Extend the futex2 interface to be aware of mempolicy.
When FUTEX2_MPOL is specified and there is a MPOL_PREFERRED or
home_node specified covering the futex address, use that hash-map.
Notably, in this case the futex will go to the global node hashtable,
even if it is a PRIVATE futex.
When FUTEX2_NUMA|FUTEX2_MPOL is specified and the user specified node
value is FUTEX_NO_NODE, the MPOL lookup (as described above) will be
tried first before reverting to setting node to the local node.
[bigeasy: add CONFIG_FUTEX_MPOL, add MPOL to FUTEX2_VALID_MASK, write
the node only to user if FUTEX_NO_NODE was supplied]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-18-bigeasy@linutronix.de
Extend the futex2 interface to be numa aware.
When FUTEX2_NUMA is specified for a futex, the user value is extended
to two words (of the same size). The first is the user value we all
know, the second one will be the node to place this futex on.
struct futex_numa_32 {
u32 val;
u32 node;
};
When node is set to ~0, WAIT will set it to the current node_id such
that WAKE knows where to find it. If userspace corrupts the node value
between WAIT and WAKE, the futex will not be found and no wakeup will
happen.
When FUTEX2_NUMA is not set, the node is simply an extension of the
hash, such that traditional futexes are still interleaved over the
nodes.
This is done to avoid having to have a separate !numa hash-table.
[bigeasy: ensure to have at least hashsize of 4 in futex_init(), add
pr_info() for size and allocation information. Cast the naddr math to
void*]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-17-bigeasy@linutronix.de
My initial testing showed that:
perf bench futex hash
reported less operations/sec with private hash. After using the same
amount of buckets in the private hash as used by the global hash then
the operations/sec were about the same.
This changed once the private hash became resizable. This feature added
an RCU section and reference counting via atomic inc+dec operation into
the hot path.
The reference counting can be avoided if the private hash is made
immutable.
Extend PR_FUTEX_HASH_SET_SLOTS by a fourth argument which denotes if the
private should be made immutable. Once set (to true) the a further
resize is not allowed (same if set to global hash).
Add PR_FUTEX_HASH_GET_IMMUTABLE which returns true if the hash can not
be changed.
Update "perf bench" suite.
For comparison, results of "perf bench futex hash -s":
- Xeon CPU E5-2650, 2 NUMA nodes, total 32 CPUs:
- Before the introducing task local hash
shared Averaged 1.487.148 operations/sec (+- 0,53%), total secs = 10
private Averaged 2.192.405 operations/sec (+- 0,07%), total secs = 10
- With the series
shared Averaged 1.326.342 operations/sec (+- 0,41%), total secs = 10
-b128 Averaged 141.394 operations/sec (+- 1,15%), total secs = 10
-Ib128 Averaged 851.490 operations/sec (+- 0,67%), total secs = 10
-b8192 Averaged 131.321 operations/sec (+- 2,13%), total secs = 10
-Ib8192 Averaged 1.923.077 operations/sec (+- 0,61%), total secs = 10
128 is the default allocation of hash buckets.
8192 was the previous amount of allocated hash buckets.
- Xeon(R) CPU E7-8890 v3, 4 NUMA nodes, total 144 CPUs:
- Before the introducing task local hash
shared Averaged 1.810.936 operations/sec (+- 0,26%), total secs = 20
private Averaged 2.505.801 operations/sec (+- 0,05%), total secs = 20
- With the series
shared Averaged 1.589.002 operations/sec (+- 0,25%), total secs = 20
-b1024 Averaged 42.410 operations/sec (+- 0,20%), total secs = 20
-Ib1024 Averaged 740.638 operations/sec (+- 1,51%), total secs = 20
-b65536 Averaged 48.811 operations/sec (+- 1,35%), total secs = 20
-Ib65536 Averaged 1.963.165 operations/sec (+- 0,18%), total secs = 20
1024 is the default allocation of hash buckets.
65536 was the previous amount of allocated hash buckets.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Link: https://lore.kernel.org/r/20250416162921.513656-16-bigeasy@linutronix.de
The mm_struct::futex_hash_lock guards the futex_hash_bucket assignment/
replacement. The futex_hash_allocate()/ PR_FUTEX_HASH_SET_SLOTS
operation can now be invoked at runtime and resize an already existing
internal private futex_hash_bucket to another size.
The reallocation is based on an idea by Thomas Gleixner: The initial
allocation of struct futex_private_hash sets the reference count
to one. Every user acquires a reference on the local hash before using
it and drops it after it enqueued itself on the hash bucket. There is no
reference held while the task is scheduled out while waiting for the
wake up.
The resize process allocates a new struct futex_private_hash and drops
the initial reference. Synchronized with mm_struct::futex_hash_lock it
is checked if the reference counter for the currently used
mm_struct::futex_phash is marked as DEAD. If so, then all users enqueued
on the current private hash are requeued on the new private hash and the
new private hash is set to mm_struct::futex_phash. Otherwise the newly
allocated private hash is saved as mm_struct::futex_phash_new and the
rehashing and reassigning is delayed to the futex_hash() caller once the
reference counter is marked DEAD.
The replacement is not performed at rcuref_put() time because certain
callers, such as futex_wait_queue(), drop their reference after changing
the task state. This change will be destroyed once the futex_hash_lock
is acquired.
The user can change the number slots with PR_FUTEX_HASH_SET_SLOTS
multiple times. An increase and decrease is allowed and request blocks
until the assignment is done.
The private hash allocated at thread creation is changed from 16 to
16 <= 4 * number_of_threads <= global_hash_size
where number_of_threads can not exceed the number of online CPUs. Should
the user PR_FUTEX_HASH_SET_SLOTS then the auto scaling is disabled.
[peterz: reorganize the code to avoid state tracking and simplify new
object handling, block the user until changes are in effect, allow
increase and decrease of the hash].
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-15-bigeasy@linutronix.de
The futex hash is system wide and shared by all tasks. Each slot
is hashed based on futex address and the VMA of the thread. Due to
randomized VMAs (and memory allocations) the same logical lock (pointer)
can end up in a different hash bucket on each invocation of the
application. This in turn means that different applications may share a
hash bucket on the first invocation but not on the second and it is not
always clear which applications will be involved. This can result in
high latency's to acquire the futex_hash_bucket::lock especially if the
lock owner is limited to a CPU and can not be effectively PI boosted.
Introduce basic infrastructure for process local hash which is shared by
all threads of process. This hash will only be used for a
PROCESS_PRIVATE FUTEX operation.
The hashmap can be allocated via:
prctl(PR_FUTEX_HASH, PR_FUTEX_HASH_SET_SLOTS, num);
A `num' of 0 means that the global hash is used instead of a private
hash.
Other values for `num' specify the number of slots for the hash and the
number must be power of two, starting with two.
The prctl() returns zero on success. This function can only be used
before a thread is created.
The current status for the private hash can be queried via:
num = prctl(PR_FUTEX_HASH, PR_FUTEX_HASH_GET_SLOTS);
which return the current number of slots. The value 0 means that the
global hash is used. Values greater than 0 indicate the number of slots
that are used. A negative number indicates an error.
For optimisation, for the private hash jhash2() uses only two arguments
the address and the offset. This omits the VMA which is always the same.
[peterz: Use 0 for global hash. A bit shuffling and renaming. ]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-13-bigeasy@linutronix.de
futex_lock_pi() and __fixup_pi_state_owner() acquire the
futex_q::lock_ptr without holding a reference assuming the previously
obtained hash bucket and the assigned lock_ptr are still valid. This
isn't the case once the private hash can be resized and becomes invalid
after the reference drop.
Introduce futex_q_lockptr_lock() to lock the hash bucket recorded in
futex_q::lock_ptr. The lock pointer is read in a RCU section to ensure
that it does not go away if the hash bucket has been replaced and the
old pointer has been observed. After locking the pointer needs to be
compared to check if it changed. If so then the hash bucket has been
replaced and the user has been moved to the new one and lock_ptr has
been updated. The lock operation needs to be redone in this case.
The locked hash bucket is not returned.
A special case is an early return in futex_lock_pi() (due to signal or
timeout) and a successful futex_wait_requeue_pi(). In both cases a valid
futex_q::lock_ptr is expected (and its matching hash bucket) but since
the waiter has been removed from the hash this can no longer be
guaranteed. Therefore before the waiter is removed and a reference is
acquired which is later dropped by the waiter to avoid a resize.
Add futex_q_lockptr_lock() and use it.
Acquire an additional reference in requeue_pi_wake_futex() and
futex_unlock_pi() while the futex_q is removed, denote this extra
reference in futex_q::drop_hb_ref and let the waiter drop the reference
in this case.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-11-bigeasy@linutronix.de
To support runtime resizing of the process private hash, it's required
to not use the obtained hash bucket once the reference count has been
dropped. The reference will be dropped after the unlock of the hash
bucket.
The amount of waiters is decremented after the unlock operation. There
is no requirement that this needs to happen after the unlock. The
increment happens before acquiring the lock to signal early that there
will be a waiter. The waiter can avoid blocking on the lock if it is
known that there will be no waiter.
There is no difference in terms of ordering if the decrement happens
before or after the unlock.
Decrease the waiter count before the unlock operation.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-10-bigeasy@linutronix.de
futex_wait_multiple_setup() changes task_struct::__state to
!TASK_RUNNING and then enqueues on multiple futexes. Every
futex_q_lock() acquires a reference on the global hash which is
dropped later.
If a rehash is in progress then the loop will block on
mm_struct::futex_hash_bucket for the rehash to complete and this will
lose the previously set task_struct::__state.
Acquire a reference on the local hash to avoiding blocking on
mm_struct::futex_hash_bucket.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-9-bigeasy@linutronix.de
futex_wait_setup() has a weird calling convention in order to return
hb to use as an argument to futex_queue().
Mostly such that requeue can have an extra test in between.
Reorder code a little to get rid of this and keep the hb usage inside
futex_wait_setup().
[bigeasy: fixes]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250416162921.513656-4-bigeasy@linutronix.de
The global hash uses futex_hashsize to save the amount of the hash
buckets that have been allocated during system boot. On each
futex_hash() invocation this number is substracted by one to get the
mask. This can be optimized by saving directly the mask avoiding the
substraction on each futex_hash() invocation.
Rename futex_hashsize to futex_hashmask and save the mask of the
allocated hash map.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Waiman Long <longman@redhat.com>
Link: https://lore.kernel.org/all/20250226091057.bX8vObR4@linutronix.de
Pull locking fix from Ingo Molnar:
"Fix a dangling pointer bug in the futex code used by the uring code.
It isn't causing problems at the moment due to uring ABI limitations
leaving it essentially unused in current usages, but is a good idea to
fix nevertheless"
* tag 'locking-urgent-2025-02-08' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
futex: Pass in task to futex_queue()
Pull non-MM updates from Andrew Morton:
"Mainly individually changelogged singleton patches. The patch series
in this pull are:
- "lib min_heap: Improve min_heap safety, testing, and documentation"
from Kuan-Wei Chiu provides various tightenings to the min_heap
library code
- "xarray: extract __xa_cmpxchg_raw" from Tamir Duberstein preforms
some cleanup and Rust preparation in the xarray library code
- "Update reference to include/asm-<arch>" from Geert Uytterhoeven
fixes pathnames in some code comments
- "Converge on using secs_to_jiffies()" from Easwar Hariharan uses
the new secs_to_jiffies() in various places where that is
appropriate
- "ocfs2, dlmfs: convert to the new mount API" from Eric Sandeen
switches two filesystems to the new mount API
- "Convert ocfs2 to use folios" from Matthew Wilcox does that
- "Remove get_task_comm() and print task comm directly" from Yafang
Shao removes now-unneeded calls to get_task_comm() in various
places
- "squashfs: reduce memory usage and update docs" from Phillip
Lougher implements some memory savings in squashfs and performs
some maintainability work
- "lib: clarify comparison function requirements" from Kuan-Wei Chiu
tightens the sort code's behaviour and adds some maintenance work
- "nilfs2: protect busy buffer heads from being force-cleared" from
Ryusuke Konishi fixes an issues in nlifs when the fs is presented
with a corrupted image
- "nilfs2: fix kernel-doc comments for function return values" from
Ryusuke Konishi fixes some nilfs kerneldoc
- "nilfs2: fix issues with rename operations" from Ryusuke Konishi
addresses some nilfs BUG_ONs which syzbot was able to trigger
- "minmax.h: Cleanups and minor optimisations" from David Laight does
some maintenance work on the min/max library code
- "Fixes and cleanups to xarray" from Kemeng Shi does maintenance
work on the xarray library code"
* tag 'mm-nonmm-stable-2025-01-24-23-16' of git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm: (131 commits)
ocfs2: use str_yes_no() and str_no_yes() helper functions
include/linux/lz4.h: add some missing macros
Xarray: use xa_mark_t in xas_squash_marks() to keep code consistent
Xarray: remove repeat check in xas_squash_marks()
Xarray: distinguish large entries correctly in xas_split_alloc()
Xarray: move forward index correctly in xas_pause()
Xarray: do not return sibling entries from xas_find_marked()
ipc/util.c: complete the kernel-doc function descriptions
gcov: clang: use correct function param names
latencytop: use correct kernel-doc format for func params
minmax.h: remove some #defines that are only expanded once
minmax.h: simplify the variants of clamp()
minmax.h: move all the clamp() definitions after the min/max() ones
minmax.h: use BUILD_BUG_ON_MSG() for the lo < hi test in clamp()
minmax.h: reduce the #define expansion of min(), max() and clamp()
minmax.h: update some comments
minmax.h: add whitespace around operators and after commas
nilfs2: do not update mtime of renamed directory that is not moved
nilfs2: handle errors that nilfs_prepare_chunk() may return
CREDITS: fix spelling mistake
...
futex_queue() -> __futex_queue() uses 'current' as the task to store in
the struct futex_q->task field. This is fine for synchronous usage of
the futex infrastructure, but it's not always correct when used by
io_uring where the task doing the initial futex_queue() might not be
available later on. This doesn't lead to any issues currently, as the
io_uring side doesn't support PI futexes, but it does leave a
potentially dangling pointer which is never a good idea.
Have futex_queue() take a task_struct argument, and have the regular
callers pass in 'current' for that. Meanwhile io_uring can just pass in
NULL, as the task should never be used off that path. In theory
req->tctx->task could be used here, but there's no point populating it
with a task field that will never be used anyway.
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/22484a23-542c-4003-b721-400688a0d055@kernel.dk
A common pattern seen when wake_qs are used to defer a wakeup
until after a lock is released is something like:
preempt_disable();
raw_spin_unlock(lock);
wake_up_q(wake_q);
preempt_enable();
So create some raw_spin_unlock*_wake() helper functions to clean
this up.
Applies on top of the fix I submitted here:
https://lore.kernel.org/lkml/20241212222138.2400498-1-jstultz@google.com/
NOTE: I recognise the unlock()/unlock_irq()/unlock_irqrestore()
variants creates its own duplication, which we could use a macro
to generate the similar functions, but I often dislike how those
generation macros making finding the actual implementation
harder, so I left the three functions as is. If folks would
prefer otherwise, let me know and I'll switch it.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241217040803.243420-1-jstultz@google.com
The powerpc user access code is special, and unlike other architectures
distinguishes between user access for reading and writing.
And commit 43a43faf53 ("futex: improve user space accesses") messed
that up. It went undetected elsewhere, but caused ppc32 to fail early
during boot, because the user access had been started with
user_read_access_begin(), but then finished off with just a plain
"user_access_end()".
Note that the address-masking user access helpers don't even have that
read-vs-write distinction, so if powerpc ever wants to do address
masking tricks, we'll have to do some extra work for it.
[ Make sure to also do it for the EFAULT case, as pointed out by
Christophe Leroy ]
Reported-by: Andreas Schwab <schwab@linux-m68k.org>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Link: https://lore.kernel.org/all/87bjxl6b0i.fsf@igel.home/
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Josh Poimboeuf reports that he got a "will-it-scale.per_process_ops 1.9%
improvement" report for his patch that changed __get_user() to use
pointer masking instead of the explicit speculation barrier. However,
that patch doesn't actually work in the general case, because some (very
bad) architecture-specific code actually depends on __get_user() also
working on kernel addresses.
A profile showed that the offending __get_user() was the futex code,
which really should be fixed up to not use that horrid legacy case.
Rewrite futex_get_value_locked() to use the modern user acccess helpers,
and inline it so that the compiler not only avoids the function call for
a few instructions, but can do CSE on the address masking.
It also turns out the x86 futex functions have unnecessary barriers in
other places, so let's fix those up too.
Link: https://lore.kernel.org/all/20241115230653.hfvzyf3aqqntgp63@jpoimboe/
Reported-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Sebastian Andrzej Siewior
Borislav Petkov (AMD)
Peter Zijlstra
Linus Torvalds
Jens Axboe
Yafang Shao
John Stultz