[RFC v1 0/2] slabinfo: sheaves statistics support

[RFC v1 0/2] slabinfo: sheaves statistics support

[Harry Yoo]

With this series, `slabinfo -r` now prints sheaves statistics (if enabled).
Written on top of slub-percpu-sheaves-v4r0 [1] branch.

Q1: Vlastimil, do you want to integrate this into the sheaves series or do
you want me to send it as a follow-up series after sheaves land mainline?

Q2: 'sheaf_swap' does not fit into the table because SLUB swaps sheaves both
at alloc and free. Probably split the counter?

[1] https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/linux.git/log/?h=slub-percpu-sheaves-v4r0

$ sudo ./slabinfo -r maple_node
Slabcache: maple_node       Aliases:  0 Order :  1 Objects: 8268

[... snip ...]

Slab Perf Counter       Alloc     Free %Al %Fr
----------------------------------------------
Fastpath               258760     5109  93   3
Slowpath                18205   126605   6  96
Page Alloc                903      520   0   0
Add partial                 0    14600   0  11
Remove partial           3472      420   1   0
Cpu partial list        12235    16577   4  12
RemoteObj/SlabFrozen        0     5969   0   4
Total                  276965   131714


Slab Sheaves Counter       Alloc    Free  %Al  %Fr  %Fa
-------------------------------------------------------
Served from sheaves     1805673   580161   86   93  N/A
Sheaves refill           236384      N/A   13  N/A  N/A
Sheaves alloc              3872     3794    0    0  N/A
Free to RCU sheaves         N/A  1255815  N/A   68    0
Main sheaf flush            N/A        0  N/A    0  N/A
Other sheaves flush         N/A   265600  N/A   14  N/A

Slab Deactivation             Occurrences %
-------------------------------------------
Slab full                           0    0%
Slab empty                        100    0%
Moved to head of partial list       0    0%
Moved to tail of partial list       0    0%
Deactivation bypass             16588   99%
Refilled from foreign frees      1595    9%
Node mismatch                       0    0%

Harry Yoo (2):
  mm/slab: add sheaves_enabled read-only sysfs file
  tools/mm/slabinfo: add support for sheaves stats

 mm/slub.c           | 10 ++++++++
 tools/mm/slabinfo.c | 62 +++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 70 insertions(+), 2 deletions(-)

-- 
2.43.0

[RFC v1 1/2] mm/slab: add sheaves_enabled read-only sysfs file

[Harry Yoo]

For user-space tools like slabinfo, it is difficult to determine whether
a cache has sheaves enabled. Expose a read-only 'sheaves_enabled' sysfs
file to indicate whether sheaves are enabled for a given cache.

Signed-off-by: Harry Yoo <harry.yoo@oracle.com>
---
 mm/slub.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/mm/slub.c b/mm/slub.c
index b5ea66cbbe68..32db440a8b05 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -8504,6 +8504,15 @@ static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
 SLAB_ATTR(remote_node_defrag_ratio);
 #endif
 
+static ssize_t sheaves_enabled_show(struct kmem_cache *s, char *buf)
+{
+	if (s->cpu_sheaves)
+		return sysfs_emit(buf, "%u\n", 1);
+	else
+		return sysfs_emit(buf, "%u\n", 0);
+}
+SLAB_ATTR_RO(sheaves_enabled);
+
 #ifdef CONFIG_SLUB_STATS
 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
 {
@@ -8656,6 +8665,7 @@ static struct attribute *slab_attrs[] = {
 #ifdef CONFIG_NUMA
 	&remote_node_defrag_ratio_attr.attr,
 #endif
+	&sheaves_enabled_attr.attr,
 #ifdef CONFIG_SLUB_STATS
 	&alloc_cpu_sheaf_attr.attr,
 	&alloc_fastpath_attr.attr,
-- 
2.43.0

[RFC v1 2/2] tools/mm/slabinfo: add support for sheaves stats

[Harry Yoo]

Add sheaves stats support to the slabinfo tool to make it easier to
investigate sheaves statistics.

"Served from sheaves" indicates the number of allocations and frees that
were served from sheaves, out of the total number of allocations and
frees (including those not using sheaves).

The other sheaves statistics reflect only the number of events among
allocations and frees that were served from sheaves.

Before:
$ sudo ./slabinfo -r maple_node
Slabcache: maple_node       Aliases:  0 Order :  1 Objects: 8433

[... snip ...]

Slab Perf Counter       Alloc     Free %Al %Fr
--------------------------------------------------
Fastpath               177539     3150  94   4
Slowpath                10841    72067   5  95
Page Alloc                903      519   0   0
Add partial                 0     8729   0  11
Remove partial           2007      420   1   0
Cpu partial list         7124     9999   3  13
RemoteObj/SlabFrozen        0     3041   0   4
Total                  188380    75217

Slab Deactivation             Occurrences %
-------------------------------------------------
Slab full                           0    0%
Slab empty                         99    0%
Moved to head of partial list       0    0%
Moved to tail of partial list       0    0%
Deactivation bypass             10012   99%
Refilled from foreign frees       807    7%
Node mismatch                       0    0%

After:
$ sudo ./slabinfo -r maple_node
Slabcache: maple_node       Aliases:  0 Order :  1 Objects: 8268

[... snip ...]

Slab Perf Counter       Alloc     Free %Al %Fr
----------------------------------------------
Fastpath               258760     5109  93   3
Slowpath                18205   126605   6  96
Page Alloc                903      520   0   0
Add partial                 0    14600   0  11
Remove partial           3472      420   1   0
Cpu partial list        12235    16577   4  12
RemoteObj/SlabFrozen        0     5969   0   4
Total                  276965   131714

Slab Sheaves Counter       Alloc    Free  %Al  %Fr  %Fa
-------------------------------------------------------
Served from sheaves     1805673   580161   86   93  N/A
Sheaves refill           236384      N/A   13  N/A  N/A
Sheaves alloc              3872     3794    0    0  N/A
Free to RCU sheaves         N/A  1255815  N/A   68    0
Main sheaf flush            N/A        0  N/A    0  N/A
Other sheaves flush         N/A   265600  N/A   14  N/A

Slab Deactivation             Occurrences %
-------------------------------------------
Slab full                           0    0%
Slab empty                        100    0%
Moved to head of partial list       0    0%
Moved to tail of partial list       0    0%
Deactivation bypass             16588   99%
Refilled from foreign frees      1595    9%
Node mismatch                       0    0%

Signed-off-by: Harry Yoo <harry.yoo@oracle.com>
---
 tools/mm/slabinfo.c | 62 +++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 60 insertions(+), 2 deletions(-)

diff --git a/tools/mm/slabinfo.c b/tools/mm/slabinfo.c
index 1433eff99feb..c930b2135b33 100644
--- a/tools/mm/slabinfo.c
+++ b/tools/mm/slabinfo.c
@@ -44,6 +44,12 @@ struct slabinfo {
 	unsigned long cmpxchg_double_cpu_fail, cmpxchg_double_fail;
 	unsigned long alloc_node_mismatch, deactivate_bypass;
 	unsigned long cpu_partial_alloc, cpu_partial_free;
+	unsigned int sheaves_enabled;
+	unsigned long alloc_cpu_sheaf, free_cpu_sheaf;
+	unsigned long free_rcu_sheaf, free_rcu_sheaf_fail;
+	unsigned long sheaf_flush_main, sheaf_flush_other;
+	unsigned long sheaf_refill, sheaf_swap;
+	unsigned long sheaf_alloc, sheaf_free;
 	int numa[MAX_NODES];
 	int numa_partial[MAX_NODES];
 } slabinfo[MAX_SLABS];
@@ -472,6 +478,7 @@ static void slab_stats(struct slabinfo *s)
 	unsigned long total_alloc;
 	unsigned long total_free;
 	unsigned long total;
+	const char *not_avail = "N/A";
 
 	if (!s->alloc_slab)
 		return;
@@ -484,7 +491,7 @@ static void slab_stats(struct slabinfo *s)
 
 	printf("\n");
 	printf("Slab Perf Counter       Alloc     Free %%Al %%Fr\n");
-	printf("--------------------------------------------------\n");
+	printf("----------------------------------------------\n");
 	printf("Fastpath             %8lu %8lu %3lu %3lu\n",
 		s->alloc_fastpath, s->free_fastpath,
 		s->alloc_fastpath * 100 / total_alloc,
@@ -525,9 +532,47 @@ static void slab_stats(struct slabinfo *s)
 	total = s->deactivate_full + s->deactivate_empty +
 			s->deactivate_to_head + s->deactivate_to_tail + s->deactivate_bypass;
 
+	if (s->sheaves_enabled) {
+		unsigned long sheaves_total_alloc = s->alloc_cpu_sheaf;
+		unsigned long sheaves_total_free = s->free_cpu_sheaf + s->free_rcu_sheaf;
+
+		total_alloc += sheaves_total_alloc;
+		total_free += sheaves_total_free;
+
+		printf("\n");
+		printf("Slab Sheaves Counter       Alloc    Free  %%Al  %%Fr  %%Fa\n");
+		printf("-------------------------------------------------------\n");
+		printf("Served from sheaves    %8lu %8lu %4lu %4lu %4s\n",
+		       s->alloc_cpu_sheaf, s->free_cpu_sheaf,
+		       total_alloc ? sheaves_total_alloc * 100 / total_alloc : 0,
+		       total_free ? sheaves_total_free * 100 / total_free : 0,
+		       not_avail);
+		printf("Sheaves refill         %8lu %8s %4lu %4s %4s\n",
+		       s->sheaf_refill, not_avail,
+		       sheaves_total_alloc ? s->sheaf_refill * 100 / sheaves_total_alloc : 0,
+		       not_avail, not_avail);
+		printf("Sheaves alloc          %8lu %8lu %4lu %4lu %4s\n",
+		       s->sheaf_alloc, s->sheaf_free,
+		       sheaves_total_alloc ? s->sheaf_alloc * 100 / sheaves_total_alloc : 0,
+		       sheaves_total_free ? s->sheaf_free * 100 / sheaves_total_free : 0,
+		       not_avail);
+		printf("Free to RCU sheaves    %8s %8lu %4s %4lu %4lu\n",
+		       not_avail, s->free_rcu_sheaf, not_avail,
+		       sheaves_total_free ? s->free_rcu_sheaf * 100 / sheaves_total_free : 0,
+		       s->free_rcu_sheaf ? s->free_rcu_sheaf_fail * 100 / s->free_rcu_sheaf : 0);
+		printf("Main sheaf flush       %8s %8lu %4s %4lu %4s\n",
+		       not_avail, s->sheaf_flush_main, not_avail,
+		       sheaves_total_free ? s->sheaf_flush_main * 100 / sheaves_total_free : 0,
+		       not_avail);
+		printf("Other sheaves flush    %8s %8lu %4s %4lu %4s\n",
+		       not_avail, s->sheaf_flush_other, not_avail,
+		       sheaves_total_free ? s->sheaf_flush_other * 100 / sheaves_total_free : 0,
+		       not_avail);
+	}
+
 	if (total) {
 		printf("\nSlab Deactivation             Occurrences %%\n");
-		printf("-------------------------------------------------\n");
+		printf("-------------------------------------------\n");
 		printf("Slab full                     %7lu  %3lu%%\n",
 			s->deactivate_full, (s->deactivate_full * 100) / total);
 		printf("Slab empty                    %7lu  %3lu%%\n",
@@ -1301,6 +1346,19 @@ static void read_slab_dir(void)
 			slab->cpu_partial_free = get_obj("cpu_partial_free");
 			slab->alloc_node_mismatch = get_obj("alloc_node_mismatch");
 			slab->deactivate_bypass = get_obj("deactivate_bypass");
+			slab->sheaves_enabled = get_obj("sheaves_enabled");
+			if (slab->sheaves_enabled) {
+				slab->alloc_cpu_sheaf = get_obj("alloc_cpu_sheaf");
+				slab->free_cpu_sheaf = get_obj("free_cpu_sheaf");
+				slab->free_rcu_sheaf = get_obj("free_rcu_sheaf");
+				slab->free_rcu_sheaf_fail = get_obj("free_rcu_sheaf_fail");
+				slab->sheaf_flush_main = get_obj("sheaf_flush_main");
+				slab->sheaf_flush_other = get_obj("sheaf_flush_other");
+				slab->sheaf_refill = get_obj("sheaf_refill");
+				slab->sheaf_swap = get_obj("sheaf_swap");
+				slab->sheaf_alloc = get_obj("sheaf_alloc");
+				slab->sheaf_free = get_obj("sheaf_free");
+			}
 			if (chdir(".."))
 				fatal("Unable to chdir from slab ../%s\n",
 				      slab->name);
-- 
2.43.0

Re: [RFC v1 0/2] slabinfo: sheaves statistics support

[Harry Yoo]

On Mon, Apr 07, 2025 at 01:18:08PM +0900, Harry Yoo wrote:
> With this series, `slabinfo -r` now prints sheaves statistics (if enabled).
> Written on top of slub-percpu-sheaves-v4r0 [1] branch.
> 
> Q1: Vlastimil, do you want to integrate this into the sheaves series or do
> you want me to send it as a follow-up series after sheaves land mainline?
> 
> Q2: 'sheaf_swap' does not fit into the table because SLUB swaps sheaves both
> at alloc and free. Probably split the counter?

Q3: Having 'Add sheaves to the barn' & 'Remove sheaves from the barn' counters
might help better understand the behavior? (just like 'Add partial' /
'Remove partial' pair in the normal SLUB path).

-- 
Cheers,
Harry (formerly known as Hyeonggon)

> [1] https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/linux.git/log/?h=slub-percpu-sheaves-v4r0
> 
> $ sudo ./slabinfo -r maple_node
> Slabcache: maple_node       Aliases:  0 Order :  1 Objects: 8268
> 
> [... snip ...]
> 
> Slab Perf Counter       Alloc     Free %Al %Fr
> ----------------------------------------------
> Fastpath               258760     5109  93   3
> Slowpath                18205   126605   6  96
> Page Alloc                903      520   0   0
> Add partial                 0    14600   0  11
> Remove partial           3472      420   1   0
> Cpu partial list        12235    16577   4  12
> RemoteObj/SlabFrozen        0     5969   0   4
> Total                  276965   131714
> 
> 
> Slab Sheaves Counter       Alloc    Free  %Al  %Fr  %Fa
> -------------------------------------------------------
> Served from sheaves     1805673   580161   86   93  N/A
> Sheaves refill           236384      N/A   13  N/A  N/A
> Sheaves alloc              3872     3794    0    0  N/A
> Free to RCU sheaves         N/A  1255815  N/A   68    0
> Main sheaf flush            N/A        0  N/A    0  N/A
> Other sheaves flush         N/A   265600  N/A   14  N/A
> 
> Slab Deactivation             Occurrences %
> -------------------------------------------
> Slab full                           0    0%
> Slab empty                        100    0%
> Moved to head of partial list       0    0%
> Moved to tail of partial list       0    0%
> Deactivation bypass             16588   99%
> Refilled from foreign frees      1595    9%
> Node mismatch                       0    0%
> 
> Harry Yoo (2):
>   mm/slab: add sheaves_enabled read-only sysfs file
>   tools/mm/slabinfo: add support for sheaves stats
> 
>  mm/slub.c           | 10 ++++++++
>  tools/mm/slabinfo.c | 62 +++++++++++++++++++++++++++++++++++++++++++--
>  2 files changed, 70 insertions(+), 2 deletions(-)
> 
> -- 
> 2.43.0
> 

Re: [RFC v1 1/2] mm/slab: add sheaves_enabled read-only sysfs file

[David Rientjes]

On Mon, 7 Apr 2025, Harry Yoo wrote:

> For user-space tools like slabinfo, it is difficult to determine whether
> a cache has sheaves enabled. Expose a read-only 'sheaves_enabled' sysfs
> file to indicate whether sheaves are enabled for a given cache.
> 

Thoughts on adding an entry to Documentation/ABI/testing/sysfs-kernel-slab 
for this one?

> Signed-off-by: Harry Yoo <harry.yoo@oracle.com>
> ---
>  mm/slub.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/mm/slub.c b/mm/slub.c
> index b5ea66cbbe68..32db440a8b05 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -8504,6 +8504,15 @@ static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
>  SLAB_ATTR(remote_node_defrag_ratio);
>  #endif
>  
> +static ssize_t sheaves_enabled_show(struct kmem_cache *s, char *buf)
> +{
> +	if (s->cpu_sheaves)
> +		return sysfs_emit(buf, "%u\n", 1);
> +	else
> +		return sysfs_emit(buf, "%u\n", 0);
> +}
> +SLAB_ATTR_RO(sheaves_enabled);
> +
>  #ifdef CONFIG_SLUB_STATS
>  static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
>  {
> @@ -8656,6 +8665,7 @@ static struct attribute *slab_attrs[] = {
>  #ifdef CONFIG_NUMA
>  	&remote_node_defrag_ratio_attr.attr,
>  #endif
> +	&sheaves_enabled_attr.attr,
>  #ifdef CONFIG_SLUB_STATS
>  	&alloc_cpu_sheaf_attr.attr,
>  	&alloc_fastpath_attr.attr,
> -- 
> 2.43.0
> 
> 

Re: [RFC v1 1/2] mm/slab: add sheaves_enabled read-only sysfs file

[Harry Yoo]

On Tue, Apr 08, 2025 at 07:45:38PM -0700, David Rientjes wrote:
> On Mon, 7 Apr 2025, Harry Yoo wrote:
> 
> > For user-space tools like slabinfo, it is difficult to determine whether
> > a cache has sheaves enabled. Expose a read-only 'sheaves_enabled' sysfs
> > file to indicate whether sheaves are enabled for a given cache.
> > 
> 
> Thoughts on adding an entry to Documentation/ABI/testing/sysfs-kernel-slab 
> for this one?

That makes sense, thanks!

New stats added by the sheaves series should also be there.

...By the way I'm a bit unsure about the name.
'sheaves', 'has_sheaves' or 'sheaves_enabled'?

-- 
Cheers,
Harry / Hyeonggon

Re: [RFC v1 1/2] mm/slab: add sheaves_enabled read-only sysfs file

[Vlastimil Babka]

On 4/11/25 00:32, Harry Yoo wrote:
> On Tue, Apr 08, 2025 at 07:45:38PM -0700, David Rientjes wrote:
>> On Mon, 7 Apr 2025, Harry Yoo wrote:
>> 
>> > For user-space tools like slabinfo, it is difficult to determine whether
>> > a cache has sheaves enabled. Expose a read-only 'sheaves_enabled' sysfs
>> > file to indicate whether sheaves are enabled for a given cache.
>> > 
>> 
>> Thoughts on adding an entry to Documentation/ABI/testing/sysfs-kernel-slab 
>> for this one?
> 
> That makes sense, thanks!
> 
> New stats added by the sheaves series should also be there.
> 
> ...By the way I'm a bit unsure about the name.
> 'sheaves', 'has_sheaves' or 'sheaves_enabled'?

How about reporting sheaf_capacity isntead?

Re: [RFC v1 0/2] slabinfo: sheaves statistics support

[Vlastimil Babka]

On 4/7/25 06:37, Harry Yoo wrote:
> On Mon, Apr 07, 2025 at 01:18:08PM +0900, Harry Yoo wrote:
>> With this series, `slabinfo -r` now prints sheaves statistics (if enabled).
>> Written on top of slub-percpu-sheaves-v4r0 [1] branch.

Thanks!

>> Q1: Vlastimil, do you want to integrate this into the sheaves series or do
>> you want me to send it as a follow-up series after sheaves land mainline?

I think it's fine to do patch 2 a follow-up. As for patch 1 I will instead
export sheaf_capacity as part of the initial support.

>> Q2: 'sheaf_swap' does not fit into the table because SLUB swaps sheaves both
>> at alloc and free. Probably split the counter?

I'm considering just dropping it instead as it's not that useful as a
measure of caching efficiency? Probably sheaf_flush_main also doesn't need
to be a separate counter as it only happens in corner cases.
Due to ABI becoming set in stone we should be conservative.

> Q3: Having 'Add sheaves to the barn' & 'Remove sheaves from the barn' counters
> might help better understand the behavior? (just like 'Add partial' /
> 'Remove partial' pair in the normal SLUB path).

Yeah that could make more sense instead I think,

Re: [RFC v1 0/2] slabinfo: sheaves statistics support

[Harry Yoo]

On Wed, Apr 23, 2025 at 04:28:59PM +0200, Vlastimil Babka wrote:
> On 4/7/25 06:37, Harry Yoo wrote:
> > On Mon, Apr 07, 2025 at 01:18:08PM +0900, Harry Yoo wrote:
> >> With this series, `slabinfo -r` now prints sheaves statistics (if enabled).
> >> Written on top of slub-percpu-sheaves-v4r0 [1] branch.
> 
> Thanks!
> 
> >> Q1: Vlastimil, do you want to integrate this into the sheaves series or do
> >> you want me to send it as a follow-up series after sheaves land mainline?
> 
> I think it's fine to do patch 2 a follow-up.

Alright.

> As for patch 1 I will instead
> export sheaf_capacity as part of the initial support.

Thanks! 'sheaf_capacity' sounds fine to me.

> >> Q2: 'sheaf_swap' does not fit into the table because SLUB swaps sheaves both
> >> at alloc and free. Probably split the counter?
> 
> I'm considering just dropping it instead as it's not that useful as a
> measure of caching efficiency? Probably sheaf_flush_main also doesn't need
> to be a separate counter as it only happens in corner cases.

No strong opinion from me.
Not adding it will be fine as we can add it later if needed.

> Due to ABI becoming set in stone we should be conservative.

Agreed.

> > Q3: Having 'Add sheaves to the barn' & 'Remove sheaves from the barn' counters
> > might help better understand the behavior? (just like 'Add partial' /
> > 'Remove partial' pair in the normal SLUB path).
> 
> Yeah that could make more sense instead I think,

-- 
Cheers,
Harry / Hyeonggon