Message-ID: <CAJuCfpEp_3Dz32fdpUaUbPaP6KZ+0fXmXBvBV1jRt9Q+LMRQAQ@mail.gmail.com>
Date: Fri, 8 Nov 2024 10:16:35 -0800
From: Suren Baghdasaryan <surenb@...gle.com>
To: Hao Ge <hao.ge@...ux.dev>
Cc: akpm@...ux-foundation.org, kent.overstreet@...ux.dev, 
	linux-kernel@...r.kernel.org, linux-mm@...ck.org, Hao Ge <gehao@...inos.cn>
Subject: Re: [PATCH] lib/alloc_tag: Remove the sysctl configuration to prevent
 users from disabling it at runtime

On Thu, Nov 7, 2024 at 11:50 PM Hao Ge <hao.ge@...ux.dev> wrote:
>
> From: Hao Ge <gehao@...inos.cn>
>
> After much consideration,I have decided to remove
> the "mem_profiling" sysctl interface to prevent
> users from dynamically enabling or disabling the
> MEMORY ALLOCATION PROFILING feature at runtime.
>
> I have taken the following actions: I set
> CONFIG_MEM_ALLOC_PROFILING_ENABLED_BY_DEFAULT=y to
> enable memory allocation profiling by default,
> and then made adjustments to mem_profiling dynamically
> during runtime.
>
> When I ran the OOM test program, I obtained useful
> information that was indeed very helpful for debugging.
>
> [ 1023.065402] Memory allocations:
> [ 1023.065407]     12.8 GiB     6546 mm/huge_memory.c:1328 func:do_huge_pmd_anonymous_page
> [ 1023.065412]      873 MiB   229985 arch/arm64/mm/fault.c:986 func:vma_alloc_zeroed_movable_folio
> [ 1023.065415]      187 MiB    29732 mm/slub.c:2412 func:alloc_slab_page
> [ 1023.065418]     99.8 MiB    25560 mm/memory.c:1065 func:folio_prealloc
> [ 1023.065421]     47.2 MiB     3189 mm/readahead.c:434 func:ra_alloc_folio
> [ 1023.065424]     30.0 MiB       15 mm/khugepaged.c:1072 func:alloc_charge_folio
> [ 1023.065428]     28.6 MiB      514 mm/compaction.c:1880 func:compaction_alloc
> [ 1023.065430]     25.8 MiB     6592 mm/page_ext.c:271 func:alloc_page_ext
> [ 1023.065433]     25.6 MiB     6546 mm/huge_memory.c:1161 func:__do_huge_pmd_anonymous_page
> [ 1023.065436]     23.5 MiB     6017 mm/shmem.c:1771 func:shmem_alloc_folio
>
> After running echo 0 > /proc/sys/vm/mem_profiling
> and then executing the same test program,
> I obtained the following results
>
> [ 1156.509699] Memory allocations:
> [ 1156.509703]      187 MiB    29645 mm/slub.c:2412 func:alloc_slab_page
> [ 1156.509707]      142 MiB     9357 mm/readahead.c:434 func:ra_alloc_folio
> [ 1156.509710]      136 MiB    41325 arch/arm64/mm/fault.c:986 func:vma_alloc_zeroed_movable_folio
> [ 1156.509713]     99.7 MiB    25531 mm/memory.c:1065 func:folio_prealloc
> [ 1156.509716]     56.0 MiB       28 mm/huge_memory.c:1328 func:do_huge_pmd_anonymous_page
> [ 1156.509719]     30.0 MiB       15 mm/khugepaged.c:1072 func:alloc_charge_folio
> [ 1156.509723]     28.6 MiB      514 mm/compaction.c:1880 func:compaction_alloc
> [ 1156.509725]     26.3 MiB     7460 mm/readahead.c:264 func:page_cache_ra_unbounded
> [ 1156.509728]     25.8 MiB     6592 mm/page_ext.c:271 func:alloc_page_ext
> [ 1156.509730]     23.5 MiB     6016 mm/shmem.c:1771 func:shmem_alloc_folio
>
> Because mem_profiling was disabled by executing
> echo 0 > /proc/sys/vm/mem_profiling,we are unable to
> record memory allocation information after the disablement.

Naturally you are unable to track the allocations after disabling it.
You disabled it as root, so I assume you know what you are doing.

> These output logs can mislead users. And similarly, the same
> applies to alloc_info.

I would understand if you made /proc/allocinfo empty after disabling
it to avoid confusing the user, but ripping out the ability to
enable/disable profiling at runtime does not make sense to me. Once
you collect required data, disabling profiling gets you back the
performance that you pay for it. There are usecases when a program on
a remote device periodically enables profiling for some time, records
the difference in allocations and then disables it. Your change breaks
such users.

>
> We already have boot parameters that allow users to
> choose whether to enable or disable.
> In order to maintain the accuracy of memory allocation
> information,I have decided to remove the runtime switch.

Well, I disagree with your decision.
NAK.

>
> Signed-off-by: Hao Ge <gehao@...inos.cn>
> ---
>  lib/alloc_tag.c | 26 --------------------------
>  1 file changed, 26 deletions(-)
>
> diff --git a/lib/alloc_tag.c b/lib/alloc_tag.c
> index 81e5f9a70f22..47fa969c23f3 100644
> --- a/lib/alloc_tag.c
> +++ b/lib/alloc_tag.c
> @@ -227,31 +227,6 @@ struct page_ext_operations page_alloc_tagging_ops = {
>  };
>  EXPORT_SYMBOL(page_alloc_tagging_ops);
>
> -#ifdef CONFIG_SYSCTL
> -static struct ctl_table memory_allocation_profiling_sysctls[] = {
> -       {
> -               .procname       = "mem_profiling",
> -               .data           = &mem_alloc_profiling_key,
> -#ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG
> -               .mode           = 0444,
> -#else
> -               .mode           = 0644,
> -#endif
> -               .proc_handler   = proc_do_static_key,
> -       },
> -};
> -
> -static void __init sysctl_init(void)
> -{
> -       if (!mem_profiling_support)
> -               memory_allocation_profiling_sysctls[0].mode = 0444;
> -
> -       register_sysctl_init("vm", memory_allocation_profiling_sysctls);
> -}
> -#else /* CONFIG_SYSCTL */
> -static inline void sysctl_init(void) {}
> -#endif /* CONFIG_SYSCTL */
> -
>  static int __init alloc_tag_init(void)
>  {
>         const struct codetag_type_desc desc = {
> @@ -264,7 +239,6 @@ static int __init alloc_tag_init(void)
>         if (IS_ERR(alloc_tag_cttype))
>                 return PTR_ERR(alloc_tag_cttype);
>
> -       sysctl_init();
>         procfs_init();
>
>         return 0;
> --
> 2.25.1
>

Powered by blists - more mailing lists