driver will not have any direct interface or
interaction with the firmware (no command interface, no capabilities,
etc.). It will operate solely over PCI to enable its DMA functionality.
Registration and un-registration of the driver are handled as part of
the mlx5_ib initialization and exit processes, as the mlx5_ib devices
will effectively be its clients.
The driver will serve as the DMA device for accessing another PCI device
to achieve optimal performance (both on the same NUMA node, P2P access,
etc.).
Upon probing, it will read its VUID over PCI to handle mlx5_ib device
registrations with the same VUID.
Upon removal, it will notify its clients to allow them to clean up the
resources that were mmaped with its DMA device.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://patch.msgid.link/b77edecfd476c3f445da96ab6aef499ae47b2829.1722512548.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This patch supports driver APIs "add_sub_dev" and "del_sub_dev", to
add and delete a plane device respectively.
A mlx5 plane device is a rdma SMI device; It provides the SMI capability
through user MAD for it's parent, the logical multi-plane aggregated
device. For a plane port:
- It supports QP0 only;
- When adding a plane device, all plane ports are added;
- For some commands like mad_ifc, both plane_index and native portnum
is needed;
- When querying or modifying a plane port context, the native portnum
must be used, as the query/modify_hca_vport_context command doesn't
support plane port.
Signed-off-by: Mark Zhang <markzhang@nvidia.com>
Link: https://lore.kernel.org/r/e933cd0562aece181f8657af2ca0f5b387d0f14e.1718553901.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
When multi-plane is supported, a logical port, which is aggregation of
multiple physical plane ports, is exposed for data transmission.
Compared with a normal mlx5 IB port, this logical port supports all
functionalities except Subnet Management.
Signed-off-by: Mark Zhang <markzhang@nvidia.com>
Link: https://lore.kernel.org/r/7e37c06c9cb243be9ac79930cd17053903785b95.1718553901.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Add UAR page index as a driver ioctl attribute to increase the number of
supported indices, previously limited to 16 bits by mlx5_ib_create_cq
struct.
Link: https://lore.kernel.org/r/0e18b34d7ec3b1ae02d694b0d545aed7413c0ef7.1719512393.git.leon@kernel.org
Signed-off-by: Akiva Goldberger <agoldberger@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Changes the create_cq verb signature by sending the entire uverbs attr
bundle as a parameter. This allows drivers to send driver specific attrs
through ioctl for the create_cq verb and access them in their driver
specific code.
Also adds a new enum value for driver specific ioctl attributes for
methods already supporting UHW.
Link: https://lore.kernel.org/r/ed147343987c0d43fd391c1b2f85e2f425747387.1719512393.git.leon@kernel.org
Signed-off-by: Akiva Goldberger <agoldberger@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Set the mkey for dmabuf at PAGE_SIZE to support any SGL
after a move operation.
ib_umem_find_best_pgsz returns 0 on error, so it is
incorrect to check the returned page_size against PAGE_SIZE
Fixes: 90da7dc820 ("RDMA/mlx5: Support dma-buf based userspace memory region")
Signed-off-by: Chiara Meiohas <cmeiohas@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Link: https://lore.kernel.org/r/1e2289b9133e89f273a4e68d459057d032cbc2ce.1718301631.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Previously, all IB dev resources are initialized on driver load. As
they are not always used, move the initialization to the time when
they are needed.
To be more specific, move PD (p0) and CQ (c0) initialization to the
time when the first SRQ is created. and move SRQs（s0 and s1)
initialization to the time first QP is created. To avoid concurrent
creations, two new mutexes are also added.
Signed-off-by: Jianbo Liu <jianbol@nvidia.com>
Link: https://lore.kernel.org/r/98c3e53a8cc0bdfeb6dec6e5bb8b037d78ab00d8.1717409369.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
UMR QP is not used in some cases, so move QP and its CQ creations from
driver load flow to the time first reg_mr occurs, that is when MR
interfaces are first called.
The initialization of dev->umrc.pd and dev->umrc.lock is still done in
driver load because pd is needed for mlx5_mkey_cache_init and the lock
is reused to protect against the concurrent creation.
When testing 4G bytes memory registration latency with rtool [1] and 8
threads in parallel, there is minor performance degradation (<5% for
the max latency) is seen for the first reg_mr with this change.
Link: https://github.com/paravmellanox/rtool [1]
Signed-off-by: Jianbo Liu <jianbol@nvidia.com>
Link: https://lore.kernel.org/r/55d3c4f8a542fd974d8a4c5816eccfb318a59b38.1717409369.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The test of write combining was added before in mlx5_ib driver. It
opens UD QP and posts NOP WQEs, and uses BlueFlame doorbell. When
BlueFlame is used, WQEs get written directly to a PCI BAR of the
device (in addition to memory) so that the device handles them without
having to access memory.
In this test, the WQEs written in memory are different from the ones
written to the BlueFlame which request CQE update. By checking the
completion reports posted on CQ, we can know if BlueFlame succeeds or
not. The write combining must be supported if BlueFlame succeeds as
its register is written using write combining.
This patch reimplements the test in the same way, but using a pair of
SQ and CQ only. It is moved to mlx5_core as a general feature used by
both mlx5_core and mlx5_ib.
Besides, save write combine test result of the PCI function, so that
its thousands of child functions such as SF can query without paying
the time and resource penalty by itself. The test function is called
only after failing to get the cached result. With this enhancement,
all thousands of SFs of the PF attached to same driver no longer need
to perform WC check explicitly, which is already done in the system.
This saves several commands per SF, thereby speeds up SF creation and
also saves completion EQ creation.
Signed-off-by: Jianbo Liu <jianbol@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/4ff5a8cc4c5b5b0d98397baa45a5019bcdbf096e.1717409369.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
umem can be NULL for user application mkeys in some cases. Therefore
umem can't be used for checking if the mkey is cacheable and it is
changed for checking a flag that indicates it. Also make sure that
all mkeys which are not returned to the cache will be destroyed.
Fixes: dd1b913fb0 ("RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow")
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Link: https://lore.kernel.org/r/2690bc5c6896bcb937f89af16a1ff0343a7ab3d0.1712140377.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
As some mkeys can't be modified with UMR due to some UMR limitations,
like the size of translation that can be updated, not all user mkeys can
be cached.
Fixes: dd1b913fb0 ("RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow")
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Link: https://lore.kernel.org/r/f2742dd934ed73b2d32c66afb8e91b823063880c.1712140377.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Currently, mkeys are managed via xarray. This implementation leads to
a degradation in cases many MRs are unregistered in parallel, due to xarray
internal implementation, for example: deregistration 1M MRs via 64 threads
is taking ~15% more time[1].
Hence, implement mkeys management via LIFO queue, which solved the
degradation.
[1]
2.8us in kernel v5.19 compare to 3.2us in kernel v6.4
Signed-off-by: Shay Drory <shayd@nvidia.com>
Link: https://lore.kernel.org/r/fde3d4cfab0f32f0ccb231cd113298256e1502c5.1695283384.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add RoCE MACsec rules when a gid is added for the MACsec netdevice and
handle their cleanup when the gid is removed or the MACsec SA is deleted.
Also support alias IP for the MACsec device, as long as we don't have
more ips than what the gid table can hold.
In addition handle the case where a gid is added but there are still no
SAs added for the MACsec device, so the rules are added later on when
the SAs are added.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Handle MACsec IP ambiguity issue, since mlx5 hw can't support
programming both the MACsec and the physical gid when they have the same
IP address, because it wouldn't know to whom to steer the traffic.
Hence in such case we delete the physical gid from the hw gid table,
which would then cause all traffic sent over it to fail, and we'll only
be able to send traffic over the MACsec gid.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Raed Salem <raeds@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAmSYzfYeHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiG/ucH/iOM/1Py/fSg0qSs
7NJ4XXlourT5zrnRMom3cm3d9gYqgTzgvKFL3kjMEexTRVYbhlcO4ZPRsiry8zxF
ToGX+V8tDMqb8WSdFHzkljRY+zDRyfEUDMlTzROAD9DunLmQtkJKyrggkeGdjkpP
OyfGqKpwlLXZRAXBil/U8Mx9MHdjJubloZwghLZr33VdUZa68+JJ9l6w163Oe/ET
K264NM0wxN/kvN57JvePgqMccQwpINylg8IhRI+XelgczjUXeJBsOA8TDv4bDN4Q
bjCLhkWbIaZtTYqvOXa/kD0T8wd7KETsMBQN8YzyDh6W0GmAlJjTawyAhA6jA5in
x3uz2W8=
=L3zp
-----END PGP SIGNATURE-----
Merge tag 'v6.4' into rdma.git for-next
Linux 6.4
Resolve conflicts between rdma rc and next in rxe_cq matching linux-next:
drivers/infiniband/sw/rxe/rxe_cq.c:
https://lore.kernel.org/r/20230622115246.365d30ad@canb.auug.org.au
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The cited commit aimed to ensure that Virtual Functions (VFs) assign a
queue affinity to a Queue Pair (QP) to distribute traffic when
the LAG master creates a hardware LAG. If the affinity was set while
the hardware was not in LAG, the firmware would ignore the affinity value.
However, this commit unintentionally assigned an affinity to QPs on the LAG
master's VPORT even if the RDMA device was not marked as LAG-enabled.
In most cases, this was not an issue because when the hardware entered
hardware LAG configuration, the RDMA device of the LAG master would be
destroyed and a new one would be created, marked as LAG-enabled.
The problem arises when a user configures Equal-Cost Multipath (ECMP).
In ECMP mode, traffic can be directed to different physical ports based on
the queue affinity, which is intended for use by VPORTS other than the
E-Switch manager. ECMP mode is supported only if both E-Switch managers are
in switchdev mode and the appropriate route is configured via IP. In this
configuration, the RDMA device is not destroyed, and we retain the RDMA
device that is not marked as LAG-enabled.
To ensure correct behavior, Send Queues (SQs) opened by the E-Switch
manager through verbs should be assigned strict affinity. This means they
will only be able to communicate through the native physical port
associated with the E-Switch manager. This will prevent the firmware from
assigning affinity and will not allow the SQs to be remapped in case of
failover.
Fixes: 802dcc7fc5 ("RDMA/mlx5: Support TX port affinity for VF drivers in LAG mode")
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Link: https://lore.kernel.org/r/425b05f4da840bc684b0f7e8ebf61aeb5cef09b0.1685960567.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
A misbehaved user can create a steering anchor that points to a kernel
flow table and then destroy the anchor without freeing the associated
STC. This creates a problem as the kernel can't destroy the flow
table since there is still a reference to it. As a result, this can
exhaust all available flow table resources, preventing other users from
using the RDMA device.
To prevent this problem, a solution is implemented where a special flow
table with two steering rules is created when a user creates a steering
anchor for the first time. The rules include one that drops all traffic
and another that points to the kernel flow table. If the steering anchor
is destroyed, only the rule pointing to the kernel's flow table is removed.
Any traffic reaching the special flow table after that is dropped.
Since the special flow table is not destroyed when the steering anchor is
destroyed, any issues are prevented from occurring. The remaining resources
are only destroyed when the RDMA device is destroyed, which happens after
all DEVX objects are freed, including the STCs, thus mitigating the issue.
Fixes: 0c6ab0ca9a ("RDMA/mlx5: Expose steering anchor to userspace")
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Link: https://lore.kernel.org/r/b4a88a871d651fa4e8f98d552553c1cfe9ba2cd6.1685960567.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This patch adds rtt_resp_dscp to the current debug controllability of
congestion control (CC) parameters.
rtt_resp_dscp can be read or written through debugfs.
If set, its value overwrites the DSCP of the generated RTT response.
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Link: https://lore.kernel.org/r/1dcc3440ee53c688f19f579a051ded81a2aaa70a.1676538714.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Synchronize the shared mlx5 branch with net:
- From Jiri: fixe a deadlock in mlx5_ib's netdev notifier unregister.
- From Mark and Patrisious: add IPsec RoCEv2 support.
- From Or: Rely on firmware to get special mkeys
* branch mlx5-next:
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
net/mlx5: Configure IPsec steering for egress RoCEv2 traffic
net/mlx5: Configure IPsec steering for ingress RoCEv2 traffic
net/mlx5: Add IPSec priorities in RDMA namespaces
net/mlx5: Implement new destination type TABLE_TYPE
net/mlx5: Introduce new destination type TABLE_TYPE
RDMA/mlx5: Track netdev to avoid deadlock during netdev notifier unregister
net/mlx5e: Propagate an internal event in case uplink netdev changes
net/mlx5e: Fix trap event handling
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Use query_sepcial_contexts to get the correct value of mkeys such as
null_mkey, terminate_scatter_list_mkey and dump_fill_mkey, as FW will
change them in certain configurations.
Link: https://lore.kernel.org/r/000236f0a9487d48809f87bcc3620a3964b2d3d3.1673960981.git.leon@kernel.org
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When removing a network namespace with mlx5 devlink instance being in
it, following callchain is performed:
cleanup_net (takes down_read(&pernet_ops_rwsem)
devlink_pernet_pre_exit()
devlink_reload()
mlx5_devlink_reload_down()
mlx5_unload_one_devl_locked()
mlx5_detach_device()
del_adev()
mlx5r_remove()
__mlx5_ib_remove()
mlx5_ib_roce_cleanup()
mlx5_remove_netdev_notifier()
unregister_netdevice_notifier (takes down_write(&pernet_ops_rwsem)
This deadlocks.
Resolve this by converting to register_netdevice_notifier_dev_net()
which does not take pernet_ops_rwsem and moves the notifier block around
according to netdev it takes as arg.
Use previously introduced netdev added/removed events to track uplink
netdev to be used for register_netdevice_notifier_dev_net() purposes.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
The non-cache mkeys are stored in the cache only to shorten restarting
application time. Don't store them longer than needed.
Configure cache entries that store non-cache MRs as temporary entries. If
30 seconds have passed and no user reclaimed the temporarily cached mkeys,
an asynchronous work will destroy the mkeys entries.
Link: https://lore.kernel.org/r/20230125222807.6921-7-michaelgur@nvidia.com
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently, when dereging an MR, if the mkey doesn't belong to a cache
entry, it will be destroyed. As a result, the restart of applications
with many non-cached mkeys is not efficient since all the mkeys are
destroyed and then recreated. This process takes a long time (for 100,000
MRs, it is ~20 seconds for dereg and ~28 seconds for re-reg).
To shorten the restart runtime, insert all cacheable mkeys to the cache.
If there is no fitting entry to the mkey properties, create a temporary
entry that fits it.
After a predetermined timeout, the cache entries will shrink to the
initial high limit.
The mkeys will still be in the cache when consuming them again after an
application restart. Therefore, the registration will be much faster
(for 100,000 MRs, it is ~4 seconds for dereg and ~5 seconds for re-reg).
The temporary cache entries created to store the non-cache mkeys are not
exposed through sysfs like the default cache entries.
Link: https://lore.kernel.org/r/20230125222807.6921-6-michaelgur@nvidia.com
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
