Leon Romanovsky says:
====================
This series fixes various corner cases in the mlx5_ib MR cache
implementation, see specific commit messages for more information.
====================
Based on the mlx5-next branch at
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux
Due to dependencies
* branch 'mlx5_mr-cache':
RDMA/mlx5: Allow MRs to be created in the cache synchronously
RDMA/mlx5: Revise how the hysteresis scheme works for cache filling
RDMA/mlx5: Fix locking in MR cache work queue
RDMA/mlx5: Lock access to ent->available_mrs/limit when doing queue_work
RDMA/mlx5: Fix MR cache size and limit debugfs
RDMA/mlx5: Always remove MRs from the cache before destroying them
RDMA/mlx5: Simplify how the MR cache bucket is located
RDMA/mlx5: Rename the tracking variables for the MR cache
RDMA/mlx5: Replace spinlock protected write with atomic var
{IB,net}/mlx5: Move asynchronous mkey creation to mlx5_ib
{IB,net}/mlx5: Assign mkey variant in mlx5_ib only
{IB,net}/mlx5: Setup mkey variant before mr create command invocation
There are many bad APIs here that are accepting a cache bucket index
instead of a bucket pointer. Many of the callers already have a bucket
pointer, so this results in a lot of confusing uses of order2idx().
Pass the struct mlx5_cache_ent into add_keys(), remove_keys(), and
alloc_cached_mr().
Once the MR is in the cache, store the cache bucket pointer directly in
the MR, replacing the 'bool allocated_from cache'.
In the end there is only one place that needs to form index from order,
alloc_mr_from_cache(). Increase the safety of this function by disallowing
it from accessing cache entries in the ODP special area.
Link: https://lore.kernel.org/r/20200310082238.239865-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Following race may occur because of the call_srcu and the placement of
the synchronize_srcu vs the xa_erase.
CPU0 CPU1
mlx5_ib_free_implicit_mr: destroy_unused_implicit_child_mr:
xa_erase(odp_mkeys)
synchronize_srcu()
xa_lock(implicit_children)
if (still in xarray)
atomic_inc()
call_srcu()
xa_unlock(implicit_children)
xa_erase(implicit_children):
xa_lock(implicit_children)
__xa_erase()
xa_unlock(implicit_children)
flush_workqueue()
[..]
free_implicit_child_mr_rcu:
(via call_srcu)
queue_work()
WARN_ON(atomic_read())
[..]
free_implicit_child_mr_work:
(via wq)
free_implicit_child_mr()
mlx5_mr_cache_invalidate()
mlx5_ib_update_xlt() <-- UMR QP fail
atomic_dec()
The wait_event() solves the race because it blocks until
free_implicit_child_mr_work() completes.
Fixes: 5256edcb98 ("RDMA/mlx5: Rework implicit ODP destroy")
Link: https://lore.kernel.org/r/20200227113918.94432-1-leon@kernel.org
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.
-----BEGIN PGP SIGNATURE-----
iHUEABYIAB0WIQT1m3YD37UfMCUQBNwp8NhrnBAZsQUCXiVO8AAKCRAp8NhrnBAZ
scTrAP9gb0d3qv0IOtHw5aGI1DAgjTUn/SzUOnsjDEn7DIoh9gEA2+ZmaEyLXKrl
+UcZb31auy5P8ueJYokRLhLAyRcOIAg=
=yaHb
-----END PGP SIGNATURE-----
Merge tag 'rds-odp-for-5.5' into rdma.git for-next
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>
Enable relaxed ordering in the mkey context when requested. As relaxed
ordering is not currently supported in UMR, disable UMR usage for relaxed
ordering MRs.
Link: https://lore.kernel.org/r/1578506740-22188-11-git-send-email-yishaih@mellanox.com
Signed-off-by: Michael Guralnik <michaelgur@mellanox.com>
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Till recently it was not possible for userspace to specify a different
IOVA, but with the new ibv_reg_mr_iova() library call this can be done.
To compute the user_va we must compute:
user_va = (iova - iova_start) + user_va_start
while being cautious of overflow and other math problems.
The iova is not reliably stored in the mmkey when the MR is created. Only
the cached creation path (the common one) set it, so it must also be set
when creating uncached MRs.
Fix the weird use of iova when computing the starting page index in the
MR. In the normal case, when iova == umem.address:
iova & (~(BIT(page_shift) - 1)) ==
ALIGN_DOWN(umem.address, odp->page_size) ==
ib_umem_start(odp)
And when iova is different using it in math with a user_va is wrong.
Finally, do not allow an implicit ODP to be created with a non-zero IOVA
as we have no support for that.
Fixes: 7bdf65d411 ("IB/mlx5: Handle page faults")
Signed-off-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
One of the steps in ODP page fault handler for WQEs is to read a WQE
from a QP send queue or receive queue buffer at a specific index.
Since the implementation of this buffer is different between kernel and
user QP the implementation of the handler needs to be aware of that and
handle it in a different way.
ODP for kernel MRs is currently supported only for RDMA_READ
and RDMA_WRITE operations so change the handler to
- read a WQE from a kernel QP send queue
- fail if access to receive queue or shared receive queue is
required for a kernel QP
Signed-off-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
So far the assumption was that ib_umem_get() and ib_umem_odp_get()
are called from flows that start in UVERBS and therefore has a user
context. This assumption restricts flows that are initiated by ULPs
and need the service that ib_umem_get() provides.
This patch changes ib_umem_get() and ib_umem_odp_get() to get IB device
directly by relying on the fact that both UVERBS and ULPs sets that
field correctly.
Reviewed-by: Guy Levi <guyle@mellanox.com>
Signed-off-by: Moni Shoua <monis@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Building MR translation table in the ODP case requires additional
flexibility, namely random access to DMA addresses. Make both direct and
indirect ODP MR use same code path, separated from the non-ODP MR code
path.
With the restructuring the correct page_shift is now used around
__mlx5_ib_populate_pas().
Fixes: d2183c6f19 ("RDMA/umem: Move page_shift from ib_umem to ib_odp_umem")
Link: https://lore.kernel.org/r/20191222124649.52300-2-leon@kernel.org
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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
-----BEGIN PGP SIGNATURE-----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=Hj8a
-----END PGP SIGNATURE-----
Merge tag 'for-linus-hmm' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
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
Replace the internal interval tree based mmu notifier with the new common
mmu_interval_notifier_insert() API. This removes a lot of code and fixes a
deadlock that can be triggered in ODP:
zap_page_range()
mmu_notifier_invalidate_range_start()
[..]
ib_umem_notifier_invalidate_range_start()
down_read(&per_mm->umem_rwsem)
unmap_single_vma()
[..]
__split_huge_page_pmd()
mmu_notifier_invalidate_range_start()
[..]
ib_umem_notifier_invalidate_range_start()
down_read(&per_mm->umem_rwsem) // DEADLOCK
mmu_notifier_invalidate_range_end()
up_read(&per_mm->umem_rwsem)
mmu_notifier_invalidate_range_end()
up_read(&per_mm->umem_rwsem)
The umem_rwsem is held across the range_start/end as the ODP algorithm for
invalidate_range_end cannot tolerate changes to the interval
tree. However, due to the nested invalidation regions the second
down_read() can deadlock if there are competing writers. The new core code
provides an alternative scheme to solve this problem.
Fixes: ca748c39ea ("RDMA/umem: Get rid of per_mm->notifier_count")
Link: https://lore.kernel.org/r/20191112202231.3856-6-jgg@ziepe.ca
Tested-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Jason Gunthorpe says:
====================
In order to hoist the interval tree code out of the drivers and into the
mmu_notifiers it is necessary for the drivers to not use the interval tree
for other things.
This series replaces the interval tree with an xarray and along the way
re-aligns all the locking to use a sensible SRCU model where the 'update'
step is done by modifying an xarray.
The result is overall much simpler and with less locking in the critical
path. Many functions were reworked for clarity and small details like
using 'imr' to refer to the implicit MR make the entire code flow here
more readable.
This also squashes at least two race bugs on its own, and quite possibily
more that haven't been identified.
====================
Merge conflicts with the odp statistics patch resolved.
* branch 'odp_rework':
RDMA/odp: Remove broken debugging call to invalidate_range
RDMA/mlx5: Do not race with mlx5_ib_invalidate_range during create and destroy
RDMA/mlx5: Do not store implicit children in the odp_mkeys xarray
RDMA/mlx5: Rework implicit ODP destroy
RDMA/mlx5: Avoid double lookups on the pagefault path
RDMA/mlx5: Reduce locking in implicit_mr_get_data()
RDMA/mlx5: Use an xarray for the children of an implicit ODP
RDMA/mlx5: Split implicit handling from pagefault_mr
RDMA/mlx5: Set the HW IOVA of the child MRs to their place in the tree
RDMA/mlx5: Lift implicit_mr_alloc() into the two routines that call it
RDMA/mlx5: Rework implicit_mr_get_data
RDMA/mlx5: Delete struct mlx5_priv->mkey_table
RDMA/mlx5: Use a dedicated mkey xarray for ODP
RDMA/mlx5: Split sig_err MR data into its own xarray
RDMA/mlx5: Use SRCU properly in ODP prefetch
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
For creation, as soon as the umem_odp is created the notifier can be
called, however the underlying MR may not have been setup yet. This would
cause problems if mlx5_ib_invalidate_range() runs. There is some
confusing/ulocked/racy code that might by trying to solve this, but
without locks it isn't going to work right.
Instead trivially solve the problem by short-circuiting the invalidation
if there are not yet any DMA mapped pages. By definition there is nothing
to invalidate in this case.
The create code will have the umem fully setup before anything is DMA
mapped, and npages is fully locked by the umem_mutex.
For destroy, invalidate the entire MR at the HW to stop DMA then DMA unmap
the pages before destroying the MR. This drives npages to zero and
prevents similar racing with invalidate while the MR is undergoing
destruction.
Arguably it would be better if the umem was created after the MR and
destroyed before, but that would require a big rework of the MR code.
Fixes: 6aec21f6a8 ("IB/mlx5: Page faults handling infrastructure")
Link: https://lore.kernel.org/r/20191009160934.3143-15-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
These mkeys are entirely internal and are never used by the HW for
page fault. They should also never be used by userspace for prefetch.
Simplify & optimize things by not including them in the xarray.
Since the prefetch path can now never see a child mkey there is no need
for the second synchronize_srcu() during imr destroy.
Link: https://lore.kernel.org/r/20191009160934.3143-14-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Use SRCU in a sensible way by removing all MRs in the implicit tree from
the two xarrays (the update operation), then a synchronize, followed by a
normal single threaded teardown.
This is only a little unusual from the normal pattern as there can still
be some work pending in the unbound wq that may also require a workqueue
flush. This is tracked with a single atomic, consolidating the redundant
existing atomics and wait queue.
For understand-ability the entire ODP implicit create/destroy flow now
largely exists in a single pair of functions within odp.c, with a few
support functions for tearing down an unused child.
Link: https://lore.kernel.org/r/20191009160934.3143-13-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that the locking is simplified combine pagefault_implicit_mr() with
implicit_mr_get_data() so that we sweep over the idx range only once,
and do the single xlt update at the end, after the child umems are
setup.
This avoids double iteration/xa_loads plus the sketchy failure path if the
xa_load() fails.
Link: https://lore.kernel.org/r/20191009160934.3143-12-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that the child MRs are stored in an xarray we can rely on the SRCU
lock to protect the xa_load and use xa_cmpxchg on the slow allocation path
to resolve races with concurrent page fault.
This reduces the scope of the critical section of umem_mutex for implicit
MRs to only cover mlx5_ib_update_xlt, and avoids taking a lock at all if
the child MR is already in the xarray. This makes it consistent with the
normal ODP MR critical section for umem_lock, and the locking approach
used for destroying an unusued implicit child MR.
The MLX5_IB_UPD_XLT_ATOMIC is no longer needed in implicit_get_child_mr()
since it is no longer called with any locks.
Link: https://lore.kernel.org/r/20191009160934.3143-11-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Currently the child leaves are stored in the shared interval tree and
every lookup for a child must be done under the interval tree rwsem.
This is further complicated by dropping the rwsem during iteration (ie the
odp_lookup(), odp_next() pattern), which requires a very tricky an
difficult to understand locking scheme with SRCU.
Instead reserve the interval tree for the exclusive use of the mmu
notifier related code in umem_odp.c and give each implicit MR a xarray
containing all the child MRs.
Since the size of each child is 1GB of VA, a 1 level xarray will index 64G
of VA, and a 2 level will index 2TB, making xarray a much better
data structure choice than an interval tree.
The locking properties of xarray will be used in the next patches to
rework the implicit ODP locking scheme into something simpler.
At this point, the xarray is locked by the implicit MR's umem_mutex, and
read can also be locked by the odp_srcu.
Link: https://lore.kernel.org/r/20191009160934.3143-10-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The single routine has a very confusing scheme to advance to the next
child MR when working on an implicit parent. This scheme can only be used
when working with an implicit parent and must not be triggered when
working on a normal MR.
Re-arrange things by directly putting all the single-MR stuff into one
function and calling it in a loop for the implicit case. Simplify some of
the error handling in the new pagefault_real_mr() to remove unneeded gotos.
Link: https://lore.kernel.org/r/20191009160934.3143-9-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Instead of rewriting all the IOVA's to 0 as things progress down the tree
make the IOVA of the children equal to placement in the tree. This makes
things easier to understand by keeping mmkey.iova == HW configuration.
Link: https://lore.kernel.org/r/20191009160934.3143-8-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This makes the routines easier to understand, particularly with respect
the locking requirements of the entire sequence. The implicit_mr_alloc()
had a lot of ifs specializing it to each of the callers, and only a very
small amount of code was actually shared.
Following patches will cause the flow in the two functions to diverge
further.
Link: https://lore.kernel.org/r/20191009160934.3143-7-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This function is intended to loop across each MTT chunk in the implicit
parent that intersects the range [io_virt, io_virt+bnct). But it is has a
confusing construction, so:
- Consistently use imr and odp_imr to refer to the implicit parent
to avoid confusion with the normal mr and odp of the child
- Directly compute the inclusive start/end indexes by shifting. This is
clearer to understand the intent and avoids any errors from unaligned
values of addr
- Iterate directly over the range of MTT indexes, do not make a loop
out of goto
- Follow 'success oriented flow', with goto error unwind
- Directly calculate the range of idx's that need update_xlt
- Ensure that any leaf MR added to the interval tree always results in an
update to the XLT
Link: https://lore.kernel.org/r/20191009160934.3143-6-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
There is a per device xarray storing mkeys that is used to store every
mkey in the system. However, this xarray is now only read by ODP for
certain ODP designated MRs (ODP, implicit ODP, MW, DEVX_INDIRECT).
Create an xarray only for use by ODP, that only contains ODP related
MKeys. This xarray is protected by SRCU and all erases are protected by a
synchronize.
This improves performance:
- All MRs in the odp_mkeys xarray are ODP MRs, so some tests for is_odp()
can be deleted. The xarray will also consume fewer nodes.
- normal MR's are never mixed with ODP MRs in a SRCU data structure so
performance sucking synchronize_srcu() on every MR destruction is not
needed.
- No smp_load_acquire(live) and xa_load() double barrier on read
Due to the SRCU locking scheme care must be taken with the placement of
the xa_store(). Once it completes the MR is immediately visible to other
threads and only through a xa_erase() & synchronize_srcu() cycle could it
be destroyed.
Link: https://lore.kernel.org/r/20191009160934.3143-4-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When working with SRCU protected xarrays the xarray itself should be the
SRCU 'update' point. Instead prefetch is using live as the SRCU update
point and this prevents switching the locking design to use the xarray
instead.
To solve this the prefetch must only read from the xarray once, and hold
on to the actual MR pointer for the duration of the async
operation. Incrementing num_pending_prefetch delays destruction of the MR,
so it is suitable.
Prefetch calls directly to the pagefault_mr using the MR pointer and only
does a single xarray lookup.
All the testing if a MR is prefetchable or not is now done only in the
prefetch code and removed from the pagefault critical path.
Link: https://lore.kernel.org/r/20191009160934.3143-2-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAl210Z8eHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGv+kIAKRpO7EuDokQL4qp
hxEEaCMJA1T055EMlNU6FVAq/ZbmapzreUyNYiRMpPWKGTWNMkhIcZQfysYeGZz5
y/KRxAiVxlcB+3v3yRmoZd/XoQmhgvJmqD4zhaGI2Utonow4f/SGSEFFZqqs9WND
4HJROjZHgQ4JBxg9Z+QMo0FxbV/DCZpEOUq51N9WJywyyDRb18zotE83stpU+pE2
fjqT7mk0NLrnYXuDRAbFC1Aau9ed4H6LlwLmxaqxq/Pt5Rz7wIKwKL9HIT4Dm/0a
qpani6phhHWL7MwUpa2wkEonFCD03rJFl3DUVJo64Ijh4up5D/jyXQ+GKV2P4WKJ
275Rb5Q=
=WiZZ
-----END PGP SIGNATURE-----
Merge tag 'v5.4-rc5' into rdma.git for-next
Linux 5.4-rc5
For dependencies in the next patches
Conflict resolved by keeping the delete of the unlock.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Provide an ODP explicit/implicit type as part of 'rdma -dd resource show
mr' dump.
For example:
$ rdma -dd resource show mr
dev mlx5_0 mrn 1 rkey 0xa99a lkey 0xa99a mrlen 50000000
pdn 9 pid 7372 comm ibv_rc_pingpong drv_odp explicit
For non-ODP MRs, we won't print "drv_odp ..." at all.
Link: https://lore.kernel.org/r/20191016062308.11886-4-leon@kernel.org
Signed-off-by: Erez Alfasi <ereza@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>
Introduce ODP diagnostic counters and count the following
per MR within IB/mlx5 driver:
1) Page faults:
Total number of faulted pages.
2) Page invalidations:
Total number of pages invalidated by the OS during all
invalidation events. The translations can be no longer
valid due to either non-present pages or mapping changes.
Link: https://lore.kernel.org/r/20191016062308.11886-2-leon@kernel.org
Signed-off-by: Erez Alfasi <ereza@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>
live is used to signal to the pagefault thread that the MR is initialized
and ready for use. It should be after the umem is assigned and all other
setup is completed. This prevents races (at least) of the form:
CPU0 CPU1
mlx5_ib_alloc_implicit_mr()
implicit_mr_alloc()
live = 1
imr->umem = umem
num_pending_prefetch_inc()
if (live)
atomic_inc(num_pending_prefetch)
atomic_set(num_pending_prefetch,0) // Overwrites other thread's store
Further, live is being used with SRCU as the 'update' in an
acquire/release fashion, so it can not be read and written raw.
Move all live = 1's to after MR initialization is completed and use
smp_store_release/smp_load_acquire() for manipulating it.
Add a missing live = 0 when an implicit MR child is deleted, before
queuing work to do synchronize_srcu().
The barriers in update_odp_mr() were some broken attempt to create a
acquire/release, but were not even applied consistently and missed the
point, delete it as well.
Fixes: 6aec21f6a8 ("IB/mlx5: Page faults handling infrastructure")
Link: https://lore.kernel.org/r/20191001153821.23621-6-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This fixes a race of the form:
CPU0 CPU1
mlx5_ib_invalidate_range() mlx5_ib_invalidate_range()
// This one actually makes npages == 0
ib_umem_odp_unmap_dma_pages()
if (npages == 0 && !dying)
// This one does nothing
ib_umem_odp_unmap_dma_pages()
if (npages == 0 && !dying)
dying = 1;
dying = 1;
schedule_work(&umem_odp->work);
// Double schedule of the same work
schedule_work(&umem_odp->work); // BOOM
npages and dying must be read and written under the umem_mutex lock.
Since whenever ib_umem_odp_unmap_dma_pages() is called mlx5 must also call
mlx5_ib_update_xlt, and both need to be done in the same locking region,
hoist the lock out of unmap.
This avoids an expensive double critical section in
mlx5_ib_invalidate_range().
Fixes: 81713d3788 ("IB/mlx5: Add implicit MR support")
Link: https://lore.kernel.org/r/20191001153821.23621-4-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
mlx5_ib_update_xlt() must be protected against parallel free of the MR it
is accessing, also it must be called single threaded while updating the
HW. Otherwise we can have races of the form:
CPU0 CPU1
mlx5_ib_update_xlt()
mlx5_odp_populate_klm()
odp_lookup() == NULL
pklm = ZAP
implicit_mr_get_data()
implicit_mr_alloc()
<update interval tree>
mlx5_ib_update_xlt
mlx5_odp_populate_klm()
odp_lookup() != NULL
pklm = VALID
mlx5_ib_post_send_wait()
mlx5_ib_post_send_wait() // Replaces VALID with ZAP
This can be solved by putting both the SRCU and the umem_mutex lock around
every call to mlx5_ib_update_xlt(). This ensures that the content of the
interval tree relavent to mlx5_odp_populate_klm() (ie mr->parent == mr)
will not change while it is running, and thus the posted WRs to update the
KLM will always reflect the correct information.
The race above will resolve by either having CPU1 wait till CPU0 completes
the ZAP or CPU0 will run after the add and instead store VALID.
The pagefault path adding children already holds the umem_mutex and SRCU,
so the only missed lock is during MR destruction.
Fixes: 81713d3788 ("IB/mlx5: Add implicit MR support")
Link: https://lore.kernel.org/r/20191001153821.23621-3-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
There is no reason to call return at the end of function which returns
void. Remove this unnecessary statement.
Link: https://lore.kernel.org/r/20191002122517.17721-3-leon@kernel.org
Signed-off-by: Erez Alfasi <ereza@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>
The removal of 'buffer' in the patch below caused free_page() to use a
value that had been offset since the wqe pointer is adjusted while the
routine runs.
The current implementation of free_pages() rounds down to a pfn,
discarding the adjustment, but this is not the right way to use the
API. Preserve the initial value and use it for free_page().
Fixes: 0f51427bd0 ("RDMA/mlx5: Cleanup WQE page fault handler")
Link: https://lore.kernel.org/r/20190916064818.19823-2-leon@kernel.org
Signed-off-by: Danit Goldberg <danitg@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@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>
To resolve dependencies in following patches
mlx5_ib.h conflict resolved by keeing both hunks
Linux 5.3-rc8
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Michael Guralnik says:
====================
The series adds support for on-demand paging for DC transport.
As DC is a mlx-only transport, the capabilities are exposed to the user
using DEVX objects and later on through mlx5dv_query_device.
====================
Based on the mlx5-next branch from
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux for
dependencies
* branch 'mlx5-odp-dc':
IB/mlx5: Add page fault handler for DC initiator WQE
IB/mlx5: Remove check of FW capabilities in ODP page fault handling
net/mlx5: Set ODP capabilities for DC transport to max
Parsing DC initiator WQEs upon page fault requires skipping an address
vector segment, as in UD WQEs.
Link: https://lore.kernel.org/r/20190819120815.21225-4-leon@kernel.org
Signed-off-by: Michael Guralnik <michaelgur@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
As page fault handling is initiated by FW, there is no need to check that
the ODP supports the operation and transport.
Link: https://lore.kernel.org/r/20190819120815.21225-3-leon@kernel.org
Signed-off-by: Michael Guralnik <michaelgur@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Jason Gunthorpe says:
====================
This is a collection of general cleanups for ODP to clarify some of the
flows around umem creation and use of the interval tree.
====================
The branch is based on v5.3-rc5 due to dependencies
* odp_fixes:
RDMA/mlx5: Use odp instead of mr->umem in pagefault_mr
RDMA/mlx5: Use ib_umem_start instead of umem.address
RDMA/core: Make invalidate_range a device operation
RDMA/odp: Use kvcalloc for the dma_list and page_list
RDMA/odp: Check for overflow when computing the umem_odp end
RDMA/odp: Provide ib_umem_odp_release() to undo the allocs
RDMA/odp: Split creating a umem_odp from ib_umem_get
RDMA/odp: Make the three ways to create a umem_odp clear
RMDA/odp: Consolidate umem_odp initialization
RDMA/odp: Make it clearer when a umem is an implicit ODP umem
RDMA/odp: Iterate over the whole rbtree directly
RDMA/odp: Use the common interval tree library instead of generic
RDMA/mlx5: Fix MR npages calculation for IB_ACCESS_HUGETLB
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
These are the same thing since mr always comes from odp->private. It is
confusing to reference the same memory via two names.
Link: https://lore.kernel.org/r/20190819111710.18440-13-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
These are subtly different, the address is the original VA requested
during umem_get, while ib_umem_start() is the version that is rounded to
the proper page size, ie is the true start of the umem's dma map.
Link: https://lore.kernel.org/r/20190819111710.18440-12-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The callback function 'invalidate_range' is implemented in a driver so the
place for it is in the ib_device_ops structure and not in ib_ucontext.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Link: https://lore.kernel.org/r/20190819111710.18440-11-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that there are allocator APIs that return the ib_umem_odp directly
it should be freed through a umem_odp free'er as well.
Link: https://lore.kernel.org/r/20190819111710.18440-8-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The three paths to build the umem_odps are kind of muddled, they are:
- As a normal ib_mr umem
- As a child in an implicit ODP umem tree
- As the root of an implicit ODP umem tree
Only the first two are actually umem's, the last is an abuse.
The implicit case can only be triggered by explicit driver request, it
should never be co-mingled with the normal case. While we are here, make
sensible function names and add some comments to make this clearer.
Link: https://lore.kernel.org/r/20190819111710.18440-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Implicit ODP umems are special, they don't have any page lists, they don't
exist in the interval tree and they are never DMA mapped.
Instead of trying to guess this based on a zero length use an explicit
flag.
Further, do not allow non-implicit umems to be 0 size.
Link: https://lore.kernel.org/r/20190819111710.18440-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Instead of intersecting a full interval, just iterate over every element
directly. This is faster and clearer.
Link: https://lore.kernel.org/r/20190819111710.18440-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
ODP depends on the several device capabilities, among them is the ability
to send UMR WQEs with that modify atomic and entity size of the MR.
Therefore, only if all conditions to send such a UMR WQE are met then
driver can report that ODP is supported. Use this check of conditions
in all places where driver needs to know about ODP support.
Also, implicit ODP support depends on ability of driver to send UMR WQEs
for an indirect mkey. Therefore, verify that all conditions to do so are
met when reporting support.
Fixes: c8d75a980f ("IB/mlx5: Respect new UMR capabilities")
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Guy Levi <guyle@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Link: https://lore.kernel.org/r/20190815083834.9245-7-leon@kernel.org
Signed-off-by: Doug Ledford <dledford@redhat.com>
Once implicit MR is being called to be released by
ib_umem_notifier_release() its leaves were marked as "dying".
However, when dereg_mr()->mlx5_ib_free_implicit_mr()->mr_leaf_free() is
called, it skips running the mr_leaf_free_action (i.e. umem_odp->work)
when those leaves were marked as "dying".
As such ib_umem_release() for the leaves won't be called and their MRs
will be leaked as well.
When an application exits/killed without calling dereg_mr we might hit the
above flow.
This fatal scenario is reported by WARN_ON() upon
mlx5_ib_dealloc_ucontext() as ibcontext->per_mm_list is not empty, the
call trace can be seen below.
Originally the "dying" mark as part of ib_umem_notifier_release() was
introduced to prevent pagefault_mr() from returning a success response
once this happened. However, we already have today the completion
mechanism so no need for that in those flows any more. Even in case a
success response will be returned the firmware will not find the pages and
an error will be returned in the following call as a released mm will
cause ib_umem_odp_map_dma_pages() to permanently fail mmget_not_zero().
Fix the above issue by dropping the "dying" from the above flows. The
other flows that are using "dying" are still needed it for their
synchronization purposes.
WARNING: CPU: 1 PID: 7218 at
drivers/infiniband/hw/mlx5/main.c:2004
mlx5_ib_dealloc_ucontext+0x84/0x90 [mlx5_ib]
CPU: 1 PID: 7218 Comm: ibv_rc_pingpong Tainted: G E
5.2.0-rc6+ #13
Call Trace:
uverbs_destroy_ufile_hw+0xb5/0x120 [ib_uverbs]
ib_uverbs_close+0x1f/0x80 [ib_uverbs]
__fput+0xbe/0x250
task_work_run+0x88/0xa0
do_exit+0x2cb/0xc30
? __fput+0x14b/0x250
do_group_exit+0x39/0xb0
get_signal+0x191/0x920
? _raw_spin_unlock_bh+0xa/0x20
? inet_csk_accept+0x229/0x2f0
do_signal+0x36/0x5e0
? put_unused_fd+0x5b/0x70
? __sys_accept4+0x1a6/0x1e0
? inet_hash+0x35/0x40
? release_sock+0x43/0x90
? _raw_spin_unlock_bh+0xa/0x20
? inet_listen+0x9f/0x120
exit_to_usermode_loop+0x5c/0xc6
do_syscall_64+0x182/0x1b0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Fixes: 81713d3788 ("IB/mlx5: Add implicit MR support")
Link: https://lore.kernel.org/r/20190805083010.21777-1-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Artemy Kovalyov <artemyko@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>
Delete DEBUG ODP dead code which is leftover from development
stage and doesn't need to be part of the upstream kernel.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Link: https://lore.kernel.org/r/20190731115627.5433-1-leon@kernel.org
Signed-off-by: Doug Ledford <dledford@redhat.com>
The device requires that memory registration work requests that update the
address translation table of a MR will be fenced if posted together. This
scenario can happen when address ranges are invalidated by the mmu in
separate concurrent calls to the invalidation callback.
We prefer to block concurrent address updates for a single MR over fencing
since making the decision if a WQE needs fencing will be more expensive
and fencing all WQEs is a too radical choice.
Further, it isn't clear that this code can even run safely concurrently,
so a lock is a safer choice.
Fixes: b4cfe447d4 ("IB/mlx5: Implement on demand paging by adding support for MMU notifiers")
Link: https://lore.kernel.org/r/20190723065733.4899-8-leon@kernel.org
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Artemy Kovalyov <artemyko@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>
Memory allocated by kvzalloc should not be freed by kfree(), use kvfree()
instead.
Fixes: 813e90b1ae ("IB/mlx5: Add advise_mr() support")
Link: https://lore.kernel.org/r/20190717082101.14196-1-hslester96@gmail.com
Signed-off-by: Chuhong Yuan <hslester96@gmail.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
From git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux
Required for dependencies in the next patches.
Resolved the conflicts:
- esw_destroy_offloads_acl_tables() use the newer mlx5_esw_for_all_vports()
version
- esw_offloads_steering_init() drop the cap test
- esw_offloads_init() drop the extra function arguments
* branch 'mlx5-next': (39 commits)
net/mlx5: Expose device definitions for object events
net/mlx5: Report EQE data upon CQ completion
net/mlx5: Report a CQ error event only when a handler was set
net/mlx5: mlx5_core_create_cq() enhancements
net/mlx5: Expose the API to register for ANY event
net/mlx5: Use event mask based on device capabilities
net/mlx5: Fix mlx5_core_destroy_cq() error flow
net/mlx5: E-Switch, Handle UC address change in switchdev mode
net/mlx5: E-Switch, Consider host PF for inline mode and vlan pop
net/mlx5: E-Switch, Use iterator for vlan and min-inline setups
net/mlx5: E-Switch, Reg/unreg function changed event at correct stage
net/mlx5: E-Switch, Consolidate eswitch function number of VFs
net/mlx5: E-Switch, Refactor eswitch SR-IOV interface
net/mlx5: Handle host PF vport mac/guid for ECPF
net/mlx5: E-Switch, Use correct flags when configuring vlan
net/mlx5: Reduce dependency on enabled_vfs counter and num_vfs
net/mlx5: Don't handle VF func change if host PF is disabled
net/mlx5: Limit scope of mlx5_get_next_phys_dev() to PCI PF devices
net/mlx5: Move pci status reg access mutex to mlx5_pci_init
net/mlx5: Rename mlx5_pci_dev_type to mlx5_coredev_type
...
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
