By default the cache size is 5 pages per CPU, but it can be disabled at
boot time by setting the rcu_min_cached_objs to zero. When that happens,
the current code will uselessly set an hrtimer to schedule refilling this
cache with zero pages. This commit therefore streamlines this process
by simply refusing the set the hrtimer when rcu_min_cached_objs is zero.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Under low-memory conditions, kvfree_rcu() will use each object's
rcu_head structure to queue objects in a singly linked list headed by
the kfree_rcu_cpu structure's ->head field. This list is passed to
call_rcu() as a unit, but there is no indication of which grace period
this list needs to wait for. This in turn prevents adding debug checks
in the kfree_rcu_work() as was done for the two page-of-pointers channels
in the kfree_rcu_cpu structure.
This commit therefore adds a ->head_free_gp_snap field to the
kfree_rcu_cpu_work structure to record this grace-period number. It also
adds a WARN_ON_ONCE() to kfree_rcu_monitor() that checks to make sure
that the required grace period has in fact elapsed.
[ paulmck: Fix kerneldoc issue raised by Stephen Rothwell. ]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The kvfree_rcu() and kfree_rcu() APIs are hazardous in that if you forget
the second argument, it works, but might sleep. This sleeping can be a
correctness bug from atomic contexts, and even in non-atomic contexts it
might introduce unacceptable latencies. This commit therefore removes the
single-argument kvfree_rcu() and kfree_rcu() macros. Code that would have
previously used these single-argument kvfree_rcu() and kfree_rcu() macros
should instead use kvfree_rcu_mightsleep() or kfree_rcu_mightsleep().
[ paulmck: Apply Joel Fernandes feedback. ]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
A global vmap_blocks-xarray array can be contented under heavy usage of
the vm_map_ram()/vm_unmap_ram() APIs. The lock_stat shows that a
"vmap_blocks.xa_lock" lock is a second in a top-list when it comes to
contentions:
<snip>
----------------------------------------
class name con-bounces contentions ...
----------------------------------------
vmap_area_lock: 2554079 2554276 ...
--------------
vmap_area_lock 1297948 [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
vmap_area_lock 1256330 [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
vmap_area_lock 1 [<00000000c95c05a7>] find_vm_area+0x16/0x70
--------------
vmap_area_lock 1738590 [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
vmap_area_lock 815688 [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
vmap_area_lock 1 [<00000000c1d619d7>] __get_vm_area_node+0xd2/0x170
vmap_blocks.xa_lock: 862689 862698 ...
-------------------
vmap_blocks.xa_lock 378418 [<00000000625a5626>] vm_map_ram+0x359/0x4a0
vmap_blocks.xa_lock 484280 [<00000000caa2ef03>] xa_erase+0xe/0x30
-------------------
vmap_blocks.xa_lock 576226 [<00000000caa2ef03>] xa_erase+0xe/0x30
vmap_blocks.xa_lock 286472 [<00000000625a5626>] vm_map_ram+0x359/0x4a0
...
<snip>
that is a result of running vm_map_ram()/vm_unmap_ram() in
a loop. The test creates 64(on 64 CPUs system) threads and
each one maps/unmaps 1 page.
After this change the "xa_lock" can be considered as a noise
in the same test condition:
<snip>
...
&xa->xa_lock#1: 10333 10394 ...
--------------
&xa->xa_lock#1 5349 [<00000000bbbc9751>] xa_erase+0xe/0x30
&xa->xa_lock#1 5045 [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
--------------
&xa->xa_lock#1 7326 [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
&xa->xa_lock#1 3068 [<00000000bbbc9751>] xa_erase+0xe/0x30
...
<snip>
Running the test_vmalloc.sh run_test_mask=1024 nr_threads=64 nr_pages=5
shows around ~8 percent of throughput improvement of vm_map_ram() and
vm_unmap_ram() APIs.
This patch does not fix vmap_area_lock/free_vmap_area_lock and
purge_vmap_area_lock bottle-necks, it is rather a separate rework.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Lorenzo Stoakes <[email protected]>
Reviewed-by: Baoquan He <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Matthew Wilcox (Oracle) <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
The kfree_rcu() and kvfree_rcu() macros' single-argument forms are
deprecated. Therefore switch to the new kfree_rcu_mightsleep() and
kvfree_rcu_mightsleep() variants. The goal is to avoid accidental use
of the single-argument forms, which can introduce functionality bugs in
atomic contexts and latency bugs in non-atomic contexts.
Acked-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kfree_rcu() and kvfree_rcu() macros' single-argument forms are
deprecated. Therefore switch to the new kfree_rcu_mightsleep() and
kvfree_rcu_mightsleep() variants. The goal is to avoid accidental use
of the single-argument forms, which can introduce functionality bugs in
atomic contexts and latency bugs in non-atomic contexts.
Cc: Theodore Ts'o <[email protected]>
Cc: Lukas Czerner <[email protected]>
Acked-by: Theodore Ts'o <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kfree_rcu() and kvfree_rcu() macros' single-argument forms are
deprecated. Therefore switch to the new kfree_rcu_mightsleep() and
kvfree_rcu_mightsleep() variants. The goal is to avoid accidental use
of the single-argument forms, which can introduce functionality bugs in
atomic contexts and latency bugs in non-atomic contexts.
Cc: Ariel Levkovich <[email protected]>
Cc: Saeed Mahameed <[email protected]>
Cc: Vlad Buslov <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kfree_rcu() and kvfree_rcu() macros' single-argument forms are
deprecated. Therefore switch to the new kfree_rcu_mightsleep() and
kvfree_rcu_mightsleep() variants. The goal is to avoid accidental use
of the single-argument forms, which can introduce functionality bugs in
atomic contexts and latency bugs in non-atomic contexts.
Acked-by: Jakub Kicinski <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: David S. Miller <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kvfree_rcu() macro's single-argument form is deprecated. Therefore
switch to the new kvfree_rcu_mightsleep() variant. The goal is to
avoid accidental use of the single-argument forms, which can introduce
functionality bugs in atomic contexts and latency bugs in non-atomic
contexts.
Acked-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kvfree_rcu() macro's single-argument form is deprecated. Therefore
switch to the new kvfree_rcu_mightsleep() variant. The goal is to
avoid accidental use of the single-argument forms, which can introduce
functionality bugs in atomic contexts and latency bugs in non-atomic
contexts.
Cc: Steven Rostedt (VMware) <[email protected]>
Acked-by: Daniel Bristot de Oliveira <[email protected]>
Acked-by: Paul E. McKenney <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kvfree_rcu() macro's single-argument form is deprecated. Therefore
switch to the new kvfree_rcu_mightsleep() variant. The goal is to
avoid accidental use of the single-argument forms, which can introduce
functionality bugs in atomic contexts and latency bugs in non-atomic
contexts.
Cc: Bryan Tan <[email protected]>
Cc: Vishnu Dasa <[email protected]>
Reviewed-by: Vishnu Dasa <[email protected]>
Acked-by: Greg Kroah-Hartman <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kvfree_rcu() macro's single-argument form is deprecated. Therefore
switch to the new kvfree_rcu_mightsleep() variant. The goal is to
avoid accidental use of the single-argument forms, which can introduce
functionality bugs in atomic contexts and latency bugs in non-atomic
contexts.
Cc: Jens Axboe <[email protected]>
Cc: Philipp Reisner <[email protected]>
Cc: Lars Ellenberg <[email protected]>
Begrudgingly-acked-by: Jens Axboe <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kfree_rcu() macro is deprecated. Rename it to its new
kfree_rcu_mightsleep() name in this documentation.
Acked-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
The kvfree_rcu() and kfree_rcu() APIs are hazardous in that if you forget
the second argument, it works, but might sleep. This sleeping can be a
correctness bug from atomic contexts, and even in non-atomic contexts
it might introduce unacceptable latencies. This commit therefore adds
kvfree_rcu_mightsleep() and kfree_rcu_mightsleep(), which will replace
the single-argument kvfree_rcu() and kfree_rcu(), respectively.
This commit enables a series of commits that switch from single-argument
kvfree_rcu() and kfree_rcu() to their _mightsleep() counterparts. Once
all of these commits land, the single-argument versions will be removed.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Currently the __vunmap() path calls __find_vmap_area() twice. Once on
entry to check that the area exists, then inside the remove_vm_area()
function which also performs a new search for the VA.
In order to improvie it from a performance point of view we split
remove_vm_area() into two new parts:
- find_unlink_vmap_area() that does a search and unlink from tree;
- __remove_vm_area() that removes without searching.
In this case there is no any functional change for remove_vm_area()
whereas vm_remove_mappings(), where a second search happens, switches to
the __remove_vm_area() variant where the already detached VA is passed as
a parameter, so there is no need to find it again.
Performance wise, i use test_vmalloc.sh with 32 threads doing alloc
free on a 64-CPUs-x86_64-box:
perf without this patch:
- 31.41% 0.50% vmalloc_test/10 [kernel.vmlinux] [k] __vunmap
- 30.92% __vunmap
- 17.67% _raw_spin_lock
native_queued_spin_lock_slowpath
- 12.33% remove_vm_area
- 11.79% free_vmap_area_noflush
- 11.18% _raw_spin_lock
native_queued_spin_lock_slowpath
0.76% free_unref_page
perf with this patch:
- 11.35% 0.13% vmalloc_test/14 [kernel.vmlinux] [k] __vunmap
- 11.23% __vunmap
- 8.28% find_unlink_vmap_area
- 7.95% _raw_spin_lock
7.44% native_queued_spin_lock_slowpath
- 1.93% free_vmap_area_noflush
- 0.56% _raw_spin_lock
0.53% native_queued_spin_lock_slowpath
0.60% __vunmap_range_noflush
__vunmap() consumes around ~20% less CPU cycles on this test.
Also, switch from find_vmap_area() to find_unlink_vmap_area() to prevent a
double access to the vmap_area_lock: one for finding area, second time is
for unlinking from a tree.
[[email protected]: switch to find_unlink_vmap_area() in vm_unmap_ram()]
Link: https://lkml.kernel.org/r/[email protected]
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reported-by: Roman Gushchin <[email protected]>
Reviewed-by: Lorenzo Stoakes <[email protected]>
Cc: Baoquan He <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox (Oracle) <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
This patch splits the lists of objects so as to avoid sending any
through RCU that have already been queued for more than one grace
period. These long-term-resident objects are immediately freed.
The remaining short-term-resident objects are queued for later freeing
using queue_rcu_work().
This change avoids delaying workqueue handlers with synchronize_rcu()
invocations. Yes, workqueue handlers are designed to handle blocking,
but avoiding blocking when unnecessary improves performance during
low-memory situations.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The schedule_delayed_monitor_work() function relies on the count of
objects queued into any given kfree_rcu_cpu structure. This count is
used to determine how quickly to schedule passing these objects to RCU.
There are three pipes where pointers can be placed. When any pipe is
offloaded, the kfree_rcu_cpu structure's ->count counter is set to zero,
which is wrong because the other pipes might still be non-empty.
This commit therefore maintains per-pipe counters, and introduces a
krc_count() helper to access the aggregate value of those counters.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The need_offload_krc() function is now lock-free, which gives the
compiler freedom to load old values from plain C-language loads from
the kfree_rcu_cpu struture's ->head pointer. This commit therefore
applied READ_ONCE() to these loads.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Currently all objects placed into a batch wait for a full grace period
to elapse after that batch is ready to send to RCU. However, this
can unnecessarily delay freeing of the first objects that were added
to the batch. After all, several RCU grace periods might have elapsed
since those objects were added, and if so, there is no point in further
deferring their freeing.
This commit therefore adds per-page grace-period snapshots which are
obtained from get_state_synchronize_rcu(). When the batch is ready
to be passed to call_rcu(), each page's snapshot is checked by passing
it to poll_state_synchronize_rcu(). If a given page's RCU grace period
has already elapsed, its objects are freed immediately by kvfree_rcu_bulk().
Otherwise, these objects are freed after a call to synchronize_rcu().
This approach requires that the pages be traversed in reverse order,
that is, the oldest ones first.
Test example:
kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \
--kconfig CONFIG_NR_CPUS=64 \
--kconfig CONFIG_RCU_NOCB_CPU=y \
--kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \
--kconfig CONFIG_RCU_LAZY=n \
--bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \
rcuscale.holdoff=20 rcuscale.kfree_loops=10000 \
torture.disable_onoff_at_boot" --trust-make
Before this commit:
Total time taken by all kfree'ers: 8535693700 ns, loops: 10000, batches: 1188, memory footprint: 2248MB
Total time taken by all kfree'ers: 8466933582 ns, loops: 10000, batches: 1157, memory footprint: 2820MB
Total time taken by all kfree'ers: 5375602446 ns, loops: 10000, batches: 1130, memory footprint: 6502MB
Total time taken by all kfree'ers: 7523283832 ns, loops: 10000, batches: 1006, memory footprint: 3343MB
Total time taken by all kfree'ers: 6459171956 ns, loops: 10000, batches: 1150, memory footprint: 6549MB
After this commit:
Total time taken by all kfree'ers: 8560060176 ns, loops: 10000, batches: 1787, memory footprint: 61MB
Total time taken by all kfree'ers: 8573885501 ns, loops: 10000, batches: 1777, memory footprint: 93MB
Total time taken by all kfree'ers: 8320000202 ns, loops: 10000, batches: 1727, memory footprint: 66MB
Total time taken by all kfree'ers: 8552718794 ns, loops: 10000, batches: 1790, memory footprint: 75MB
Total time taken by all kfree'ers: 8601368792 ns, loops: 10000, batches: 1724, memory footprint: 62MB
The reduction in memory footprint is well in excess of an order of
magnitude.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The need_offload_krc() function currently holds the krcp->lock in order
to safely check krcp->head. This commit removes the need for this lock
in that function by updating the krcp->head pointer using WRITE_ONCE()
macro so that readers can carry out lockless loads of that pointer.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The kvfree_rcu() code maintains lists of pages of pointers, but also a
singly linked list, with the latter being used when memory allocation
fails. Traversal of these two types of lists is currently open coded.
This commit simplifies the code by providing kvfree_rcu_bulk() and
kvfree_rcu_list() functions, respectively, to traverse these two types
of lists. This patch does not introduce any functional change.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
This commit improves the readability and maintainability of the
kvfree_rcu() code by switching from an open-coded linked list to
the standard Linux-kernel circular doubly linked list. This patch
does not introduce any functional change.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Currently a kvfree_call_rcu() takes an offset within a structure as
a second parameter, so a helper such as a kvfree_rcu_arg_2() has to
convert rcu_head and a freed ptr to an offset in order to pass it. That
leads to an extra conversion on macro entry.
Instead of converting, refactor the code in way that a pointer that has
to be freed is passed directly to the kvfree_call_rcu().
This patch does not make any functional change and is transparent to
all kvfree_rcu() users.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Earlier commits in this series allow battery-powered systems to build
their kernels with the default-disabled CONFIG_RCU_LAZY=y Kconfig option.
This Kconfig option causes call_rcu() to delay its callbacks in order
to batch them. This means that a given RCU grace period covers more
callbacks, thus reducing the number of grace periods, in turn reducing
the amount of energy consumed, which increases battery lifetime which
can be a very good thing. This is not a subtle effect: In some important
use cases, the battery lifetime is increased by more than 10%.
This CONFIG_RCU_LAZY=y option is available only for CPUs that offload
callbacks, for example, CPUs mentioned in the rcu_nocbs kernel boot
parameter passed to kernels built with CONFIG_RCU_NOCB_CPU=y.
Delaying callbacks is normally not a problem because most callbacks do
nothing but free memory. If the system is short on memory, a shrinker
will kick all currently queued lazy callbacks out of their laziness,
thus freeing their memory in short order. Similarly, the rcu_barrier()
function, which blocks until all currently queued callbacks are invoked,
will also kick lazy callbacks, thus enabling rcu_barrier() to complete
in a timely manner.
However, there are some cases where laziness is not a good option.
For example, synchronize_rcu() invokes call_rcu(), and blocks until
the newly queued callback is invoked. It would not be a good for
synchronize_rcu() to block for ten seconds, even on an idle system.
Therefore, synchronize_rcu() invokes call_rcu_hurry() instead of
call_rcu(). The arrival of a non-lazy call_rcu_hurry() callback on a
given CPU kicks any lazy callbacks that might be already queued on that
CPU. After all, if there is going to be a grace period, all callbacks
might as well get full benefit from it.
Yes, this could be done the other way around by creating a
call_rcu_lazy(), but earlier experience with this approach and
feedback at the 2022 Linux Plumbers Conference shifted the approach
to call_rcu() being lazy with call_rcu_hurry() for the few places
where laziness is inappropriate.
And another call_rcu() instance that cannot be lazy is the one
in queue_rcu_work(), given that callers to queue_rcu_work() are
not necessarily OK with long delays.
Therefore, make queue_rcu_work() use call_rcu_hurry() in order to revert
to the old behavior.
[ paulmck: Apply s/call_rcu_flush/call_rcu_hurry/ feedback from Tejun Heo. ]
Signed-off-by: Uladzislau Rezki <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Acked-by: Tejun Heo <[email protected]>
Cc: Lai Jiangshan <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Earlier commits in this series allow battery-powered systems to build
their kernels with the default-disabled CONFIG_RCU_LAZY=y Kconfig option.
This Kconfig option causes call_rcu() to delay its callbacks in order
to batch them. This means that a given RCU grace period covers more
callbacks, thus reducing the number of grace periods, in turn reducing
the amount of energy consumed, which increases battery lifetime which
can be a very good thing. This is not a subtle effect: In some important
use cases, the battery lifetime is increased by more than 10%.
This CONFIG_RCU_LAZY=y option is available only for CPUs that offload
callbacks, for example, CPUs mentioned in the rcu_nocbs kernel boot
parameter passed to kernels built with CONFIG_RCU_NOCB_CPU=y.
Delaying callbacks is normally not a problem because most callbacks do
nothing but free memory. If the system is short on memory, a shrinker
will kick all currently queued lazy callbacks out of their laziness,
thus freeing their memory in short order. Similarly, the rcu_barrier()
function, which blocks until all currently queued callbacks are invoked,
will also kick lazy callbacks, thus enabling rcu_barrier() to complete
in a timely manner.
However, there are some cases where laziness is not a good option.
For example, synchronize_rcu() invokes call_rcu(), and blocks until
the newly queued callback is invoked. It would not be a good for
synchronize_rcu() to block for ten seconds, even on an idle system.
Therefore, synchronize_rcu() invokes call_rcu_hurry() instead of
call_rcu(). The arrival of a non-lazy call_rcu_hurry() callback on a
given CPU kicks any lazy callbacks that might be already queued on that
CPU. After all, if there is going to be a grace period, all callbacks
might as well get full benefit from it.
Yes, this could be done the other way around by creating a
call_rcu_lazy(), but earlier experience with this approach and
feedback at the 2022 Linux Plumbers Conference shifted the approach
to call_rcu() being lazy with call_rcu_hurry() for the few places
where laziness is inappropriate.
And another call_rcu() instance that cannot be lazy is the one in the
scsi_eh_scmd_add() function. Leaving this instance lazy results in
unacceptably slow boot times.
Therefore, make scsi_eh_scmd_add() use call_rcu_hurry() in order to
revert to the old behavior.
[ paulmck: Apply s/call_rcu_flush/call_rcu_hurry/ feedback from Tejun Heo. ]
Tested-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Cc: "James E.J. Bottomley" <[email protected]>
Cc: <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Acked-by: Martin K. Petersen <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
This is for debug purpose and is called when an allocation attempt occurs.
This event gives some information about:
- start address of allocated area;
- size that is requested;
- alignment that is required;
- vstart/vend restriction;
- if an allocation fails.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Steven Rostedt (Google) <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox (Oracle) <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Currently the monitor work is scheduled with a fixed interval of HZ/20,
which is roughly 50 milliseconds. The drawback of this approach is
low utilization of the 512 page slots in scenarios with infrequence
kvfree_rcu() calls. For example on an Android system:
<snip>
kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6
kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1
kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9
kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48
kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3
kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76
kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1
kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5
kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102
<snip>
In many cases, out of 512 slots, fewer than 10 were actually used.
In order to improve batching and make utilization more efficient this
commit sets a drain interval to a fixed 5-seconds interval. Floods are
detected when a page fills quickly, and in that case, the reclaim work
is re-scheduled for the next scheduling-clock tick (jiffy).
After this change:
<snip>
kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121
kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47
kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510
kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169
kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510
kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123
kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322
kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185
kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510
<snip>
Here, all but one of the cases, more than one hundreds slots were used,
representing an order-of-magnitude improvement.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
This commit introduces a RCU_NOCB_CPU_CB_BOOST Kconfig option that
prevents rcuo kthreads from running at real-time priority, even in
kernels built with RCU_BOOST. This capability is important to devices
needing low-latency (as in a few milliseconds) response from expedited
RCU grace periods, but which are not running a classic real-time workload.
On such devices, permitting the rcuo kthreads to run at real-time priority
results in unacceptable latencies imposed on the application tasks,
which run as SCHED_OTHER.
See for example the following trace output:
<snip>
<...>-60 [006] d..1 2979.028717: rcu_batch_start: rcu_preempt CBs=34619 bl=270
<snip>
If that rcuop kthread were permitted to run at real-time SCHED_FIFO
priority, it would monopolize its CPU for hundreds of milliseconds
while invoking those 34619 RCU callback functions, which would cause an
unacceptably long latency spike for many application stacks on Android
platforms.
However, some existing real-time workloads require that callback
invocation run at SCHED_FIFO priority, for example, those running on
systems with heavy SCHED_OTHER background loads. (It is the real-time
system's administrator's responsibility to make sure that important
real-time tasks run at a higher priority than do RCU's kthreads.)
Therefore, this new RCU_NOCB_CPU_CB_BOOST Kconfig option defaults to
"y" on kernels built with PREEMPT_RT and defaults to "n" otherwise.
The effect is to preserve current behavior for real-time systems, but for
other systems to allow expedited RCU grace periods to run with real-time
priority while continuing to invoke RCU callbacks as SCHED_OTHER.
As you would expect, this RCU_NOCB_CPU_CB_BOOST Kconfig option has no
effect except on CPUs with offloaded RCU callbacks.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Acked-by: Joel Fernandes (Google) <[email protected]>
Reviewed-by: Neeraj Upadhyay <[email protected]>
A get_random_bytes() function can cause a high contention if it is called
across CPUs simultaneously. Because it shares one lock per all CPUs:
<snip>
class name con-bounces contentions waittime-min waittime-max waittime-total waittime-avg acq-bounces acquisitions holdtime-min holdtime-max holdtime-total holdtime-avg
&crng->lock: 663145 665886 0.05 8.85 261966.66 0.39 7188152 13731279 0.04 11.89 2181582.30 0.16
-----------
&crng->lock 307835 [<00000000acba59cd>] _extract_crng+0x48/0x90
&crng->lock 358051 [<00000000f0075abc>] _crng_backtrack_protect+0x32/0x90
-----------
&crng->lock 234241 [<00000000f0075abc>] _crng_backtrack_protect+0x32/0x90
&crng->lock 431645 [<00000000acba59cd>] _extract_crng+0x48/0x90
<snip>
Switch from the get_random_bytes() to prandom_u32() that does not have any
internal contention when a random value is needed for the tests.
The reason is to minimize CPU cycles introduced by the test-suite itself
from the vmalloc performance metrics.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
A vmap_area can travel between different places. For example
attached/detached to/from different rb-trees. In order to prevent fancy
bugs, initialize a VA's list node after it is removed from the list, so it
pairs with VA's rb_node which is also initialized.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Baoquan He <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
It implies that __alloc_vmap_area() allocates only from the global vmap
space, therefore a list-head and rb-tree, which represent a free vmap
space, are not passed as parameters to this function and are accessed
directly from this function.
Extend the __alloc_vmap_area() and other dependent functions to have a
possibility to allocate from different trees making an interface common
and not specific.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Baoquan He <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Patch series "Reduce a vmalloc internal lock contention preparation work".
This small serias is preparation work to implement per-cpu vmalloc
allocation in order to reduce a high internal lock contention. This
series does not introduce any functional changes, it is only about
preparation.
This patch (of 5):
Currently link_va() and unlik_va(), in order to figure out a tree type,
compares a passed root value with a global free_vmap_area_root variable to
distinguish the augmented rb-tree from a regular one. It is hard coded
since such functions can manipulate only with specific
"free_vmap_area_root" tree that represents a global free vmap space.
Make it common by introducing "_augment" versions of both internal
functions, so it is possible to deal with different trees.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/[email protected]
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Baoquan He <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
That extra variable has been introduced just for keeping an original
passed gfp_mask because it is updated with __GFP_NOWARN on entry, thus
error handling messages were broken.
Instead we can keep an original gfp_mask without modifying it and add an
extra __GFP_NOWARN flag together with gfp_mask as a parameter to the
vm_area_alloc_pages() function. It will make it less confused.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Cc: Vasily Averin <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Uladzislau Rezki <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
A caller initiates the drain procces from its context once the
drain threshold is reached or passed. There are at least two
drawbacks of doing so:
a) a caller can be a high-prio or RT task. In that case it can
stuck in doing the actual drain of all lazily freed areas.
This is not optimal because such tasks usually are latency
sensitive where the control should be returned back as soon
as possible in order to drive such workloads in time. See
96e2db4561 ("mm/vmalloc: rework the drain logic")
b) It is not safe to call vfree() during holding a spinlock due
to the vmap_purge_lock mutex. The was a report about this from
Zeal Robot <[email protected]> here:
https://lore.kernel.org/all/[email protected]
Moving the drain to the separate work context addresses those
issues.
v1->v2:
- Added prefix "_work" to the drain worker function.
v2->v3:
- Remove the drain_vmap_work_in_progress. Extra queuing
is expectable under heavy load but it can be disregarded
because a work will bail out if nothing to be done.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Uladzislau Rezki <[email protected]>
Cc: Vasily Averin <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The kvfree_rcu() header comment's description of the "ptr" parameter
is unclear, therefore rephrase it to make it clear that it is a pointer
to the memory to eventually be passed to kvfree().
Reported-by: Steven Rostedt <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Instead of invoking a synchronize_rcu() to free a pointer
after a grace period we can directly make use of new API
that does the same but in more efficient way.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Instead of invoking a synchronize_rcu() to free a pointer
after a grace period we can directly make use of new API
that does the same but in more efficient way.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Steven Rostedt (VMware) <[email protected]>
We used to include an alignment overhead into a search length, in that
case we guarantee that a found area will definitely fit after applying a
specific alignment that user specifies. From the other hand we do not
guarantee that an area has the lowest address if an alignment is >=
PAGE_SIZE.
It means that, when a user specifies a special alignment together with a
range that corresponds to an exact requested size then an allocation
will fail. This is what happens to KASAN, it wants the free block that
exactly matches a specified range during onlining memory banks:
[root@vm-0 fedora]# echo online > /sys/devices/system/memory/memory82/state
[root@vm-0 fedora]# echo online > /sys/devices/system/memory/memory83/state
[root@vm-0 fedora]# echo online > /sys/devices/system/memory/memory85/state
[root@vm-0 fedora]# echo online > /sys/devices/system/memory/memory84/state
vmap allocation for size 16777216 failed: use vmalloc=<size> to increase size
bash: vmalloc: allocation failure: 16777216 bytes, mode:0x6000c0(GFP_KERNEL), nodemask=(null),cpuset=/,mems_allowed=0
CPU: 4 PID: 1644 Comm: bash Kdump: loaded Not tainted 4.18.0-339.el8.x86_64+debug #1
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
Call Trace:
dump_stack+0x8e/0xd0
warn_alloc.cold.90+0x8a/0x1b2
? zone_watermark_ok_safe+0x300/0x300
? slab_free_freelist_hook+0x85/0x1a0
? __get_vm_area_node+0x240/0x2c0
? kfree+0xdd/0x570
? kmem_cache_alloc_node_trace+0x157/0x230
? notifier_call_chain+0x90/0x160
__vmalloc_node_range+0x465/0x840
? mark_held_locks+0xb7/0x120
Fix it by making sure that find_vmap_lowest_match() returns lowest start
address with any given alignment value, i.e. for alignments bigger then
PAGE_SIZE the algorithm rolls back toward parent nodes checking right
sub-trees if the most left free block did not fit due to alignment
overhead.
Link: https://lkml.kernel.org/r/[email protected]
Fixes: 68ad4a3304 ("mm/vmalloc.c: keep track of free blocks for vmap allocation")
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reported-by: Ping Fang <[email protected]>
Tested-by: David Hildenbrand <[email protected]>
Reviewed-by: David Hildenbrand <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
In case of simultaneous vmalloc allocations, for example it is 1GB and 12
CPUs my system is able to hit "BUG: soft lockup" for !CONFIG_PREEMPT
kernel.
RIP: 0010:__alloc_pages_bulk+0xa9f/0xbb0
Call Trace:
__vmalloc_node_range+0x11c/0x2d0
__vmalloc_node+0x4b/0x70
fix_size_alloc_test+0x44/0x60 [test_vmalloc]
test_func+0xe7/0x1f0 [test_vmalloc]
kthread+0x11a/0x140
ret_from_fork+0x22/0x30
To address this issue invoke a bulk-allocator many times until all pages
are obtained, i.e. do batched page requests adding cond_resched()
meanwhile to reschedule. Batched value is hard-coded and is 100 pages per
call.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently for order-0 pages we use a bulk-page allocator to get set of
pages. From the other hand not allocating all pages is something that
might occur. In that case we should fallbak to the single-page allocator
trying to get missing pages, because it is more permissive(direct reclaim,
etc).
Introduce a vm_area_alloc_pages() function where the described logic is
implemented.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Matthew Wilcox (Oracle) <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Recently there has been introduced a page bulk allocator for users which
need to get number of pages per one call request.
For order-0 pages switch to an alloc_pages_bulk_array_node() instead of
alloc_pages_node(), the reason is the former is not capable of allocating
set of pages, thus a one call is per one page.
Second, according to my tests the bulk allocator uses less cycles even for
scenarios when only one page is requested. Running the "perf" on same
test case shows below difference:
<default>
- 45.18% __vmalloc_node
- __vmalloc_node_range
- 35.60% __alloc_pages
- get_page_from_freelist
3.36% __list_del_entry_valid
3.00% check_preemption_disabled
1.42% prep_new_page
<default>
<patch>
- 31.00% __vmalloc_node
- __vmalloc_node_range
- 14.48% __alloc_pages_bulk
3.22% __list_del_entry_valid
- 0.83% __alloc_pages
get_page_from_freelist
<patch>
The "test_vmalloc.sh" also shows performance improvements:
fix_size_alloc_test_4MB loops: 1000000 avg: 89105095 usec
fix_size_alloc_test loops: 1000000 avg: 513672 usec
full_fit_alloc_test loops: 1000000 avg: 748900 usec
long_busy_list_alloc_test loops: 1000000 avg: 8043038 usec
random_size_alloc_test loops: 1000000 avg: 4028582 usec
fix_align_alloc_test loops: 1000000 avg: 1457671 usec
fix_size_alloc_test_4MB loops: 1000000 avg: 62083711 usec
fix_size_alloc_test loops: 1000000 avg: 449207 usec
full_fit_alloc_test loops: 1000000 avg: 735985 usec
long_busy_list_alloc_test loops: 1000000 avg: 5176052 usec
random_size_alloc_test loops: 1000000 avg: 2589252 usec
fix_align_alloc_test loops: 1000000 avg: 1365009 usec
For example 4MB allocations illustrates ~30% gain, all the
rest is also better.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Acked-by: Mel Gorman <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Nicholas Piggin <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently we have three functions which depend on each other.
Two of them are quite tiny and the last one where the most
work is done. All of them are related to queuing RCU batches
to reclaim objects after a GP.
1. kfree_rcu_monitor(). It consist of few lines. It acquires a spin-lock
and calls kfree_rcu_drain_unlock().
2. kfree_rcu_drain_unlock(). It also consists of few lines of code. It
calls queue_kfree_rcu_work() to queue the batch. If this fails,
it rearms the monitor work to try again later.
3. queue_kfree_rcu_work(). This provides the bulk of the functionality,
attempting to start a new batch to free objects after a GP.
Since there are no external users of functions [2] and [3], both
can eliminated by moving all logic directly into [1], which both
shrinks and simplifies the code.
Also replace comments which start with "/*" to "//" format to make it
unified across the file.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The kvfree_rcu() function now defers allocations in the common
case due to the fact that there is no lockless access to the
memory-allocator caches/pools. In addition, in CONFIG_PREEMPT_NONE=y
and in CONFIG_PREEMPT_VOLUNTARY=y kernels, there is no reliable way to
determine if spinlocks are held. As a result, allocation is deferred in
the common case, and the two-argument form of kvfree_rcu() thus uses the
"channel 3" queue through all the rcu_head structures. This channel
is called referred to as the emergency case in comments, and these
comments are now obsolete.
This commit therefore updates these comments to reflect the new
common-case nature of such emergencies.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Replace an open-coded version of the kfree_rcu_monitor() function body
with a call to that function.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Before attempting to start a new batch the "monitor_todo" variable is
set to "false" and set back to "true" when a previous RCU batch is still
in progress. This is at best confusing.
Thus change this variable to "false" only when a new batch has been
successfully queued, otherwise, just leave it be.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The rcu_scheduler_active flag is set to RCU_SCHEDULER_RUNNING once the
scheduler is up and running. That signal is used in order to check
and queue a "monitor work" to reclaim freed objects (if there are any)
during early boot. This flag is used by kvfree_rcu() to determine when
work can safely be queued, at which point memory passed to earlier
invocations of kvfree_rcu() can be processed.
However, only "krcp->head" is checked for objects that need to be
released, and there are now two more, namely, "krcp->bkvhead[0]" and
"krcp->bkvhead[1]". Therefore, check these two additional channels.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
nr_bkv_objs is a count of the objects in the kvfree_rcu page cache.
Accessing it requires holding the ->lock. Switch to READ_ONCE() and
WRITE_ONCE() macros to provide lockless access to this counter.
This lockless access is used for the shrinker.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
A 'single_cpu_test' parameter is odd and it does not exist anymore.
Instead there was introduced a 'nr_threads' one. If it is not set it
behaves as the former parameter.
That is why update a "stress mode" according to this change specifying
number of workers which are equal to number of CPUs. Also update an
output of help message based on a new interface.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Cc: Shuah Khan <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The single-argument variant of kfree_rcu() is currently not
tested by any member of the rcutoture test suite. This
commit therefore adds rcuscale code to test it. This
testing is controlled by two new boolean module parameters,
kfree_rcu_test_single and kfree_rcu_test_double. If one
is set and the other not, only the corresponding variant
is tested, otherwise both are tested, with the variant to
be tested determined randomly on each invocation.
Both of these module parameters are initialized to false,
so setting either to true will test only that variant.
Suggested-by: Paul E. McKenney <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Running an rcuscale stress-suite can lead to "Out of memory" of a
system. This can happen under high memory pressure with a small amount
of physical memory.
For example, a KVM test configuration with 64 CPUs and 512 megabytes
can result in OOM when running rcuscale with below parameters:
../kvm.sh --torture rcuscale --allcpus --duration 10 --kconfig CONFIG_NR_CPUS=64 \
--bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 rcuscale.holdoff=20 \
rcuscale.kfree_loops=10000 torture.disable_onoff_at_boot" --trust-make
<snip>
[ 12.054448] kworker/1:1H invoked oom-killer: gfp_mask=0x2cc0(GFP_KERNEL|__GFP_NOWARN), order=0, oom_score_adj=0
[ 12.055303] CPU: 1 PID: 377 Comm: kworker/1:1H Not tainted 5.11.0-rc3+ #510
[ 12.055416] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.12.0-1 04/01/2014
[ 12.056485] Workqueue: events_highpri fill_page_cache_func
[ 12.056485] Call Trace:
[ 12.056485] dump_stack+0x57/0x6a
[ 12.056485] dump_header+0x4c/0x30a
[ 12.056485] ? del_timer_sync+0x20/0x30
[ 12.056485] out_of_memory.cold.47+0xa/0x7e
[ 12.056485] __alloc_pages_slowpath.constprop.123+0x82f/0xc00
[ 12.056485] __alloc_pages_nodemask+0x289/0x2c0
[ 12.056485] __get_free_pages+0x8/0x30
[ 12.056485] fill_page_cache_func+0x39/0xb0
[ 12.056485] process_one_work+0x1ed/0x3b0
[ 12.056485] ? process_one_work+0x3b0/0x3b0
[ 12.060485] worker_thread+0x28/0x3c0
[ 12.060485] ? process_one_work+0x3b0/0x3b0
[ 12.060485] kthread+0x138/0x160
[ 12.060485] ? kthread_park+0x80/0x80
[ 12.060485] ret_from_fork+0x22/0x30
[ 12.062156] Mem-Info:
[ 12.062350] active_anon:0 inactive_anon:0 isolated_anon:0
[ 12.062350] active_file:0 inactive_file:0 isolated_file:0
[ 12.062350] unevictable:0 dirty:0 writeback:0
[ 12.062350] slab_reclaimable:2797 slab_unreclaimable:80920
[ 12.062350] mapped:1 shmem:2 pagetables:8 bounce:0
[ 12.062350] free:10488 free_pcp:1227 free_cma:0
...
[ 12.101610] Out of memory and no killable processes...
[ 12.102042] Kernel panic - not syncing: System is deadlocked on memory
[ 12.102583] CPU: 1 PID: 377 Comm: kworker/1:1H Not tainted 5.11.0-rc3+ #510
[ 12.102600] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.12.0-1 04/01/2014
<snip>
Because kvfree_rcu() has a fallback path, memory allocation failure is
not the end of the world. Furthermore, the added overhead of aggressive
GFP settings must be balanced against the overhead of the fallback path,
which is a cache miss for double-argument kvfree_rcu() and a call to
synchronize_rcu() for single-argument kvfree_rcu(). The current choice
of GFP_KERNEL|__GFP_NOWARN can result in longer latencies than a call
to synchronize_rcu(), so less-tenacious GFP flags would be helpful.
Here is the tradeoff that must be balanced:
a) Minimize use of the fallback path,
b) Avoid pushing the system into OOM,
c) Bound allocation latency to that of synchronize_rcu(), and
d) Leave the emergency reserves to use cases lacking fallbacks.
This commit therefore changes GFP flags from GFP_KERNEL|__GFP_NOWARN to
GFP_KERNEL|__GFP_NORETRY|__GFP_NOMEMALLOC|__GFP_NOWARN. This combination
leaves the emergency reserves alone and can initiate reclaim, but will
not invoke the OOM killer.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
__GFP_RETRY_MAYFAIL can spend quite a bit of time reclaiming, and this
can be wasted effort given that there is a fallback code path in case
memory allocation fails.
__GFP_NORETRY does perform some light-weight reclaim, but it will fail
under OOM conditions, allowing the fallback to be taken as an alternative
to hard-OOMing the system.
There is a four-way tradeoff that must be balanced:
1) Minimize use of the fallback path;
2) Avoid full-up OOM;
3) Do a light-wait allocation request;
4) Avoid dipping into the emergency reserves.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Single-argument kvfree_rcu() must be invoked from sleepable contexts,
so we can directly allocate pages. Furthermmore, the fallback in case
of page-allocation failure is the high-latency synchronize_rcu(), so it
makes sense to do these page allocations from the fastpath, and even to
permit limited sleeping within the allocator.
This commit therefore allocates if needed on the fastpath using
GFP_KERNEL|__GFP_RETRY_MAYFAIL. This also has the beneficial effect
of leaving kvfree_rcu()'s per-CPU caches to the double-argument variant
of kvfree_rcu(), given that the double-argument variant cannot directly
invoke the allocator.
[ paulmck: Add add_ptr_to_bulk_krc_lock header comment per Michal Hocko. ]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
This commit adds self tests for early-boot use of RCU-tasks grace periods.
It tests all three variants (Rude, Tasks, and Tasks Trace) and covers
both synchronous (e.g., synchronize_rcu_tasks()) and asynchronous (e.g.,
call_rcu_tasks()) grace-period APIs.
Self-tests are run only in kernels built with CONFIG_PROVE_RCU=y.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
[ paulmck: Handle CONFIG_PROVE_RCU=n and identify test cases' callbacks. ]
Signed-off-by: Paul E. McKenney <[email protected]>
This commit open-codes the __kvfree_rcu() macro, thus saving a
few lines of code and improving readability.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
There is a kvfree_rcu() single argument macro that handles pointers
returned by kvmalloc(). Even though it also handles pointer returned by
kmalloc(), readability suffers.
This commit therefore updates the kfree_rcu() macro to explicitly pair
with kmalloc(), thus improving readability.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
A current "lazy drain" model suffers from at least two issues.
First one is related to the unsorted list of vmap areas, thus in order to
identify the [min:max] range of areas to be drained, it requires a full
list scan. What is a time consuming if the list is too long.
Second one and as a next step is about merging all fragments with a free
space. What is also a time consuming because it has to iterate over
entire list which holds outstanding lazy areas.
See below the "preemptirqsoff" tracer that illustrates a high latency. It
is ~24676us. Our workloads like audio and video are effected by such long
latency:
<snip>
tracer: preemptirqsoff
preemptirqsoff latency trace v1.1.5 on 4.9.186-perf+
--------------------------------------------------------------------
latency: 24676 us, #4/4, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 P:8)
-----------------
| task: crtc_commit:112-261 (uid:0 nice:0 policy:1 rt_prio:16)
-----------------
=> started at: __purge_vmap_area_lazy
=> ended at: __purge_vmap_area_lazy
_------=> CPU#
/ _-----=> irqs-off
| / _----=> need-resched
|| / _---=> hardirq/softirq
||| / _--=> preempt-depth
|||| / delay
cmd pid ||||| time | caller
\ / ||||| \ | /
crtc_com-261 1...1 1us*: _raw_spin_lock <-__purge_vmap_area_lazy
[...]
crtc_com-261 1...1 24675us : _raw_spin_unlock <-__purge_vmap_area_lazy
crtc_com-261 1...1 24677us : trace_preempt_on <-__purge_vmap_area_lazy
crtc_com-261 1...1 24683us : <stack trace>
=> free_vmap_area_noflush
=> remove_vm_area
=> __vunmap
=> vfree
=> drm_property_free_blob
=> drm_mode_object_unreference
=> drm_property_unreference_blob
=> __drm_atomic_helper_crtc_destroy_state
=> sde_crtc_destroy_state
=> drm_atomic_state_default_clear
=> drm_atomic_state_clear
=> drm_atomic_state_free
=> complete_commit
=> _msm_drm_commit_work_cb
=> kthread_worker_fn
=> kthread
=> ret_from_fork
<snip>
To address those two issues we can redesign a purging of the outstanding
lazy areas. Instead of queuing vmap areas to the list, we replace it by
the separate rb-tree. In hat case an area is located in the tree/list in
ascending order. It will give us below advantages:
a) Outstanding vmap areas are merged creating bigger coalesced blocks,
thus it becomes less fragmented.
b) It is possible to calculate a flush range [min:max] without scanning
all elements. It is O(1) access time or complexity;
c) The final merge of areas with the rb-tree that represents a free
space is faster because of (a). As a result the lock contention is
also reduced.
Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: huang ying <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
PowerPC testing encountered boot failures due to RCU Tasks not being
fully initialized until core_initcall() time. This commit therefore
initializes RCU Tasks (along with Rude RCU and RCU Tasks Trace) just
before early_initcall() time, thus allowing waiting on RCU Tasks grace
periods from early_initcall() handlers.
Link: https://lore.kernel.org/rcu/[email protected]/
Fixes: 36dadef23f ("kprobes: Init kprobes in early_initcall")
Tested-by: Daniel Axtens <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The current memmory-allocation interface causes the following difficulties
for kvfree_rcu():
a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will
complain about violation of the nesting rules, as in "BUG: Invalid
wait context". This Kconfig option checks for proper raw_spinlock
vs. spinlock nesting, in particular, it is not legal to acquire a
spinlock_t while holding a raw_spinlock_t.
This is a problem because kfree_rcu() uses raw_spinlock_t whereas the
"page allocator" internally deals with spinlock_t to access to its
zones. The code also can be broken from higher level of view:
<snip>
raw_spin_lock(&some_lock);
kfree_rcu(some_pointer, some_field_offset);
<snip>
b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into
sleeplock. This means that invoking the page allocator from atomic
contexts results in "BUG: scheduling while atomic".
c) Please note that call_rcu() is already invoked from raw atomic context,
so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu()
will also be called from atomic raw context.
This commit therefore defers page allocation to a clean context using the
combination of an hrtimer and a workqueue. The hrtimer stage is required
in order to avoid deadlocks with the scheduler. This deferred allocation
is required only when kvfree_rcu()'s per-CPU page cache is empty.
Link: https://lore.kernel.org/lkml/[email protected]/
Fixes: 3042f83f19 ("rcu: Support reclaim for head-less object")
Reported-by: Sebastian Andrzej Siewior <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Get rid of BUG() macro, that should be used only when a critical situation
happens and a system is not able to function anymore.
Replace it with WARN() macro instead, dump some extra information about
start/end addresses of both VAs which overlap. Such overlap data can help
to figure out what happened making further analysis easier. For example
if both areas are identical it could mean a double free.
A recovery process consists of declining all further steps regarding
inserting of conflicting overlap range. In that sense find_va_links() now
can return NULL, so its return value has to be checked by callers.
Side effect of such process is it can leak memory, but it is better than
just killing a machine for no good reason. Apart of that a debugging
process can be done on alive system.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Linus Torvalds <[email protected]>
An augment_tree_propagate_from() function uses its own implementation that
populates a tree from the specified node toward a root node.
On the other hand the RB_DECLARE_CALLBACKS_MAX macro provides the
"propagate()" callback that does exactly the same. Having two similar
functions does not make sense and is redundant.
Reuse "built in" functionality to the macros. So the code size gets
reduced.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Linus Torvalds <[email protected]>
This function is for debug purpose only. Currently it uses recursion for
tree traversal, checking an augmented value of each node to find out if it
is valid or not.
The recursion can corrupt the stack because the tree can be huge if
synthetic tests are applied. To prevent it, navigate the tree from bottom
to upper levels using a regular list instead, because nodes are linked
among each other also. It is faster and without recursion.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Linus Torvalds <[email protected]>
Currently when a VA is deallocated and is about to be placed back to the
tree, it can be either: merged with next/prev neighbors or inserted if not
coalesced.
On those steps the tree can be populated several times. For example when
both neighbors are merged. It can be avoided and simplified in fact.
Therefore do it only once when VA points to final merged area, after all
manipulations: merging/removing/inserting.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Linus Torvalds <[email protected]>
Introduce four new test cases for testing the kvfree_rcu()
interface. Two of them belong to single argument functionality
and another two for 2-argument functionality.
The aim is to stress and check how kvfree_rcu() behaves under
different load and memory conditions and analyze its performance
throughput.
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Make kvfree_rcu() capable of freeing objects that will not
embed an rcu_head within it. This saves storage overhead in
such objects. Reclaiming headless objects this way requires
only a single argument (pointer to the object).
After this patch, there are two ways to use kvfree_rcu():
a) kvfree_rcu(ptr, rhf);
struct X {
struct rcu_head rhf;
unsigned char data[100];
};
void *ptr = kvmalloc(sizeof(struct X), GFP_KERNEL);
if (ptr)
kvfree_rcu(ptr, rhf);
b) kvfree_rcu(ptr);
void *ptr = kvmalloc(some_bytes, GFP_KERNEL);
if (ptr)
kvfree_rcu(ptr);
Note that the headless usage (example b) can only be used in a code
that can sleep. This is enforced by the CONFIG_DEBUG_ATOMIC_SLEEP
option.
Co-developed-by: Joel Fernandes (Google) <[email protected]>
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Update the kvfree_call_rcu() function with head-less support.
This allows RCU to reclaim objects without an embedded rcu_head.
tree-RCU:
We introduce two chains of arrays to store SLAB-backed and vmalloc
pointers, each. Storage in either of these arrays does not require
embedding an rcu_head within the object.
Maintaining the arrays may become impossible due to high memory
pressure. For such cases there is an emergency path. Objects with
rcu_head inside are just queued on a backup rcu_head list. Later on
that list is drained. As for the head-less variant, as the current
context can sleep, the following emergency measures are applied:
a) Synchronously wait until a grace period has elapsed.
b) Call kvfree().
tiny-RCU:
For double argument calls, there are no new changes in behavior. For
single argument call, kvfree() is directly inlined on the current
stack after a synchronize_rcu() call. Note that for tiny-RCU, any
call to synchronize_rcu() is actually a quiescent state, therefore
it does nothing.
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Co-developed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
kvmalloc() can allocate two types of objects: SLAB backed
and vmalloc backed. How it behaves depends on requested
object's size and memory pressure.
Add a kvfree_rcu() interface that can free memory allocated
via kvmalloc(). It is a simple alias to kfree_rcu() which
can now handle either type of object.
<snip>
struct test_kvfree_rcu {
struct rcu_head rcu;
unsigned char array[100];
};
struct test_kvfree_rcu *p;
p = kvmalloc(10 * PAGE_SIZE);
if (p)
kvfree_rcu(p, rcu);
<snip>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Co-developed-by: Joel Fernandes (Google) <[email protected]>
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Rename kvfree_rcu() function to the kvfree_rcu_local() one.
The purpose is to prevent a conflict of two same function
declarations. The kvfree_rcu() will be globally visible
what would lead to a build error. No functional change.
Cc: [email protected]
Cc: [email protected]
Cc: Andrew Morton <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The following changes are introduced:
1. Rename rcu_invoke_kfree_callback() to rcu_invoke_kvfree_callback(),
as well as the associated trace events, so the rcu_kfree_callback(),
becomes rcu_kvfree_callback(). The reason is to be aligned with kvfree()
notation.
2. Rename __is_kfree_rcu_offset to __is_kvfree_rcu_offset. All RCU
paths use kvfree() now instead of kfree(), thus rename it.
3. Rename kfree_call_rcu() to the kvfree_call_rcu(). The reason is,
it is capable of freeing vmalloc() memory now. Do the same with
__kfree_rcu() macro, it becomes __kvfree_rcu(), the goal is the
same.
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Co-developed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Replace kfree() with kvfree() in rcu_reclaim_tiny().
This makes it possible to release either SLAB or vmalloc
objects after a GP.
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
To do so, we use an array of kvfree_rcu_bulk_data structures.
It consists of two elements:
- index number 0 corresponds to slab pointers.
- index number 1 corresponds to vmalloc pointers.
Keeping vmalloc pointers separated from slab pointers makes
it possible to invoke the right freeing API for the right
kind of pointer.
It also prepares us for future headless support for vmalloc
and SLAB objects. Such objects cannot be queued on a linked
list and are instead directly into an array.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Co-developed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
In order to reduce the dynamic need for pages in kfree_rcu(),
pre-allocate a configurable number of pages per CPU and link
them in a list. When kfree_rcu() reclaims objects, the object's
container page is cached into a list instead of being released
to the low-level page allocator.
Such an approach provides O(1) access to free pages while also
reducing the number of requests to the page allocator. It also
makes the kfree_rcu() code to have free pages available during
a low memory condition.
A read-only sysfs parameter (rcu_min_cached_objs) reflects the
minimum number of allowed cached pages per CPU.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
Introduce helpers to lock and unlock per-cpu "kfree_rcu_cpu"
structures. That will make kfree_call_rcu() more readable
and prevent programming errors.
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
We can simplify KFREE_BULK_MAX_ENTR macro and get rid of
magic numbers which were used to make the structure to be
exactly one page.
Suggested-by: Boqun Feng <[email protected]>
Reviewed-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
It is possible that one of the channels cannot be detached
because its free channel is busy and previously queued data
has not been processed yet. On the other hand, another
channel can be successfully detached causing the monitor
work to stop.
Prevent that by rescheduling the monitor work if there are
any channels in the pending state after a detach attempt.
Fixes: 34c8817455 ("rcu: Support kfree_bulk() interface in kfree_rcu()")
Acked-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The event is given three parameters, first one is the name
of RCU flavour, second one is the number of elements in array
for free and last one is an address of the array holding
pointers to be freed by the kfree_bulk() function.
To enable the trace event your kernel has to be build with
CONFIG_RCU_TRACE=y, after that it is possible to track the
events using ftrace subsystem.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
The kfree_rcu() logic can be improved further by using kfree_bulk()
interface along with "basic batching support" introduced earlier.
The are at least two advantages of using "bulk" interface:
- in case of large number of kfree_rcu() requests kfree_bulk()
reduces the per-object overhead caused by calling kfree()
per-object.
- reduces the number of cache-misses due to "pointer chasing"
between objects which can be far spread between each other.
This approach defines a new kfree_rcu_bulk_data structure that
stores pointers in an array with a specific size. Number of entries
in that array depends on PAGE_SIZE making kfree_rcu_bulk_data
structure to be exactly one page.
Since it deals with "block-chain" technique there is an extra
need in dynamic allocation when a new block is required. Memory
is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that
allows to skip direct reclaim under low memory condition to
prevent stalling and fails silently under high memory pressure.
The "emergency path" gets maintained when a system is run out of
memory. In that case objects are linked into regular list.
The "rcuperf" was run to analyze this change in terms of memory
consumption and kfree_bulk() throughput.
1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs
with following parameters:
kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1
dev.2020.01.10a branch
Default / CONFIG_SLAB
53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB
53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB
53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB
Patch / CONFIG_SLAB
23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB
23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB
24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB
Default / CONFIG_SLUB
51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB
51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB
51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB
Patch / CONFIG_SLUB
50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB
50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB
50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB
in case of CONFIG_SLAB there is double increase in performance and
slightly higher memory usage. As for CONFIG_SLUB, the performance
figures are better together with lower memory usage.
2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters:
CONFIG_SLAB=y
kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1
102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB
89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB
rcuperf shows approximately ~12% better throughput in case of
using "bulk" interface. The "drain logic" or its RCU callback
does the work faster that leads to better throughput.
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Tested-by: Joel Fernandes (Google) <[email protected]>
Signed-off-by: Paul E. McKenney <[email protected]>
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Uladzislau Rezki (Sony) <[email protected]>
Acked-by: Andrew Morton <[email protected]>
Cc: Hillf Danton <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Oleksiy Avramchenko <[email protected]>
Cc: Steven Rostedt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
