gcc-16 has gained some more advanced inter-procedual optimization
techniques that enable it to inline the dummy_tlb_add_page() and
dummy_tlb_flush() function pointers into a specialized version of
__arm_v7s_unmap:
WARNING: modpost: vmlinux: section mismatch in reference: __arm_v7s_unmap+0x2cc (section: .text) -> dummy_tlb_add_page (section: .init.text)
ERROR: modpost: Section mismatches detected.
>From what I can tell, the transformation is correct, as this is only
called when __arm_v7s_unmap() is called from arm_v7s_do_selftests(),
which is also __init. Since __arm_v7s_unmap() however is not __init,
gcc cannot inline the inner function calls directly.
In debug_objects_selftest(), the same thing happens. Both the
caller and the leaf function are __init, but the IPA pulls
it into a non-init one:
WARNING: modpost: vmlinux: section mismatch in reference: lookup_object_or_alloc+0x7c (section: .text.lookup_object_or_alloc) -> is_static_object (section: .init.text)
Marking the affected functions as not "__init" would reliably avoid this
issue but is not a good solution because it removes an otherwise correct
annotation. I tried marking the functions as 'noinline', but that ended
up not covering all the affected configurations.
With some more experimenting, I found that marking these functions as
__attribute__((noipa)) is both logical and reliable.
In order to keep the syntax readable, add a custom macro for this in
include/linux/compiler_attributes.h next to other related macros and
use it to annotate both files.
Link: https://lore.kernel.org/all/abRB6g-48ZX6Yl2r@willie-the-truck/
Cc: Will Deacon <will@kernel.org>
Cc: Thomas Gleixner <tglx@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Miguel Ojeda <ojeda@kernel.org>
Cc: linux-kbuild@vger.kernel.org
Cc: stable@vger.kernel.org
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Will Deacon <will@kernel.org>
Acked-by: Thomas Gleixner <tglx@kernel.org>
Acked-by: Miguel Ojeda <ojeda@kernel.org>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
debugobjects uses __GFP_HIGH for allocations as it might be invoked
within locked regions. That worked perfectly fine until v6.18. It still
works correctly when deferred page initialization is disabled and works
by chance when no page allocation is required before deferred page
initialization has completed.
Since v6.18 allocations w/o a reclaim flag cause new_slab() to end up in
alloc_frozen_pages_nolock_noprof(), which returns early when deferred
page initialization has not yet completed. As the deferred page
initialization takes quite a while the debugobject pool is depleted and
debugobjects are disabled.
This can be worked around when PREEMPT_COUNT is enabled as that allows
debugobjects to add __GFP_KSWAPD_RECLAIM to the GFP flags when the context
is preemtible. When PREEMPT_COUNT is disabled the context is unknown and
the reclaim bit can't be set because the caller might hold locks which
might deadlock in the allocator.
In preemptible context the reclaim bit is harmless and not a performance
issue as that's usually invoked from slow path initialization context.
That makes debugobjects depend on PREEMPT_COUNT || !DEFERRED_STRUCT_PAGE_INIT.
Fixes: af92793e52 ("slab: Introduce kmalloc_nolock() and kfree_nolock().")
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Link: https://patch.msgid.link/87pl6gznti.ffs@tglx
In a vain attempt to consolidate the email zoo switch everything to the
kernel.org account.
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fill_pool_map is used to suppress nesting violations caused by acquiring
a spinlock_t (from within the memory allocator) while holding a
raw_spinlock_t. The used annotation is wrong.
LD_WAIT_SLEEP is for always sleeping lock types such as mutex_t.
LD_WAIT_CONFIG is for lock type which are sleeping while spinning on
PREEMPT_RT such as spinlock_t.
Use LD_WAIT_CONFIG as override.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://patch.msgid.link/20251127153652.291697-3-bigeasy@linutronix.de
The pool of free objects is refilled on several occasions such as object
initialisation. On PREEMPT_RT refilling is limited to preemptible
sections due to sleeping locks used by the memory allocator. The system
boots with disabled interrupts so the pool can not be refilled.
If too many objects are initialized and the pool gets empty then
debugobjects disables itself.
Refiling can also happen early in the boot with disabled interrupts as
long as the scheduler is not operational. If the scheduler can not
preempt a task then a sleeping lock can not be contended.
Allow to additionally refill the pool if the scheduler is not
operational.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://patch.msgid.link/20251127153652.291697-2-bigeasy@linutronix.de
The freelist is freed at a constant rate independent of the actual usage
requirements. That's bad in scenarios where usage comes in bursts. The end
of a burst puts the objects on the free list and freeing proceeds even when
the next burst which requires objects started again.
Keep track of the usage with a exponentially wheighted moving average and
take that into account in the worker function which frees objects from the
free list.
This further reduces the kmem_cache allocation/free rate for a full kernel
compile:
kmem_cache_alloc() kmem_cache_free()
Baseline: 225k 173k
Usage: 170k 117k
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/87bjznhme2.ffs@tglx
Right now the per CPU pools are only refilled when they become
empty. That's suboptimal especially when there are still non-freed objects
in the to free list.
Check whether an allocation from the per CPU pool emptied a batch and try
to allocate from the free pool if that still has objects available.
kmem_cache_alloc() kmem_cache_free()
Baseline: 295k 245k
Refill: 225k 173k
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164914.439053085@linutronix.de
In situations where objects are rapidly allocated from the pool and handed
back, the size of the per CPU pool turns out to be too small.
Double the size of the per CPU pool.
This reduces the kmem cache allocation and free operations during a kernel compile:
alloc free
Baseline: 380k 330k
Double size: 295k 245k
Especially the reduction of allocations is important because that happens
in the hot path when objects are initialized.
The maximum increase in per CPU pool memory consumption is about 2.5K per
online CPU, which is acceptable.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164914.378676302@linutronix.de
Adding and removing single objects in a loop is bad in terms of lock
contention and cache line accesses.
To implement batching, record the last object in a batch in the object
itself. This is trivialy possible as hlists are strictly stacks. At a batch
boundary, when the first object is added to the list the object stores a
pointer to itself in debug_obj::batch_last. When the next object is added
to the list then the batch_last pointer is retrieved from the first object
in the list and stored in the to be added one.
That means for batch processing the first object always has a pointer to
the last object in a batch, which allows to move batches in a cache line
efficient way and reduces the lock held time.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164914.258995000@linutronix.de
Move the debug_obj::object pointer into a union and add a pointer to the
last node in a batch. That allows to implement batch processing efficiently
by utilizing the stack property of hlist:
When the first object of a batch is added to the list, then the batch
pointer is set to the hlist node of the object itself. Any subsequent add
retrieves the pointer to the last node from the first object in the list
and uses that for storing the last node pointer in the newly added object.
Add the pointer to the data structure and ensure that all relevant pool
sizes are strictly batch sized. The actual batching implementation follows
in subsequent changes.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164914.139204961@linutronix.de
__free_object() is uncomprehensibly complex. The same can be achieved by:
1) Adding the object to the per CPU pool
2) If that pool is full, move a batch of objects into the global pool
or if the global pool is full into the to free pool
This also prepares for batch processing.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164913.955542307@linutronix.de
The current allocation scheme tries to allocate from the per CPU pool
first. If that fails it allocates one object from the global pool and then
refills the per CPU pool from the global pool.
That is in the way of switching the pool management to batch mode as the
global pool needs to be a strict stack of batches, which does not allow
to allocate single objects.
Rework the code to refill the per CPU pool first and then allocate the
object from the refilled batch. Also try to allocate from the to free pool
first to avoid freeing and reallocating objects.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164913.893554162@linutronix.de
The contention on the global pool lock can be reduced by strict batch
processing where batches of objects are moved from one list head to another
instead of moving them object by object. This also reduces the cache
footprint because it avoids the list walk and dirties at maximum three
cache lines instead of potentially up to eighteen.
To prepare for that, move the hlist head and related counters into a
struct.
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164913.646171170@linutronix.de
The contention on the global pool_lock can be massive when the global pool
needs to be refilled and many CPUs try to handle this.
Address this by:
- splitting the refill from free list and allocation.
Refill from free list has no constraints vs. the context on RT, so
it can be tried outside of the RT specific preemptible() guard
- Let only one CPU handle the free list
- Let only one CPU do allocations unless the pool level is below
half of the minimum fill level.
Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/20240911083521.2257-4-thunder.leizhen@huawei.com-
Link: https://lore.kernel.org/all/20241007164913.582118421@linutronix.de
--
lib/debugobjects.c | 84 +++++++++++++++++++++++++++++++++++++----------------
1 file changed, 59 insertions(+), 25 deletions(-)
Linus Torvalds
Arnd BergmannandJoerg Roedel
Helen KoikeandThomas Gleixner
Philipp HahnandThomas Gleixner
Sebastian Andrzej SiewiorandThomas Gleixner
Zhen LeiandThomas Gleixner