lkml.org 
[lkml]   [2019]   [Dec]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v16 11/25] mm: pagewalk: Add p4d_entry() and pgd_entry()
From
Date
On 12/12/19 2:15 PM, Steven Price wrote:
> On 12/12/2019 11:33, Thomas Hellström (VMware) wrote:
>> On 12/12/19 12:23 PM, Thomas Hellström (VMware) wrote:
>>> On 12/6/19 2:53 PM, Steven Price wrote:
>>>> pgd_entry() and pud_entry() were removed by commit 0b1fbfe50006c410
>>>> ("mm/pagewalk: remove pgd_entry() and pud_entry()") because there were
>>>> no users. We're about to add users so reintroduce them, along with
>>>> p4d_entry() as we now have 5 levels of tables.
>>>>
>>>> Note that commit a00cc7d9dd93d66a ("mm, x86: add support for
>>>> PUD-sized transparent hugepages") already re-added pud_entry() but
>>>> with
>>>> different semantics to the other callbacks. Since there have never
>>>> been upstream users of this, revert the semantics back to match the
>>>> other callbacks. This means pud_entry() is called for all entries, not
>>>> just transparent huge pages.
>
> When I wrote that there were no upstream users, which sadly shows how
> long ago that was :(
>
>>> Actually, there are two users of pud_entry(), in hmm.c and since
>>> 5.5rc1 also mapping_dirty_helpers.c. The latter one is unproblematic
>>> and requires no attention but the one in hmm.c is probably largely
>>> untested, and seems to assume it was called outside of the spinlock.
>>>
>>> The problem with the current patch is that the hmm pud_entry will
>>> traverse also pmds, so that will be done twice now.
>>>
>>> In another thread we were discussing a means of rerunning the level
>>> (in case of a race), or continuing after a level, based on the
>>> return value after the callback. The change was fairly invasive,
>>>
>> Hmm. Forgot to remove the above text that appears twice. :(. The
>> correct one is inline below.
>>
>>>
>>>> Tested-by: Zong Li <zong.li@sifive.com>
>>>> Signed-off-by: Steven Price <steven.price@arm.com>
>>>> ---
>>>>   include/linux/pagewalk.h | 19 +++++++++++++------
>>>>   mm/pagewalk.c            | 27 ++++++++++++++++-----------
>>>>   2 files changed, 29 insertions(+), 17 deletions(-)
>>>>
>>>> diff --git a/include/linux/pagewalk.h b/include/linux/pagewalk.h
>>>> index 6ec82e92c87f..06790f23957f 100644
>>>> --- a/include/linux/pagewalk.h
>>>> +++ b/include/linux/pagewalk.h
>>>> @@ -8,15 +8,15 @@ struct mm_walk;
>>>>     /**
>>>>    * mm_walk_ops - callbacks for walk_page_range
>>>> - * @pud_entry:        if set, called for each non-empty PUD
>>>> (2nd-level) entry
>>>> - *            this handler should only handle pud_trans_huge() puds.
>>>> - *            the pmd_entry or pte_entry callbacks will be used for
>>>> - *            regular PUDs.
>>>> - * @pmd_entry:        if set, called for each non-empty PMD
>>>> (3rd-level) entry
>>>> + * @pgd_entry:        if set, called for each non-empty PGD
>>>> (top-level) entry
>>>> + * @p4d_entry:        if set, called for each non-empty P4D entry
>>>> + * @pud_entry:        if set, called for each non-empty PUD entry
>>>> + * @pmd_entry:        if set, called for each non-empty PMD entry
>>>>    *            this handler is required to be able to handle
>>>>    *            pmd_trans_huge() pmds.  They may simply choose to
>>>>    *            split_huge_page() instead of handling it explicitly.
>>>> - * @pte_entry:        if set, called for each non-empty PTE
>>>> (4th-level) entry
>>>> + * @pte_entry:        if set, called for each non-empty PTE
>>>> (lowest-level)
>>>> + *            entry
>>>>    * @pte_hole:        if set, called for each hole at all levels
>>>>    * @hugetlb_entry:    if set, called for each hugetlb entry
>>>>    * @test_walk:        caller specific callback function to
>>>> determine whether
>>>> @@ -27,8 +27,15 @@ struct mm_walk;
>>>>    * @pre_vma:            if set, called before starting walk on a
>>>> non-null vma.
>>>>    * @post_vma:           if set, called after a walk on a non-null
>>>> vma, provided
>>>>    *                      that @pre_vma and the vma walk succeeded.
>>>> + *
>>>> + * p?d_entry callbacks are called even if those levels are folded
>>>> on a
>>>> + * particular architecture/configuration.
>>>>    */
>>>>   struct mm_walk_ops {
>>>> +    int (*pgd_entry)(pgd_t *pgd, unsigned long addr,
>>>> +             unsigned long next, struct mm_walk *walk);
>>>> +    int (*p4d_entry)(p4d_t *p4d, unsigned long addr,
>>>> +             unsigned long next, struct mm_walk *walk);
>>>>       int (*pud_entry)(pud_t *pud, unsigned long addr,
>>>>                unsigned long next, struct mm_walk *walk);
>>>>       int (*pmd_entry)(pmd_t *pmd, unsigned long addr,
>>>> diff --git a/mm/pagewalk.c b/mm/pagewalk.c
>>>> index ea0b9e606ad1..c089786e7a7f 100644
>>>> --- a/mm/pagewalk.c
>>>> +++ b/mm/pagewalk.c
>>>> @@ -94,15 +94,9 @@ static int walk_pud_range(p4d_t *p4d, unsigned
>>>> long addr, unsigned long end,
>>>>           }
>>>>             if (ops->pud_entry) {
>>>> -            spinlock_t *ptl = pud_trans_huge_lock(pud, walk->vma);
>>>> -
>>>> -            if (ptl) {
>>>> -                err = ops->pud_entry(pud, addr, next, walk);
>>>> -                spin_unlock(ptl);
>>>> -                if (err)
>>>> -                    break;
>>>> -                continue;
>>>> -            }
>>>> +            err = ops->pud_entry(pud, addr, next, walk);
>>>> +            if (err)
>>>> +                break;
>>>
>>> Actually, there are two current users of pud_entry(), in hmm.c and
>>> since 5.5rc1 also mapping_dirty_helpers.c. The latter one is
>>> unproblematic and requires no attention but the one in hmm.c is
>>> probably largely untested, and seems to assume it was called outside
>>> of the spinlock.
>
> Thanks for pointing that out, I guess the simplest fix would be to
> squash in something like the below which should restore the old
> behaviour for hmm.c without affecting others.
>
> Steve

I'm not fully sure that the old behaviour is the correct one, but definitely hmm's pud_entry needs some fixing.
I'm more concerned with the pagewalk code. With your patch it actually splits all huge puds present in the page-table
on each page walk which is not what we want.

One idea would be to add a new member to struct_mm_walk:

enum page_walk_ret_action {
ACTION_SUBTREE = 0,
ACTION_CONTINUE = 1,
ACTION_AGAIN = 2 /* Only for levels that thave p?d_unstable */
};

struct mm_walk {
...
enum page_walk_ret_action action; /* or perhaps as an enum */
};


if (ops->pud_entry) {
walk->action = ACTION_SUBTREE;
...
...
...
if (walk->action == ACTION_AGAIN) /* Callback tried to split huge entry, but failed */
goto again;
else if (walk->action == ACTION_CONTINUE) /* Done with this subtree. Probably huge entry handled. */
continue;
/* ACTION_SUBTREE falls through */
}

we discussed something similar before on linux-mm, but the idea then was to redefine
the positive return value of the callback to the action, but that meant changing those existing callbacks that relied on
a positive return value. The above would be helpful also for pmd_entry.

/Thomas




\
 
 \ /
  Last update: 2019-12-12 15:05    [W:0.259 / U:0.164 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site