This patch is not the full fix and still causes to call traces
during mlx5_ib_dereg_mr().
This reverts commit f0ae4afe3d.
Fixes: f0ae4afe3d ("RDMA/mlx5: Fix releasing unallocated memory in dereg MR flow")
Link: https://lore.kernel.org/r/20211222101312.1358616-1-maorg@nvidia.com
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Acked-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
For the case of IB_MR_TYPE_DM the mr does doesn't have a umem, even though
it is a user MR. This causes function mlx5_free_priv_descs() to think that
it is a kernel MR, leading to wrongly accessing mr->descs that will get
wrong values in the union which leads to attempt to release resources that
were not allocated in the first place.
For example:
DMA-API: mlx5_core 0000:08:00.1: device driver tries to free DMA memory it has not allocated [device address=0x0000000000000000] [size=0 bytes]
WARNING: CPU: 8 PID: 1021 at kernel/dma/debug.c:961 check_unmap+0x54f/0x8b0
RIP: 0010:check_unmap+0x54f/0x8b0
Call Trace:
debug_dma_unmap_page+0x57/0x60
mlx5_free_priv_descs+0x57/0x70 [mlx5_ib]
mlx5_ib_dereg_mr+0x1fb/0x3d0 [mlx5_ib]
ib_dereg_mr_user+0x60/0x140 [ib_core]
uverbs_destroy_uobject+0x59/0x210 [ib_uverbs]
uobj_destroy+0x3f/0x80 [ib_uverbs]
ib_uverbs_cmd_verbs+0x435/0xd10 [ib_uverbs]
? uverbs_finalize_object+0x50/0x50 [ib_uverbs]
? lock_acquire+0xc4/0x2e0
? lock_acquired+0x12/0x380
? lock_acquire+0xc4/0x2e0
? lock_acquire+0xc4/0x2e0
? ib_uverbs_ioctl+0x7c/0x140 [ib_uverbs]
? lock_release+0x28a/0x400
ib_uverbs_ioctl+0xc0/0x140 [ib_uverbs]
? ib_uverbs_ioctl+0x7c/0x140 [ib_uverbs]
__x64_sys_ioctl+0x7f/0xb0
do_syscall_64+0x38/0x90
Fix it by reorganizing the dereg flow and mlx5_ib_mr structure:
- Move the ib_umem field into the user MRs structure in the union as it's
applicable only there.
- Function mlx5_ib_dereg_mr() will now call mlx5_free_priv_descs() only
in case there isn't udata, which indicates that this isn't a user MR.
Fixes: f18ec42231 ("RDMA/mlx5: Use a union inside mlx5_ib_mr")
Link: https://lore.kernel.org/r/66bb1dd253c1fd7ceaa9fc411061eefa457b86fb.1637581144.git.leonro@nvidia.com
Signed-off-by: Alaa Hleihel <alaa@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAmF/AjYeHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiG1hkIAJ6sFDbvb4M4LMwf
Slh2NVL9o5sLMBDzVwnVlyMSKDbMn1WBKreGssaLgZjGDc74lxsdSmw5l9MZm0JN
xlq95Q6XFiuu+0qDHPWwfDz3JFO4TqW2ZLLPWk9NnkNbRXqccSrlVRi1RpgE1t3/
NUtS8CQLu6A2BYMc6mkk3aV6IwSNKOkWbM5eBHSvU4j8B6lLbNQop0AfO/wyY1xB
U6LiVE1RpN/b7Yv+75ITtNzuHzVIBx6305FvSnOlKbMKKvIClt96Vd2OeuoEkK+6
wGU8JraB1+fc0GckAhynNrjWQWdvi0MAhFWWEJxjS20OGcV1rXDduNfkVNauO1Zn
+dNyJ3s=
=g9fz
-----END PGP SIGNATURE-----
Merge tag 'v5.15' into rdma.git for-next
Pull in the accepted for-rc patches as the next merge needs a newer base.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
A small series to clean up the mlx5 mkey code across the mlx5_core and
InfiniBand.
* branch 'mlx5_mkey':
RDMA/mlx5: Attach ndescs to mlx5_ib_mkey
RDMA/mlx5: Move struct mlx5_core_mkey to mlx5_ib
RDMA/mlx5: Replace struct mlx5_core_mkey by u32 key
RDMA/mlx5: Remove pd from struct mlx5_core_mkey
RDMA/mlx5: Remove size from struct mlx5_core_mkey
RDMA/mlx5: Remove iova from struct mlx5_core_mkey
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Generalize the use of ndescs by adding it to mlx5_ib_mkey.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Move mlx5_core_mkey struct to mlx5_ib, as the mlx5_core doesn't use it
at this point.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
In mlx5_core and vdpa there is no use of mlx5_core_mkey members except
for the key itself.
As preparation for moving mlx5_core_mkey to mlx5_ib, the occurrences of
struct mlx5_core_mkey in all modules except for mlx5_ib are replaced by
a u32 key.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
There is no read of mkey->pd, only write. Remove it.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
mkey->size is already stored in ibmr->length, no need to store it here.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
iova is already stored in ibmr->iova, no need to store it here.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Shay Drory <shayd@nvidia.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Currently, if a cache entry is empty, the driver will try to take MRs
from larger cache entries. This behavior consumes a lot of memory.
In addition, when searching for an mkey in an entry, the entry is locked.
When using a multithreaded application with the old behavior, the threads
will block each other more often, which can hurt performance as can be
seen in the table below.
Therefore, avoid it by creating a new mkey when the requested cache entry
is empty.
The test was performed on a machine with
Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz 44 cores.
Here are the time measures for allocating MRs of 2^6 pages. The search in
the cache started from entry 6.
+------------+---------------------+---------------------+
| | Old behavior | New behavior |
| +----------+----------+----------+----------+
| | 1 thread | 5 thread | 1 thread | 5 thread |
+============+==========+==========+==========+==========+
| 1,000 MRs | 14 ms | 30 ms | 14 ms | 80 ms |
+------------+----------+----------+----------+----------+
| 10,000 MRs | 135 ms | 6 sec | 173 ms | 880 ms |
+------------+----------+----------+----------+----------+
|100,000 MRs | 11.2 sec | 57 sec | 1.74 sec | 8.8 sec |
+------------+----------+----------+----------+----------+
Link: https://lore.kernel.org/r/71af2770c737b936f7b10f457f0ef303ffcf7ad7.1632644527.git.leonro@nvidia.com
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The XLT chunk alignment depends on ent_size not sizeof(ent_size) aka
sizeof(size_t). The incoming ent_size is either 8 or 16, so the
miscalculation when 16 is required is only an over-alignment and
functional harmless.
Fixes: 8010d74b99 ("RDMA/mlx5: Split the WR setup out of mlx5_ib_update_xlt()")
Link: https://lore.kernel.org/r/20210908081849.7948-2-schnelle@linux.ibm.com
Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In commit 8010d74b99 ("RDMA/mlx5: Split the WR setup out of
mlx5_ib_update_xlt()") the allocation logic was split out of
mlx5_ib_update_xlt() and the logic was changed to enable better OOM
handling. Sadly this change introduced a miscalculation of the number of
entries that were actually allocated when under memory pressure where it
can actually become 0 which on s390 lets dma_map_single() fail.
It can also lead to corruption of the free pages list when the wrong
number of entries is used in the calculation of sg->length which is used
as argument for free_pages().
Fix this by using the allocation size instead of misusing get_order(size).
Cc: stable@vger.kernel.org
Fixes: 8010d74b99 ("RDMA/mlx5: Split the WR setup out of mlx5_ib_update_xlt()")
Link: https://lore.kernel.org/r/20210908081849.7948-1-schnelle@linux.ibm.com
Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This allows using the normal sg_table APIs and makes all the code
cleaner. Remove sgt, nents and nmapd from ib_umem.
Link: https://lore.kernel.org/r/20210824142531.3877007-4-maorg@nvidia.com
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Fixing a typo that causes a cache entry to shrink immediately after adding
to it new MRs if the entry size exceeds the high limit. In doing so, the
cache misses its purpose to prevent the creation of new mkeys on the
runtime by using the cached ones.
Fixes: b9358bdbc7 ("RDMA/mlx5: Fix locking in MR cache work queue")
Link: https://lore.kernel.org/r/fcb546986be346684a016f5ca23a0567399145fa.1627370131.git.leonro@nvidia.com
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAmDPuyMeHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGvxgH/RKvSuRPwkJ2Jcp9
VLi5kCbqtJlYLq6tB6peSJ8otKgxkcRwC0pIY4LlYIAWYboktLQ5RKp/9nB2h2FN
aMZUMu6AI/lVJyFMI5MnKnJIUiUq+WXR3lSSlw68vwFLFdzqUZFNq+bqeiVvnIy1
yqA6naj24Tu/RbYffQoPvdSJcU2SLXRMxwD8HRGiU2d51RaFsOvsZvF+P5TVcsEV
ZmttJeER21CaI/A809eqaFmyGrUOcZZK9roZEbMwanTZOMw18biEsLu/UH4kBX01
JC4+RlGxcWjQ5YNZgChsgoOK/CHzc6ITztTntdeDWAvwZjQFzV7pCy4/3BWne3O+
5178yHM=
=o8cN
-----END PGP SIGNATURE-----
Merge tag 'v5.13-rc7' into rdma.git for-next
Linux 5.13-rc7
Needed for dependencies in following patches. Merge conflict in rxe_cmop.c
resolved by compining both patches.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Relaxed Ordering is a capability that can only benefit users that support
it. All kernel ULPs should support Relaxed Ordering, as they are designed
to read data only after observing the CQE and use the DMA API correctly.
Hence, implicitly enable Relaxed Ordering by default for MR transfers in
kernel ULPs.
Link: https://lore.kernel.org/r/b7e820aab7402b8efa63605f4ea465831b3b1e5e.1623236426.git.leonro@nvidia.com
Signed-off-by: Avihai Horon <avihaih@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The old version of ib_umem_get() need these udata as a parameter but now
they are unnecessary.
Fixes: c320e527e1 ("IB: Allow calls to ib_umem_get from kernel ULPs")
Link: https://lore.kernel.org/r/1620807142-39157-4-git-send-email-liweihang@huawei.com
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Acked-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
mlx5_core_dev holds pointer to static profile, hence when the
log_max_qp of the profile is override by some device, then it
effect all other mlx5 devices that share the same profile.
Fix it by having a profile instance for every mlx5 device.
Fixes: 883371c453 ("net/mlx5: Check FW limitations on log_max_qp before setting it")
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Maor Gottlieb says:
====================
This series from Maor extends MEMIC to support atomic operations from the
host in addition to already supported regular read/write.
====================
* 'memic_ops':
RDMA/mlx5: Expose UAPI to query DM
RDMA/mlx5: Add support in MEMIC operations
RDMA/mlx5: Add support to MODIFY_MEMIC command
RDMA/mlx5: Re-organize the DM code
RDMA/mlx5: Move all DM logic to separate file
RDMA/uverbs: Make UVERBS_OBJECT_METHODS to consider line number
net/mlx5: Add MEMIC operations related bits
Move all device memory related code to a separate file.
Link: https://lore.kernel.org/r/20210411122924.60230-4-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
To update xlt (during mlx5_ib_reg_user_mr()), the driver can request up to
1 MB (order-8) memory, depending on the size of the MR. This costly
allocation can sometimes take very long to return (a few seconds). This
causes the calling application to hang for a long time, especially when
the system is fragmented. To avoid these long latency spikes, the calls
the higher order allocations need to fail faster in case they are not
available.
In order to acheive this we need __GFP_NORETRY flag in the gfp_mask before
during fetching the free pages. Allow the algorithm to automatically fall
back to smaller page sizes.
Link: https://lore.kernel.org/r/1617425635-35631-1-git-send-email-praveen.kannoju@oracle.com
Signed-off-by: Praveen Kumar Kannoju <praveen.kannoju@oracle.com>
Acked-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
There is no need to create the ODP EQ if the user doesn't use ODP MRs.
Hence, create it only when the first ODP MR is created. This EQ will be
destroyed only when the device is unloaded.
This will decrease the number of EQs created per device. for example: If
we creates 1K devices (SF/VF/etc'), than we will decrease the num of EQs
by 1K.
Link: https://lore.kernel.org/r/20210314125418.179716-1-leon@kernel.org
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Fix the following smatch error:
drivers/infiniband/hw/mlx5/mr.c:1950 mlx5_ib_dereg_mr() error: uninitialized symbol 'rc'.
Fixes: e6fb246cca ("RDMA/mlx5: Consolidate MR destruction to mlx5_ib_dereg_mr()")
Link: https://lore.kernel.org/r/20210314082250.10143-1-leon@kernel.org
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Now that this is only used in a few places in mr.c give it a sensible
name. It has nothing to do with the cache and can be invoked on any
MR. DMA is stopped and the user cannot touch the MR any further once it
completes.
Link: https://lore.kernel.org/r/20210304120745.1090751-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Now that the SRCU stuff has been removed the entire MR destroy logic can
be made a lot simpler. Currently there are many different ways to destroy a
MR and it makes it really hard to do this task correctly. Route all
destruction through mlx5_ib_dereg_mr() and make it work for all
situations.
Since it turns out all the different MR types do basically the same thing
this removes a lot of knowledge of MR internals from ODP and leaves ODP
just exporting an operation to clean up children.
This fixes a few weird corner cases bugs and firmly uses the correct
ordering of the MR destruction:
- Stop parallel access to the mkey via the ODP xarray
- Stop DMA
- Release the umem
- Clean up ODP children
- Free/Recycle the MR
Link: https://lore.kernel.org/r/20210304120745.1090751-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The struct mlx5_ib_mr can be used for three different things, but only one
at a time:
- In the user MR cache
- As a kernel MR
- As a user MR
Overlay the three things into a single union with the following rules:
- If the mr is found on the cache_ent->head list then it is a cache MR
and umem == NULL. The entire union is zero after the MR is removed from
the cache.
- If umem != NULL or type == IB_MR_TYPE_USER then it is a user MR.
- If umem == NULL then it is a kernel MR
This reduces the size of struct mlx5_ib_mr to 552 bytes from 702.
The only place the three flows overlap in the code is during dereg, so add
a few extra checks along there.
Link: https://lore.kernel.org/r/20210304120745.1090751-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
All of the ODP code assumes when it calls mlx5_mr_cache_alloc() the ODP
related fields are zero'd. This is true if the MR was just allocated, but
if the MR is recycled through the cache then the values are never zero'd.
This causes a bug in the odp_stats, they don't reset when the MR is
reallocated, also is_odp_implicit is never 0'd.
So we can use memset on a block of the mlx5_ib_mr reorganize the structure
to put all the data that can be zero'd by the cache at the end.
It is organized as an anonymous struct because the next patch will make
this a union.
Delete the unused smr_info. Don't set the kernel only desc_size on the
user path. No longer any need to zero mr->parent before freeing it, the
memset() will get it now.
Fixes: a3de94e3d6 ("IB/mlx5: Introduce ODP diagnostic counters")
Link: https://lore.kernel.org/r/20210304120745.1090751-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Cleanup the synchronize_srcu() from the ODP flow as it was found to be a
very heavy time consumer as part of dereg_mr.
For example de-registration of 10000 ODP MRs each with size of 2M hugepage
took 19.6 sec comparing de-registration of same number of non ODP MRs that
took 172 ms.
The new locking scheme uses the wait_event() mechanism which follows the
use count of the MR instead of using synchronize_srcu().
By that change, the time required for the above test took 95 ms which is
even better than the non ODP flow.
Once fully dropped the srcu usage, had to come with a lock to protect the
XA access.
As part of using the above mechanism we could also clean the
num_deferred_work stuff and follow the use count instead.
Link: https://lore.kernel.org/r/20210202071309.2057998-1-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Implement the new driver method 'reg_user_mr_dmabuf'. Utilize the core
functions to import dma-buf based memory region and update the mappings.
Add code to handle dma-buf related page fault.
Link: https://lore.kernel.org/r/1608067636-98073-5-git-send-email-jianxin.xiong@intel.com
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
If the MR cache entry invalidation failed, then we detach this entry from
the cache, therefore we must to free the memory as well.
Allcation backtrace for the leaker:
[<00000000d8e423b0>] alloc_cache_mr+0x23/0xc0 [mlx5_ib]
[<000000001f21304c>] create_cache_mr+0x3f/0xf0 [mlx5_ib]
[<000000009d6b45dc>] mlx5_ib_alloc_implicit_mr+0x41/0×210 [mlx5_ib]
[<00000000879d0d68>] mlx5_ib_reg_user_mr+0x9e/0×6e0 [mlx5_ib]
[<00000000be74bf89>] create_qp+0x2fc/0xf00 [ib_uverbs]
[<000000001a532d22>] ib_uverbs_handler_UVERBS_METHOD_COUNTERS_READ+0x1d9/0×230 [ib_uverbs]
[<0000000070f46001>] rdma_alloc_commit_uobject+0xb5/0×120 [ib_uverbs]
[<000000006d8a0b38>] uverbs_alloc+0x2b/0xf0 [ib_uverbs]
[<00000000075217c9>] ksysioctl+0x234/0×7d0
[<00000000eb5c120b>] __x64_sys_ioctl+0x16/0×20
[<00000000db135b48>] do_syscall_64+0x59/0×2e0
Fixes: 1769c4c575 ("RDMA/mlx5: Always remove MRs from the cache before destroying them")
Link: https://lore.kernel.org/r/20201213132940.345554-2-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently, DM MR registration flow doesn't set the mlx5_ib_dev pointer and
can cause a NULL pointer dereference if userspace dumps the MR via rdma
tool.
Assign the IB device together with the other fields and remove the
redundant reference of mlx5_ib_dev from mlx5_ib_mr.
Cc: stable@vger.kernel.org
Fixes: 6c29f57ea4 ("IB/mlx5: Device memory mr registration support")
Link: https://lore.kernel.org/r/20201203190807.127189-1-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This is all a giant train wreck of error handling, in many cases the MR is
left in some corrupted state where continuing on is going to lead to
chaos, or various unwinds/order is missed.
rereg had three possible completely different actions, depending on flags
and various details about the MR. Split the three actions into three
functions, and call the right action from the start.
For each action carefully design the error handling to fit the action:
- UMR access/PD update is a simple UMR, if it fails the MR isn't changed,
so do nothing
- PAS update over UMR is multiple UMR operations. To keep everything sane
revoke access to the MKey while it is being changed and restore it once
the MR is correct.
- Recreating the mkey should completely build a parallel MR with a fully
loaded PAS then swap and destroy the old one. If it fails the original
should be left untouched. This is handled in the core code. Directly
call the normal MR creation functions, possibly re-using the existing
umem.
Add support for working with ODP MRs. The READ/WRITE access flags can be
changed by UMR and we can trivially convert to/from ODP MRs using the
logic to build a completely new MR.
This new logic also fixes various problems with MRs continuing to work
while their PAS lists are no longer valid, eg during a page size change.
Link: https://lore.kernel.org/r/20201130075839.278575-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This function handles an ODP and regular MR flow all mushed together, even
though the two flows are quite different. Split them into two dedicated
functions.
Link: https://lore.kernel.org/r/20201130075839.278575-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
mlx5 has an ugly flow where it tries to allocate a new MR and replace the
existing MR in the same memory during rereg. This is very complicated and
buggy. Instead of trying to replace in-place inside the driver, provide
support from uverbs to change the entire HW object assigned to a handle
during rereg_mr.
Since destroying a MR is allowed to fail (ie if a MW is pointing at it)
and can't be detected in advance, the algorithm creates a completely new
uobject to hold the new MR and swaps the IDR entries of the two objects.
The old MR in the temporary IDR entry is destroyed, and if it fails
rereg_mr succeeds and destruction is deferred to FD release. This
complexity is why this cannot live in a driver safely.
Link: https://lore.kernel.org/r/20201130075839.278575-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
DMA operation of the IB device is done using ib_device->dma_device.
Instead of accessing parent of the IB device, use the PCI dma device which
is setup to ib_device->dma_device during IB device registration.
Link: https://lore.kernel.org/r/20201125064628.8431-1-leon@kernel.org
Signed-off-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Now that all the PAS arrays or UMR XLT's for mkcs are filled using
rdma_for_each_block() we can use the common ib_umem_find_best_pgsz()
algorithm.
Link: https://lore.kernel.org/r/20201026132314.1336717-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Mixing these together is just a mess, make a dedicated version,
mlx5_ib_update_mr_pas(), which directly loads the whole MTT for a non-ODP
MR.
The split out version can trivially use a simple loop with
rdma_for_each_block() which allows using the core code to compute the MR
pages and avoids seeking in the SGL list after each chunk as the
__mlx5_ib_populate_pas() call required.
Significantly speeds loading large MTTs.
Link: https://lore.kernel.org/r/20201026132314.1336717-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The memory allocation is quite complicated, and makes this function hard
to understand. Refactor things so that a function call sets up the WR, SG,
DMA mapping and buffer, further splitting that into buffer and DMA/wr.
This also slightly changes the buffer allocation logic to try an order 0
page allocation (with OOM warnings on) before going to the emergency page.
Link: https://lore.kernel.org/r/20201026132314.1336717-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This routine converts the umem SGL into a list of fixed pages for DMA,
which is exactly what rdma_umem_for_each_dma_block() is for, use the
common code directly.
Link: https://lore.kernel.org/r/20201026132314.1336717-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This is the same as ib_umem_num_dma_blocks(umem, 1UL << page_shift), have
the callers compute it directly.
Link: https://lore.kernel.org/r/20201026131936.1335664-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Only alloc_mr_from_cache() needs order and can trivially compute it, so
lift it to the one call site and remove the NULL arguments.
Link: https://lore.kernel.org/r/20201026131936.1335664-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
For the user MR path, instead of calling this after getting the umem, call
it as part of creating the struct mlx5_ib_mr and distill its output to a
single page_shift stored inside the mr.
This avoids passing around the tuple of its output. Based on the umem and
page_shift, the output arguments can be computed using:
count == ib_umem_num_pages(mr->umem)
shift == mr->page_shift
ncont == ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift)
order == order_base_2(ncont)
And since mr->page_shift == umem_odp->page_shift then ncont ==
ib_umem_num_dma_blocks() == ib_umem_odp_num_pages() for ODP umems.
Link: https://lore.kernel.org/r/20201026131936.1335664-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
reg_pages should always contain mr->npage since when the mr is finally
de-reg'd it is always subtracted out.
If there were any error exits then mlx5_ib_rereg_user_mr() would leave the
reg_pages adjusted and this will cause it to be double subtracted
eventually.
The manipulation of reg_pages is inherently connected to the umem, so lift
it out of set_mr_fields() and only adjust it around creating/destroying a
umem.
reg_pages is only used for diagnostics in sysfs.
Fixes: 7d0cc6edcc ("IB/mlx5: Add MR cache for large UMR regions")
Link: https://lore.kernel.org/r/20201026131936.1335664-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The is only ever set to non-zero if the MR is from the cache, and if it is
cached then the order is in cached_ent->order.
Make it clearer that use_umr_mtt_update() only returns true for cached MRs
and remove the redundant data.
Link: https://lore.kernel.org/r/20201026131936.1335664-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
