* Re: [PATCH v13 2/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Mathieu Desnoyers @ 2026-01-13 0:47 UTC (permalink / raw)
To: Michal Hocko
Cc: Andrew Morton, linux-kernel, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <aWVP9sPqG8VC8Oq_@tiehlicka>
On 2026-01-12 14:48, Michal Hocko wrote:
> On Mon 12-01-26 14:37:49, Mathieu Desnoyers wrote:
>> On 2026-01-12 03:42, Michal Hocko wrote:
>>> Hi,
>>> sorry to jump in this late but the timing of previous versions didn't
>>> really work well for me.
>>>
>>> On Sun 11-01-26 14:49:57, Mathieu Desnoyers wrote:
>>> [...]
>>>> Here is a (possibly incomplete) list of the prior approaches that were
>>>> used or proposed, along with their downside:
>>>>
>>>> 1) Per-thread rss tracking: large error on many-thread processes.
>>>>
>>>> 2) Per-CPU counters: up to 12% slower for short-lived processes and 9%
>>>> increased system time in make test workloads [1]. Moreover, the
>>>> inaccuracy increases with O(n^2) with the number of CPUs.
>>>>
>>>> 3) Per-NUMA-node counters: requires atomics on fast-path (overhead),
>>>> error is high with systems that have lots of NUMA nodes (32 times
>>>> the number of NUMA nodes).
>>>>
>>>> The approach proposed here is to replace this by the hierarchical
>>>> per-cpu counters, which bounds the inaccuracy based on the system
>>>> topology with O(N*logN).
>>>
>>> The concept of hierarchical pcp counter is interesting and I am
>>> definitely not opposed if there are more users that would benefit.
>>>
>>> From the OOM POV, IIUC the primary problem is that get_mm_counter
>>> (percpu_counter_read_positive) is too imprecise on systems when the task
>>> is moving around a large number of cpus. In the list of alternative
>>> solutions I do not see percpu_counter_sum_positive to be mentioned.
>>> oom_badness() is a really slow path and taking the slow path to
>>> calculate a much more precise value seems acceptable. Have you
>>> considered that option?
>> I must admit I assumed that since there was already a mechanism in place
>> to ensure it's not necessary to sum per-cpu counters when the oom killer
>> is trying to select tasks, it must be because this
>>
>> O(nr_possible_cpus * nr_processes)
>>
>> operation must be too slow for the oom killer requirements.
>>
>> AFAIU, the oom killer is executed when the memory allocator fails to
>> allocate memory, which can be within code paths which need to progress
>> eventually. So even though it's a slow path compared to the allocator
>> fast path, there must be at least _some_ expectations about it
>> completing within a decent amount of time. What would that ballpark be ?
>
> I do not think we have ever promissed more than the oom killer will try
> to unlock the system blocked on memory shortage.
>
>> To give an order of magnitude, I've tried modifying the upstream
>> oom killer to use percpu_counter_sum_positive and compared it to
>> the hierarchical approach:
>>
>> AMD EPYC 9654 96-Core (2 sockets)
>> Within a KVM, configured with 256 logical cpus.
>>
>> nr_processes=40 nr_processes=10000
>> Counter sum: 0.4 ms 81.0 ms
>> HPCC with 2-pass: 0.3 ms 9.3 ms
>
> These are peanuts for the global oom situations. We have had situations
> when soft lockup detector triggered because of the process tree
> traversal so adding 100ms is not really critical.
>
>> So as we scale up the number of processes on large SMP systems,
>> the latency caused by the oom killer task selection greatly
>> increases with the counter sums compared with the hierarchical
>> approach.
>
> Yes, I am not really questioning the hierarchical approach will perform
> much better but I am thinking of a good enough solution and calculating
> the number might be just that stop gap solution (that would be also
> suitable for stable tree backports). I am not ruling out improving on
> top of that by a more clever solution like your hierarchical counters
> approach. Especially if there are more benefits from that elsewhere.
>
Would you be OK with introducing changes in the following order ?
1) Fix the OOM killer inaccuracy by using counter sum (iteration on all
cpu counters) in task selection. This may slow down the oom killer,
but would at least fix its current inaccuracy issues. This could be
backported to stable kernels.
2) Introduce the hierarchical percpu counters on top, as a oom killer
task selection performance optimization (reduce latency of oom kill).
This way, (2) becomes purely a performance optimization, so it's easy
to bissect and revert if it causes issues.
I agree that bringing a fix along with a performance optimization within
a single commit makes it hard to backport to stable, and tricky to
revert if it causes problems.
As for finding other users of the hpcc, I have ideas, but not so much
time available to try them out, as I'm pretty much doing this in my
spare time.
One possibility I see would be to track memory access patterns by
sampling with page fault handlers for NUMA locality tracking.
Using hierarchical counters could help make quick migration decisions
(either task migration or moving memory pages across NUMA nodes)
based on hpcc approximations, including using intermediate counters
within the tree levels.
Another category of use-cases would be tracking resource limits
(e.g. cgroups, network traffic control) through HPCC, and then
using approximated counter values to validate whether the current
usage goes beyond the limit threshold.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
^ permalink raw reply
* Re: [PATCH v2 1/2] tracing: Make the backup instance readonly
From: Steven Rostedt @ 2026-01-13 0:47 UTC (permalink / raw)
To: Masami Hiramatsu (Google)
Cc: Mathieu Desnoyers, linux-kernel, linux-trace-kernel
In-Reply-To: <20260112194551.2446d041@gandalf.local.home>
On Mon, 12 Jan 2026 19:45:51 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:
> Let's not add any flags to the eventfs_entry. That is created for every
> event, and events are already too big in size. I don't want to increase the
> size of each event for this feature.
Actually, there's not many of these. But they are all static variables. How
are you going to differentiate them?
-- Steve
^ permalink raw reply
* Re: [PATCH v2 1/2] tracing: Make the backup instance readonly
From: Steven Rostedt @ 2026-01-13 0:45 UTC (permalink / raw)
To: Masami Hiramatsu (Google)
Cc: Mathieu Desnoyers, linux-kernel, linux-trace-kernel
In-Reply-To: <20260113093100.e91c5be8b8fa5f1e64899ba4@kernel.org>
On Tue, 13 Jan 2026 09:31:00 +0900
Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:
> > + static struct eventfs_entry event_ro_entries[] = {
> > + {
> > + .name = "format",
> > + .callback = event_callback,
> > + },
> > +#ifdef CONFIG_PERF_EVENTS
> > + {
> > + .name = "id",
> > + .callback = event_callback,
> > + },
> > +#endif
> > + };
>
> Thinking about this hack again, it would be easier to maintain if I add
> a readonly flag to each eventfs_entry and specify readonly when creating
> the eventfs top directory.
>
> Let me update it.
Let's not add any flags to the eventfs_entry. That is created for every
event, and events are already too big in size. I don't want to increase the
size of each event for this feature.
-- Steve
^ permalink raw reply
* Re: [PATCH v2] rtla: Fix parse_cpu_set() bug introduced by strtoi()
From: Masami Hiramatsu @ 2026-01-13 0:43 UTC (permalink / raw)
To: Costa Shulyupin
Cc: Steven Rostedt, Tomas Glozar, Wander Lairson Costa, Ivan Pravdin,
Tiezhu Yang, linux-trace-kernel, linux-kernel
In-Reply-To: <20260112192642.212848-2-costa.shul@redhat.com>
On Mon, 12 Jan 2026 21:26:41 +0200
Costa Shulyupin <costa.shul@redhat.com> wrote:
> The patch 'Replace atoi() with a robust strtoi()' introduced a bug
> in parse_cpu_set(), which relies on partial parsing of the input string.
>
> The function parses CPU specifications like '0-3,5' by incrementing
> a pointer through the string. strtoi() rejects strings with trailing
> characters, causing parse_cpu_set() to fail on any CPU list with
> multiple entries.
>
> Restore the original use of atoi() in parse_cpu_set().
>
> Fixes: 7e9dfccf8f11 ("rtla: Replace atoi() with a robust strtoi()")
> Signed-off-by: Costa Shulyupin <costa.shul@redhat.com>
Looks good to me.
Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Thanks,
> ---
>
> v2:
> - Split the patch.
>
> ---
> tools/tracing/rtla/src/utils.c | 10 ++++------
> 1 file changed, 4 insertions(+), 6 deletions(-)
>
> diff --git a/tools/tracing/rtla/src/utils.c b/tools/tracing/rtla/src/utils.c
> index 18986a5aed3c..0da3b2470c31 100644
> --- a/tools/tracing/rtla/src/utils.c
> +++ b/tools/tracing/rtla/src/utils.c
> @@ -128,18 +128,16 @@ int parse_cpu_set(char *cpu_list, cpu_set_t *set)
> nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
>
> for (p = cpu_list; *p; ) {
> - if (strtoi(p, &cpu))
> - goto err;
> - if (cpu < 0 || cpu >= nr_cpus)
> + cpu = atoi(p);
> + if (cpu < 0 || (!cpu && *p != '0') || cpu >= nr_cpus)
> goto err;
>
> while (isdigit(*p))
> p++;
> if (*p == '-') {
> p++;
> - if (strtoi(p, &end_cpu))
> - goto err;
> - if (end_cpu < cpu || end_cpu >= nr_cpus)
> + end_cpu = atoi(p);
> + if (end_cpu < cpu || (!end_cpu && *p != '0') || end_cpu >= nr_cpus)
> goto err;
> while (isdigit(*p))
> p++;
> --
> 2.52.0
>
>
--
Masami Hiramatsu (Google) <mhiramat@kernel.org>
^ permalink raw reply
* Re: [PATCH v2 1/2] tracing: Make the backup instance readonly
From: Masami Hiramatsu @ 2026-01-13 0:31 UTC (permalink / raw)
To: Masami Hiramatsu (Google)
Cc: Steven Rostedt, Mathieu Desnoyers, linux-kernel,
linux-trace-kernel
In-Reply-To: <176788218294.1398317.13118277798168167277.stgit@mhiramat.tok.corp.google.com>
On Thu, 8 Jan 2026 23:23:03 +0900
"Masami Hiramatsu (Google)" <mhiramat@kernel.org> wrote:
> diff --git a/kernel/trace/trace_events.c b/kernel/trace/trace_events.c
> index 9b07ad9eb284..741b16b54d90 100644
> --- a/kernel/trace/trace_events.c
> +++ b/kernel/trace/trace_events.c
> @@ -1379,6 +1379,9 @@ static int __ftrace_set_clr_event(struct trace_array *tr, const char *match,
> {
> int ret;
>
> + if (trace_array_is_readonly(tr))
> + return -EPERM;
> +
> mutex_lock(&event_mutex);
> ret = __ftrace_set_clr_event_nolock(tr, match, sub, event, set, mod);
> mutex_unlock(&event_mutex);
> @@ -2817,8 +2820,8 @@ event_subsystem_dir(struct trace_array *tr, const char *name,
> } else
> __get_system(system);
>
> - /* ftrace only has directories no files */
> - if (strcmp(name, "ftrace") == 0)
> + /* ftrace only has directories no files, readonly instance too. */
> + if (strcmp(name, "ftrace") == 0 || trace_array_is_readonly(tr))
> nr_entries = 0;
> else
> nr_entries = ARRAY_SIZE(system_entries);
> @@ -2979,7 +2982,6 @@ event_create_dir(struct eventfs_inode *parent, struct trace_event_file *file)
> struct eventfs_inode *e_events;
> struct eventfs_inode *ei;
> const char *name;
> - int nr_entries;
> int ret;
> static struct eventfs_entry event_entries[] = {
> {
> @@ -3024,6 +3026,18 @@ event_create_dir(struct eventfs_inode *parent, struct trace_event_file *file)
> },
> #endif
> };
> + static struct eventfs_entry event_ro_entries[] = {
> + {
> + .name = "format",
> + .callback = event_callback,
> + },
> +#ifdef CONFIG_PERF_EVENTS
> + {
> + .name = "id",
> + .callback = event_callback,
> + },
> +#endif
> + };
Thinking about this hack again, it would be easier to maintain if I add
a readonly flag to each eventfs_entry and specify readonly when creating
the eventfs top directory.
Let me update it.
Thank you,
--
Masami Hiramatsu (Google) <mhiramat@kernel.org>
^ permalink raw reply
* Re: [PATCH v3] ftrace: Do not over-allocate ftrace memory
From: Guenter Roeck @ 2026-01-13 0:15 UTC (permalink / raw)
To: Steven Rostedt
Cc: Masami Hiramatsu, Mark Rutland, Mathieu Desnoyers, linux-kernel,
linux-trace-kernel
In-Reply-To: <20260107022221.2907717-1-linux@roeck-us.net>
Hi Steven,
On 1/6/26 18:22, Guenter Roeck wrote:
> The pg_remaining calculation in ftrace_process_locs() assumes that
> ENTRIES_PER_PAGE multiplied by 2^order equals the actual capacity of the
> allocated page group. However, ENTRIES_PER_PAGE is PAGE_SIZE / ENTRY_SIZE
> (integer division). When PAGE_SIZE is not a multiple of ENTRY_SIZE (e.g.
> 4096 / 24 = 170 with remainder 16), high-order allocations (like 256 pages)
> have significantly more capacity than 256 * 170. This leads to pg_remaining
> being underestimated, which in turn makes skip (derived from skipped -
> pg_remaining) larger than expected, causing the WARN(skip != remaining)
> to trigger.
>
> Extra allocated pages for ftrace: 2 with 654 skipped
> WARNING: CPU: 0 PID: 0 at kernel/trace/ftrace.c:7295 ftrace_process_locs+0x5bf/0x5e0
>
> A similar problem in ftrace_allocate_records() can result in allocating
> too many pages. This can trigger the second warning in
> ftrace_process_locs().
>
> Extra allocated pages for ftrace
> WARNING: CPU: 0 PID: 0 at kernel/trace/ftrace.c:7276 ftrace_process_locs+0x548/0x580
>
> Use the actual capacity of a page group to determine the number of pages
> to allocate. Have ftrace_allocate_pages() return the number of allocated
> pages to avoid having to calculate it. Use the actual page group capacity
> when validating the number of unused pages due to skipped entries.
> Drop the definition of ENTRIES_PER_PAGE since it is no longer used.
>
> Fixes: 4a3efc6baff93 ("ftrace: Update the mcount_loc check of skipped entries")
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---
> v3: Restore the code calculating the number of unused pages due to skipped
> entries. Use the actual number of entries per page group when
> calculating the number of entries in unused page groups.
> v2: Have ftrace_allocate_pages() return the number of allocated pages,
> and drop the page count calculation code as well as the associated
> warnings from ftrace_process_locs().
Any additional feedback / comments / suggestions ?
Thanks,
Guenter
>
> kernel/trace/ftrace.c | 29 +++++++++++++++--------------
> 1 file changed, 15 insertions(+), 14 deletions(-)
>
> diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
> index ef2d5dca6f70..2c9692b45215 100644
> --- a/kernel/trace/ftrace.c
> +++ b/kernel/trace/ftrace.c
> @@ -1148,7 +1148,6 @@ struct ftrace_page {
> };
>
> #define ENTRY_SIZE sizeof(struct dyn_ftrace)
> -#define ENTRIES_PER_PAGE (PAGE_SIZE / ENTRY_SIZE)
>
> static struct ftrace_page *ftrace_pages_start;
> static struct ftrace_page *ftrace_pages;
> @@ -3834,7 +3833,8 @@ static int ftrace_update_code(struct module *mod, struct ftrace_page *new_pgs)
> return 0;
> }
>
> -static int ftrace_allocate_records(struct ftrace_page *pg, int count)
> +static int ftrace_allocate_records(struct ftrace_page *pg, int count,
> + unsigned long *num_pages)
> {
> int order;
> int pages;
> @@ -3844,7 +3844,7 @@ static int ftrace_allocate_records(struct ftrace_page *pg, int count)
> return -EINVAL;
>
> /* We want to fill as much as possible, with no empty pages */
> - pages = DIV_ROUND_UP(count, ENTRIES_PER_PAGE);
> + pages = DIV_ROUND_UP(count * ENTRY_SIZE, PAGE_SIZE);
> order = fls(pages) - 1;
>
> again:
> @@ -3859,6 +3859,7 @@ static int ftrace_allocate_records(struct ftrace_page *pg, int count)
> }
>
> ftrace_number_of_pages += 1 << order;
> + *num_pages += 1 << order;
> ftrace_number_of_groups++;
>
> cnt = (PAGE_SIZE << order) / ENTRY_SIZE;
> @@ -3887,12 +3888,14 @@ static void ftrace_free_pages(struct ftrace_page *pages)
> }
>
> static struct ftrace_page *
> -ftrace_allocate_pages(unsigned long num_to_init)
> +ftrace_allocate_pages(unsigned long num_to_init, unsigned long *pages)
> {
> struct ftrace_page *start_pg;
> struct ftrace_page *pg;
> int cnt;
>
> + *pages = 0;
> +
> if (!num_to_init)
> return NULL;
>
> @@ -3906,7 +3909,7 @@ ftrace_allocate_pages(unsigned long num_to_init)
> * waste as little space as possible.
> */
> for (;;) {
> - cnt = ftrace_allocate_records(pg, num_to_init);
> + cnt = ftrace_allocate_records(pg, num_to_init, pages);
> if (cnt < 0)
> goto free_pages;
>
> @@ -7192,8 +7195,6 @@ static int ftrace_process_locs(struct module *mod,
> if (!count)
> return 0;
>
> - pages = DIV_ROUND_UP(count, ENTRIES_PER_PAGE);
> -
> /*
> * Sorting mcount in vmlinux at build time depend on
> * CONFIG_BUILDTIME_MCOUNT_SORT, while mcount loc in
> @@ -7206,7 +7207,7 @@ static int ftrace_process_locs(struct module *mod,
> test_is_sorted(start, count);
> }
>
> - start_pg = ftrace_allocate_pages(count);
> + start_pg = ftrace_allocate_pages(count, &pages);
> if (!start_pg)
> return -ENOMEM;
>
> @@ -7305,27 +7306,27 @@ static int ftrace_process_locs(struct module *mod,
> /* We should have used all pages unless we skipped some */
> if (pg_unuse) {
> unsigned long pg_remaining, remaining = 0;
> - unsigned long skip;
> + long skip;
>
> /* Count the number of entries unused and compare it to skipped. */
> - pg_remaining = (ENTRIES_PER_PAGE << pg->order) - pg->index;
> + pg_remaining = (PAGE_SIZE << pg->order) / ENTRY_SIZE - pg->index;
>
> if (!WARN(skipped < pg_remaining, "Extra allocated pages for ftrace")) {
>
> skip = skipped - pg_remaining;
>
> - for (pg = pg_unuse; pg; pg = pg->next)
> + for (pg = pg_unuse; pg && skip > 0; pg = pg->next) {
> remaining += 1 << pg->order;
> + skip -= (PAGE_SIZE << pg->order) / ENTRY_SIZE;
> + }
>
> pages -= remaining;
>
> - skip = DIV_ROUND_UP(skip, ENTRIES_PER_PAGE);
> -
> /*
> * Check to see if the number of pages remaining would
> * just fit the number of entries skipped.
> */
> - WARN(skip != remaining, "Extra allocated pages for ftrace: %lu with %lu skipped",
> + WARN(pg || skip > 0, "Extra allocated pages for ftrace: %lu with %lu skipped",
> remaining, skipped);
> }
> /* Need to synchronize with ftrace_location_range() */
^ permalink raw reply
* Re: [PATCH bpf-next 2/4] x86/fgraph,bpf: Switch kprobe_multi program stack unwind to hw_regs path
From: Steven Rostedt @ 2026-01-12 22:07 UTC (permalink / raw)
To: Jiri Olsa
Cc: Masami Hiramatsu, Josh Poimboeuf, Mahe Tardy, Peter Zijlstra, bpf,
linux-trace-kernel, x86, Yonghong Song, Song Liu, Andrii Nakryiko
In-Reply-To: <20260112214940.1222115-3-jolsa@kernel.org>
On Mon, 12 Jan 2026 22:49:38 +0100
Jiri Olsa <jolsa@kernel.org> wrote:
> To recreate same stack setup for return probe as we have for entry
> probe, we set the instruction pointer to the attached function address,
> which gets us the same unwind setup and same stack trace.
>
> With the fix, entry probe:
>
> # bpftrace -e 'kprobe:__x64_sys_newuname* { print(kstack)}'
> Attaching 1 probe...
>
> __x64_sys_newuname+9
> do_syscall_64+134
> entry_SYSCALL_64_after_hwframe+118
>
> return probe:
>
> # bpftrace -e 'kretprobe:__x64_sys_newuname* { print(kstack)}'
> Attaching 1 probe...
>
> __x64_sys_newuname+4
> do_syscall_64+134
> entry_SYSCALL_64_after_hwframe+118
But is this really correct?
The stack trace of the return from __x86_sys_newuname is from offset "+4".
The stack trace from entry is offset "+9". Isn't it confusing that the
offset is likely not from the return portion of that function?
-- Steve
>
> Fixes: 6d08340d1e35 ("Revert "perf/x86: Always store regs->ip in perf_callchain_kernel()"")
> Reported-by: Mahe Tardy <mahe.tardy@gmail.com>
> Signed-off-by: Jiri Olsa <jolsa@kernel.org>
^ permalink raw reply
* [PATCH bpf-next 4/4] selftests/bpf: Allow to benchmark trigger with stacktrace
From: Jiri Olsa @ 2026-01-12 21:49 UTC (permalink / raw)
To: Masami Hiramatsu, Steven Rostedt, Josh Poimboeuf
Cc: Peter Zijlstra, bpf, linux-trace-kernel, x86, Yonghong Song,
Song Liu, Andrii Nakryiko, Mahe Tardy
In-Reply-To: <20260112214940.1222115-1-jolsa@kernel.org>
Adding support to call bpf_get_stackid helper from trigger programs,
so far added for kprobe multi.
Adding the --stacktrace/-g option to enable it.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
tools/testing/selftests/bpf/bench.c | 4 ++++
tools/testing/selftests/bpf/bench.h | 1 +
.../selftests/bpf/benchs/bench_trigger.c | 1 +
.../selftests/bpf/progs/trigger_bench.c | 18 ++++++++++++++++++
4 files changed, 24 insertions(+)
diff --git a/tools/testing/selftests/bpf/bench.c b/tools/testing/selftests/bpf/bench.c
index bd29bb2e6cb5..8dadd9c928ec 100644
--- a/tools/testing/selftests/bpf/bench.c
+++ b/tools/testing/selftests/bpf/bench.c
@@ -265,6 +265,7 @@ static const struct argp_option opts[] = {
{ "verbose", 'v', NULL, 0, "Verbose debug output"},
{ "affinity", 'a', NULL, 0, "Set consumer/producer thread affinity"},
{ "quiet", 'q', NULL, 0, "Be more quiet"},
+ { "stacktrace", 'g', NULL, 0, "Get stack trace"},
{ "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0,
"Set of CPUs for producer threads; implies --affinity"},
{ "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0,
@@ -350,6 +351,9 @@ static error_t parse_arg(int key, char *arg, struct argp_state *state)
case 'q':
env.quiet = true;
break;
+ case 'g':
+ env.stacktrace = true;
+ break;
case ARG_PROD_AFFINITY_SET:
env.affinity = true;
if (parse_num_list(arg, &env.prod_cpus.cpus,
diff --git a/tools/testing/selftests/bpf/bench.h b/tools/testing/selftests/bpf/bench.h
index bea323820ffb..7cf21936e7ed 100644
--- a/tools/testing/selftests/bpf/bench.h
+++ b/tools/testing/selftests/bpf/bench.h
@@ -26,6 +26,7 @@ struct env {
bool list;
bool affinity;
bool quiet;
+ bool stacktrace;
int consumer_cnt;
int producer_cnt;
int nr_cpus;
diff --git a/tools/testing/selftests/bpf/benchs/bench_trigger.c b/tools/testing/selftests/bpf/benchs/bench_trigger.c
index 34018fc3927f..aeec9edd3851 100644
--- a/tools/testing/selftests/bpf/benchs/bench_trigger.c
+++ b/tools/testing/selftests/bpf/benchs/bench_trigger.c
@@ -146,6 +146,7 @@ static void setup_ctx(void)
bpf_program__set_autoload(ctx.skel->progs.trigger_driver, true);
ctx.skel->rodata->batch_iters = args.batch_iters;
+ ctx.skel->rodata->stacktrace = env.stacktrace;
}
static void load_ctx(void)
diff --git a/tools/testing/selftests/bpf/progs/trigger_bench.c b/tools/testing/selftests/bpf/progs/trigger_bench.c
index 2898b3749d07..479400d96fa4 100644
--- a/tools/testing/selftests/bpf/progs/trigger_bench.c
+++ b/tools/testing/selftests/bpf/progs/trigger_bench.c
@@ -25,6 +25,23 @@ static __always_inline void inc_counter(void)
__sync_add_and_fetch(&hits[cpu & CPU_MASK].value, 1);
}
+volatile const int stacktrace;
+
+typedef __u64 stack_trace_t[128];
+
+struct {
+ __uint(type, BPF_MAP_TYPE_STACK_TRACE);
+ __uint(max_entries, 16384);
+ __type(key, __u32);
+ __type(value, stack_trace_t);
+} stackmap SEC(".maps");
+
+static __always_inline void do_stacktrace(void *ctx)
+{
+ if (stacktrace)
+ bpf_get_stackid(ctx, &stackmap, 0);
+}
+
SEC("?uprobe")
int bench_trigger_uprobe(void *ctx)
{
@@ -96,6 +113,7 @@ SEC("?kprobe.multi/bpf_get_numa_node_id")
int bench_trigger_kprobe_multi(void *ctx)
{
inc_counter();
+ do_stacktrace(ctx);
return 0;
}
--
2.52.0
^ permalink raw reply related
* [PATCH bpf-next 3/4] selftests/bpf: Fix kprobe multi stacktrace_ips test
From: Jiri Olsa @ 2026-01-12 21:49 UTC (permalink / raw)
To: Masami Hiramatsu, Steven Rostedt, Josh Poimboeuf
Cc: Peter Zijlstra, bpf, linux-trace-kernel, x86, Yonghong Song,
Song Liu, Andrii Nakryiko, Mahe Tardy
In-Reply-To: <20260112214940.1222115-1-jolsa@kernel.org>
We now include the attached function in the stack trace,
fixing the test accordingly.
Fixes: c9e208fa93cd ("selftests/bpf: Add stacktrace ips test for kprobe_multi/kretprobe_multi")
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c b/tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c
index c9efdd2a5b18..43781a67051b 100644
--- a/tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c
+++ b/tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c
@@ -74,7 +74,8 @@ static void test_stacktrace_ips_kprobe_multi(bool retprobe)
load_kallsyms();
- check_stacktrace_ips(bpf_map__fd(skel->maps.stackmap), skel->bss->stack_key, 4,
+ check_stacktrace_ips(bpf_map__fd(skel->maps.stackmap), skel->bss->stack_key, 5,
+ ksym_get_addr("bpf_testmod_stacktrace_test"),
ksym_get_addr("bpf_testmod_stacktrace_test_3"),
ksym_get_addr("bpf_testmod_stacktrace_test_2"),
ksym_get_addr("bpf_testmod_stacktrace_test_1"),
--
2.52.0
^ permalink raw reply related
* [PATCH bpf-next 2/4] x86/fgraph,bpf: Switch kprobe_multi program stack unwind to hw_regs path
From: Jiri Olsa @ 2026-01-12 21:49 UTC (permalink / raw)
To: Masami Hiramatsu, Steven Rostedt, Josh Poimboeuf
Cc: Mahe Tardy, Peter Zijlstra, bpf, linux-trace-kernel, x86,
Yonghong Song, Song Liu, Andrii Nakryiko
In-Reply-To: <20260112214940.1222115-1-jolsa@kernel.org>
Mahe reported missing function from stack trace on top of kprobe
multi program. The missing function is the very first one in the
stacktrace, the one that the bpf program is attached to.
# bpftrace -e 'kprobe:__x64_sys_newuname* { print(kstack)}'
Attaching 1 probe...
do_syscall_64+134
entry_SYSCALL_64_after_hwframe+118
('*' is used for kprobe_multi attachment)
The reason is that the previous change (the Fixes commit) fixed
stack unwind for tracepoint, but removed attached function address
from the stack trace on top of kprobe multi programs, which I also
overlooked in the related test (check following patch).
The tracepoint and kprobe_multi have different stack setup, but use
same unwind path. I think it's better to keep the previous change,
which fixed tracepoint unwind and instead change the kprobe multi
unwind as explained below.
The bpf program stack unwind calls perf_callchain_kernel for kernel
portion and it follows two unwind paths based on X86_EFLAGS_FIXED
bit in pt_regs.flags.
When the bit set we unwind from stack represented by pt_regs argument,
otherwise we unwind currently executed stack up to 'first_frame'
boundary.
The 'first_frame' value is taken from regs.rsp value, but ftrace_caller
and ftrace_regs_caller (ftrace trampoline) functions set the regs.rsp
to the previous stack frame, so we skip the attached function entry.
If we switch kprobe_multi unwind to use the X86_EFLAGS_FIXED bit,
we can control the start of the unwind and get back the attached
function address. As another benefit we also cut extra unwind cycles
needed to reach the 'first_frame' boundary.
The speedup can be meassured with trigger bench for kprobe_multi
program and stacktrace support.
- without bpf_get_stackid call:
# ./bench -w2 -d5 -a -p1 trig-kprobe-multi
Summary: hits 0.857 ± 0.003M/s ( 0.857M/prod), drops 0.000 ± 0.000M/s, total operations 0.857 ± 0.003M/s
- with bpf_get_stackid call:
# ./bench -w2 -d5 -a -g -p1 trig-kprobe-multi
Summary: hits 1.302 ± 0.002M/s ( 1.302M/prod), drops 0.000 ± 0.000M/s, total operations 1.302 ± 0.002M/s
Note the '-g' option for stacktrace added in following change.
To recreate same stack setup for return probe as we have for entry
probe, we set the instruction pointer to the attached function address,
which gets us the same unwind setup and same stack trace.
With the fix, entry probe:
# bpftrace -e 'kprobe:__x64_sys_newuname* { print(kstack)}'
Attaching 1 probe...
__x64_sys_newuname+9
do_syscall_64+134
entry_SYSCALL_64_after_hwframe+118
return probe:
# bpftrace -e 'kretprobe:__x64_sys_newuname* { print(kstack)}'
Attaching 1 probe...
__x64_sys_newuname+4
do_syscall_64+134
entry_SYSCALL_64_after_hwframe+118
Fixes: 6d08340d1e35 ("Revert "perf/x86: Always store regs->ip in perf_callchain_kernel()"")
Reported-by: Mahe Tardy <mahe.tardy@gmail.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
arch/x86/include/asm/ftrace.h | 2 +-
kernel/trace/fgraph.c | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/ftrace.h b/arch/x86/include/asm/ftrace.h
index b08c95872eed..c56e1e63b893 100644
--- a/arch/x86/include/asm/ftrace.h
+++ b/arch/x86/include/asm/ftrace.h
@@ -57,7 +57,7 @@ arch_ftrace_get_regs(struct ftrace_regs *fregs)
}
#define arch_ftrace_partial_regs(regs) do { \
- regs->flags &= ~X86_EFLAGS_FIXED; \
+ regs->flags |= X86_EFLAGS_FIXED; \
regs->cs = __KERNEL_CS; \
} while (0)
diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c
index cc48d16be43e..6279e0a753cf 100644
--- a/kernel/trace/fgraph.c
+++ b/kernel/trace/fgraph.c
@@ -825,7 +825,7 @@ __ftrace_return_to_handler(struct ftrace_regs *fregs, unsigned long frame_pointe
}
if (fregs)
- ftrace_regs_set_instruction_pointer(fregs, ret);
+ ftrace_regs_set_instruction_pointer(fregs, trace.func);
bit = ftrace_test_recursion_trylock(trace.func, ret);
/*
--
2.52.0
^ permalink raw reply related
* [PATCH bpf-next 1/4] x86/fgraph: Fix return_to_handler regs.rsp value
From: Jiri Olsa @ 2026-01-12 21:49 UTC (permalink / raw)
To: Masami Hiramatsu, Steven Rostedt, Josh Poimboeuf
Cc: Peter Zijlstra, bpf, linux-trace-kernel, x86, Yonghong Song,
Song Liu, Andrii Nakryiko, Mahe Tardy
In-Reply-To: <20260112214940.1222115-1-jolsa@kernel.org>
The previous change (Fixes commit) messed up the rsp register value,
which is wrong because it's already adjusted with FRAME_SIZE, we need
the original value (after UNWIND_HINT_FUNC hint).
Fixes: 20a0bc10272f ("x86/fgraph,bpf: Fix stack ORC unwind from kprobe_multi return probe")
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
arch/x86/kernel/ftrace_64.S | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kernel/ftrace_64.S b/arch/x86/kernel/ftrace_64.S
index a132608265f6..613be81b6e88 100644
--- a/arch/x86/kernel/ftrace_64.S
+++ b/arch/x86/kernel/ftrace_64.S
@@ -368,13 +368,15 @@ SYM_CODE_START(return_to_handler)
subq $8, %rsp
UNWIND_HINT_FUNC
+ movq %rsp, %rdi
+
/* Save ftrace_regs for function exit context */
subq $(FRAME_SIZE), %rsp
movq %rax, RAX(%rsp)
movq %rdx, RDX(%rsp)
movq %rbp, RBP(%rsp)
- movq %rsp, RSP(%rsp)
+ movq %rdi, RSP(%rsp)
movq %rsp, %rdi
call ftrace_return_to_handler
--
2.52.0
^ permalink raw reply related
* [PATCH bpf-next 0/4] x86/fgraph,bpf: Fix ORC stack unwind from kprobe_multi
From: Jiri Olsa @ 2026-01-12 21:49 UTC (permalink / raw)
To: Masami Hiramatsu, Steven Rostedt, Josh Poimboeuf
Cc: Peter Zijlstra, bpf, linux-trace-kernel, x86, Yonghong Song,
Song Liu, Andrii Nakryiko, Mahe Tardy
hi,
Mahe reported missing function from stack trace on top of kprobe multi
program. It turned out the latest fix [1] needs some more fixing.
thanks,
jirka
[1] https://lore.kernel.org/bpf/20251104215405.168643-1-jolsa@kernel.org/
---
Jiri Olsa (4):
x86/fgraph: Fix return_to_handler regs.rsp value
x86/fgraph,bpf: Switch kprobe_multi program stack unwind to hw_regs path
selftests/bpf: Fix kprobe multi stacktrace_ips test
selftests/bpf: Allow to benchmark trigger with stacktrace
arch/x86/include/asm/ftrace.h | 2 +-
arch/x86/kernel/ftrace_64.S | 4 +++-
kernel/trace/fgraph.c | 2 +-
tools/testing/selftests/bpf/bench.c | 4 ++++
tools/testing/selftests/bpf/bench.h | 1 +
tools/testing/selftests/bpf/benchs/bench_trigger.c | 1 +
tools/testing/selftests/bpf/prog_tests/stacktrace_ips.c | 3 ++-
tools/testing/selftests/bpf/progs/trigger_bench.c | 18 ++++++++++++++++++
8 files changed, 31 insertions(+), 4 deletions(-)
^ permalink raw reply
* Re: [PATCHv6 bpf-next 7/9] bpf: Add trampoline ip hash table
From: Jiri Olsa @ 2026-01-12 21:27 UTC (permalink / raw)
To: Andrii Nakryiko, Alan Maguire
Cc: Steven Rostedt, Florent Revest, Mark Rutland, bpf, linux-kernel,
linux-trace-kernel, linux-arm-kernel, Alexei Starovoitov,
Daniel Borkmann, Andrii Nakryiko, Menglong Dong, Song Liu
In-Reply-To: <CAEf4BzYgqWXoKTffa5Y6Xm-nPbL9aFgrStR0GfUs4-88f10EgQ@mail.gmail.com>
On Fri, Jan 09, 2026 at 04:36:41PM -0800, Andrii Nakryiko wrote:
> On Tue, Dec 30, 2025 at 6:51 AM Jiri Olsa <jolsa@kernel.org> wrote:
> >
> > Following changes need to lookup trampoline based on its ip address,
> > adding hash table for that.
> >
> > Signed-off-by: Jiri Olsa <jolsa@kernel.org>
> > ---
> > include/linux/bpf.h | 7 +++++--
> > kernel/bpf/trampoline.c | 30 +++++++++++++++++++-----------
> > 2 files changed, 24 insertions(+), 13 deletions(-)
> >
> > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > index 4e7d72dfbcd4..c85677aae865 100644
> > --- a/include/linux/bpf.h
> > +++ b/include/linux/bpf.h
> > @@ -1325,14 +1325,17 @@ struct bpf_tramp_image {
> > };
> >
> > struct bpf_trampoline {
> > - /* hlist for trampoline_table */
> > - struct hlist_node hlist;
> > + /* hlist for trampoline_key_table */
> > + struct hlist_node hlist_key;
> > + /* hlist for trampoline_ip_table */
> > + struct hlist_node hlist_ip;
> > struct ftrace_ops *fops;
> > /* serializes access to fields of this trampoline */
> > struct mutex mutex;
> > refcount_t refcnt;
> > u32 flags;
> > u64 key;
> > + unsigned long ip;
> > struct {
> > struct btf_func_model model;
> > void *addr;
> > diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
> > index 789ff4e1f40b..bdac9d673776 100644
> > --- a/kernel/bpf/trampoline.c
> > +++ b/kernel/bpf/trampoline.c
> > @@ -24,9 +24,10 @@ const struct bpf_prog_ops bpf_extension_prog_ops = {
> > #define TRAMPOLINE_HASH_BITS 10
> > #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
> >
> > -static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
> > +static struct hlist_head trampoline_key_table[TRAMPOLINE_TABLE_SIZE];
> > +static struct hlist_head trampoline_ip_table[TRAMPOLINE_TABLE_SIZE];
> >
> > -/* serializes access to trampoline_table */
> > +/* serializes access to trampoline tables */
> > static DEFINE_MUTEX(trampoline_mutex);
> >
> > #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
> > @@ -135,15 +136,15 @@ void bpf_image_ksym_del(struct bpf_ksym *ksym)
> > PAGE_SIZE, true, ksym->name);
> > }
> >
> > -static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
> > +static struct bpf_trampoline *bpf_trampoline_lookup(u64 key, unsigned long ip)
> > {
> > struct bpf_trampoline *tr;
> > struct hlist_head *head;
> > int i;
> >
> > mutex_lock(&trampoline_mutex);
> > - head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
> > - hlist_for_each_entry(tr, head, hlist) {
> > + head = &trampoline_key_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
> > + hlist_for_each_entry(tr, head, hlist_key) {
> > if (tr->key == key) {
> > refcount_inc(&tr->refcnt);
> > goto out;
> > @@ -164,8 +165,12 @@ static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
> > #endif
> >
> > tr->key = key;
> > - INIT_HLIST_NODE(&tr->hlist);
> > - hlist_add_head(&tr->hlist, head);
> > + tr->ip = ftrace_location(ip);
> > + INIT_HLIST_NODE(&tr->hlist_key);
> > + INIT_HLIST_NODE(&tr->hlist_ip);
> > + hlist_add_head(&tr->hlist_key, head);
> > + head = &trampoline_ip_table[hash_64(tr->ip, TRAMPOLINE_HASH_BITS)];
>
> For key lookups we check that there is no existing trampoline for the
> given key. Can it happen that we have two trampolines at the same IP
> but using two different keys?
so multiple keys (different static functions with same name) resolving to
the same ip happened in past and we should now be able to catch those in
pahole, right? CC-ing Alan ;-)
however, that should fail the attachment on ftrace/direct layer
say we have already registered and attached trampoline key1-ip1,
follow-up attachment of trampoline with key2-ip1 will fail on:
bpf_trampoline_update
register_fentry
direct_ops_add
update_ftrace_direct_add
...
/* Make sure requested entries are not already registered. */
fails, because ip1 is already in direct_functions
...
jirka
>
>
>
> > + hlist_add_head(&tr->hlist_ip, head);
> > refcount_set(&tr->refcnt, 1);
> > mutex_init(&tr->mutex);
> > for (i = 0; i < BPF_TRAMP_MAX; i++)
> > @@ -846,7 +851,7 @@ void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
> > prog->aux->attach_btf_id);
> >
> > bpf_lsm_find_cgroup_shim(prog, &bpf_func);
> > - tr = bpf_trampoline_lookup(key);
> > + tr = bpf_trampoline_lookup(key, 0);
> > if (WARN_ON_ONCE(!tr))
> > return;
> >
> > @@ -866,7 +871,7 @@ struct bpf_trampoline *bpf_trampoline_get(u64 key,
> > {
> > struct bpf_trampoline *tr;
> >
> > - tr = bpf_trampoline_lookup(key);
> > + tr = bpf_trampoline_lookup(key, tgt_info->tgt_addr);
> > if (!tr)
> > return NULL;
> >
> > @@ -902,7 +907,8 @@ void bpf_trampoline_put(struct bpf_trampoline *tr)
> > * fexit progs. The fentry-only trampoline will be freed via
> > * multiple rcu callbacks.
> > */
> > - hlist_del(&tr->hlist);
> > + hlist_del(&tr->hlist_key);
> > + hlist_del(&tr->hlist_ip);
> > if (tr->fops) {
> > ftrace_free_filter(tr->fops);
> > kfree(tr->fops);
> > @@ -1175,7 +1181,9 @@ static int __init init_trampolines(void)
> > int i;
> >
> > for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
> > - INIT_HLIST_HEAD(&trampoline_table[i]);
> > + INIT_HLIST_HEAD(&trampoline_key_table[i]);
> > + for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
> > + INIT_HLIST_HEAD(&trampoline_ip_table[i]);
> > return 0;
> > }
> > late_initcall(init_trampolines);
> > --
> > 2.52.0
> >
^ permalink raw reply
* Re: [PATCH v14 0/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Andrew Morton @ 2026-01-12 21:22 UTC (permalink / raw)
To: Mathieu Desnoyers
Cc: linux-kernel, Paul E. McKenney, Steven Rostedt, Masami Hiramatsu,
Dennis Zhou, Tejun Heo, Christoph Lameter, Martin Liu,
David Rientjes, christian.koenig, Shakeel Butt, SeongJae Park,
Michal Hocko, Johannes Weiner, Sweet Tea Dorminy, Lorenzo Stoakes,
Liam R . Howlett, Mike Rapoport, Suren Baghdasaryan,
Vlastimil Babka, Christian Brauner, Wei Yang, David Hildenbrand,
Miaohe Lin, Al Viro, linux-mm, linux-trace-kernel, Yu Zhao,
Roman Gushchin, Mateusz Guzik, Matthew Wilcox, Baolin Wang,
Aboorva Devarajan
In-Reply-To: <20260112200056.1250404-1-mathieu.desnoyers@efficios.com>
On Mon, 12 Jan 2026 15:00:53 -0500 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> Introduce hierarchical per-cpu counters and use them for RSS tracking to
> fix the per-mm RSS tracking which has become too inaccurate for OOM
> killer purposes on large many-core systems.
Thanks, I'll throw this back into the testing pot while monitoring the
discussion with Michal.
^ permalink raw reply
* [PATCH] tracing: Have show_event_trigger/filter format a bit more in columns
From: Steven Rostedt @ 2026-01-12 20:34 UTC (permalink / raw)
To: LKML, Linux Trace Kernel
Cc: Masami Hiramatsu, Mathieu Desnoyers, Aaron Tomlin
From: Steven Rostedt <rostedt@goodmis.org>
By doing:
# trace-cmd sqlhist -e -n futex_wait select TIMESTAMP_DELTA_USECS as lat from sys_enter_futex as start join sys_exit_futex as end on start.common_pid = end.common_pid
and
# trace-cmd start -e futex_wait -f 'lat > 100' -e page_pool_state_release -f 'pfn == 1'
The output of the show_event_trigger and show_event_filter files are well
aligned because of the inconsistent 'tab' spacing:
~# cat /sys/kernel/tracing/show_event_triggers
syscalls:sys_exit_futex hist:keys=common_pid:vals=hitcount:__lat_12046_2=common_timestamp.usecs-$__arg_12046_1:sort=hitcount:size=2048:clock=global:onmatch(syscalls.sys_enter_futex).trace(futex_wait,$__lat_12046_2) [active]
syscalls:sys_enter_futex hist:keys=common_pid:vals=hitcount:__arg_12046_1=common_timestamp.usecs:sort=hitcount:size=2048:clock=global [active]
~# cat /sys/kernel/tracing/show_event_filters
synthetic:futex_wait (lat > 100)
page_pool:page_pool_state_release (pfn == 1)
This makes it not so easy to read. Instead, force the spacing to be at
least 32 bytes from the beginning (one space if the system:event is longer
than 30 bytes):
~# cat /sys/kernel/tracing/show_event_triggers
syscalls:sys_exit_futex hist:keys=common_pid:vals=hitcount:__lat_8125_2=common_timestamp.usecs-$__arg_8125_1:sort=hitcount:size=2048:clock=global:onmatch(syscalls.sys_enter_futex).trace(futex_wait,$__lat_8125_2) [active]
syscalls:sys_enter_futex hist:keys=common_pid:vals=hitcount:__arg_8125_1=common_timestamp.usecs:sort=hitcount:size=2048:clock=global [active]
~# cat /sys/kernel/tracing/show_event_filters
synthetic:futex_wait (lat > 100)
page_pool:page_pool_state_release (pfn == 1)
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
---
[
This is based on top of these patches:
https://lore.kernel.org/linux-trace-kernel/20260105142939.2655342-1-atomlin@atomlin.com/
]
kernel/trace/trace_events.c | 26 ++++++++++++++++++++++----
1 file changed, 22 insertions(+), 4 deletions(-)
diff --git a/kernel/trace/trace_events.c b/kernel/trace/trace_events.c
index 36936697fa2a..f372a6374164 100644
--- a/kernel/trace/trace_events.c
+++ b/kernel/trace/trace_events.c
@@ -1662,6 +1662,18 @@ static void t_stop(struct seq_file *m, void *p)
mutex_unlock(&event_mutex);
}
+static int get_call_len(struct trace_event_call *call)
+{
+ int len;
+
+ /* Get the length of "<system>:<event>" */
+ len = strlen(call->class->system) + 1;
+ len += strlen(trace_event_name(call));
+
+ /* Set the index to 32 bytes to separate event from data */
+ return len >= 32 ? 1 : 32 - len;
+}
+
/**
* t_show_filters - seq_file callback to display active event filters
* @m: The seq_file interface for formatted output
@@ -1676,14 +1688,17 @@ static int t_show_filters(struct seq_file *m, void *v)
struct trace_event_file *file = v;
struct trace_event_call *call = file->event_call;
struct event_filter *filter;
+ int len;
guard(rcu)();
filter = rcu_dereference(file->filter);
if (!filter || !filter->filter_string)
return 0;
- seq_printf(m, "%s:%s\t%s\n", call->class->system,
- trace_event_name(call), filter->filter_string);
+ len = get_call_len(call);
+
+ seq_printf(m, "%s:%s%*.s%s\n", call->class->system,
+ trace_event_name(call), len, "", filter->filter_string);
return 0;
}
@@ -1702,6 +1717,7 @@ static int t_show_triggers(struct seq_file *m, void *v)
struct trace_event_file *file = v;
struct trace_event_call *call = file->event_call;
struct event_trigger_data *data;
+ int len;
/*
* The event_mutex is held by t_start(), protecting the
@@ -1710,9 +1726,11 @@ static int t_show_triggers(struct seq_file *m, void *v)
if (list_empty(&file->triggers))
return 0;
+ len = get_call_len(call);
+
list_for_each_entry_rcu(data, &file->triggers, list) {
- seq_printf(m, "%s:%s\t", call->class->system,
- trace_event_name(call));
+ seq_printf(m, "%s:%s%*.s", call->class->system,
+ trace_event_name(call), len, "");
data->cmd_ops->print(m, data);
}
--
2.51.0
^ permalink raw reply related
* [PATCH v14 3/3] mm: Implement precise OOM killer task selection
From: Mathieu Desnoyers @ 2026-01-12 20:00 UTC (permalink / raw)
To: Andrew Morton
Cc: linux-kernel, Mathieu Desnoyers, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Michal Hocko, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <20260112200056.1250404-1-mathieu.desnoyers@efficios.com>
Use the hierarchical tree counter approximation to implement the OOM
killer task selection with a 2-pass algorithm. The first pass selects
the process that has the highest badness points approximation, and the
second pass compares each process using the current max badness points
approximation.
The second pass uses an approximate comparison to eliminate all processes
which are below the current max badness points approximation accuracy
range.
Summing the per-CPU counters to calculate the precise badness of tasks
is only required for tasks with an approximate badness within the
accuracy range of the current max points value.
Limit to 16 the maximum number of badness sums allowed for an OOM killer
task selection before falling back to the approximated comparison. This
ensures bounded execution time for scenarios where many tasks have
badness within the accuracy of the maximum badness approximation.
Tested with the following script:
#!/bin/sh
for a in $(seq 1 10); do (tail /dev/zero &); done
sleep 5
for a in $(seq 1 10); do (tail /dev/zero &); done
sleep 2
for a in $(seq 1 10); do (tail /dev/zero &); done
echo "Waiting for tasks to finish"
wait
Results: OOM kill order on a 128GB memory system
================================================
* systemd and sd-pam are chosen first due to their oom_score_ajd:100:
Out of memory: Killed process 3502 (systemd) total-vm:20096kB, anon-rss:0kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:72kB oom_score_adj:100
Out of memory: Killed process 3503 ((sd-pam)) total-vm:21432kB, anon-rss:0kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:76kB oom_score_adj:100
* The first batch of 10 processes are gradually killed, consecutively
picking the one that uses the most memory. The fact that we are
freeing memory from the previous processes increases the threshold
at which the remaining processes of that group are killed. Processes
from the second and third batches of 10 processes have time to start
before we complete killing the first 10 processes:
Out of memory: Killed process 3703 (tail) total-vm:6591280kB, anon-rss:6578176kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:12936kB oom_score_adj:0
Out of memory: Killed process 3705 (tail) total-vm:6731716kB, anon-rss:6709248kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:13212kB oom_score_adj:0
Out of memory: Killed process 3707 (tail) total-vm:6977216kB, anon-rss:6946816kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:13692kB oom_score_adj:0
Out of memory: Killed process 3699 (tail) total-vm:7205640kB, anon-rss:7184384kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:14136kB oom_score_adj:0
Out of memory: Killed process 3713 (tail) total-vm:7463204kB, anon-rss:7438336kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:14644kB oom_score_adj:0
Out of memory: Killed process 3701 (tail) total-vm:7739204kB, anon-rss:7716864kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:15180kB oom_score_adj:0
Out of memory: Killed process 3709 (tail) total-vm:8050176kB, anon-rss:8028160kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:15792kB oom_score_adj:0
Out of memory: Killed process 3711 (tail) total-vm:8362236kB, anon-rss:8339456kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:16404kB oom_score_adj:0
Out of memory: Killed process 3715 (tail) total-vm:8649360kB, anon-rss:8634368kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:16972kB oom_score_adj:0
Out of memory: Killed process 3697 (tail) total-vm:8951788kB, anon-rss:8929280kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:17560kB oom_score_adj:0
* Even though there is a 2 seconds delay between the 2nd and 3rd batches
those appear to execute in mixed order. Therefore, let's consider them
as a single batch of 20 processes. We are hitting oom at a lower
memory threshold because at this point the 20 remaining proceses are
running rather than the previous 10. The process with highest memory
usage is selected for oom, thus making room for the remaining
processes so they can use more memory before they fill the available
memory, thus explaining why the memory use for selected processes
gradually increases, until all system memory is used by the last one:
Out of memory: Killed process 3731 (tail) total-vm:7089868kB, anon-rss:7077888kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:13912kB oom_score_adj:0
Out of memory: Killed process 3721 (tail) total-vm:7417248kB, anon-rss:7405568kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:14556kB oom_score_adj:0
Out of memory: Killed process 3729 (tail) total-vm:7795864kB, anon-rss:7766016kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:15300kB oom_score_adj:0
Out of memory: Killed process 3723 (tail) total-vm:8259620kB, anon-rss:8224768kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:16208kB oom_score_adj:0
Out of memory: Killed process 3737 (tail) total-vm:8695984kB, anon-rss:8667136kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:17060kB oom_score_adj:0
Out of memory: Killed process 3735 (tail) total-vm:9295980kB, anon-rss:9265152kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:18240kB oom_score_adj:0
Out of memory: Killed process 3727 (tail) total-vm:9907900kB, anon-rss:9895936kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:19428kB oom_score_adj:0
Out of memory: Killed process 3719 (tail) total-vm:10631248kB, anon-rss:10600448kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:20844kB oom_score_adj:0
Out of memory: Killed process 3733 (tail) total-vm:11341720kB, anon-rss:11321344kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:22232kB oom_score_adj:0
Out of memory: Killed process 3725 (tail) total-vm:12348124kB, anon-rss:12320768kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:24204kB oom_score_adj:0
Out of memory: Killed process 3759 (tail) total-vm:12978888kB, anon-rss:12967936kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:25440kB oom_score_adj:0
Out of memory: Killed process 3751 (tail) total-vm:14386412kB, anon-rss:14352384kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:28196kB oom_score_adj:0
Out of memory: Killed process 3741 (tail) total-vm:16153168kB, anon-rss:16130048kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:31652kB oom_score_adj:0
Out of memory: Killed process 3753 (tail) total-vm:18414856kB, anon-rss:18391040kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:36076kB oom_score_adj:0
Out of memory: Killed process 3745 (tail) total-vm:21389456kB, anon-rss:21356544kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:41904kB oom_score_adj:0
Out of memory: Killed process 3747 (tail) total-vm:25659348kB, anon-rss:25632768kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:50260kB oom_score_adj:0
Out of memory: Killed process 3755 (tail) total-vm:32030820kB, anon-rss:32006144kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:62720kB oom_score_adj:0
Out of memory: Killed process 3743 (tail) total-vm:42648456kB, anon-rss:42614784kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:83504kB oom_score_adj:0
Out of memory: Killed process 3757 (tail) total-vm:63971028kB, anon-rss:63938560kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:125228kB oom_score_adj:0
Out of memory: Killed process 3749 (tail) total-vm:127799660kB, anon-rss:127778816kB, file-rss:0kB, shmem-rss:0kB, UID:1000 pgtables:250140kB oom_score_adj:0
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Martin Liu <liumartin@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: christian.koenig@amd.com
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R . Howlett" <liam.howlett@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-mm@kvack.org
Cc: linux-trace-kernel@vger.kernel.org
Cc: Yu Zhao <yuzhao@google.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Aboorva Devarajan <aboorvad@linux.ibm.com>
---
Changes since v12:
- The oom_task_origin case needs to set points_min rather than points
variable now, otherwise points_min is used without being initialized.
Changes since v11:
- get_mm_counter_sum() returns a precise sum.
- Use unsigned long type rather than unsigned int for accuracy.
- Use precise sum min/max calculation to compare the chosen vs
current points.
- The first pass finds the maximum task's min points. The
second pass eliminates all tasks for which the max points
are below the currently chosen min points, and uses a precise
sum to validate the candidates which are possibly in range.
---
fs/proc/base.c | 2 +-
include/linux/mm.h | 34 +++++++++++++++---
include/linux/oom.h | 11 +++++-
mm/oom_kill.c | 84 +++++++++++++++++++++++++++++++++++++--------
4 files changed, 109 insertions(+), 22 deletions(-)
diff --git a/fs/proc/base.c b/fs/proc/base.c
index 4eec684baca9..d75d0ce97032 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -589,7 +589,7 @@ static int proc_oom_score(struct seq_file *m, struct pid_namespace *ns,
unsigned long points = 0;
long badness;
- badness = oom_badness(task, totalpages);
+ badness = oom_badness(task, totalpages, false, NULL, NULL);
/*
* Special case OOM_SCORE_ADJ_MIN for all others scale the
* badness value into [0, 2000] range which we have been
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 6d938b3e3709..680f2811702e 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2855,14 +2855,32 @@ static inline struct percpu_counter_tree_level_item *get_rss_stat_items(struct m
/*
* per-process(per-mm_struct) statistics.
*/
+static inline unsigned long __get_mm_counter(struct mm_struct *mm, int member, bool approximate,
+ unsigned long *accuracy_under, unsigned long *accuracy_over)
+{
+ if (approximate) {
+ if (accuracy_under && accuracy_over) {
+ unsigned long under, over;
+
+ percpu_counter_tree_approximate_accuracy_range(&mm->rss_stat[member], &under, &over);
+ *accuracy_under += under;
+ *accuracy_over += over;
+ }
+ return percpu_counter_tree_approximate_sum_positive(&mm->rss_stat[member]);
+ } else {
+ return percpu_counter_tree_precise_sum_positive(&mm->rss_stat[member]);
+ }
+}
+
static inline unsigned long get_mm_counter(struct mm_struct *mm, int member)
{
- return percpu_counter_tree_approximate_sum_positive(&mm->rss_stat[member]);
+ return __get_mm_counter(mm, member, true, NULL, NULL);
}
+
static inline unsigned long get_mm_counter_sum(struct mm_struct *mm, int member)
{
- return percpu_counter_tree_precise_sum_positive(&mm->rss_stat[member]);
+ return __get_mm_counter(mm, member, false, NULL, NULL);
}
void mm_trace_rss_stat(struct mm_struct *mm, int member);
@@ -2903,11 +2921,17 @@ static inline int mm_counter(struct folio *folio)
return mm_counter_file(folio);
}
+static inline unsigned long __get_mm_rss(struct mm_struct *mm, bool approximate,
+ unsigned long *accuracy_under, unsigned long *accuracy_over)
+{
+ return __get_mm_counter(mm, MM_FILEPAGES, approximate, accuracy_under, accuracy_over) +
+ __get_mm_counter(mm, MM_ANONPAGES, approximate, accuracy_under, accuracy_over) +
+ __get_mm_counter(mm, MM_SHMEMPAGES, approximate, accuracy_under, accuracy_over);
+}
+
static inline unsigned long get_mm_rss(struct mm_struct *mm)
{
- return get_mm_counter(mm, MM_FILEPAGES) +
- get_mm_counter(mm, MM_ANONPAGES) +
- get_mm_counter(mm, MM_SHMEMPAGES);
+ return __get_mm_rss(mm, true, NULL, NULL);
}
static inline unsigned long get_mm_hiwater_rss(struct mm_struct *mm)
diff --git a/include/linux/oom.h b/include/linux/oom.h
index 7b02bc1d0a7e..f8e5bfaf7b39 100644
--- a/include/linux/oom.h
+++ b/include/linux/oom.h
@@ -48,6 +48,12 @@ struct oom_control {
unsigned long totalpages;
struct task_struct *chosen;
long chosen_points;
+ bool approximate;
+ /*
+ * Number of precise badness points sums performed by this task
+ * selection.
+ */
+ int nr_precise;
/* Used to print the constraint info. */
enum oom_constraint constraint;
@@ -97,7 +103,10 @@ static inline vm_fault_t check_stable_address_space(struct mm_struct *mm)
}
long oom_badness(struct task_struct *p,
- unsigned long totalpages);
+ unsigned long totalpages,
+ bool approximate,
+ unsigned long *accuracy_under,
+ unsigned long *accuracy_over);
extern bool out_of_memory(struct oom_control *oc);
diff --git a/mm/oom_kill.c b/mm/oom_kill.c
index 5eb11fbba704..12a1419f518a 100644
--- a/mm/oom_kill.c
+++ b/mm/oom_kill.c
@@ -53,6 +53,14 @@
#define CREATE_TRACE_POINTS
#include <trace/events/oom.h>
+/*
+ * Maximum number of badness sums allowed before using an approximated
+ * comparison. This ensures bounded execution time for scenarios where
+ * many tasks have badness within the accuracy of the maximum badness
+ * approximation.
+ */
+static int max_precise_badness_sums = 16;
+
static int sysctl_panic_on_oom;
static int sysctl_oom_kill_allocating_task;
static int sysctl_oom_dump_tasks = 1;
@@ -194,12 +202,16 @@ static bool should_dump_unreclaim_slab(void)
* oom_badness - heuristic function to determine which candidate task to kill
* @p: task struct of which task we should calculate
* @totalpages: total present RAM allowed for page allocation
+ * @approximate: whether the value can be approximated
+ * @accuracy_under: accuracy of the badness value approximation (under value)
+ * @accuracy_over: accuracy of the badness value approximation (over value)
*
* The heuristic for determining which task to kill is made to be as simple and
* predictable as possible. The goal is to return the highest value for the
* task consuming the most memory to avoid subsequent oom failures.
*/
-long oom_badness(struct task_struct *p, unsigned long totalpages)
+long oom_badness(struct task_struct *p, unsigned long totalpages, bool approximate,
+ unsigned long *accuracy_under, unsigned long *accuracy_over)
{
long points;
long adj;
@@ -228,7 +240,8 @@ long oom_badness(struct task_struct *p, unsigned long totalpages)
* The baseline for the badness score is the proportion of RAM that each
* task's rss, pagetable and swap space use.
*/
- points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
+ points = __get_mm_rss(p->mm, approximate, accuracy_under, accuracy_over) +
+ __get_mm_counter(p->mm, MM_SWAPENTS, approximate, accuracy_under, accuracy_over) +
mm_pgtables_bytes(p->mm) / PAGE_SIZE;
task_unlock(p);
@@ -309,7 +322,8 @@ static enum oom_constraint constrained_alloc(struct oom_control *oc)
static int oom_evaluate_task(struct task_struct *task, void *arg)
{
struct oom_control *oc = arg;
- long points;
+ unsigned long accuracy_under = 0, accuracy_over = 0;
+ long points, points_min, points_max;
if (oom_unkillable_task(task))
goto next;
@@ -335,20 +349,47 @@ static int oom_evaluate_task(struct task_struct *task, void *arg)
* killed first if it triggers an oom, then select it.
*/
if (oom_task_origin(task)) {
- points = LONG_MAX;
+ points_min = LONG_MAX;
goto select;
}
- points = oom_badness(task, oc->totalpages);
- if (points == LONG_MIN || points < oc->chosen_points)
- goto next;
+ points = oom_badness(task, oc->totalpages, true, &accuracy_under, &accuracy_over);
+ if (points != LONG_MIN) {
+ percpu_counter_tree_approximate_min_max_range(points,
+ accuracy_under, accuracy_over,
+ &points_min, &points_max);
+ }
+ if (oc->approximate) {
+ /*
+ * Keep the process which has the highest minimum
+ * possible points value based on approximation.
+ */
+ if (points == LONG_MIN || points_min < oc->chosen_points)
+ goto next;
+ } else {
+ /*
+ * Eliminate processes which are certainly below the
+ * chosen points minimum possible value with an
+ * approximation.
+ */
+ if (points == LONG_MIN || (long)(points_max - oc->chosen_points) < 0)
+ goto next;
+
+ if (oc->nr_precise < max_precise_badness_sums) {
+ oc->nr_precise++;
+ /* Precise evaluation. */
+ points_min = points_max = points = oom_badness(task, oc->totalpages, false, NULL, NULL);
+ if (points == LONG_MIN || (long)(points - oc->chosen_points) < 0)
+ goto next;
+ }
+ }
select:
if (oc->chosen)
put_task_struct(oc->chosen);
get_task_struct(task);
oc->chosen = task;
- oc->chosen_points = points;
+ oc->chosen_points = points_min;
next:
return 0;
abort:
@@ -358,14 +399,8 @@ static int oom_evaluate_task(struct task_struct *task, void *arg)
return 1;
}
-/*
- * Simple selection loop. We choose the process with the highest number of
- * 'points'. In case scan was aborted, oc->chosen is set to -1.
- */
-static void select_bad_process(struct oom_control *oc)
+static void select_bad_process_iter(struct oom_control *oc)
{
- oc->chosen_points = LONG_MIN;
-
if (is_memcg_oom(oc))
mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
else {
@@ -379,6 +414,25 @@ static void select_bad_process(struct oom_control *oc)
}
}
+/*
+ * Simple selection loop. We choose the process with the highest number of
+ * 'points'. In case scan was aborted, oc->chosen is set to -1.
+ */
+static void select_bad_process(struct oom_control *oc)
+{
+ oc->chosen_points = LONG_MIN;
+ oc->nr_precise = 0;
+
+ /* Approximate scan. */
+ oc->approximate = true;
+ select_bad_process_iter(oc);
+ if (oc->chosen == (void *)-1UL)
+ return;
+ /* Precise scan. */
+ oc->approximate = false;
+ select_bad_process_iter(oc);
+}
+
static int dump_task(struct task_struct *p, void *arg)
{
struct oom_control *oc = arg;
--
2.39.5
^ permalink raw reply related
* [PATCH v14 1/3] lib: Introduce hierarchical per-cpu counters
From: Mathieu Desnoyers @ 2026-01-12 20:00 UTC (permalink / raw)
To: Andrew Morton
Cc: linux-kernel, Mathieu Desnoyers, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Michal Hocko, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <20260112200056.1250404-1-mathieu.desnoyers@efficios.com>
* Motivation
The purpose of this hierarchical split-counter scheme is to:
- Minimize contention when incrementing and decrementing counters,
- Provide fast access to a sum approximation,
- Provide a sum approximation with an acceptable accuracy level when
scaling to many-core systems.
- Provide approximate and precise comparison of two counters, and
between a counter and a value.
It aims at fixing the per-mm RSS tracking which has become too
inaccurate for OOM killer purposes on large many-core systems [1].
* Design
The hierarchical per-CPU counters propagate a sum approximation through
a N-way tree. When reaching the batch size, the carry is propagated
through a binary tree which consists of logN(nr_cpu_ids) levels. The
batch size for each level is twice the batch size of the prior level.
Example propagation diagram with 8 cpus through a binary tree:
Level 0: 0 1 2 3 4 5 6 7
| / | / | / | /
| / | / | / | /
| / | / | / | /
Level 1: 0 1 2 3
| / | /
| / | /
| / | /
Level 2: 0 1
| /
| /
| /
Level 3: 0
For a binary tree, the maximum inaccuracy is bound by:
batch_size * log2(nr_cpus) * nr_cpus
which evolves with O(n*log(n)) as the number of CPUs increases.
For a N-way tree, the maximum inaccuracy can be pre-calculated
based on the the N-arity of each level and the batch size.
Link: https://lore.kernel.org/lkml/20250331223516.7810-2-sweettea-kernel@dorminy.me/ # [1]
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Martin Liu <liumartin@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: christian.koenig@amd.com
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R . Howlett" <liam.howlett@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-mm@kvack.org
Cc: linux-trace-kernel@vger.kernel.org
Cc: Yu Zhao <yuzhao@google.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Aboorva Devarajan <aboorvad@linux.ibm.com>
---
Changes since v12:
- Use atomic_long_set for percpu_counter_tree_set in UP build.
- percpu_counter_tree_precise_sum returns long type.
Changes since v11:
- Reduce level0 per-cpu memory allocation to the size required by those
percpu counters.
- Use unsigned long type rather than unsigned int.
- Introduce functions to return the min/max range of possible
precise sums.
Changes since v10:
- Fold "mm: Take into account hierarchical percpu tree items for static
mm_struct definitions".
Changes since v9:
- Introduce kerneldoc documentation.
- Document structure fields.
- Remove inline from percpu_counter_tree_add().
- Reject batch_size==1 which makes no sense.
- Fix copy-paste bug in percpu_counter_tree_precise_compare()
(wrong inequality comparison).
- Rename "inaccuracy" to "accuracy", which makes it easier to
document.
- Track accuracy limits more precisely. In two's complement
signed integers, the range before a n-bit underflow is one
unit larger than the range before a n-bit overflow. This sums
for all the counters within the tree. Therefore the "under"
vs "over" accuracy is not symmetrical.
Changes since v8:
- Remove migrate guard from the fast path. It does not
matter through which path the carry is propagated up
the tree.
- Rebase on top of v6.18-rc6.
- Introduce percpu_counter_tree_init_many and
percpu_counter_tree_destroy_many APIs.
- Move tree items allocation to the caller.
- Introduce percpu_counter_tree_items_size().
- Move percpu_counter_tree_subsystem_init() call before mm_core_init()
so percpu_counter_tree_items_size() is initialized before it is used.
Changes since v7:
- Explicitly initialize the subsystem from start_kernel() right
after mm_core_init() so it is up and running before the creation of
the first mm at boot.
- Remove the module.h include which is not needed with the explicit
initialization.
- Only consider levels>0 items for order={0,1} nr_items. No
functional change except to allocate only the amount of memory
which is strictly needed.
- Introduce positive precise sum API to handle a scenario where an
unlucky precise sum iteration would hit negative counter values
concurrently with counter updates.
Changes since v5:
- Introduce percpu_counter_tree_approximate_sum_positive.
- Introduce !CONFIG_SMP static inlines for UP build.
- Remove percpu_counter_tree_set_bias from the public API and make it
static.
Changes since v3:
- Add gfp flags to init function.
Changes since v2:
- Introduce N-way tree to reduce tree depth on larger systems.
Changes since v1:
- Remove percpu_counter_tree_precise_sum_unbiased from public header,
make this function static,
- Introduce precise and approximate comparisons between two counters,
- Reorder the struct percpu_counter_tree fields,
- Introduce approx_sum field, which points to the approximate sum
for the percpu_counter_tree_approximate_sum() fast path.
---
include/linux/mm_types.h | 6 +-
include/linux/percpu_counter_tree.h | 344 ++++++++++++++
init/main.c | 2 +
lib/Makefile | 1 +
lib/percpu_counter_tree.c | 702 ++++++++++++++++++++++++++++
5 files changed, 1052 insertions(+), 3 deletions(-)
create mode 100644 include/linux/percpu_counter_tree.h
create mode 100644 lib/percpu_counter_tree.c
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index aa4639888f89..9f861ceabe61 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -1366,9 +1366,9 @@ static inline void __mm_flags_set_mask_bits_word(struct mm_struct *mm,
MT_FLAGS_USE_RCU)
extern struct mm_struct init_mm;
-#define MM_STRUCT_FLEXIBLE_ARRAY_INIT \
-{ \
- [0 ... sizeof(cpumask_t) + MM_CID_STATIC_SIZE - 1] = 0 \
+#define MM_STRUCT_FLEXIBLE_ARRAY_INIT \
+{ \
+ [0 ... sizeof(cpumask_t) + MM_CID_STATIC_SIZE + PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE - 1] = 0 \
}
/* Pointer magic because the dynamic array size confuses some compilers. */
diff --git a/include/linux/percpu_counter_tree.h b/include/linux/percpu_counter_tree.h
new file mode 100644
index 000000000000..cc02b32d5d22
--- /dev/null
+++ b/include/linux/percpu_counter_tree.h
@@ -0,0 +1,344 @@
+/* SPDX-License-Identifier: GPL-2.0+ OR MIT */
+/* SPDX-FileCopyrightText: 2025 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> */
+
+#ifndef _PERCPU_COUNTER_TREE_H
+#define _PERCPU_COUNTER_TREE_H
+
+#include <linux/preempt.h>
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+#ifdef CONFIG_SMP
+
+#if NR_CPUS == (1U << 0)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 0
+#elif NR_CPUS <= (1U << 1)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 1
+#elif NR_CPUS <= (1U << 2)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 3
+#elif NR_CPUS <= (1U << 3)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 7
+#elif NR_CPUS <= (1U << 4)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 7
+#elif NR_CPUS <= (1U << 5)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 11
+#elif NR_CPUS <= (1U << 6)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 21
+#elif NR_CPUS <= (1U << 7)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 21
+#elif NR_CPUS <= (1U << 8)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 37
+#elif NR_CPUS <= (1U << 9)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 73
+#elif NR_CPUS <= (1U << 10)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 149
+#elif NR_CPUS <= (1U << 11)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 293
+#elif NR_CPUS <= (1U << 12)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 585
+#elif NR_CPUS <= (1U << 13)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 1173
+#elif NR_CPUS <= (1U << 14)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 2341
+#elif NR_CPUS <= (1U << 15)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 4681
+#elif NR_CPUS <= (1U << 16)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 4681
+#elif NR_CPUS <= (1U << 17)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 8777
+#elif NR_CPUS <= (1U << 18)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 17481
+#elif NR_CPUS <= (1U << 19)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 34953
+#elif NR_CPUS <= (1U << 20)
+# define PERCPU_COUNTER_TREE_STATIC_NR_ITEMS 69905
+#else
+# error "Unsupported number of CPUs."
+#endif
+
+struct percpu_counter_tree_level_item {
+ atomic_long_t count; /*
+ * Count the number of carry for this tree item.
+ * The carry counter is kept at the order of the
+ * carry accounted for at this tree level.
+ */
+} ____cacheline_aligned_in_smp;
+
+#define PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE \
+ (PERCPU_COUNTER_TREE_STATIC_NR_ITEMS * sizeof(struct percpu_counter_tree_level_item))
+
+struct percpu_counter_tree {
+ /* Fast-path fields. */
+ unsigned long __percpu *level0; /* Pointer to per-CPU split counters (tree level 0). */
+ unsigned long level0_bit_mask; /* Bit mask to apply to detect carry propagation from tree level 0. */
+ union {
+ unsigned long *i; /* Approximate sum for single-CPU topology. */
+ atomic_long_t *a; /* Approximate sum for SMP topology. */
+ } approx_sum;
+ long bias; /* Bias to apply to counter precise and approximate values. */
+
+ /* Slow-path fields. */
+ struct percpu_counter_tree_level_item *items; /* Array of tree items for levels 1 to N. */
+ unsigned long batch_size; /*
+ * The batch size is the increment step at level 0 which
+ * triggers a carry propagation. The batch size is required
+ * to be greater than 1, and a power of 2.
+ */
+ /*
+ * The tree approximate sum is guaranteed to be within this accuracy range:
+ * (precise_sum - approx_accuracy_range.under) <= approx_sum <= (precise_sum + approx_accuracy_range.over).
+ * This accuracy is derived from the hardware topology and the tree batch_size.
+ * The "under" accuracy is larger than the "over" accuracy because the negative range of a
+ * two's complement signed integer is one unit larger than the positive range. This delta
+ * is summed for each tree item, which leads to a significantly larger "under" accuracy range
+ * compared to the "over" accuracy range.
+ */
+ struct {
+ unsigned long under;
+ unsigned long over;
+ } approx_accuracy_range;
+};
+
+size_t percpu_counter_tree_items_size(void);
+int percpu_counter_tree_init_many(struct percpu_counter_tree *counters, struct percpu_counter_tree_level_item *items,
+ unsigned int nr_counters, unsigned long batch_size, gfp_t gfp_flags);
+int percpu_counter_tree_init(struct percpu_counter_tree *counter, struct percpu_counter_tree_level_item *items,
+ unsigned long batch_size, gfp_t gfp_flags);
+void percpu_counter_tree_destroy_many(struct percpu_counter_tree *counter, unsigned int nr_counters);
+void percpu_counter_tree_destroy(struct percpu_counter_tree *counter);
+void percpu_counter_tree_add(struct percpu_counter_tree *counter, long inc);
+long percpu_counter_tree_precise_sum(struct percpu_counter_tree *counter);
+int percpu_counter_tree_approximate_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b);
+int percpu_counter_tree_approximate_compare_value(struct percpu_counter_tree *counter, long v);
+int percpu_counter_tree_precise_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b);
+int percpu_counter_tree_precise_compare_value(struct percpu_counter_tree *counter, long v);
+void percpu_counter_tree_set(struct percpu_counter_tree *counter, long v);
+void percpu_counter_tree_approximate_accuracy_range(struct percpu_counter_tree *counter,
+ unsigned long *under, unsigned long *over);
+int percpu_counter_tree_subsystem_init(void);
+
+/**
+ * percpu_counter_tree_approximate_sum() - Return approximate counter sum.
+ * @counter: The counter to sum.
+ *
+ * Querying the approximate sum is fast, but it is only accurate within
+ * the bounds delimited by percpu_counter_tree_approximate_accuracy_range().
+ * This is meant to be used when speed is preferred over accuracy.
+ *
+ * Return: The current approximate counter sum.
+ */
+static inline
+long percpu_counter_tree_approximate_sum(struct percpu_counter_tree *counter)
+{
+ unsigned long v;
+
+ if (!counter->level0_bit_mask)
+ v = READ_ONCE(*counter->approx_sum.i);
+ else
+ v = atomic_long_read(counter->approx_sum.a);
+ return (long) (v + (unsigned long)READ_ONCE(counter->bias));
+}
+
+#else /* !CONFIG_SMP */
+
+#define PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE 0
+
+struct percpu_counter_tree_level_item;
+
+struct percpu_counter_tree {
+ atomic_long_t count;
+};
+
+static inline
+size_t percpu_counter_tree_items_size(void)
+{
+ return 0;
+}
+
+static inline
+int percpu_counter_tree_init_many(struct percpu_counter_tree *counters, struct percpu_counter_tree_level_item *items,
+ unsigned int nr_counters, unsigned long batch_size, gfp_t gfp_flags)
+{
+ for (unsigned int i = 0; i < nr_counters; i++)
+ atomic_long_set(&counters[i].count, 0);
+ return 0;
+}
+
+static inline
+int percpu_counter_tree_init(struct percpu_counter_tree *counter, struct percpu_counter_tree_level_item *items,
+ unsigned long batch_size, gfp_t gfp_flags)
+{
+ return percpu_counter_tree_init_many(counter, items, 1, batch_size, gfp_flags);
+}
+
+static inline
+void percpu_counter_tree_destroy_many(struct percpu_counter_tree *counter, unsigned int nr_counters)
+{
+}
+
+static inline
+void percpu_counter_tree_destroy(struct percpu_counter_tree *counter)
+{
+}
+
+static inline
+long percpu_counter_tree_precise_sum(struct percpu_counter_tree *counter)
+{
+ return atomic_long_read(&counter->count);
+}
+
+static inline
+int percpu_counter_tree_precise_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b)
+{
+ long count_a = percpu_counter_tree_precise_sum(a),
+ count_b = percpu_counter_tree_precise_sum(b);
+
+ if (count_a == count_b)
+ return 0;
+ if (count_a < count_b)
+ return -1;
+ return 1;
+}
+
+static inline
+int percpu_counter_tree_precise_compare_value(struct percpu_counter_tree *counter, long v)
+{
+ long count = percpu_counter_tree_precise_sum(counter);
+
+ if (count == v)
+ return 0;
+ if (count < v)
+ return -1;
+ return 1;
+}
+
+static inline
+int percpu_counter_tree_approximate_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b)
+{
+ return percpu_counter_tree_precise_compare(a, b);
+}
+
+static inline
+int percpu_counter_tree_approximate_compare_value(struct percpu_counter_tree *counter, long v)
+{
+ return percpu_counter_tree_precise_compare_value(counter, v);
+}
+
+static inline
+void percpu_counter_tree_set(struct percpu_counter_tree *counter, long v)
+{
+ atomic_long_set(&counter->count, v);
+}
+
+static inline
+void percpu_counter_tree_approximate_accuracy_range(struct percpu_counter_tree *counter,
+ unsigned long *under, unsigned long *over)
+{
+ *under = 0;
+ *over = 0;
+}
+
+static inline
+void percpu_counter_tree_add(struct percpu_counter_tree *counter, long inc)
+{
+ atomic_long_add(inc, &counter->count);
+}
+
+static inline
+long percpu_counter_tree_approximate_sum(struct percpu_counter_tree *counter)
+{
+ return percpu_counter_tree_precise_sum(counter);
+}
+
+static inline
+int percpu_counter_tree_subsystem_init(void)
+{
+ return 0;
+}
+
+#endif /* CONFIG_SMP */
+
+/**
+ * percpu_counter_tree_approximate_sum_positive() - Return a positive approximate counter sum.
+ * @counter: The counter to sum.
+ *
+ * Return an approximate counter sum which is guaranteed to be greater
+ * or equal to 0.
+ *
+ * Return: The current positive approximate counter sum.
+ */
+static inline
+long percpu_counter_tree_approximate_sum_positive(struct percpu_counter_tree *counter)
+{
+ long v = percpu_counter_tree_approximate_sum(counter);
+ return v > 0 ? v : 0;
+}
+
+/**
+ * percpu_counter_tree_precise_sum_positive() - Return a positive precise counter sum.
+ * @counter: The counter to sum.
+ *
+ * Return a precise counter sum which is guaranteed to be greater
+ * or equal to 0.
+ *
+ * Return: The current positive precise counter sum.
+ */
+static inline
+long percpu_counter_tree_precise_sum_positive(struct percpu_counter_tree *counter)
+{
+ long v = percpu_counter_tree_precise_sum(counter);
+ return v > 0 ? v : 0;
+}
+
+/**
+ * percpu_counter_tree_approximate_min_max_range() - Return the approximation min and max precise values.
+ * @approx_sum: Approximated sum.
+ * @under: Tree accuracy range (under).
+ * @over: Tree accuracy range (over).
+ * @precise_min: Minimum possible value for precise sum (output).
+ * @precise_max: Maximum possible value for precise sum (output).
+ *
+ * Calculate the minimum and maximum precise values for a given
+ * approximation and (under, over) accuracy range.
+ *
+ * The range of the approximation as a function of the precise sum is expressed as:
+ *
+ * approx_sum >= precise_sum - approx_accuracy_range.under
+ * approx_sum <= precise_sum + approx_accuracy_range.over
+ *
+ * Therefore, the range of the precise sum as a function of the approximation is expressed as:
+ *
+ * precise_sum <= approx_sum + approx_accuracy_range.under
+ * precise_sum >= approx_sum - approx_accuracy_range.over
+ */
+static inline
+void percpu_counter_tree_approximate_min_max_range(long approx_sum, unsigned long under, unsigned long over,
+ long *precise_min, long *precise_max)
+{
+ *precise_min = approx_sum - over;
+ *precise_max = approx_sum + under;
+}
+
+/**
+ * percpu_counter_tree_approximate_min_max() - Return the tree approximation, min and max possible precise values.
+ * @counter: The counter to sum.
+ * @approx_sum: Approximate sum (output).
+ * @precise_min: Minimum possible value for precise sum (output).
+ * @precise_max: Maximum possible value for precise sum (output).
+ *
+ * Return the approximate sum, minimum and maximum precise values for
+ * a counter.
+ */
+static inline
+void percpu_counter_tree_approximate_min_max(struct percpu_counter_tree *counter,
+ long *approx_sum, long *precise_min, long *precise_max)
+{
+ unsigned long under, over;
+ long v = percpu_counter_tree_approximate_sum(counter);
+
+ percpu_counter_tree_approximate_accuracy_range(counter, &under, &over);
+ percpu_counter_tree_approximate_min_max_range(v, under, over, precise_min, precise_max);
+ *approx_sum = v;
+}
+
+#endif /* _PERCPU_COUNTER_TREE_H */
diff --git a/init/main.c b/init/main.c
index b84818ad9685..85bcc7c1ccd0 100644
--- a/init/main.c
+++ b/init/main.c
@@ -104,6 +104,7 @@
#include <linux/pidfs.h>
#include <linux/ptdump.h>
#include <linux/time_namespace.h>
+#include <linux/percpu_counter_tree.h>
#include <net/net_namespace.h>
#include <asm/io.h>
@@ -1063,6 +1064,7 @@ void start_kernel(void)
vfs_caches_init_early();
sort_main_extable();
trap_init();
+ percpu_counter_tree_subsystem_init();
mm_core_init();
maple_tree_init();
poking_init();
diff --git a/lib/Makefile b/lib/Makefile
index aaf677cf4527..bff117556424 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -181,6 +181,7 @@ obj-$(CONFIG_TEXTSEARCH_KMP) += ts_kmp.o
obj-$(CONFIG_TEXTSEARCH_BM) += ts_bm.o
obj-$(CONFIG_TEXTSEARCH_FSM) += ts_fsm.o
obj-$(CONFIG_SMP) += percpu_counter.o
+obj-$(CONFIG_SMP) += percpu_counter_tree.o
obj-$(CONFIG_AUDIT_GENERIC) += audit.o
obj-$(CONFIG_AUDIT_COMPAT_GENERIC) += compat_audit.o
diff --git a/lib/percpu_counter_tree.c b/lib/percpu_counter_tree.c
new file mode 100644
index 000000000000..cf8feffe06f5
--- /dev/null
+++ b/lib/percpu_counter_tree.c
@@ -0,0 +1,702 @@
+// SPDX-License-Identifier: GPL-2.0+ OR MIT
+// SPDX-FileCopyrightText: 2025 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+
+/*
+ * Split Counters With Tree Approximation Propagation
+ *
+ * * Propagation diagram when reaching batch size thresholds (± batch size):
+ *
+ * Example diagram for 8 CPUs:
+ *
+ * log2(8) = 3 levels
+ *
+ * At each level, each pair propagates its values to the next level when
+ * reaching the batch size thresholds.
+ *
+ * Counters at levels 0, 1, 2 can be kept on a single byte ([-128 .. +127] range),
+ * although it may be relevant to keep them on 64-bit counters for
+ * simplicity. (complexity vs memory footprint tradeoff)
+ *
+ * Counter at level 3 can be kept on a 32-bit counter.
+ *
+ * Level 0: 0 1 2 3 4 5 6 7
+ * | / | / | / | /
+ * | / | / | / | /
+ * | / | / | / | /
+ * Level 1: 0 1 2 3
+ * | / | /
+ * | / | /
+ * | / | /
+ * Level 2: 0 1
+ * | /
+ * | /
+ * | /
+ * Level 3: 0
+ *
+ * * Approximation accuracy:
+ *
+ * BATCH(level N): Level N batch size.
+ *
+ * Example for BATCH(level 0) = 32.
+ *
+ * BATCH(level 0) = 32
+ * BATCH(level 1) = 64
+ * BATCH(level 2) = 128
+ * BATCH(level N) = BATCH(level 0) * 2^N
+ *
+ * per-counter global
+ * accuracy accuracy
+ * Level 0: [ -32 .. +31] ±256 (8 * 32)
+ * Level 1: [ -64 .. +63] ±256 (4 * 64)
+ * Level 2: [-128 .. +127] ±256 (2 * 128)
+ * Total: ------ ±768 (log2(nr_cpu_ids) * BATCH(level 0) * nr_cpu_ids)
+ *
+ * Note that the global accuracy can be calculated more precisely
+ * by taking into account that the positive accuracy range is
+ * 31 rather than 32.
+ *
+ * -----
+ *
+ * Approximate Sum Carry Propagation
+ *
+ * Let's define a number of counter bits for each level, e.g.:
+ *
+ * log2(BATCH(level 0)) = log2(32) = 5
+ *
+ * nr_bit value_mask range
+ * Level 0: 5 bits v 0 .. +31
+ * Level 1: 1 bit (v & ~((1UL << 5) - 1)) 0 .. +63
+ * Level 2: 1 bit (v & ~((1UL << 6) - 1)) 0 .. +127
+ * Level 3: 25 bits (v & ~((1UL << 7) - 1)) 0 .. 2^32-1
+ *
+ * Note: Use a full 64-bit per-cpu counter at level 0 to allow precise sum.
+ *
+ * Note: Use cacheline aligned counters at levels above 0 to prevent false sharing.
+ * If memory footprint is an issue, a specialized allocator could be used
+ * to eliminate padding.
+ *
+ * Example with expanded values:
+ *
+ * counter_add(counter, inc):
+ *
+ * if (!inc)
+ * return;
+ *
+ * res = percpu_add_return(counter @ Level 0, inc);
+ * orig = res - inc;
+ * if (inc < 0) {
+ * inc = -(-inc & ~0b00011111); // Clear used bits
+ * // xor bit 5: underflow
+ * if ((inc ^ orig ^ res) & 0b00100000)
+ * inc -= 0b00100000;
+ * } else {
+ * inc &= ~0b00011111; // Clear used bits
+ * // xor bit 5: overflow
+ * if ((inc ^ orig ^ res) & 0b00100000)
+ * inc += 0b00100000;
+ * }
+ * if (!inc)
+ * return;
+ *
+ * res = atomic_long_add_return(counter @ Level 1, inc);
+ * orig = res - inc;
+ * if (inc < 0) {
+ * inc = -(-inc & ~0b00111111); // Clear used bits
+ * // xor bit 6: underflow
+ * if ((inc ^ orig ^ res) & 0b01000000)
+ * inc -= 0b01000000;
+ * } else {
+ * inc &= ~0b00111111; // Clear used bits
+ * // xor bit 6: overflow
+ * if ((inc ^ orig ^ res) & 0b01000000)
+ * inc += 0b01000000;
+ * }
+ * if (!inc)
+ * return;
+ *
+ * res = atomic_long_add_return(counter @ Level 2, inc);
+ * orig = res - inc;
+ * if (inc < 0) {
+ * inc = -(-inc & ~0b01111111); // Clear used bits
+ * // xor bit 7: underflow
+ * if ((inc ^ orig ^ res) & 0b10000000)
+ * inc -= 0b10000000;
+ * } else {
+ * inc &= ~0b01111111; // Clear used bits
+ * // xor bit 7: overflow
+ * if ((inc ^ orig ^ res) & 0b10000000)
+ * inc += 0b10000000;
+ * }
+ * if (!inc)
+ * return;
+ *
+ * atomic_long_add(counter @ Level 3, inc);
+ */
+
+#include <linux/percpu_counter_tree.h>
+#include <linux/cpumask.h>
+#include <linux/percpu.h>
+#include <linux/atomic.h>
+#include <linux/errno.h>
+#include <linux/slab.h>
+#include <linux/math.h>
+
+#define MAX_NR_LEVELS 5
+
+/*
+ * The counter configuration is selected at boot time based on the
+ * hardware topology.
+ */
+struct counter_config {
+ unsigned int nr_items; /*
+ * nr_items is the number of items in the tree for levels 1
+ * up to and including the final level (approximate sum).
+ * It excludes the level 0 per-CPU counters.
+ */
+ unsigned char nr_levels; /*
+ * nr_levels is the number of hierarchical counter tree levels.
+ * It excludes the final level (approximate sum).
+ */
+ unsigned char n_arity_order[MAX_NR_LEVELS]; /*
+ * n-arity of tree nodes for each level from
+ * 0 to (nr_levels - 1).
+ */
+};
+
+static const struct counter_config per_nr_cpu_order_config[] = {
+ [0] = { .nr_items = 0, .nr_levels = 0, .n_arity_order = { 0 } },
+ [1] = { .nr_items = 1, .nr_levels = 1, .n_arity_order = { 1 } },
+ [2] = { .nr_items = 3, .nr_levels = 2, .n_arity_order = { 1, 1 } },
+ [3] = { .nr_items = 7, .nr_levels = 3, .n_arity_order = { 1, 1, 1 } },
+ [4] = { .nr_items = 7, .nr_levels = 3, .n_arity_order = { 2, 1, 1 } },
+ [5] = { .nr_items = 11, .nr_levels = 3, .n_arity_order = { 2, 2, 1 } },
+ [6] = { .nr_items = 21, .nr_levels = 3, .n_arity_order = { 2, 2, 2 } },
+ [7] = { .nr_items = 21, .nr_levels = 3, .n_arity_order = { 3, 2, 2 } },
+ [8] = { .nr_items = 37, .nr_levels = 3, .n_arity_order = { 3, 3, 2 } },
+ [9] = { .nr_items = 73, .nr_levels = 3, .n_arity_order = { 3, 3, 3 } },
+ [10] = { .nr_items = 149, .nr_levels = 4, .n_arity_order = { 3, 3, 2, 2 } },
+ [11] = { .nr_items = 293, .nr_levels = 4, .n_arity_order = { 3, 3, 3, 2 } },
+ [12] = { .nr_items = 585, .nr_levels = 4, .n_arity_order = { 3, 3, 3, 3 } },
+ [13] = { .nr_items = 1173, .nr_levels = 5, .n_arity_order = { 3, 3, 3, 2, 2 } },
+ [14] = { .nr_items = 2341, .nr_levels = 5, .n_arity_order = { 3, 3, 3, 3, 2 } },
+ [15] = { .nr_items = 4681, .nr_levels = 5, .n_arity_order = { 3, 3, 3, 3, 3 } },
+ [16] = { .nr_items = 4681, .nr_levels = 5, .n_arity_order = { 4, 3, 3, 3, 3 } },
+ [17] = { .nr_items = 8777, .nr_levels = 5, .n_arity_order = { 4, 4, 3, 3, 3 } },
+ [18] = { .nr_items = 17481, .nr_levels = 5, .n_arity_order = { 4, 4, 4, 3, 3 } },
+ [19] = { .nr_items = 34953, .nr_levels = 5, .n_arity_order = { 4, 4, 4, 4, 3 } },
+ [20] = { .nr_items = 69905, .nr_levels = 5, .n_arity_order = { 4, 4, 4, 4, 4 } },
+};
+
+static const struct counter_config *counter_config; /* Hierarchical counter configuration for the hardware topology. */
+static unsigned int nr_cpus_order; /* Order of nr_cpu_ids. */
+static unsigned long accuracy_multiplier; /* Calculate accuracy for a given batch size (multiplication factor). */
+
+static
+int __percpu_counter_tree_init(struct percpu_counter_tree *counter,
+ unsigned long batch_size, gfp_t gfp_flags,
+ unsigned long __percpu *level0,
+ struct percpu_counter_tree_level_item *items)
+{
+ /* Batch size must be greater than 1, and a power of 2. */
+ if (WARN_ON(batch_size <= 1 || (batch_size & (batch_size - 1))))
+ return -EINVAL;
+ counter->batch_size = batch_size;
+ counter->bias = 0;
+ counter->level0 = level0;
+ counter->items = items;
+ if (!nr_cpus_order) {
+ counter->approx_sum.i = per_cpu_ptr(counter->level0, 0);
+ counter->level0_bit_mask = 0;
+ } else {
+ counter->approx_sum.a = &counter->items[counter_config->nr_items - 1].count;
+ counter->level0_bit_mask = 1UL << get_count_order(batch_size);
+ }
+ /*
+ * Each tree item signed integer has a negative range which is
+ * one unit greater than the positive range.
+ */
+ counter->approx_accuracy_range.under = batch_size * accuracy_multiplier;
+ counter->approx_accuracy_range.over = (batch_size - 1) * accuracy_multiplier;
+ return 0;
+}
+
+/**
+ * percpu_counter_tree_init_many() - Initialize many per-CPU counter trees.
+ * @counters: An array of @nr_counters counters to initialize.
+ * Their memory is provided by the caller.
+ * @items: Pointer to memory area where to store tree items.
+ * This memory is provided by the caller.
+ * Its size needs to be at least @nr_counters * percpu_counter_tree_items_size().
+ * @nr_counters: The number of counter trees to initialize
+ * @batch_size: The batch size is the increment step at level 0 which triggers a
+ * carry propagation.
+ * The batch size is required to be greater than 1, and a power of 2.
+ * @gfp_flags: gfp flags to pass to the per-CPU allocator.
+ *
+ * Initialize many per-CPU counter trees using a single per-CPU
+ * allocator invocation for @nr_counters counters.
+ *
+ * Return:
+ * * %0: Success
+ * * %-EINVAL: - Invalid @batch_size argument
+ * * %-ENOMEM: - Out of memory
+ */
+int percpu_counter_tree_init_many(struct percpu_counter_tree *counters, struct percpu_counter_tree_level_item *items,
+ unsigned int nr_counters, unsigned long batch_size, gfp_t gfp_flags)
+{
+ void __percpu *level0, *level0_iter;
+ size_t counter_size = sizeof(*counters->level0),
+ items_size = percpu_counter_tree_items_size();
+ void *items_iter;
+ unsigned int i;
+ int ret;
+
+ memset(items, 0, items_size * nr_counters);
+ level0 = __alloc_percpu_gfp(nr_counters * counter_size,
+ __alignof__(*counters->level0), gfp_flags);
+ if (!level0)
+ return -ENOMEM;
+ level0_iter = level0;
+ items_iter = items;
+ for (i = 0; i < nr_counters; i++) {
+ ret = __percpu_counter_tree_init(&counters[i], batch_size, gfp_flags, level0_iter, items_iter);
+ if (ret)
+ goto free_level0;
+ level0_iter += counter_size;
+ items_iter += items_size;
+ }
+ return 0;
+
+free_level0:
+ free_percpu(level0);
+ return ret;
+}
+
+/**
+ * percpu_counter_tree_init() - Initialize one per-CPU counter tree.
+ * @counter: Counter to initialize.
+ * Its memory is provided by the caller.
+ * @items: Pointer to memory area where to store tree items.
+ * This memory is provided by the caller.
+ * Its size needs to be at least percpu_counter_tree_items_size().
+ * @batch_size: The batch size is the increment step at level 0 which triggers a
+ * carry propagation.
+ * The batch size is required to be greater than 1, and a power of 2.
+ * @gfp_flags: gfp flags to pass to the per-CPU allocator.
+ *
+ * Initialize one per-CPU counter tree.
+ *
+ * Return:
+ * * %0: Success
+ * * %-EINVAL: - Invalid @batch_size argument
+ * * %-ENOMEM: - Out of memory
+ */
+int percpu_counter_tree_init(struct percpu_counter_tree *counter, struct percpu_counter_tree_level_item *items,
+ unsigned long batch_size, gfp_t gfp_flags)
+{
+ return percpu_counter_tree_init_many(counter, items, 1, batch_size, gfp_flags);
+}
+
+/**
+ * percpu_counter_tree_destroy_many() - Destroy many per-CPU counter trees.
+ * @counters: Array of counters trees to destroy.
+ * @nr_counters: The number of counter trees to destroy.
+ *
+ * Release internal resources allocated for @nr_counters per-CPU counter trees.
+ */
+
+void percpu_counter_tree_destroy_many(struct percpu_counter_tree *counters, unsigned int nr_counters)
+{
+ free_percpu(counters->level0);
+}
+
+/**
+ * percpu_counter_tree_destroy() - Destroy one per-CPU counter tree.
+ * @counter: Counter to destroy.
+ *
+ * Release internal resources allocated for one per-CPU counter tree.
+ */
+void percpu_counter_tree_destroy(struct percpu_counter_tree *counter)
+{
+ return percpu_counter_tree_destroy_many(counter, 1);
+}
+
+static
+long percpu_counter_tree_carry(long orig, long res, long inc, unsigned long bit_mask)
+{
+ if (inc < 0) {
+ inc = -(-inc & ~(bit_mask - 1));
+ /*
+ * xor bit_mask: underflow.
+ *
+ * If inc has bit set, decrement an additional bit if
+ * there is _no_ bit transition between orig and res.
+ * Else, inc has bit cleared, decrement an additional
+ * bit if there is a bit transition between orig and
+ * res.
+ */
+ if ((inc ^ orig ^ res) & bit_mask)
+ inc -= bit_mask;
+ } else {
+ inc &= ~(bit_mask - 1);
+ /*
+ * xor bit_mask: overflow.
+ *
+ * If inc has bit set, increment an additional bit if
+ * there is _no_ bit transition between orig and res.
+ * Else, inc has bit cleared, increment an additional
+ * bit if there is a bit transition between orig and
+ * res.
+ */
+ if ((inc ^ orig ^ res) & bit_mask)
+ inc += bit_mask;
+ }
+ return inc;
+}
+
+/*
+ * It does not matter through which path the carry propagates up the
+ * tree, therefore there is no need to disable preemption because the
+ * cpu number is only used to favor cache locality.
+ */
+static
+void percpu_counter_tree_add_slowpath(struct percpu_counter_tree *counter, long inc)
+{
+ unsigned int level_items, nr_levels = counter_config->nr_levels,
+ level, n_arity_order;
+ unsigned long bit_mask;
+ struct percpu_counter_tree_level_item *item = counter->items;
+ unsigned int cpu = raw_smp_processor_id();
+
+ WARN_ON_ONCE(!nr_cpus_order); /* Should never be called for 1 cpu. */
+
+ n_arity_order = counter_config->n_arity_order[0];
+ bit_mask = counter->level0_bit_mask << n_arity_order;
+ level_items = 1U << (nr_cpus_order - n_arity_order);
+
+ for (level = 1; level < nr_levels; level++) {
+ atomic_long_t *count = &item[cpu & (level_items - 1)].count;
+ unsigned long orig, res;
+
+ res = atomic_long_add_return_relaxed(inc, count);
+ orig = res - inc;
+ inc = percpu_counter_tree_carry(orig, res, inc, bit_mask);
+ if (likely(!inc))
+ return;
+ item += level_items;
+ n_arity_order = counter_config->n_arity_order[level];
+ level_items >>= n_arity_order;
+ bit_mask <<= n_arity_order;
+ }
+ atomic_long_add(inc, counter->approx_sum.a);
+}
+
+/**
+ * percpu_counter_tree_add() - Add to a per-CPU counter tree.
+ * @counter: Counter added to.
+ * @inc: Increment value (either positive or negative).
+ *
+ * Add @inc to a per-CPU counter tree. This is a fast-path which will
+ * typically increment per-CPU counters as long as there is no carry
+ * greater or equal to the counter tree batch size.
+ */
+void percpu_counter_tree_add(struct percpu_counter_tree *counter, long inc)
+{
+ unsigned long bit_mask = counter->level0_bit_mask, orig, res;
+
+ res = this_cpu_add_return(*counter->level0, inc);
+ orig = res - inc;
+ inc = percpu_counter_tree_carry(orig, res, inc, bit_mask);
+ if (likely(!inc))
+ return;
+ percpu_counter_tree_add_slowpath(counter, inc);
+}
+
+
+static
+long percpu_counter_tree_precise_sum_unbiased(struct percpu_counter_tree *counter)
+{
+ unsigned long sum = 0;
+ int cpu;
+
+ for_each_possible_cpu(cpu)
+ sum += *per_cpu_ptr(counter->level0, cpu);
+ return (long) sum;
+}
+
+/**
+ * percpu_counter_tree_precise_sum() - Return precise counter sum.
+ * @counter: The counter to sum.
+ *
+ * Querying the precise sum is relatively expensive because it needs to
+ * iterate over all CPUs.
+ * This is meant to be used when accuracy is preferred over speed.
+ *
+ * Return: The current precise counter sum.
+ */
+long percpu_counter_tree_precise_sum(struct percpu_counter_tree *counter)
+{
+ return percpu_counter_tree_precise_sum_unbiased(counter) + READ_ONCE(counter->bias);
+}
+
+static
+int compare_delta(long delta, unsigned long accuracy_neg, unsigned long accuracy_pos)
+{
+ if (delta >= 0) {
+ if (delta <= accuracy_pos)
+ return 0;
+ else
+ return 1;
+ } else {
+ if (-delta <= accuracy_neg)
+ return 0;
+ else
+ return -1;
+ }
+}
+
+/**
+ * percpu_counter_tree_approximate_compare - Approximated comparison of two counter trees.
+ * @a: First counter to compare.
+ * @b: Second counter to compare.
+ *
+ * Evaluate an approximate comparison of two counter trees.
+ * This approximation comparison is fast, and provides an accurate
+ * answer if the counters are found to be either less than or greater
+ * than the other. However, if the approximated comparison returns
+ * 0, the counters respective sums are found to be within the two
+ * counters accuracy range.
+ *
+ * Return:
+ * * %0 - Counters @a and @b do not differ by more than the sum of their respective
+ * accuracy ranges.
+ * * %-1 - Counter @a less than counter @b.
+ * * %1 - Counter @a is greater than counter @b.
+ */
+int percpu_counter_tree_approximate_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b)
+{
+ return compare_delta(percpu_counter_tree_approximate_sum(a) - percpu_counter_tree_approximate_sum(b),
+ a->approx_accuracy_range.over + b->approx_accuracy_range.under,
+ a->approx_accuracy_range.under + b->approx_accuracy_range.over);
+}
+
+/**
+ * percpu_counter_tree_approximate_compare_value - Approximated comparison of a counter tree against a given value.
+ * @counter: Counter to compare.
+ * @v: Value to compare.
+ *
+ * Evaluate an approximate comparison of a counter tree against a given value.
+ * This approximation comparison is fast, and provides an accurate
+ * answer if the counter is found to be either less than or greater
+ * than the value. However, if the approximated comparison returns
+ * 0, the value is within the counter accuracy range.
+ *
+ * Return:
+ * * %0 - The value @v is within the accuracy range of the counter.
+ * * %-1 - The value @v is less than the counter.
+ * * %1 - The value @v is greater than the counter.
+ */
+int percpu_counter_tree_approximate_compare_value(struct percpu_counter_tree *counter, long v)
+{
+ return compare_delta(v - percpu_counter_tree_approximate_sum(counter),
+ counter->approx_accuracy_range.under,
+ counter->approx_accuracy_range.over);
+}
+
+/**
+ * percpu_counter_tree_precise_compare - Precise comparison of two counter trees.
+ * @a: First counter to compare.
+ * @b: Second counter to compare.
+ *
+ * Evaluate a precise comparison of two counter trees.
+ * As an optimization, it uses the approximate counter comparison
+ * to quickly compare counters which are far apart. Only cases where
+ * counter sums are within the accuracy range require precise counter
+ * sums.
+ *
+ * Return:
+ * * %0 - Counters are equal.
+ * * %-1 - Counter @a less than counter @b.
+ * * %1 - Counter @a is greater than counter @b.
+ */
+int percpu_counter_tree_precise_compare(struct percpu_counter_tree *a, struct percpu_counter_tree *b)
+{
+ long count_a = percpu_counter_tree_approximate_sum(a),
+ count_b = percpu_counter_tree_approximate_sum(b);
+ unsigned long accuracy_a, accuracy_b;
+ long delta = count_a - count_b;
+ int res;
+
+ res = compare_delta(delta,
+ a->approx_accuracy_range.over + b->approx_accuracy_range.under,
+ a->approx_accuracy_range.under + b->approx_accuracy_range.over);
+ /* The values are distanced enough for an accurate approximated comparison. */
+ if (res)
+ return res;
+
+ /*
+ * The approximated comparison is within the accuracy range, therefore at least one
+ * precise sum is needed. Sum the counter which has the largest accuracy first.
+ */
+ if (delta >= 0) {
+ accuracy_a = a->approx_accuracy_range.under;
+ accuracy_b = b->approx_accuracy_range.over;
+ } else {
+ accuracy_a = a->approx_accuracy_range.over;
+ accuracy_b = b->approx_accuracy_range.under;
+ }
+ if (accuracy_b < accuracy_a) {
+ count_a = percpu_counter_tree_precise_sum(a);
+ res = compare_delta(count_a - count_b,
+ b->approx_accuracy_range.under,
+ b->approx_accuracy_range.over);
+ if (res)
+ return res;
+ /* Precise sum of second counter is required. */
+ count_b = percpu_counter_tree_precise_sum(b);
+ } else {
+ count_b = percpu_counter_tree_precise_sum(b);
+ res = compare_delta(count_a - count_b,
+ a->approx_accuracy_range.over,
+ a->approx_accuracy_range.under);
+ if (res)
+ return res;
+ /* Precise sum of second counter is required. */
+ count_a = percpu_counter_tree_precise_sum(a);
+ }
+ if (count_a - count_b < 0)
+ return -1;
+ if (count_a - count_b > 0)
+ return 1;
+ return 0;
+}
+
+/**
+ * percpu_counter_tree_precise_compare_value - Precise comparison of a counter tree against a given value.
+ * @counter: Counter to compare.
+ * @v: Value to compare.
+ *
+ * Evaluate a precise comparison of a counter tree against a given value.
+ * As an optimization, it uses the approximate counter comparison
+ * to quickly identify whether the counter and value are far apart.
+ * Only cases where the value is within the counter accuracy range
+ * require a precise counter sum.
+ *
+ * Return:
+ * * %0 - The value @v is equal to the counter.
+ * * %-1 - The value @v is less than the counter.
+ * * %1 - The value @v is greater than the counter.
+ */
+int percpu_counter_tree_precise_compare_value(struct percpu_counter_tree *counter, long v)
+{
+ long count = percpu_counter_tree_approximate_sum(counter);
+ int res;
+
+ res = compare_delta(v - count,
+ counter->approx_accuracy_range.under,
+ counter->approx_accuracy_range.over);
+ /* The values are distanced enough for an accurate approximated comparison. */
+ if (res)
+ return res;
+
+ /* Precise sum is required. */
+ count = percpu_counter_tree_precise_sum(counter);
+ if (v - count < 0)
+ return -1;
+ if (v - count > 0)
+ return 1;
+ return 0;
+}
+
+static
+void percpu_counter_tree_set_bias(struct percpu_counter_tree *counter, long bias)
+{
+ WRITE_ONCE(counter->bias, bias);
+}
+
+/**
+ * percpu_counter_tree_set - Set the counter tree sum to a given value.
+ * @counter: Counter to set.
+ * @v: Value to set.
+ *
+ * Set the counter sum to a given value. It can be useful for instance
+ * to reset the counter sum to 0. Note that even after setting the
+ * counter sum to a given value, the counter sum approximation can
+ * return any value within the accuracy range around that value.
+ */
+void percpu_counter_tree_set(struct percpu_counter_tree *counter, long v)
+{
+ percpu_counter_tree_set_bias(counter,
+ v - percpu_counter_tree_precise_sum_unbiased(counter));
+}
+
+/**
+ * percpu_counter_tree_approximate_accuracy_range - Query the accuracy range for a counter tree.
+ * @counter: Counter to query.
+ * @under: Pointer where to store the to the accuracy range below the approximation.
+ * @over: Pointer where to store the to the accuracy range above the approximation.
+ *
+ * Query the accuracy range limits for the counter.
+ * Because of two's complement binary representation, the "under" range is typically
+ * slightly larger than the "over" range.
+ * Those values are derived from the hardware topology and the counter tree batch size.
+ * They are invariant for a given counter tree.
+ * Using this function should not be typically required, see the following functions instead:
+ * * percpu_counter_tree_approximate_compare(),
+ * * percpu_counter_tree_approximate_compare_value(),
+ * * percpu_counter_tree_precise_compare(),
+ * * percpu_counter_tree_precise_compare_value().
+ */
+void percpu_counter_tree_approximate_accuracy_range(struct percpu_counter_tree *counter,
+ unsigned long *under, unsigned long *over)
+{
+ *under = counter->approx_accuracy_range.under;
+ *over = counter->approx_accuracy_range.over;
+}
+
+/*
+ * percpu_counter_tree_items_size - Query the size required for counter tree items.
+ *
+ * Query the size of the memory area required to hold the counter tree
+ * items. This depends on the hardware topology and is invariant after
+ * boot.
+ *
+ * Return: Size required to hold tree items.
+ */
+size_t percpu_counter_tree_items_size(void)
+{
+ if (!nr_cpus_order)
+ return 0;
+ return counter_config->nr_items * sizeof(struct percpu_counter_tree_level_item);
+}
+
+static void __init calculate_accuracy_topology(void)
+{
+ unsigned int nr_levels = counter_config->nr_levels, level;
+ unsigned int level_items = 1U << nr_cpus_order;
+ unsigned long batch_size = 1;
+
+ for (level = 0; level < nr_levels; level++) {
+ unsigned int n_arity_order = counter_config->n_arity_order[level];
+
+ /*
+ * The accuracy multiplier is derived from a batch size of 1
+ * to speed up calculating the accuracy at tree initialization.
+ */
+ accuracy_multiplier += batch_size * level_items;
+ batch_size <<= n_arity_order;
+ level_items >>= n_arity_order;
+ }
+}
+
+int __init percpu_counter_tree_subsystem_init(void)
+{
+ nr_cpus_order = get_count_order(nr_cpu_ids);
+ if (WARN_ON_ONCE(nr_cpus_order >= ARRAY_SIZE(per_nr_cpu_order_config))) {
+ printk(KERN_ERR "Unsupported number of CPUs (%u)\n", nr_cpu_ids);
+ return -1;
+ }
+ counter_config = &per_nr_cpu_order_config[nr_cpus_order];
+ calculate_accuracy_topology();
+ return 0;
+}
--
2.39.5
^ permalink raw reply related
* [PATCH v14 2/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Mathieu Desnoyers @ 2026-01-12 20:00 UTC (permalink / raw)
To: Andrew Morton
Cc: linux-kernel, Mathieu Desnoyers, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Michal Hocko, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <20260112200056.1250404-1-mathieu.desnoyers@efficios.com>
Use hierarchical per-cpu counters for rss tracking to fix the per-mm RSS
tracking which has become too inaccurate for OOM killer purposes on
large many-core systems.
The following rss tracking issues were noted by Sweet Tea Dorminy [1],
which lead to picking wrong tasks as OOM kill target:
Recently, several internal services had an RSS usage regression as part of a
kernel upgrade. Previously, they were on a pre-6.2 kernel and were able to
read RSS statistics in a backup watchdog process to monitor and decide if
they'd overrun their memory budget. Now, however, a representative service
with five threads, expected to use about a hundred MB of memory, on a 250-cpu
machine had memory usage tens of megabytes different from the expected amount
-- this constituted a significant percentage of inaccuracy, causing the
watchdog to act.
This was a result of commit f1a7941243c1 ("mm: convert mm's rss stats
into percpu_counter") [1]. Previously, the memory error was bounded by
64*nr_threads pages, a very livable megabyte. Now, however, as a result of
scheduler decisions moving the threads around the CPUs, the memory error could
be as large as a gigabyte.
This is a really tremendous inaccuracy for any few-threaded program on a
large machine and impedes monitoring significantly. These stat counters are
also used to make OOM killing decisions, so this additional inaccuracy could
make a big difference in OOM situations -- either resulting in the wrong
process being killed, or in less memory being returned from an OOM-kill than
expected.
Here is a (possibly incomplete) list of the prior approaches that were
used or proposed, along with their downside:
1) Per-thread rss tracking: large error on many-thread processes.
2) Per-CPU counters: up to 12% slower for short-lived processes and 9%
increased system time in make test workloads [1]. Moreover, the
inaccuracy increases with O(n^2) with the number of CPUs.
3) Per-NUMA-node counters: requires atomics on fast-path (overhead),
error is high with systems that have lots of NUMA nodes (32 times
the number of NUMA nodes).
The approach proposed here is to replace this by the hierarchical
per-cpu counters, which bounds the inaccuracy based on the system
topology with O(N*logN).
commit 82241a83cd15 ("mm: fix the inaccurate memory statistics issue for
users") introduced get_mm_counter_sum() for precise /proc memory status
queries. Implement it with percpu_counter_tree_precise_sum() since it
is not a fast path and precision is preferred over speed.
* Testing results:
Test hardware: 2 sockets AMD EPYC 9654 96-Core Processor (384 logical CPUs total)
Methodology:
Comparing the current upstream implementation with the hierarchical
counters is done by keeping both implementations wired up in parallel,
and running a single-process, single-threaded program which hops
randomly across CPUs in the system, calling mmap(2) and munmap(2) on
random CPUs, keeping track of an array of allocated mappings, randomly
choosing entries to either map or unmap.
get_mm_counter() is instrumented to compare the upstream counter
approximation to the precise value, and print the delta when going over
a given threshold. The delta of the hierarchical counter approximation
to the precise value is also printed for comparison.
After a few minutes running this test, the upstream implementation
counter approximation reaches a 1GB delta from the
precise value, compared to 80MB delta with the hierarchical counter.
The hierarchical counter provides a guaranteed maximum approximation
inaccuracy of 192MB on that hardware topology.
* Fast path implementation comparison
The new inline percpu_counter_tree_add() uses a this_cpu_add_return()
for the fast path (under a certain allocation size threshold). Above
that, it calls a slow path which "trickles up" the carry to upper level
counters with atomic_add_return.
In comparison, the upstream counters implementation calls
percpu_counter_add_batch which uses this_cpu_try_cmpxchg() on the fast
path, and does a raw_spin_lock_irqsave above a certain threshold.
The hierarchical implementation is therefore expected to have less
contention on mid-sized allocations than the upstream counters because
the atomic counters tracking those bits are only shared across nearby
CPUs. In comparison, the upstream counters immediately use a global
spinlock when reaching the threshold.
* Benchmarks
Using will-it-scale page_fault1 benchmarks to compare the upstream
counters to the hierarchical counters. This is done with hyperthreading
disabled. The speedup is within the standard deviation of the upstream
runs, so the overhead is not significant.
upstream hierarchical speedup
page_fault1_processes -s 100 -t 1 614783 615558 +0.1%
page_fault1_threads -s 100 -t 1 612788 612447 -0.1%
page_fault1_processes -s 100 -t 96 37994977 37932035 -0.2%
page_fault1_threads -s 100 -t 96 2484130 2504860 +0.8%
page_fault1_processes -s 100 -t 192 71262917 71118830 -0.2%
page_fault1_threads -s 100 -t 192 2446437 2469296 +0.1%
Link: https://lore.kernel.org/lkml/20250331223516.7810-2-sweettea-kernel@dorminy.me/ # [1]
Link: https://lore.kernel.org/lkml/20250704150226.47980-1-mathieu.desnoyers@efficios.com/
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Martin Liu <liumartin@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: christian.koenig@amd.com
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R . Howlett" <liam.howlett@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-mm@kvack.org
Cc: linux-trace-kernel@vger.kernel.org
Cc: Yu Zhao <yuzhao@google.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Aboorva Devarajan <aboorvad@linux.ibm.com>
---
Changes since v13:
- Change check_mm print format from %d to %ld.
Changes since v10:
- Rebase on top of mm_struct static init fixes.
- Change the alignment of mm_struct flexible array to the alignment of
the rss counters items (which are cacheline aligned on SMP).
- Move the rss counters items to first position within the flexible
array at the end of the mm_struct to place content in decreasing
alignment requirement order.
Changes since v8:
- Use percpu_counter_tree_init_many and
percpu_counter_tree_destroy_many APIs.
- Remove percpu tree items allocation. Extend mm_struct size to include
rss items. Those are handled through the new helpers
get_rss_stat_items() and get_rss_stat_items_size() and passed
as parameter to percpu_counter_tree_init_many().
Changes since v7:
- Use precise sum positive API to handle a scenario where an unlucky
precise sum iteration would observe negative counter values due to
concurrent updates.
Changes since v6:
- Rebased on v6.18-rc3.
- Implement get_mm_counter_sum as percpu_counter_tree_precise_sum for
/proc virtual files memory state queries.
Changes since v5:
- Use percpu_counter_tree_approximate_sum_positive.
Change since v4:
- get_mm_counter needs to return 0 or a positive value.
Revert "fixup"
This reverts commit 46f29b72bfaaab4c6ee5a1ac8d29543b230fa293.
---
include/linux/mm.h | 19 ++++++++++----
include/linux/mm_types.h | 50 +++++++++++++++++++++++++++----------
include/trace/events/kmem.h | 2 +-
kernel/fork.c | 22 +++++++++-------
4 files changed, 65 insertions(+), 28 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 6f959d8ca4b4..6d938b3e3709 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2843,38 +2843,47 @@ static inline bool get_user_page_fast_only(unsigned long addr,
{
return get_user_pages_fast_only(addr, 1, gup_flags, pagep) == 1;
}
+
+static inline struct percpu_counter_tree_level_item *get_rss_stat_items(struct mm_struct *mm)
+{
+ unsigned long ptr = (unsigned long)mm;
+
+ ptr += offsetof(struct mm_struct, flexible_array);
+ return (struct percpu_counter_tree_level_item *)ptr;
+}
+
/*
* per-process(per-mm_struct) statistics.
*/
static inline unsigned long get_mm_counter(struct mm_struct *mm, int member)
{
- return percpu_counter_read_positive(&mm->rss_stat[member]);
+ return percpu_counter_tree_approximate_sum_positive(&mm->rss_stat[member]);
}
static inline unsigned long get_mm_counter_sum(struct mm_struct *mm, int member)
{
- return percpu_counter_sum_positive(&mm->rss_stat[member]);
+ return percpu_counter_tree_precise_sum_positive(&mm->rss_stat[member]);
}
void mm_trace_rss_stat(struct mm_struct *mm, int member);
static inline void add_mm_counter(struct mm_struct *mm, int member, long value)
{
- percpu_counter_add(&mm->rss_stat[member], value);
+ percpu_counter_tree_add(&mm->rss_stat[member], value);
mm_trace_rss_stat(mm, member);
}
static inline void inc_mm_counter(struct mm_struct *mm, int member)
{
- percpu_counter_inc(&mm->rss_stat[member]);
+ percpu_counter_tree_add(&mm->rss_stat[member], 1);
mm_trace_rss_stat(mm, member);
}
static inline void dec_mm_counter(struct mm_struct *mm, int member)
{
- percpu_counter_dec(&mm->rss_stat[member]);
+ percpu_counter_tree_add(&mm->rss_stat[member], -1);
mm_trace_rss_stat(mm, member);
}
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 9f861ceabe61..c3e8f0ce3112 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -18,7 +18,7 @@
#include <linux/page-flags-layout.h>
#include <linux/workqueue.h>
#include <linux/seqlock.h>
-#include <linux/percpu_counter.h>
+#include <linux/percpu_counter_tree.h>
#include <linux/types.h>
#include <linux/rseq_types.h>
#include <linux/bitmap.h>
@@ -1070,6 +1070,19 @@ typedef struct {
DECLARE_BITMAP(__mm_flags, NUM_MM_FLAG_BITS);
} __private mm_flags_t;
+/*
+ * The alignment of the mm_struct flexible array is based on the largest
+ * alignment of its content:
+ * __alignof__(struct percpu_counter_tree_level_item) provides a
+ * cacheline aligned alignment on SMP systems, else alignment on
+ * unsigned long on UP systems.
+ */
+#ifdef CONFIG_SMP
+# define __mm_struct_flexible_array_aligned __aligned(__alignof__(struct percpu_counter_tree_level_item))
+#else
+# define __mm_struct_flexible_array_aligned __aligned(__alignof__(unsigned long))
+#endif
+
struct kioctx_table;
struct iommu_mm_data;
struct mm_struct {
@@ -1215,7 +1228,7 @@ struct mm_struct {
unsigned long saved_e_flags;
#endif
- struct percpu_counter rss_stat[NR_MM_COUNTERS];
+ struct percpu_counter_tree rss_stat[NR_MM_COUNTERS];
struct linux_binfmt *binfmt;
@@ -1326,10 +1339,13 @@ struct mm_struct {
} __randomize_layout;
/*
- * The mm_cpumask needs to be at the end of mm_struct, because it
- * is dynamically sized based on nr_cpu_ids.
+ * The rss hierarchical counter items, mm_cpumask, and mm_cid
+ * masks need to be at the end of mm_struct, because they are
+ * dynamically sized based on nr_cpu_ids.
+ * The content of the flexible array needs to be placed in
+ * decreasing alignment requirement order.
*/
- char flexible_array[] __aligned(__alignof__(unsigned long));
+ char flexible_array[] __mm_struct_flexible_array_aligned;
};
/* Copy value to the first system word of mm flags, non-atomically. */
@@ -1368,22 +1384,28 @@ extern struct mm_struct init_mm;
#define MM_STRUCT_FLEXIBLE_ARRAY_INIT \
{ \
- [0 ... sizeof(cpumask_t) + MM_CID_STATIC_SIZE + PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE - 1] = 0 \
+ [0 ... PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE + sizeof(cpumask_t) + MM_CID_STATIC_SIZE - 1] = 0 \
}
-/* Pointer magic because the dynamic array size confuses some compilers. */
-static inline void mm_init_cpumask(struct mm_struct *mm)
+static inline size_t get_rss_stat_items_size(void)
{
- unsigned long cpu_bitmap = (unsigned long)mm;
-
- cpu_bitmap += offsetof(struct mm_struct, flexible_array);
- cpumask_clear((struct cpumask *)cpu_bitmap);
+ return percpu_counter_tree_items_size() * NR_MM_COUNTERS;
}
/* Future-safe accessor for struct mm_struct's cpu_vm_mask. */
static inline cpumask_t *mm_cpumask(struct mm_struct *mm)
{
- return (struct cpumask *)&mm->flexible_array;
+ unsigned long ptr = (unsigned long)mm;
+
+ ptr += offsetof(struct mm_struct, flexible_array);
+ /* Skip RSS stats counters. */
+ ptr += get_rss_stat_items_size();
+ return (struct cpumask *)ptr;
+}
+
+static inline void mm_init_cpumask(struct mm_struct *mm)
+{
+ cpumask_clear((struct cpumask *)mm_cpumask(mm));
}
#ifdef CONFIG_LRU_GEN
@@ -1475,6 +1497,8 @@ static inline cpumask_t *mm_cpus_allowed(struct mm_struct *mm)
unsigned long bitmap = (unsigned long)mm;
bitmap += offsetof(struct mm_struct, flexible_array);
+ /* Skip RSS stats counters. */
+ bitmap += get_rss_stat_items_size();
/* Skip cpu_bitmap */
bitmap += cpumask_size();
return (struct cpumask *)bitmap;
diff --git a/include/trace/events/kmem.h b/include/trace/events/kmem.h
index 7f93e754da5c..91c81c44f884 100644
--- a/include/trace/events/kmem.h
+++ b/include/trace/events/kmem.h
@@ -442,7 +442,7 @@ TRACE_EVENT(rss_stat,
__entry->mm_id = mm_ptr_to_hash(mm);
__entry->curr = !!(current->mm == mm);
__entry->member = member;
- __entry->size = (percpu_counter_sum_positive(&mm->rss_stat[member])
+ __entry->size = (percpu_counter_tree_approximate_sum_positive(&mm->rss_stat[member])
<< PAGE_SHIFT);
),
diff --git a/kernel/fork.c b/kernel/fork.c
index b1f3915d5f8e..8b56d81af734 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -133,6 +133,11 @@
*/
#define MAX_THREADS FUTEX_TID_MASK
+/*
+ * Batch size of rss stat approximation
+ */
+#define RSS_STAT_BATCH_SIZE 32
+
/*
* Protected counters by write_lock_irq(&tasklist_lock)
*/
@@ -626,14 +631,12 @@ static void check_mm(struct mm_struct *mm)
"Please make sure 'struct resident_page_types[]' is updated as well");
for (i = 0; i < NR_MM_COUNTERS; i++) {
- long x = percpu_counter_sum(&mm->rss_stat[i]);
-
- if (unlikely(x)) {
+ if (unlikely(percpu_counter_tree_precise_compare_value(&mm->rss_stat[i], 0) != 0))
pr_alert("BUG: Bad rss-counter state mm:%p type:%s val:%ld Comm:%s Pid:%d\n",
- mm, resident_page_types[i], x,
+ mm, resident_page_types[i],
+ percpu_counter_tree_precise_sum(&mm->rss_stat[i]),
current->comm,
task_pid_nr(current));
- }
}
if (mm_pgtables_bytes(mm))
@@ -731,7 +734,7 @@ void __mmdrop(struct mm_struct *mm)
put_user_ns(mm->user_ns);
mm_pasid_drop(mm);
mm_destroy_cid(mm);
- percpu_counter_destroy_many(mm->rss_stat, NR_MM_COUNTERS);
+ percpu_counter_tree_destroy_many(mm->rss_stat, NR_MM_COUNTERS);
free_mm(mm);
}
@@ -1123,8 +1126,9 @@ static struct mm_struct *mm_init(struct mm_struct *mm, struct task_struct *p,
if (mm_alloc_cid(mm, p))
goto fail_cid;
- if (percpu_counter_init_many(mm->rss_stat, 0, GFP_KERNEL_ACCOUNT,
- NR_MM_COUNTERS))
+ if (percpu_counter_tree_init_many(mm->rss_stat, get_rss_stat_items(mm),
+ NR_MM_COUNTERS, RSS_STAT_BATCH_SIZE,
+ GFP_KERNEL_ACCOUNT))
goto fail_pcpu;
mm->user_ns = get_user_ns(user_ns);
@@ -3006,7 +3010,7 @@ void __init mm_cache_init(void)
* dynamically sized based on the maximum CPU number this system
* can have, taking hotplug into account (nr_cpu_ids).
*/
- mm_size = sizeof(struct mm_struct) + cpumask_size() + mm_cid_size();
+ mm_size = sizeof(struct mm_struct) + cpumask_size() + mm_cid_size() + get_rss_stat_items_size();
mm_cachep = kmem_cache_create_usercopy("mm_struct",
mm_size, ARCH_MIN_MMSTRUCT_ALIGN,
--
2.39.5
^ permalink raw reply related
* [PATCH v14 0/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Mathieu Desnoyers @ 2026-01-12 20:00 UTC (permalink / raw)
To: Andrew Morton
Cc: linux-kernel, Mathieu Desnoyers, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Michal Hocko, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
Introduce hierarchical per-cpu counters and use them for RSS tracking to
fix the per-mm RSS tracking which has become too inaccurate for OOM
killer purposes on large many-core systems.
The following rss tracking issues were noted by Sweet Tea Dorminy [1],
which lead to picking wrong tasks as OOM kill target:
Recently, several internal services had an RSS usage regression as part of a
kernel upgrade. Previously, they were on a pre-6.2 kernel and were able to
read RSS statistics in a backup watchdog process to monitor and decide if
they'd overrun their memory budget. Now, however, a representative service
with five threads, expected to use about a hundred MB of memory, on a 250-cpu
machine had memory usage tens of megabytes different from the expected amount
-- this constituted a significant percentage of inaccuracy, causing the
watchdog to act.
This was a result of commit f1a7941243c1 ("mm: convert mm's rss stats
into percpu_counter") [1]. Previously, the memory error was bounded by
64*nr_threads pages, a very livable megabyte. Now, however, as a result of
scheduler decisions moving the threads around the CPUs, the memory error could
be as large as a gigabyte.
This is a really tremendous inaccuracy for any few-threaded program on a
large machine and impedes monitoring significantly. These stat counters are
also used to make OOM killing decisions, so this additional inaccuracy could
make a big difference in OOM situations -- either resulting in the wrong
process being killed, or in less memory being returned from an OOM-kill than
expected.
The approach proposed here is to replace this by the hierarchical
per-cpu counters, which bounds the inaccuracy based on the system
topology with O(N*logN).
Notable changes for v14:
- Change check_mm print format from %d to %ld (was folded into the wrong
patch).
I've done moderate testing of this series on a 256-core VM with 128GB
RAM. Figuring out whether this indeed helps solve issues with real-life
workloads will require broader feedback from the community.
This series is based on v6.19-rc4, on top of the following two
preparation series:
https://lore.kernel.org/linux-mm/20251224173358.647691-1-mathieu.desnoyers@efficios.com/T/#t
https://lore.kernel.org/linux-mm/20251224173810.648699-1-mathieu.desnoyers@efficios.com/T/#t
Andrew, this series replaces v13, for testing in mm-new.
Thanks!
Mathieu
Link: https://lore.kernel.org/lkml/20250331223516.7810-2-sweettea-kernel@dorminy.me/ # [1]
To: Andrew Morton <akpm@linux-foundation.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Martin Liu <liumartin@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: christian.koenig@amd.com
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R . Howlett" <liam.howlett@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-mm@kvack.org
Cc: linux-trace-kernel@vger.kernel.org
Cc: Yu Zhao <yuzhao@google.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Aboorva Devarajan <aboorvad@linux.ibm.com>
Mathieu Desnoyers (3):
lib: Introduce hierarchical per-cpu counters
mm: Fix OOM killer inaccuracy on large many-core systems
mm: Implement precise OOM killer task selection
fs/proc/base.c | 2 +-
include/linux/mm.h | 49 +-
include/linux/mm_types.h | 54 ++-
include/linux/oom.h | 11 +-
include/linux/percpu_counter_tree.h | 344 ++++++++++++++
include/trace/events/kmem.h | 2 +-
init/main.c | 2 +
kernel/fork.c | 22 +-
lib/Makefile | 1 +
lib/percpu_counter_tree.c | 702 ++++++++++++++++++++++++++++
mm/oom_kill.c | 84 +++-
11 files changed, 1223 insertions(+), 50 deletions(-)
create mode 100644 include/linux/percpu_counter_tree.h
create mode 100644 lib/percpu_counter_tree.c
--
2.39.5
^ permalink raw reply
* Re: [PATCH v13 2/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Michal Hocko @ 2026-01-12 19:48 UTC (permalink / raw)
To: Mathieu Desnoyers
Cc: Andrew Morton, linux-kernel, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <f2c04264-17ca-418f-bc43-e8aa6fa6cd0d@efficios.com>
On Mon 12-01-26 14:37:49, Mathieu Desnoyers wrote:
> On 2026-01-12 03:42, Michal Hocko wrote:
> > Hi,
> > sorry to jump in this late but the timing of previous versions didn't
> > really work well for me.
> >
> > On Sun 11-01-26 14:49:57, Mathieu Desnoyers wrote:
> > [...]
> > > Here is a (possibly incomplete) list of the prior approaches that were
> > > used or proposed, along with their downside:
> > >
> > > 1) Per-thread rss tracking: large error on many-thread processes.
> > >
> > > 2) Per-CPU counters: up to 12% slower for short-lived processes and 9%
> > > increased system time in make test workloads [1]. Moreover, the
> > > inaccuracy increases with O(n^2) with the number of CPUs.
> > >
> > > 3) Per-NUMA-node counters: requires atomics on fast-path (overhead),
> > > error is high with systems that have lots of NUMA nodes (32 times
> > > the number of NUMA nodes).
> > >
> > > The approach proposed here is to replace this by the hierarchical
> > > per-cpu counters, which bounds the inaccuracy based on the system
> > > topology with O(N*logN).
> >
> > The concept of hierarchical pcp counter is interesting and I am
> > definitely not opposed if there are more users that would benefit.
> >
> > From the OOM POV, IIUC the primary problem is that get_mm_counter
> > (percpu_counter_read_positive) is too imprecise on systems when the task
> > is moving around a large number of cpus. In the list of alternative
> > solutions I do not see percpu_counter_sum_positive to be mentioned.
> > oom_badness() is a really slow path and taking the slow path to
> > calculate a much more precise value seems acceptable. Have you
> > considered that option?
> I must admit I assumed that since there was already a mechanism in place
> to ensure it's not necessary to sum per-cpu counters when the oom killer
> is trying to select tasks, it must be because this
>
> O(nr_possible_cpus * nr_processes)
>
> operation must be too slow for the oom killer requirements.
>
> AFAIU, the oom killer is executed when the memory allocator fails to
> allocate memory, which can be within code paths which need to progress
> eventually. So even though it's a slow path compared to the allocator
> fast path, there must be at least _some_ expectations about it
> completing within a decent amount of time. What would that ballpark be ?
I do not think we have ever promissed more than the oom killer will try
to unlock the system blocked on memory shortage.
> To give an order of magnitude, I've tried modifying the upstream
> oom killer to use percpu_counter_sum_positive and compared it to
> the hierarchical approach:
>
> AMD EPYC 9654 96-Core (2 sockets)
> Within a KVM, configured with 256 logical cpus.
>
> nr_processes=40 nr_processes=10000
> Counter sum: 0.4 ms 81.0 ms
> HPCC with 2-pass: 0.3 ms 9.3 ms
These are peanuts for the global oom situations. We have had situations
when soft lockup detector triggered because of the process tree
traversal so adding 100ms is not really critical.
> So as we scale up the number of processes on large SMP systems,
> the latency caused by the oom killer task selection greatly
> increases with the counter sums compared with the hierarchical
> approach.
Yes, I am not really questioning the hierarchical approach will perform
much better but I am thinking of a good enough solution and calculating
the number might be just that stop gap solution (that would be also
suitable for stable tree backports). I am not ruling out improving on
top of that by a more clever solution like your hierarchical counters
approach. Especially if there are more benefits from that elsewhere.
--
Michal Hocko
SUSE Labs
^ permalink raw reply
* Re: [PATCH v13 2/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Mathieu Desnoyers @ 2026-01-12 19:37 UTC (permalink / raw)
To: Michal Hocko
Cc: Andrew Morton, linux-kernel, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, Shakeel Butt,
SeongJae Park, Johannes Weiner, Sweet Tea Dorminy,
Lorenzo Stoakes, Liam R . Howlett, Mike Rapoport,
Suren Baghdasaryan, Vlastimil Babka, Christian Brauner, Wei Yang,
David Hildenbrand, Miaohe Lin, Al Viro, linux-mm,
linux-trace-kernel, Yu Zhao, Roman Gushchin, Mateusz Guzik,
Matthew Wilcox, Baolin Wang, Aboorva Devarajan
In-Reply-To: <aWSz5OHBwtfYrruu@tiehlicka>
On 2026-01-12 03:42, Michal Hocko wrote:
> Hi,
> sorry to jump in this late but the timing of previous versions didn't
> really work well for me.
>
> On Sun 11-01-26 14:49:57, Mathieu Desnoyers wrote:
> [...]
>> Here is a (possibly incomplete) list of the prior approaches that were
>> used or proposed, along with their downside:
>>
>> 1) Per-thread rss tracking: large error on many-thread processes.
>>
>> 2) Per-CPU counters: up to 12% slower for short-lived processes and 9%
>> increased system time in make test workloads [1]. Moreover, the
>> inaccuracy increases with O(n^2) with the number of CPUs.
>>
>> 3) Per-NUMA-node counters: requires atomics on fast-path (overhead),
>> error is high with systems that have lots of NUMA nodes (32 times
>> the number of NUMA nodes).
>>
>> The approach proposed here is to replace this by the hierarchical
>> per-cpu counters, which bounds the inaccuracy based on the system
>> topology with O(N*logN).
>
> The concept of hierarchical pcp counter is interesting and I am
> definitely not opposed if there are more users that would benefit.
>
> From the OOM POV, IIUC the primary problem is that get_mm_counter
> (percpu_counter_read_positive) is too imprecise on systems when the task
> is moving around a large number of cpus. In the list of alternative
> solutions I do not see percpu_counter_sum_positive to be mentioned.
> oom_badness() is a really slow path and taking the slow path to
> calculate a much more precise value seems acceptable. Have you
> considered that option?
I must admit I assumed that since there was already a mechanism in place
to ensure it's not necessary to sum per-cpu counters when the oom killer
is trying to select tasks, it must be because this
O(nr_possible_cpus * nr_processes)
operation must be too slow for the oom killer requirements.
AFAIU, the oom killer is executed when the memory allocator fails to
allocate memory, which can be within code paths which need to progress
eventually. So even though it's a slow path compared to the allocator
fast path, there must be at least _some_ expectations about it
completing within a decent amount of time. What would that ballpark be ?
To give an order of magnitude, I've tried modifying the upstream
oom killer to use percpu_counter_sum_positive and compared it to
the hierarchical approach:
AMD EPYC 9654 96-Core (2 sockets)
Within a KVM, configured with 256 logical cpus.
nr_processes=40 nr_processes=10000
Counter sum: 0.4 ms 81.0 ms
HPCC with 2-pass: 0.3 ms 9.3 ms
So as we scale up the number of processes on large SMP systems,
the latency caused by the oom killer task selection greatly
increases with the counter sums compared with the hierarchical
approach.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
^ permalink raw reply
* [PATCH v2] rtla: Fix parse_cpu_set() bug introduced by strtoi()
From: Costa Shulyupin @ 2026-01-12 19:26 UTC (permalink / raw)
To: Steven Rostedt, Tomas Glozar, Costa Shulyupin,
Wander Lairson Costa, Ivan Pravdin, Tiezhu Yang,
linux-trace-kernel, linux-kernel
The patch 'Replace atoi() with a robust strtoi()' introduced a bug
in parse_cpu_set(), which relies on partial parsing of the input string.
The function parses CPU specifications like '0-3,5' by incrementing
a pointer through the string. strtoi() rejects strings with trailing
characters, causing parse_cpu_set() to fail on any CPU list with
multiple entries.
Restore the original use of atoi() in parse_cpu_set().
Fixes: 7e9dfccf8f11 ("rtla: Replace atoi() with a robust strtoi()")
Signed-off-by: Costa Shulyupin <costa.shul@redhat.com>
---
v2:
- Split the patch.
---
tools/tracing/rtla/src/utils.c | 10 ++++------
1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/tools/tracing/rtla/src/utils.c b/tools/tracing/rtla/src/utils.c
index 18986a5aed3c..0da3b2470c31 100644
--- a/tools/tracing/rtla/src/utils.c
+++ b/tools/tracing/rtla/src/utils.c
@@ -128,18 +128,16 @@ int parse_cpu_set(char *cpu_list, cpu_set_t *set)
nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
for (p = cpu_list; *p; ) {
- if (strtoi(p, &cpu))
- goto err;
- if (cpu < 0 || cpu >= nr_cpus)
+ cpu = atoi(p);
+ if (cpu < 0 || (!cpu && *p != '0') || cpu >= nr_cpus)
goto err;
while (isdigit(*p))
p++;
if (*p == '-') {
p++;
- if (strtoi(p, &end_cpu))
- goto err;
- if (end_cpu < cpu || end_cpu >= nr_cpus)
+ end_cpu = atoi(p);
+ if (end_cpu < cpu || (!end_cpu && *p != '0') || end_cpu >= nr_cpus)
goto err;
while (isdigit(*p))
p++;
--
2.52.0
^ permalink raw reply related
* Re: [PATCH bpf-next 2/3] bpf: Introduce BPF_BRANCH_SNAPSHOT_F_COPY flag for bpf_get_branch_snapshot helper
From: kernel test robot @ 2026-01-12 19:22 UTC (permalink / raw)
To: Leon Hwang, bpf
Cc: llvm, oe-kbuild-all, Alexei Starovoitov, Daniel Borkmann,
Andrii Nakryiko, Martin KaFai Lau, Eduard Zingerman, Song Liu,
Yonghong Song, John Fastabend, KP Singh, Stanislav Fomichev,
Hao Luo, Jiri Olsa, David S . Miller, David Ahern,
Thomas Gleixner, Ingo Molnar, Borislav Petkov, Dave Hansen, x86,
H . Peter Anvin, Matt Bobrowski, Steven Rostedt, Masami Hiramatsu,
Mathieu Desnoyers, Shuah Khan, Leon Hwang, netdev, linux-kernel,
linux-trace-kernel
In-Reply-To: <20260109153420.32181-3-leon.hwang@linux.dev>
Hi Leon,
kernel test robot noticed the following build errors:
[auto build test ERROR on bpf-next/master]
url: https://github.com/intel-lab-lkp/linux/commits/Leon-Hwang/bpf-x64-Call-perf_snapshot_branch_stack-in-trampoline/20260109-234435
base: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master
patch link: https://lore.kernel.org/r/20260109153420.32181-3-leon.hwang%40linux.dev
patch subject: [PATCH bpf-next 2/3] bpf: Introduce BPF_BRANCH_SNAPSHOT_F_COPY flag for bpf_get_branch_snapshot helper
config: x86_64-kexec (https://download.01.org/0day-ci/archive/20260112/202601122013.hmoeIXXs-lkp@intel.com/config)
compiler: clang version 20.1.8 (https://github.com/llvm/llvm-project 87f0227cb60147a26a1eeb4fb06e3b505e9c7261)
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20260112/202601122013.hmoeIXXs-lkp@intel.com/reproduce)
If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202601122013.hmoeIXXs-lkp@intel.com/
All errors (new ones prefixed by >>):
>> ld.lld: error: undefined symbol: bpf_branch_snapshot
>>> referenced by bpf_trace.c:1182 (kernel/trace/bpf_trace.c:1182)
>>> vmlinux.o:(bpf_get_branch_snapshot)
>>> referenced by bpf_trace.c:0 (kernel/trace/bpf_trace.c:0)
>>> vmlinux.o:(bpf_get_branch_snapshot)
--
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
^ permalink raw reply
* Re: [PATCH v12 0/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Steven Rostedt @ 2026-01-12 19:18 UTC (permalink / raw)
To: Mathieu Desnoyers
Cc: Andrew Morton, linux-kernel, Paul E. McKenney, Masami Hiramatsu,
Dennis Zhou, Tejun Heo, Christoph Lameter, Martin Liu,
David Rientjes, christian.koenig, Shakeel Butt, SeongJae Park,
Michal Hocko, Johannes Weiner, Sweet Tea Dorminy, Lorenzo Stoakes,
Liam R . Howlett, Mike Rapoport, Suren Baghdasaryan,
Vlastimil Babka, Christian Brauner, Wei Yang, David Hildenbrand,
Miaohe Lin, Al Viro, linux-mm, linux-trace-kernel, Yu Zhao,
Roman Gushchin, Mateusz Guzik, Matthew Wilcox, Baolin Wang,
Aboorva Devarajan
In-Reply-To: <171c5a7a-f12a-46a0-b791-18e25a3c7877@efficios.com>
On Mon, 12 Jan 2026 13:36:48 -0500
Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> That being said, I think this AI review can lead to interesting
> discussions where others can pitch in, so having it in the open
> is worthwhile.
Agreed. I'm interested in seeing what AI catches.
-- Steve
^ permalink raw reply
* Re: [PATCH v13 2/3] mm: Fix OOM killer inaccuracy on large many-core systems
From: Mathieu Desnoyers @ 2026-01-12 18:46 UTC (permalink / raw)
To: Shakeel Butt
Cc: Andrew Morton, linux-kernel, Paul E. McKenney, Steven Rostedt,
Masami Hiramatsu, Dennis Zhou, Tejun Heo, Christoph Lameter,
Martin Liu, David Rientjes, christian.koenig, SeongJae Park,
Michal Hocko, Johannes Weiner, Sweet Tea Dorminy, Lorenzo Stoakes,
Liam R . Howlett, Mike Rapoport, Suren Baghdasaryan,
Vlastimil Babka, Christian Brauner, Wei Yang, David Hildenbrand,
Miaohe Lin, Al Viro, linux-mm, linux-trace-kernel, Yu Zhao,
Roman Gushchin, Mateusz Guzik, Matthew Wilcox, Baolin Wang,
Aboorva Devarajan
In-Reply-To: <tzgfnxf3tii4s65inlzghhrzgc5gkp2stugnuzuqnjraqbp57k@khqklb37lci2>
On 2026-01-12 12:29, Shakeel Butt wrote:
> On Sun, Jan 11, 2026 at 02:49:57PM -0500, Mathieu Desnoyers wrote:
> [...]
>> @@ -626,14 +631,12 @@ static void check_mm(struct mm_struct *mm)
>> "Please make sure 'struct resident_page_types[]' is updated as well");
>>
>> for (i = 0; i < NR_MM_COUNTERS; i++) {
>> - long x = percpu_counter_sum(&mm->rss_stat[i]);
>> -
>> - if (unlikely(x)) {
>> - pr_alert("BUG: Bad rss-counter state mm:%p type:%s val:%ld Comm:%s Pid:%d\n",
>> - mm, resident_page_types[i], x,
>> + if (unlikely(percpu_counter_tree_precise_compare_value(&mm->rss_stat[i], 0) != 0))
>> + pr_alert("BUG: Bad rss-counter state mm:%p type:%s val:%d Comm:%s Pid:%d\n",
>> + mm, resident_page_types[i],
>> + percpu_counter_tree_precise_sum(&mm->rss_stat[i]),
>
> I haven't gone through the whole series yet but this "val:%ld" to
> "val:%d" conversion seems like unintended. Next patch fixed it but it
> should be fixed here.
My mistake, will fix for v14.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
^ permalink raw reply
page: next (older) | prev (newer) | latest
- recent:[subjects (threaded)|topics (new)|topics (active)]
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox