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>
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>
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>
This is a significant simplification, no extra list is kept per FD, and
the interval tree is now shared between all the ucontexts, reducing
overhead if there are multiple ucontexts active.
Link: https://lore.kernel.org/r/20190806231548.25242-7-jgg@ziepe.ca
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Since the page size can be extended in the ODP case by IB_ACCESS_HUGETLB
the existing overflow checks done by ib_umem_get() are not
sufficient. Check for overflow again.
Further, remove the unchecked math from the inlines and just use the
precomputed value stored in the interval_tree_node.
Link: https://lore.kernel.org/r/20190819111710.18440-9-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This is the last creation API that is overloaded for both, there is very
little code sharing and a driver has to be specifically ready for a
umem_odp to be created to use the odp version.
Link: https://lore.kernel.org/r/20190819111710.18440-7-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>
ODP is working with userspace VA's in the interval tree which always fit
into an unsigned long, so we can use the common code.
This comes at a cost of a 16 byte increase in ib_umem_odp struct size due
to storing the interval tree start/last in addition to the umem
addr/length. However these values were computed and are performance
critical for the interval lookup, so this seems like a worthwhile trade
off.
Removes 2k of .text from the kernel.
Link: https://lore.kernel.org/r/20190819111710.18440-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This value has always been set to PAGE_SHIFT in the core code, the only
thing that does differently was the ODP path. Move the value into the ODP
struct and still use it for ODP, but change all the non-ODP things to just
use PAGE_SHIFT/PAGE_SIZE/PAGE_MASK directly.
Reviewed-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Combine contiguous regions of PAGE_SIZE pages into single scatter list
entry while building the scatter table for a umem. This minimizes the
number of the entries in the scatter list and reduces the DMA mapping
overhead, particularly with the IOMMU.
Set default max_seg_size in core for IB devices to 2G and do not combine
if we exceed this limit.
Also, purge npages in struct ib_umem as we now DMA map the umem SGL with
sg_nents and npage computation is not needed. Drivers should now be using
ib_umem_num_pages(), so fix the last stragglers.
Move npages tracking to ib_umem_odp as ODP drivers still need it.
Suggested-by: Jason Gunthorpe <jgg@ziepe.ca>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Acked-by: Adit Ranadive <aditr@vmware.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
Tested-by: Gal Pressman <galpress@amazon.com>
Tested-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When CONFIG_INFINIBAND_ON_DEMAND_PAGING is not set there is no caller to
ib_alloc_odp_umem() so let's remove it.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
A sub-range in ODP implicit MR should take its write permission from the
MR and not be set always to allow.
Fixes: d07d1d70ce ("IB/umem: Update on demand page (ODP) support")
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Convert various places to more readable code, which embeds
CONFIG_INFINIBAND_ON_DEMAND_PAGING into the code flow.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
synchronize_rcu is slow enough that it should be avoided on the syscall
path when user space is destroying MRs. After all the rework we can now
trivially do this by having call_srcu kfree the per_mm.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
mmu_notifier_unregister() can race between a invalidate_start/end and
cause the invalidate_end to be skipped. This causes an imbalance in the
locking, which lockdep complains about.
This is not actually a bug, as we immediately kfree the memory holding the
lock, but it simple enough to fix.
Mark when the notifier is being destroyed and abort the start callback.
This can be done under the lock we already obtained, and can re-purpose
the invalidate_range test we already have.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This is intrinsically racy and the scheme is simply unnecessary. New MR
registration can wait for any on going invalidation to fully complete.
CPU0 CPU1
if (atomic_read())
if (atomic_dec_and_test() &&
!list_empty())
{ /* not taken */ }
list_add()
Putting the new UMEM into some kind of purgatory until another invalidate
rolls through..
Instead hold the read side of the umem_rwsem across the pair'd start/end
and get rid of the racy 'deferred add' approach.
Since all umem's in the rbt are always ready to go, also get rid of the
mn_counters_active stuff.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Since ODP had a single struct mmu_notifier located in the ucontext it
could only handle a single MM at a time, and this prevented it from using
the new owning_mm system.
With the prior rework it is now simple to let ODP track multiple MMs per
ucontext, finish the job so that the per_mm is allocated on a mm by mm
basis, and freed when the last umem is dropped from the ucontext.
As a side effect the new saner locking removes the lockdep splat about
nesting the umem_rwsem between mmu_notifier_unregister and
ib_umem_odp_release.
It also makes ODP work with multiple processes, across, fork, etc.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This is the first step to make ODP use the owning_mm that is now part of
struct ib_umem.
Each ODP umem is linked to a single per_mm structure, which in turn, is
linked to a single mm, via the embedded mmu_notifier. This first patch
introduces the structure and reworks eveything to use it.
This also needs to introduce tgid into the ib_ucontext_per_mm, as
get_user_pages_remote() requires the originating task for statistics
tracking.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
These two structures are linked together, use the container_of pattern
instead of a double allocation to make the code simpler and easier to
follow.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
All of these functions already require the ODP version of the umem struct,
make this very clear by having the signature require it. This paves the
way to using the container_of() pattern to link umem_odp and umem
together.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
There are several blockable mmu notifiers which might sleep in
mmu_notifier_invalidate_range_start and that is a problem for the
oom_reaper because it needs to guarantee a forward progress so it cannot
depend on any sleepable locks.
Currently we simply back off and mark an oom victim with blockable mmu
notifiers as done after a short sleep. That can result in selecting a new
oom victim prematurely because the previous one still hasn't torn its
memory down yet.
We can do much better though. Even if mmu notifiers use sleepable locks
there is no reason to automatically assume those locks are held. Moreover
majority of notifiers only care about a portion of the address space and
there is absolutely zero reason to fail when we are unmapping an unrelated
range. Many notifiers do really block and wait for HW which is harder to
handle and we have to bail out though.
This patch handles the low hanging fruit.
__mmu_notifier_invalidate_range_start gets a blockable flag and callbacks
are not allowed to sleep if the flag is set to false. This is achieved by
using trylock instead of the sleepable lock for most callbacks and
continue as long as we do not block down the call chain.
I think we can improve that even further because there is a common pattern
to do a range lookup first and then do something about that. The first
part can be done without a sleeping lock in most cases AFAICS.
The oom_reaper end then simply retries if there is at least one notifier
which couldn't make any progress in !blockable mode. A retry loop is
already implemented to wait for the mmap_sem and this is basically the
same thing.
The simplest way for driver developers to test this code path is to wrap
userspace code which uses these notifiers into a memcg and set the hard
limit to hit the oom. This can be done e.g. after the test faults in all
the mmu notifier managed memory and set the hard limit to something really
small. Then we are looking for a proper process tear down.
[akpm@linux-foundation.org: coding style fixes]
[akpm@linux-foundation.org: minor code simplification]
Link: http://lkml.kernel.org/r/20180716115058.5559-1-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Christian König <christian.koenig@amd.com> # AMD notifiers
Acked-by: Leon Romanovsky <leonro@mellanox.com> # mlx and umem_odp
Reported-by: David Rientjes <rientjes@google.com>
Cc: "David (ChunMing) Zhou" <David1.Zhou@amd.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: David Airlie <airlied@linux.ie>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Mike Marciniszyn <mike.marciniszyn@intel.com>
Cc: Dennis Dalessandro <dennis.dalessandro@intel.com>
Cc: Sudeep Dutt <sudeep.dutt@intel.com>
Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
Cc: Dimitri Sivanich <sivanich@sgi.com>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: "Jérôme Glisse" <jglisse@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The RDMA/umem uses generic RB-trees macros to generate various ib_umem
access functions. The generation is performed with INTERVAL_TREE_DEFINE
macro, which allows one of two modes: declare all functions as static or
declare none of the function to be static.
The second mode of operation produces the following sparse errors:
drivers/infiniband/core/umem_rbtree.c:69:1:
warning: symbol 'rbt_ib_umem_iter_first' was not declared.
Should it be static?
drivers/infiniband/core/umem_rbtree.c:69:1:
warning: symbol 'rbt_ib_umem_iter_next' was not declared.
Should it be static?
Code relocation together with declaration of such functions to be
"static" solves the issue.
Because there is no need to have separate file for two functions,
let's consolidate umem_rtree.c and umem_odp.c into one file.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Allow interval trees to quickly check for overlaps to avoid unnecesary
tree lookups in interval_tree_iter_first().
As of this patch, all interval tree flavors will require using a
'rb_root_cached' such that we can have the leftmost node easily
available. While most users will make use of this feature, those with
special functions (in addition to the generic insert, delete, search
calls) will avoid using the cached option as they can do funky things
with insertions -- for example, vma_interval_tree_insert_after().
[jglisse@redhat.com: fix deadlock from typo vm_lock_anon_vma()]
Link: http://lkml.kernel.org/r/20170808225719.20723-1-jglisse@redhat.com
Link: http://lkml.kernel.org/r/20170719014603.19029-12-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Doug Ledford <dledford@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Cc: David Airlie <airlied@linux.ie>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Christian Benvenuti <benve@cisco.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add IB_ACCESS_HUGETLB ib_reg_mr flag.
Hugetlb region registered with this flag
will use single translation entry per huge page.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Currently ODP MR may explicitly register virtual address space area
of limited length.
This change allows MR to cover entire process virtual address space
dynamicaly adding/removing translation entries to device MTT.
Add following changes to support implicit MR:
* Allow umem to be zero size to back-up implicit MR.
* Add new function ib_alloc_odp_umem() to add virtual memory regions
to implicit MR dynamically on demand.
* Add new function rbt_ib_umem_lookup() to find dynamically added
virtual memory regions.
* Expose function rbt_ib_umem_for_each_in_range() to other modules and
make it safe
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
* Add an interval tree implementation for ODP umems. Create an
interval tree for each ucontext (including a count of the number of
ODP MRs in this context, semaphore, etc.), and register ODP umems in
the interval tree.
* Add MMU notifiers handling functions, using the interval tree to
notify only the relevant umems and underlying MRs.
* Register to receive MMU notifier events from the MM subsystem upon
ODP MR registration (and unregister accordingly).
* Add a completion object to synchronize the destruction of ODP umems.
* Add mechanism to abort page faults when there's a concurrent invalidation.
The way we synchronize between concurrent invalidations and page
faults is by keeping a counter of currently running invalidations, and
a sequence number that is incremented whenever an invalidation is
caught. The page fault code checks the counter and also verifies that
the sequence number hasn't progressed before it updates the umem's
page tables. This is similar to what the kvm module does.
In order to prevent the case where we register a umem in the middle of
an ongoing notifier, we also keep a per ucontext counter of the total
number of active mmu notifiers. We only enable new umems when all the
running notifiers complete.
Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Yuval Dagan <yuvalda@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
* Extend the umem struct to keep the ODP related data.
* Allocate and initialize the ODP related information in the umem
(page_list, dma_list) and freeing as needed in the end of the run.
* Store a reference to the process PID struct in the ucontext. Used to
safely obtain the task_struct and the mm during fault handling,
without preventing the task destruction if needed.
* Add 2 helper functions: ib_umem_odp_map_dma_pages and
ib_umem_odp_unmap_dma_pages. These functions get the DMA addresses
of specific pages of the umem (and, currently, pin them).
* Support for page faults only - IB core will keep the reference on
the pages used and call put_page when freeing an ODP umem
area. Invalidations support will be added in a later patch.
Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
