Do we still need SLAB_MEM_SPREAD (and possibly others)?

Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Steven Rostedt]

I was looking into moving eventfs_inode into a slab, and after cutting and
pasting the tracefs allocator:

	tracefs_inode_cachep = kmem_cache_create("tracefs_inode_cache",
						 sizeof(struct tracefs_inode),
						 0, (SLAB_RECLAIM_ACCOUNT|
						     SLAB_MEM_SPREAD|
						     SLAB_ACCOUNT),
						 init_once);

I figured I should know what those slab flags mean. I also looked at what
others in fs use for their slabs. The above is rather common (which I
probably just copied from another file system), but I wanted to know what
they are for.

When I got to SLAB_MEM_SPREAD, I found that it's a common flag and there's
a lot of caches that just set that and nothing else.

But I couldn't find how it was used.

Then I found this commit:

 16a1d968358a ("mm/slab: remove mm/slab.c and slab_def.h")

Which I think removed the only use case of SLAB_MEM_SPREAD.

 $ git grep SLAB_MEM_SPREAD mm
mm/slab.h:                            SLAB_MEM_SPREAD | \

That's all I find in the mm directory.

Is it obsolete now? Can we delete it? Maybe there's other SLAB_* flags that
are no longer used. I don't know, I haven't audited them.

-- Steve

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Yosry Ahmed]

On Wed, Jan 31, 2024 at 2:20 PM Steven Rostedt <rostedt@goodmis.org> wrote:
>
> I was looking into moving eventfs_inode into a slab, and after cutting and
> pasting the tracefs allocator:
>
>         tracefs_inode_cachep = kmem_cache_create("tracefs_inode_cache",
>                                                  sizeof(struct tracefs_inode),
>                                                  0, (SLAB_RECLAIM_ACCOUNT|
>                                                      SLAB_MEM_SPREAD|
>                                                      SLAB_ACCOUNT),
>                                                  init_once);
>
> I figured I should know what those slab flags mean. I also looked at what
> others in fs use for their slabs. The above is rather common (which I
> probably just copied from another file system), but I wanted to know what
> they are for.
>
> When I got to SLAB_MEM_SPREAD, I found that it's a common flag and there's
> a lot of caches that just set that and nothing else.
>
> But I couldn't find how it was used.
>
> Then I found this commit:
>
>  16a1d968358a ("mm/slab: remove mm/slab.c and slab_def.h")
>
> Which I think removed the only use case of SLAB_MEM_SPREAD.
>
>  $ git grep SLAB_MEM_SPREAD mm
> mm/slab.h:                            SLAB_MEM_SPREAD | \
>
> That's all I find in the mm directory.
>
> Is it obsolete now? Can we delete it? Maybe there's other SLAB_* flags that
> are no longer used. I don't know, I haven't audited them.

Perhaps cpuset_do_slab_mem_spread() as well.

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Vlastimil Babka]

On 1/31/24 23:25, Yosry Ahmed wrote:
> On Wed, Jan 31, 2024 at 2:20 PM Steven Rostedt <rostedt@goodmis.org> wrote:
>>
>> I was looking into moving eventfs_inode into a slab, and after cutting and
>> pasting the tracefs allocator:
>>
>>         tracefs_inode_cachep = kmem_cache_create("tracefs_inode_cache",
>>                                                  sizeof(struct tracefs_inode),
>>                                                  0, (SLAB_RECLAIM_ACCOUNT|
>>                                                      SLAB_MEM_SPREAD|
>>                                                      SLAB_ACCOUNT),
>>                                                  init_once);
>>
>> I figured I should know what those slab flags mean. I also looked at what
>> others in fs use for their slabs. The above is rather common (which I
>> probably just copied from another file system), but I wanted to know what
>> they are for.
>>
>> When I got to SLAB_MEM_SPREAD, I found that it's a common flag and there's
>> a lot of caches that just set that and nothing else.
>>
>> But I couldn't find how it was used.
>>
>> Then I found this commit:
>>
>>  16a1d968358a ("mm/slab: remove mm/slab.c and slab_def.h")
>>
>> Which I think removed the only use case of SLAB_MEM_SPREAD.
>>
>>  $ git grep SLAB_MEM_SPREAD mm
>> mm/slab.h:                            SLAB_MEM_SPREAD | \
>>
>> That's all I find in the mm directory.
>>
>> Is it obsolete now? Can we delete it? Maybe there's other SLAB_* flags that
>> are no longer used. I don't know, I haven't audited them.
> 
> Perhaps cpuset_do_slab_mem_spread() as well.

Yep, good find. Show how obscure mm/slab.c was in the end :)

CCing a few more new people who did slab changes recently, who'd like some
low hanging fruit of negative diffcount? :)

Thanks,
Vlastimil

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Chengming Zhou]

On 2024/2/1 06:40, Vlastimil Babka wrote:
> On 1/31/24 23:25, Yosry Ahmed wrote:
>> On Wed, Jan 31, 2024 at 2:20 PM Steven Rostedt <rostedt@goodmis.org> wrote:
>>>
>>> I was looking into moving eventfs_inode into a slab, and after cutting and
>>> pasting the tracefs allocator:
>>>
>>>         tracefs_inode_cachep = kmem_cache_create("tracefs_inode_cache",
>>>                                                  sizeof(struct tracefs_inode),
>>>                                                  0, (SLAB_RECLAIM_ACCOUNT|
>>>                                                      SLAB_MEM_SPREAD|
>>>                                                      SLAB_ACCOUNT),
>>>                                                  init_once);
>>>
>>> I figured I should know what those slab flags mean. I also looked at what
>>> others in fs use for their slabs. The above is rather common (which I
>>> probably just copied from another file system), but I wanted to know what
>>> they are for.
>>>
>>> When I got to SLAB_MEM_SPREAD, I found that it's a common flag and there's
>>> a lot of caches that just set that and nothing else.
>>>
>>> But I couldn't find how it was used.
>>>
>>> Then I found this commit:
>>>
>>>  16a1d968358a ("mm/slab: remove mm/slab.c and slab_def.h")
>>>
>>> Which I think removed the only use case of SLAB_MEM_SPREAD.
>>>
>>>  $ git grep SLAB_MEM_SPREAD mm
>>> mm/slab.h:                            SLAB_MEM_SPREAD | \
>>>
>>> That's all I find in the mm directory.
>>>
>>> Is it obsolete now? Can we delete it? Maybe there's other SLAB_* flags that
>>> are no longer used. I don't know, I haven't audited them.
>>
>> Perhaps cpuset_do_slab_mem_spread() as well.
> 
> Yep, good find. Show how obscure mm/slab.c was in the end :)
> 
> CCing a few more new people who did slab changes recently, who'd like some
> low hanging fruit of negative diffcount? :)

Thanks for CCing, I can prepare the patch to do it. IIUC, what I need to do is:

1. delete SLAB_MEM_SPREAD and all its uses.

2. cpuset_do_slab_mem_spread() is not used anymore, should we keep the interface?
   Since it's the interface exported by cgroup-v1 "cpuset.memory_spread_slab".

RE: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Song, Xiongwei]

> -----Original Message-----
> From: Chengming Zhou <chengming.zhou@linux.dev>
> Sent: Thursday, February 1, 2024 2:27 PM
> To: Vlastimil Babka <vbabka@suse.cz>; Yosry Ahmed
> <yosryahmed@google.com>; Steven Rostedt <rostedt@goodmis.org>
> Cc: LKML <linux-kernel@vger.kernel.org>; linux-mm@kvack.org; Andrew
> Morton <akpm@linux-foundation.org>; Linus Torvalds <torvalds@linux-
> foundation.org>; Kees Cook <keescook@chromium.org>; Christoph Lameter
> <cl@linux.com>; David Rientjes <rientjes@google.com>; Hyeonggon Yoo
> <42.hyeyoo@gmail.com>; Song, Xiongwei <Xiongwei.Song@windriver.com>;
> Chengming Zhou <zhouchengming@bytedance.com>; Zheng Yejian
> <zhengyejian1@huawei.com>
> Subject: Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?
> 
> 
> On 2024/2/1 06:40, Vlastimil Babka wrote:
> > On 1/31/24 23:25, Yosry Ahmed wrote:
> >> On Wed, Jan 31, 2024 at 2:20 PM Steven Rostedt <rostedt@goodmis.org>
> wrote:
> >>>
> >>> I was looking into moving eventfs_inode into a slab, and after cutting and
> >>> pasting the tracefs allocator:
> >>>
> >>>         tracefs_inode_cachep = kmem_cache_create("tracefs_inode_cache",
> >>>                                                  sizeof(struct tracefs_inode),
> >>>                                                  0, (SLAB_RECLAIM_ACCOUNT|
> >>>                                                      SLAB_MEM_SPREAD|
> >>>                                                      SLAB_ACCOUNT),
> >>>                                                  init_once);
> >>>
> >>> I figured I should know what those slab flags mean. I also looked at what
> >>> others in fs use for their slabs. The above is rather common (which I
> >>> probably just copied from another file system), but I wanted to know
> what
> >>> they are for.
> >>>
> >>> When I got to SLAB_MEM_SPREAD, I found that it's a common flag and
> there's
> >>> a lot of caches that just set that and nothing else.
> >>>
> >>> But I couldn't find how it was used.
> >>>
> >>> Then I found this commit:
> >>>
> >>>  16a1d968358a ("mm/slab: remove mm/slab.c and slab_def.h")
> >>>
> >>> Which I think removed the only use case of SLAB_MEM_SPREAD.
> >>>
> >>>  $ git grep SLAB_MEM_SPREAD mm
> >>> mm/slab.h:                            SLAB_MEM_SPREAD | \
> >>>
> >>> That's all I find in the mm directory.
> >>>
> >>> Is it obsolete now? Can we delete it? Maybe there's other SLAB_* flags
> that
> >>> are no longer used. I don't know, I haven't audited them.
> >>
> >> Perhaps cpuset_do_slab_mem_spread() as well.
> >
> > Yep, good find. Show how obscure mm/slab.c was in the end :)
> >
> > CCing a few more new people who did slab changes recently, who'd like
> some
> > low hanging fruit of negative diffcount? :)
> 
> Thanks for CCing, I can prepare the patch to do it. IIUC, what I need to do is:
> 
> 1. delete SLAB_MEM_SPREAD and all its uses.
> 
> 2. cpuset_do_slab_mem_spread() is not used anymore, should we keep the
> interface?
>    Since it's the interface exported by cgroup-v1
> "cpuset.memory_spread_slab".

Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is useless.

Regards,
Xiongwei

RE: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Christoph Lameter (Ampere)]

On Sun, 4 Feb 2024, Song, Xiongwei wrote:

> Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is useless.

SLAB_MEM_SPREAD does not do anything anymore. SLUB relies on the 
"spreading" via the page allocator memory policies instead of doing its 
own like SLAB used to do.

What does FILE_SPREAD_SLAB do? Dont see anything there either.

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Waiman Long]

On 2/5/24 20:46, Song, Xiongwei wrote:
> Adding the maintainers of cpuset of cgroup.
>
>> On Sun, 4 Feb 2024, Song, Xiongwei wrote:
>>
>>> Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is useless.
>> SLAB_MEM_SPREAD does not do anything anymore. SLUB relies on the
>> "spreading" via the page allocator memory policies instead of doing its
>> own like SLAB used to do.
>>
>> What does FILE_SPREAD_SLAB do? Dont see anything there either.
> The FILE_SPREAD_SLAB flag is used by cpuset.memory_spread_slab with read/write operations:
>
> In kernel/cgroup/cpuset.c,
> static struct cftype legacy_files[] = {
> ... snip ...
>          {
>                  .name = "memory_spread_slab",
>                  .read_u64 = cpuset_read_u64,
>                  .write_u64 = cpuset_write_u64,
>                  .private = FILE_SPREAD_SLAB,
>          },
> ... snip ...
> };

It looks like that memory_spread_slab may have effect only on the slab 
allocator. With the removal of the slab allocator, memory_spread_slab is 
now a no-op. However, the memory_spread_slab cgroupfs file is an 
externally visible API. So we can't just remove it as it may break 
existing applications. We can certainly deprecate it and advise users 
not to use it.

Cheers,
Longman

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Waiman Long]

On 2/5/24 22:16, Waiman Long wrote:
>
> On 2/5/24 20:46, Song, Xiongwei wrote:
>> Adding the maintainers of cpuset of cgroup.
>>
>>> On Sun, 4 Feb 2024, Song, Xiongwei wrote:
>>>
>>>> Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is 
>>>> useless.
>>> SLAB_MEM_SPREAD does not do anything anymore. SLUB relies on the
>>> "spreading" via the page allocator memory policies instead of doing its
>>> own like SLAB used to do.
>>>
>>> What does FILE_SPREAD_SLAB do? Dont see anything there either.
>> The FILE_SPREAD_SLAB flag is used by cpuset.memory_spread_slab with 
>> read/write operations:
>>
>> In kernel/cgroup/cpuset.c,
>> static struct cftype legacy_files[] = {
>> ... snip ...
>>          {
>>                  .name = "memory_spread_slab",
>>                  .read_u64 = cpuset_read_u64,
>>                  .write_u64 = cpuset_write_u64,
>>                  .private = FILE_SPREAD_SLAB,
>>          },
>> ... snip ...
>> };
>
> It looks like that memory_spread_slab may have effect only on the slab 
> allocator. With the removal of the slab allocator, memory_spread_slab 
> is now a no-op. However, the memory_spread_slab cgroupfs file is an 
> externally visible API. So we can't just remove it as it may break 
> existing applications. We can certainly deprecate it and advise users 
> not to use it.

BTW, cpuset doesn't use SLAB_MEM_SPREAD directly. Instead it set the 
task's PFA_SPREAD_SLAB and let other subsystems test it to act 
appropriately. Other than cpuset, the latest upstream kernel doesn't 
check or use this flag at all.

Cheers,
Longman

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Chengming Zhou]

On 2024/2/6 11:20, Waiman Long wrote:
> 
> On 2/5/24 22:16, Waiman Long wrote:
>>
>> On 2/5/24 20:46, Song, Xiongwei wrote:
>>> Adding the maintainers of cpuset of cgroup.
>>>
>>>> On Sun, 4 Feb 2024, Song, Xiongwei wrote:
>>>>
>>>>> Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is useless.
>>>> SLAB_MEM_SPREAD does not do anything anymore. SLUB relies on the
>>>> "spreading" via the page allocator memory policies instead of doing its
>>>> own like SLAB used to do.
>>>>
>>>> What does FILE_SPREAD_SLAB do? Dont see anything there either.
>>> The FILE_SPREAD_SLAB flag is used by cpuset.memory_spread_slab with read/write operations:
>>>
>>> In kernel/cgroup/cpuset.c,
>>> static struct cftype legacy_files[] = {
>>> ... snip ...
>>>          {
>>>                  .name = "memory_spread_slab",
>>>                  .read_u64 = cpuset_read_u64,
>>>                  .write_u64 = cpuset_write_u64,
>>>                  .private = FILE_SPREAD_SLAB,
>>>          },
>>> ... snip ...
>>> };
>>
>> It looks like that memory_spread_slab may have effect only on the slab allocator. With the removal of the slab allocator, memory_spread_slab is now a no-op. However, the memory_spread_slab cgroupfs file is an externally visible API. So we can't just remove it as it may break existing applications. We can certainly deprecate it and advise users not to use it.
> 
> BTW, cpuset doesn't use SLAB_MEM_SPREAD directly. Instead it set the task's PFA_SPREAD_SLAB and let other subsystems test it to act appropriately. Other than cpuset, the latest upstream kernel doesn't check or use this flag at all.
> 

Ok, get it. So cpuset_do_slab_mem_spread() can be removed, but
this cpuset file interface and PFA_SPREAD_SLAB will be keeped.

Thanks.

Re: Do we still need SLAB_MEM_SPREAD (and possibly others)?

[Waiman Long]

On 2/5/24 22:25, Chengming Zhou wrote:
> On 2024/2/6 11:20, Waiman Long wrote:
>> On 2/5/24 22:16, Waiman Long wrote:
>>> On 2/5/24 20:46, Song, Xiongwei wrote:
>>>> Adding the maintainers of cpuset of cgroup.
>>>>
>>>>> On Sun, 4 Feb 2024, Song, Xiongwei wrote:
>>>>>
>>>>>> Once SLAB_MEM_SPREAD is removed, IMO, cpuset.memory_spread_slab is useless.
>>>>> SLAB_MEM_SPREAD does not do anything anymore. SLUB relies on the
>>>>> "spreading" via the page allocator memory policies instead of doing its
>>>>> own like SLAB used to do.
>>>>>
>>>>> What does FILE_SPREAD_SLAB do? Dont see anything there either.
>>>> The FILE_SPREAD_SLAB flag is used by cpuset.memory_spread_slab with read/write operations:
>>>>
>>>> In kernel/cgroup/cpuset.c,
>>>> static struct cftype legacy_files[] = {
>>>> ... snip ...
>>>>           {
>>>>                   .name = "memory_spread_slab",
>>>>                   .read_u64 = cpuset_read_u64,
>>>>                   .write_u64 = cpuset_write_u64,
>>>>                   .private = FILE_SPREAD_SLAB,
>>>>           },
>>>> ... snip ...
>>>> };
>>> It looks like that memory_spread_slab may have effect only on the slab allocator. With the removal of the slab allocator, memory_spread_slab is now a no-op. However, the memory_spread_slab cgroupfs file is an externally visible API. So we can't just remove it as it may break existing applications. We can certainly deprecate it and advise users not to use it.
>> BTW, cpuset doesn't use SLAB_MEM_SPREAD directly. Instead it set the task's PFA_SPREAD_SLAB and let other subsystems test it to act appropriately. Other than cpuset, the latest upstream kernel doesn't check or use this flag at all.
>>
> Ok, get it. So cpuset_do_slab_mem_spread() can be removed, but
> this cpuset file interface and PFA_SPREAD_SLAB will be keeped.

Yes, for now.

Cheers,
Longman