There is a concurrency bug that may cause the wrong value to be loaded
when a CPU is modifying the maple tree.
CPU1:
mtree_insert_range()
mas_insert()
mas_store_root()
...
mas_root_expand()
...
rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
ma_set_meta(node, maple_leaf_64, 0, slot); <---IP
CPU2:
mtree_load()
mtree_lookup_walk()
ma_data_end();
When CPU1 is about to execute the instruction pointed to by IP, the
ma_data_end() executed by CPU2 may return the wrong end position, which
will cause the value loaded by mtree_load() to be wrong.
An example of triggering the bug:
Add mdelay(100) between rcu_assign_pointer() and ma_set_meta() in
mas_root_expand().
static DEFINE_MTREE(tree);
int work(void *p) {
unsigned long val;
for (int i = 0 ; i< 30; ++i) {
val = (unsigned long)mtree_load(&tree, 8);
mdelay(5);
pr_info("%lu",val);
}
return 0;
}
mt_init_flags(&tree, MT_FLAGS_USE_RCU);
mtree_insert(&tree, 0, (void*)12345, GFP_KERNEL);
run_thread(work)
mtree_insert(&tree, 1, (void*)56789, GFP_KERNEL);
In RCU mode, mtree_load() should always return the value before or after
the data structure is modified, and in this example mtree_load(&tree, 8)
may return 56789 which is not expected, it should always return NULL. Fix
it by put ma_set_meta() before rcu_assign_pointer().
Link: https://lkml.kernel.org/r/20230314124203.91572-4-zhangpeng.00@bytedance.com
Fixes: 54a611b605 ("Maple Tree: add new data structure")
Signed-off-by: Peng Zhang <zhangpeng.00@bytedance.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Dereferencing RCU objects within the RCU callback without the RCU check
has caused lockdep to complain. Fix the RCU dereferencing by using the
RCU callback lock to ensure the operation is safe.
Also stop creating a new lock to use for dereferencing during destruction
of the tree or subtree. Instead, pass through a pointer to the tree that
has the lock that is held for RCU dereferencing checking. It also does
not make sense to use the maple state in the freeing scenario as the tree
walk is a special case where the tree no longer has the normal encodings
and parent pointers.
Link: https://lkml.kernel.org/r/20230227173632.3292573-8-surenb@google.com
Fixes: 54a611b605 ("Maple Tree: add new data structure")
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reported-by: Suren Baghdasaryan <surenb@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
During the development of the maple tree, the strategy of freeing multiple
nodes changed and, in the process, the pivots were reused to store
pointers to dead nodes. To ensure the readers see accurate pivots, the
writers need to mark the nodes as dead and call smp_wmb() to ensure any
readers can identify the node as dead before using the pivot values.
There were two places where the old method of marking the node as dead
without smp_wmb() were being used, which resulted in RCU readers seeing
the wrong pivot value before seeing the node was dead. Fix this race
condition by using mte_set_node_dead() which has the smp_wmb() call to
ensure the race is closed.
Add a WARN_ON() to the ma_free_rcu() call to ensure all nodes being freed
are marked as dead to ensure there are no other call paths besides the two
updated paths.
This is necessary for the RCU mode of the maple tree.
Link: https://lkml.kernel.org/r/20230227173632.3292573-6-surenb@google.com
Fixes: 54a611b605 ("Maple Tree: add new data structure")
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Conflicts:
drivers/net/ethernet/mediatek/mtk_ppe.c
3fbe4d8c0e ("net: ethernet: mtk_eth_soc: ppe: add support for flow accounting")
924531326e ("net: ethernet: mtk_eth_soc: add missing ppe cache flush when deleting a flow")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Saeed Mahameed says:
====================
mlx5-updates-2023-03-20
mlx5 dynamic msix
This patch series adds support for dynamic msix vectors allocation in mlx5.
Eli Cohen Says:
================
The following series of patches modifies mlx5_core to work with the
dynamic MSIX API. Currently, mlx5_core allocates all the interrupt
vectors it needs and distributes them amongst the consumers. With the
introduction of dynamic MSIX support, which allows for allocation of
interrupts more than once, we now allocate vectors as we need them.
This allows other drivers running on top of mlx5_core to allocate
interrupt vectors for their own use. An example for this is mlx5_vdpa,
which uses these vectors to propagate interrupts directly from the
hardware to the vCPU [1].
As a preparation for using this series, a use after free issue is fixed
in lib/cpu_rmap.c and the allocator for rmap entries has been modified.
A complementary API for irq_cpu_rmap_add() has also been introduced.
[1] https://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux.git/patch/?id=0f2bf1fcae96a83b8c5581854713c9fc3407556e
================
* tag 'mlx5-updates-2023-03-20' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux:
net/mlx5: Provide external API for allocating vectors
net/mlx5: Use one completion vector if eth is disabled
net/mlx5: Refactor calculation of required completion vectors
net/mlx5: Move devlink registration before mlx5_load
net/mlx5: Use dynamic msix vectors allocation
net/mlx5: Refactor completion irq request/release code
net/mlx5: Improve naming of pci function vectors
net/mlx5: Use newer affinity descriptor
net/mlx5: Modify struct mlx5_irq to use struct msi_map
net/mlx5: Fix wrong comment
net/mlx5e: Coding style fix, add empty line
lib: cpu_rmap: Add irq_cpu_rmap_remove to complement irq_cpu_rmap_add
lib: cpu_rmap: Use allocator for rmap entries
lib: cpu_rmap: Avoid use after free on rmap->obj array entries
====================
Link: https://lore.kernel.org/r/20230324231341.29808-1-saeed@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
atomic_t based reference counting, including refcount_t, uses
atomic_inc_not_zero() for acquiring a reference. atomic_inc_not_zero() is
implemented with a atomic_try_cmpxchg() loop. High contention of the
reference count leads to retry loops and scales badly. There is nothing to
improve on this implementation as the semantics have to be preserved.
Provide rcuref as a scalable alternative solution which is suitable for RCU
managed objects. Similar to refcount_t it comes with overflow and underflow
detection and mitigation.
rcuref treats the underlying atomic_t as an unsigned integer and partitions
this space into zones:
0x00000000 - 0x7FFFFFFF valid zone (1 .. (INT_MAX + 1) references)
0x80000000 - 0xBFFFFFFF saturation zone
0xC0000000 - 0xFFFFFFFE dead zone
0xFFFFFFFF no reference
rcuref_get() unconditionally increments the reference count with
atomic_add_negative_relaxed(). rcuref_put() unconditionally decrements the
reference count with atomic_add_negative_release().
This unconditional increment avoids the inc_not_zero() problem, but
requires a more complex implementation on the put() side when the count
drops from 0 to -1.
When this transition is detected then it is attempted to mark the reference
count dead, by setting it to the midpoint of the dead zone with a single
atomic_cmpxchg_release() operation. This operation can fail due to a
concurrent rcuref_get() elevating the reference count from -1 to 0 again.
If the unconditional increment in rcuref_get() hits a reference count which
is marked dead (or saturated) it will detect it after the fact and bring
back the reference count to the midpoint of the respective zone. The zones
provide enough tolerance which makes it practically impossible to escape
from a zone.
The racy implementation of rcuref_put() requires to protect rcuref_put()
against a grace period ending in order to prevent a subtle use after
free. As RCU is the only mechanism which allows to protect against that, it
is not possible to fully replace the atomic_inc_not_zero() based
implementation of refcount_t with this scheme.
The final drop is slightly more expensive than the atomic_dec_return()
counterpart, but that's not the case which this is optimized for. The
optimization is on the high frequeunt get()/put() pairs and their
scalability.
The performance of an uncontended rcuref_get()/put() pair where the put()
is not dropping the last reference is still on par with the plain atomic
operations, while at the same time providing overflow and underflow
detection and mitigation.
The performance of rcuref compared to plain atomic_inc_not_zero() and
atomic_dec_return() based reference counting under contention:
- Micro benchmark: All CPUs running a increment/decrement loop on an
elevated reference count, which means the 0 to -1 transition never
happens.
The performance gain depends on microarchitecture and the number of
CPUs and has been observed in the range of 1.3X to 4.7X
- Conversion of dst_entry::__refcnt to rcuref and testing with the
localhost memtier/memcached benchmark. That benchmark shows the
reference count contention prominently.
The performance gain depends on microarchitecture and the number of
CPUs and has been observed in the range of 1.1X to 2.6X over the
previous fix for the false sharing issue vs. struct
dst_entry::__refcnt.
When memtier is run over a real 1Gb network connection, there is a
small gain on top of the false sharing fix. The two changes combined
result in a 2%-5% total gain for that networked test.
Reported-by: Wangyang Guo <wangyang.guo@intel.com>
Reported-by: Arjan Van De Ven <arjan.van.de.ven@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230323102800.158429195@linutronix.de
Pull xfs percpu counter fixes from Darrick Wong:
"We discovered a filesystem summary counter corruption problem that was
traced to cpu hot-remove racing with the call to percpu_counter_sum
that sets the free block count in the superblock when writing it to
disk. The root cause is that percpu_counter_sum doesn't cull from
dying cpus and hence misses those counter values if the cpu shutdown
hooks have not yet run to merge the values.
I'm hoping this is a fairly painless fix to the problem, since the
dying cpu mask should generally be empty. It's been in for-next for a
week without any complaints from the bots.
- Fix a race in the percpu counters summation code where the
summation failed to add in the values for any CPUs that were dying
but not yet dead. This fixes some minor discrepancies and incorrect
assertions when running generic/650"
* tag 'xfs-6.3-fixes-4' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
pcpcntr: remove percpu_counter_sum_all()
fork: remove use of percpu_counter_sum_all
pcpcntrs: fix dying cpu summation race
cpumask: introduce for_each_cpu_or
Add a function to complement irq_cpu_rmap_add(). It removes the irq from
the reverse mapping by setting the notifier to NULL. The function calls
irq_set_affinity_notifier() with NULL at the notify argument which then
cancel any pending notifier work and decrement reference on the
notifier. When ref count reaches zero, the glue pointer is kfree and the
rmap entry is set to NULL serving both to avoid second attempt to
release it and also making the rmap entry available for subsequent
mapping.
It should be noted the drivers usually creates the reverse mapping at
initialization time and remove it at unload time so we do not expect
failures in allocating rmap due to kref holding the glue entry.
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Eli Cohen <elic@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Use a proper allocator for rmap entries using a naive for loop. The
allocator relies on whether an entry is NULL to be considered free.
Remove the used field of rmap which is not needed.
Also, avoid crashing the kernel if an entry is not available.
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Eli Cohen <elic@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
When calling irq_set_affinity_notifier() with NULL at the notify
argument, it will cause freeing of the glue pointer in the
corresponding array entry but will leave the pointer in the array. A
subsequent call to free_irq_cpu_rmap() will try to free this entry again
leading to possible use after free.
Fix that by setting NULL to the array entry and checking that we have
non-zero at the array entry when iterating over the array in
free_irq_cpu_rmap().
The current code does not suffer from this since there are no cases
where irq_set_affinity_notifier(irq, NULL) (note the NULL passed for the
notify arg) is called, followed by a call to free_irq_cpu_rmap() so we
don't hit and issue. Subsequent patches in this series excersize this
flow, hence the required fix.
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Eli Cohen <elic@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Patch series "Fix mas_skip_node() for mas_empty_area()", v2.
mas_empty_area() was incorrectly returning an error when there was room.
The issue was tracked down to mas_skip_node() using the incorrect
end-of-slot count. Instead of using the nodes hard limit, the limit of
data should be used.
mas_skip_node() was also setting the min and max to that of the child
node, which was unnecessary. Within these limits being set, there was
also a bug that corrupted the maple state's max if the offset was set to
the maximum node pivot. The bug was without consequence unless there was
a sufficient gap in the next child node which would cause an error to be
returned.
This patch set fixes these errors by removing the limit setting from
mas_skip_node() and uses the mas_data_end() for slot limits, and adds
tests for all failures discovered.
This patch (of 2):
mas_skip_node() is used to move the maple state to the node with a higher
limit. It does this by walking up the tree and increasing the slot count.
Since slot count may not be able to be increased, it may need to walk up
multiple times to find room to walk right to a higher limit node. The
limit of slots that was being used was the node limit and not the last
location of data in the node. This would cause the maple state to be
shifted outside actual data and enter an error state, thus returning
-EBUSY.
The result of the incorrect error state means that mas_awalk() would
return an error instead of finding the allocation space.
The fix is to use mas_data_end() in mas_skip_node() to detect the nodes
data end point and continue walking the tree up until it is safe to move
to a node with a higher limit.
The walk up the tree also sets the maple state limits so remove the buggy
code from mas_skip_node(). Setting the limits had the unfortunate side
effect of triggering another bug if the parent node was full and the there
was no suitable gap in the second last child, but room in the next child.
mas_skip_node() may also be passed a maple state in an error state from
mas_anode_descend() when no allocations are available. Return on such an
error state immediately.
Link: https://lkml.kernel.org/r/20230307180247.2220303-1-Liam.Howlett@oracle.com
Link: https://lkml.kernel.org/r/20230307180247.2220303-2-Liam.Howlett@oracle.com
Fixes: 54a611b605 ("Maple Tree: add new data structure")
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reported-by: Snild Dolkow <snild@sony.com>
Link: https://lore.kernel.org/linux-mm/cb8dc31a-fef2-1d09-f133-e9f7b9f9e77a@sony.com/
Tested-by: Snild Dolkow <snild@sony.com>
Cc: Peng Zhang <zhangpeng.00@bytedance.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
percpu_counter_sum_all() is now redundant as the race condition it
was invented to handle is now dealt with by percpu_counter_sum()
directly and all users of percpu_counter_sum_all() have been
removed.
Remove it.
This effectively reverts the changes made in f689054aac
("percpu_counter: add percpu_counter_sum_all interface") except for
the cpumask iteration that fixes percpu_counter_sum() made earlier
in this series.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
In commit f689054aac ("percpu_counter: add percpu_counter_sum_all
interface") a race condition between a cpu dying and
percpu_counter_sum() iterating online CPUs was identified. The
solution was to iterate all possible CPUs for summation via
percpu_counter_sum_all().
We recently had a percpu_counter_sum() call in XFS trip over this
same race condition and it fired a debug assert because the
filesystem was unmounting and the counter *should* be zero just
before we destroy it. That was reported here:
https://lore.kernel.org/linux-kernel/20230314090649.326642-1-yebin@huaweicloud.com/
likely as a result of running generic/648 which exercises
filesystems in the presence of CPU online/offline events.
The solution to use percpu_counter_sum_all() is an awful one. We
use percpu counters and percpu_counter_sum() for accurate and
reliable threshold detection for space management, so a summation
race condition during these operations can result in overcommit of
available space and that may result in filesystem shutdowns.
As percpu_counter_sum_all() iterates all possible CPUs rather than
just those online or even those present, the mask can include CPUs
that aren't even installed in the machine, or in the case of
machines that can hot-plug CPU capable nodes, even have physical
sockets present in the machine.
Fundamentally, this race condition is caused by the CPU being
offlined being removed from the cpu_online_mask before the notifier
that cleans up per-cpu state is run. Hence percpu_counter_sum() will
not sum the count for a cpu currently being taken offline,
regardless of whether the notifier has run or not. This is
the root cause of the bug.
The percpu counter notifier iterates all the registered counters,
locks the counter and moves the percpu count to the global sum.
This is serialised against other operations that move the percpu
counter to the global sum as well as percpu_counter_sum() operations
that sum the percpu counts while holding the counter lock.
Hence the notifier is safe to run concurrently with sum operations,
and the only thing we actually need to care about is that
percpu_counter_sum() iterates dying CPUs. That's trivial to do,
and when there are no CPUs dying, it has no addition overhead except
for a cpumask_or() operation.
This change makes percpu_counter_sum() always do the right thing in
the presence of CPU hot unplug events and makes
percpu_counter_sum_all() unnecessary. This, in turn, means that
filesystems like XFS, ext4, and btrfs don't have to work out when
they should use percpu_counter_sum() vs percpu_counter_sum_all() in
their space accounting algorithms
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
