Re: [PATCH 3/3] mm: shmem: override mTHP shmem default with a kernel parameter

["Maíra Canal" <mcanal@igalia.com>]

Hi Barry,

On 28/10/24 19:35, Barry Song wrote:
> On Mon, Oct 28, 2024 at 7:34 PM Maíra Canal <mcanal@igalia.com> wrote:
>>
>> Hi Barry,
>>
>> On 28/10/24 08:09, Barry Song wrote:
>>> On Mon, Oct 28, 2024 at 6:10 PM Maíra Canal <mcanal@igalia.com> wrote:
>>>>
>>>> Hi Barry,
>>>>
>>>> On 27/10/24 18:54, Barry Song wrote:
>>>>> On Mon, Oct 28, 2024 at 6:58 AM Maíra Canal <mcanal@igalia.com> wrote:
>>>>>>
>>>>>> Add the ``thp_shmem=`` kernel command line to allow specifying the
>>>>>> default policy of each supported shmem hugepage size. The kernel parameter
>>>>>> accepts the following format:
>>>>>>
>>>>>> thp_shmem=<size>[KMG],<size>[KMG]:<policy>;<size>[KMG]-<size>[KMG]:<policy>
>>>>>>
>>>>>> For example,
>>>>>>
>>>>>> thp_shmem=16K-64K:always;128K,512K:inherit;256K:advise;1M-2M:never;4M-8M:within_size
>>>>>>
>>>>>> By configuring the default policy of several shmem huge pages, the user
>>>>>> can take advantage of mTHP before it's been configured through sysfs.
>>>>>>
>>>>>> Signed-off-by: Maíra Canal <mcanal@igalia.com>
>>>>>> ---
>>>>>>     .../admin-guide/kernel-parameters.txt         |  10 ++
>>>>>>     Documentation/admin-guide/mm/transhuge.rst    |  17 +++
>>>>>>     mm/shmem.c                                    | 109 +++++++++++++++++-
>>>>>>     3 files changed, 135 insertions(+), 1 deletion(-)
>>>>>>
>>>>>
>>>>> Hi Maíra,
>>>>>
>>>>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
>>>>>> index acabb04d0dd4..595fa096e28b 100644
>>>>>> --- a/Documentation/admin-guide/kernel-parameters.txt
>>>>>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>>>>>> @@ -6700,6 +6700,16 @@
>>>>>>                            Force threading of all interrupt handlers except those
>>>>>>                            marked explicitly IRQF_NO_THREAD.
>>>>>>
>>>>>> +       shmem_anon=     [KNL]
>>>>>> +                       Format: <size>[KMG],<size>[KMG]:<policy>;<size>[KMG]-<size>[KMG]:<policy>
>>>>>> +                       Control the default policy of each hugepage size for the
>>>>>> +                       internal shmem mount. <policy> is one of policies available
>>>>>> +                       for the shmem mount ("always", "inherit", "never", "within_size",
>>>>>> +                       and "advise").
>>>>>> +                       It can be used multiple times for multiple shmem THP sizes.
>>>>>> +                       See Documentation/admin-guide/mm/transhuge.rst for more
>>>>>> +                       details.
>>>>>
>>>>> I'm not sure this is the right name. How about "thp_shmem"?
>>>>
>>>> Oops, sorry about that.
>>>>
>>>>>
>>>>>> +
>>>>>>            topology=       [S390,EARLY]
>>>>>>                            Format: {off | on}
>>>>>>                            Specify if the kernel should make use of the cpu
>>>>>> diff --git a/Documentation/admin-guide/mm/transhuge.rst b/Documentation/admin-guide/mm/transhuge.rst
>>>>>> index 9b5b02c4d1ab..47e7fc30e22d 100644
>>>>>> --- a/Documentation/admin-guide/mm/transhuge.rst
>>>>>> +++ b/Documentation/admin-guide/mm/transhuge.rst
>>>>>> @@ -332,6 +332,23 @@ allocation policy for the internal shmem mount by using the kernel parameter
>>>>>>     seven valid policies for shmem (``always``, ``within_size``, ``advise``,
>>>>>>     ``never``, ``deny``, and ``force``).
>>>>>>
>>>>>> +In the same manner as ``thp_anon`` controls each supported anonymous THP
>>>>>> +size, ``thp_shmem`` controls each supported shmem THP size. ``thp_shmem``
>>>>>> +has the same format as ``thp_anon``, but also supports the policy
>>>>>> +``within_size``.
>>>>>> +
>>>>>> +``thp_shmem=`` may be specified multiple times to configure all THP sizes
>>>>>> +as required. If ``thp_shmem=`` is specified at least once, any shmem THP
>>>>>> +sizes not explicitly configured on the command line are implicitly set to
>>>>>> +``never``.
>>>>>> +
>>>>>> +``transparent_hugepage_shmem`` setting only affects the global toggle. If
>>>>>> +``thp_shmem`` is not specified, PMD_ORDER hugepage will default to
>>>>>> +``inherit``. However, if a valid ``thp_shmem`` setting is provided by the
>>>>>> +user, the PMD_ORDER hugepage policy will be overridden. If the policy for
>>>>>> +PMD_ORDER is not defined within a valid ``thp_shmem``, its policy will
>>>>>> +default to ``never``.
>>>>>> +
>>>>>>     Hugepages in tmpfs/shmem
>>>>>>     ========================
>>>>>>
>>>>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>>>>> index 24cdeafd8260..0a7a7d04f725 100644
>>>>>> --- a/mm/shmem.c
>>>>>> +++ b/mm/shmem.c
>>
>> [...]
>>
>>>>>>     static int __init setup_transparent_hugepage_shmem(char *str)
>>>>>>     {
>>>>>>            int huge, ret = 0;
>>>>>> @@ -5206,6 +5228,91 @@ static int __init setup_transparent_hugepage_shmem(char *str)
>>>>>>     }
>>>>>>     __setup("transparent_hugepage_shmem=", setup_transparent_hugepage_shmem);
>>>>>>
>>>>>> +static char str_dup[PAGE_SIZE] __initdata;
>>>>>> +static int __init setup_thp_shmem(char *str)
>>>>>> +{
>>>>>> +       char *token, *range, *policy, *subtoken;
>>>>>> +       unsigned long always, inherit, madvise, within_size;
>>>>>> +       char *start_size, *end_size;
>>>>>> +       int start, end, nr;
>>>>>> +       char *p;
>>>>>> +
>>>>>> +       if (!str || strlen(str) + 1 > PAGE_SIZE)
>>>>>> +               goto err;
>>>>>> +       strcpy(str_dup, str);
>>>>>> +
>>>>>> +       always = huge_shmem_orders_always;
>>>>>> +       inherit = huge_shmem_orders_inherit;
>>>>>> +       madvise = huge_shmem_orders_madvise;
>>>>>> +       within_size = huge_shmem_orders_within_size;
>>>>>> +       p = str_dup;
>>>>>> +       while ((token = strsep(&p, ";")) != NULL) {
>>>>>> +               range = strsep(&token, ":");
>>>>>> +               policy = token;
>>>>>> +
>>>>>> +               if (!policy)
>>>>>> +                       goto err;
>>>>>> +
>>>>>> +               while ((subtoken = strsep(&range, ",")) != NULL) {
>>>>>> +                       if (strchr(subtoken, '-')) {
>>>>>> +                               start_size = strsep(&subtoken, "-");
>>>>>> +                               end_size = subtoken;
>>>>>> +
>>>>>> +                               start = get_order_from_str(start_size);
>>>>>> +                               end = get_order_from_str(end_size);
>>>>>> +                       } else {
>>>>>> +                               start = end = get_order_from_str(subtoken);
>>>>>> +                       }
>>>>>> +
>>>>>> +                       if (start < 0 || end < 0 || start > end)
>>>>>> +                               goto err;
>>>>>> +
>>>>>> +                       nr = end - start + 1;
>>>>>> +                       if (!strcmp(policy, "always")) {
>>>>>> +                               bitmap_set(&always, start, nr);
>>>>>> +                               bitmap_clear(&inherit, start, nr);
>>>>>> +                               bitmap_clear(&madvise, start, nr);
>>>>>> +                               bitmap_clear(&within_size, start, nr);
>>>>>> +                       } else if (!strcmp(policy, "advise")) {
>>>>>> +                               bitmap_set(&madvise, start, nr);
>>>>>> +                               bitmap_clear(&inherit, start, nr);
>>>>>> +                               bitmap_clear(&always, start, nr);
>>>>>> +                               bitmap_clear(&within_size, start, nr);
>>>>>> +                       } else if (!strcmp(policy, "inherit")) {
>>>>>> +                               bitmap_set(&inherit, start, nr);
>>>>>> +                               bitmap_clear(&madvise, start, nr);
>>>>>> +                               bitmap_clear(&always, start, nr);
>>>>>> +                               bitmap_clear(&within_size, start, nr);
>>>>>> +                       } else if (!strcmp(policy, "within_size")) {
>>>>>> +                               bitmap_set(&within_size, start, nr);
>>>>>> +                               bitmap_clear(&inherit, start, nr);
>>>>>> +                               bitmap_clear(&madvise, start, nr);
>>>>>> +                               bitmap_clear(&always, start, nr);
>>>>>> +                       } else if (!strcmp(policy, "never")) {
>>>>>> +                               bitmap_clear(&inherit, start, nr);
>>>>>> +                               bitmap_clear(&madvise, start, nr);
>>>>>> +                               bitmap_clear(&always, start, nr);
>>>>>> +                               bitmap_clear(&within_size, start, nr);
>>>>>> +                       } else {
>>>>>> +                               pr_err("invalid policy %s in thp_shmem boot parameter\n", policy);
>>>>>> +                               goto err;
>>>>>> +                       }
>>>>>> +               }
>>>>>> +       }
>>>>>> +
>>>>>> +       huge_shmem_orders_always = always;
>>>>>> +       huge_shmem_orders_madvise = madvise;
>>>>>> +       huge_shmem_orders_inherit = inherit;
>>>>>> +       huge_shmem_orders_within_size = within_size;
>>>>>> +       shmem_orders_configured = true;
>>>>>> +       return 1;
>>>>>> +
>>>>>> +err:
>>>>>> +       pr_warn("thp_shmem=%s: error parsing string, ignoring setting\n", str);
>>>>>> +       return 0;
>>>>>> +}
>>>>>
>>>>> Can we share source code with thp_anon since there's a lot of duplication?
>>>>
>>>> I'm not a regular mm contributor and I'm most usually around drivers, so
>>>> I don't know exactly here I could add shared code. Should I add the
>>>> headers to "internal.h"?
>>>
>>> My comment isn't related to drivers or memory management. It's solely about
>>> avoiding code duplication. For example, we could create a shared function to
>>> handle both controls, reducing redundant code :-)
>>
>> Let me rephrase it.
>>
>> I completely agree that we should avoid code duplication. I'm asking
>> where is the best place to add the headers of the shared functions.
>> "linux/shmem_fs.h" doesn't look appropriate to me, so I believe the
>> remaining options would be "linux/huge_mm.h" or "internal.h".
> 
> Both locations seem quite odd. I have a feeling that these boot command
> elements are purely internal, yet internal.h contains something that is
> actually 'external' to mm. The shared code isn't 'external' enough to belong
> in internal.h.
> 
> I didn't realize that shmem has placed these controls in its own file;
> I thought they
> were also located in mm/huge_memory.c. Given the current situation, I would
> prefer to keep the code as it is and tolerate the code duplication.
> 
> Unless we are going to place controls for shmem and other thp controls in
> one place, I feel your code is better than having a shared function either in
> internal.h or linux/huge_mm.h.

Sorry, I only catch your e-mail after sending v2. If possible, please,
take a look on v2 [1] and let me know if you still prefer to duplicate
the code.

[1] 
https://lore.kernel.org/linux-mm/20241029002324.1062723-1-mcanal@igalia.com/T/

Best Regards,
- Maíra

> 
>>
>> I would like to know your opinion about those two options.
>>
>> Best Regards,
>> - Maíra
>>
>>>
>>>>
>>>> Best Regards,
>>>> - Maíra
>>>>
>>>>>
>>>>>> +__setup("thp_shmem=", setup_thp_shmem);
>>>>>> +
>>>>>>     #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
>>>>>>
>>>>>>     #else /* !CONFIG_SHMEM */
>>>>>> --
>>>>>> 2.46.2
>>>>>>
>>>>>
> 
> Thanks
> barry