Re: [PATCH 4/4] mm: Make most walk page paths with pmd_trans_unstable() to retry

From: Yang Shi
Date: Mon Jun 05 2023 - 14:46:29 EST

On Fri, Jun 2, 2023 at 4:06 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> For most of the page walk paths, logically it'll always be good to have the
> pmd retries if hit pmd_trans_unstable() race. We can treat it as none
> pmd (per comment above pmd_trans_unstable()), but in most cases we're not
> even treating that as a none pmd. If to fix it anyway, a retry will be the
> most accurate.
>
> I've went over all the pmd_trans_unstable() special cases and this patch
> should cover all the rest places where we should retry properly with
> unstable pmd. With the newly introduced ACTION_AGAIN since 2020 we can
> easily achieve that.
>
> These are the call sites that I think should be fixed with it:
>
> *** fs/proc/task_mmu.c:
> smaps_pte_range[634] if (pmd_trans_unstable(pmd))
> clear_refs_pte_range[1194] if (pmd_trans_unstable(pmd))
> pagemap_pmd_range[1542] if (pmd_trans_unstable(pmdp))
> gather_pte_stats[1891] if (pmd_trans_unstable(pmd))
> *** mm/memcontrol.c:
> mem_cgroup_count_precharge_pte_range[6024] if (pmd_trans_unstable(pmd))
> mem_cgroup_move_charge_pte_range[6244] if (pmd_trans_unstable(pmd))
> *** mm/memory-failure.c:
> hwpoison_pte_range[794] if (pmd_trans_unstable(pmdp))
> *** mm/mempolicy.c:
> queue_folios_pte_range[517] if (pmd_trans_unstable(pmd))
> *** mm/madvise.c:
> madvise_cold_or_pageout_pte_range[425] if (pmd_trans_unstable(pmd))
> madvise_free_pte_range[625] if (pmd_trans_unstable(pmd))
>
> IIUC most of them may or may not be a big issue even without a retry,
> either because they're already not strict (smaps, pte_stats, MADV_COLD,
> .. it can mean e.g. the statistic may be inaccurate or one less 2M chunk to
> cold worst case), but some of them could have functional error without the
> retry afaiu (e.g. pagemap, where we can have the output buffer shifted over
> the unstable pmd range.. so IIUC the pagemap result can be wrong).
>
> While these call sites all look fine, and don't need any change:
>
> *** include/linux/pgtable.h:
> pmd_devmap_trans_unstable[1418] return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
> *** mm/gup.c:
> follow_pmd_mask[695] if (pmd_trans_unstable(pmd))
> *** mm/mapping_dirty_helpers.c:
> wp_clean_pmd_entry[131] if (!pmd_trans_unstable(&pmdval))
> *** mm/memory.c:
> do_anonymous_page[4060] if (unlikely(pmd_trans_unstable(vmf->pmd)))
> *** mm/migrate_device.c:
> migrate_vma_insert_page[616] if (unlikely(pmd_trans_unstable(pmdp)))
> *** mm/mincore.c:
> mincore_pte_range[116] if (pmd_trans_unstable(pmd)) {
>
> Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> ---
> fs/proc/task_mmu.c | 17 +++++++++++++----
> mm/madvise.c | 8 ++++++--
> mm/memcontrol.c | 8 ++++++--
> mm/memory-failure.c | 4 +++-
> mm/mempolicy.c | 4 +++-
> 5 files changed, 31 insertions(+), 10 deletions(-)
>
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index 6259dd432eeb..823eaba5c6bf 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -631,8 +631,11 @@ static int smaps_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
> goto out;
> }
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> goto out;
> + }
> +
> /*
> * The mmap_lock held all the way back in m_start() is what
> * keeps khugepaged out of here and from collapsing things
> @@ -1191,8 +1194,10 @@ static int clear_refs_pte_range(pmd_t *pmd, unsigned long addr,
> return 0;
> }
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }
>
> pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> for (; addr != end; pte++, addr += PAGE_SIZE) {
> @@ -1539,8 +1544,10 @@ static int pagemap_pmd_range(pmd_t *pmdp, unsigned long addr, unsigned long end,
> return err;
> }
>
> - if (pmd_trans_unstable(pmdp))
> + if (pmd_trans_unstable(pmdp)) {
> + walk->action = ACTION_AGAIN;
> return 0;

Had a quick look at the pagemap code, I agree with your analysis,
"returning 0" may mess up pagemap, retry should be fine. But I'm
wondering whether we should just fill in empty entries. Anyway I don't
have a strong opinion on this, just a little bit concerned by
potential indefinite retry.

> + }
> #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
>
> /*
> @@ -1888,8 +1895,10 @@ static int gather_pte_stats(pmd_t *pmd, unsigned long addr,
> return 0;
> }
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }
> #endif
> orig_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
> do {
> diff --git a/mm/madvise.c b/mm/madvise.c
> index 78cd12581628..0fd81712022c 100644
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -424,8 +424,10 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
> }
>
> regular_folio:
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }
> #endif
> tlb_change_page_size(tlb, PAGE_SIZE);
> orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> @@ -626,8 +628,10 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
> if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
> goto next;
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }
>
> tlb_change_page_size(tlb, PAGE_SIZE);
> orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 6ee433be4c3b..15e50f033e41 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -6021,8 +6021,10 @@ static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
> return 0;
> }
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;

Either retry or keep as is is fine to me. I'm not aware of anyone
complaining about noticeable inaccurate charges due to this. But we
may have potential indefinite retry anyway, so if you don't want to
risk this, it may be better just keep it as is.

> + }
> pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> for (; addr != end; pte++, addr += PAGE_SIZE)
> if (get_mctgt_type(vma, addr, *pte, NULL))
> @@ -6241,8 +6243,10 @@ static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
> return 0;
> }
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }
> retry:
> pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> for (; addr != end; addr += PAGE_SIZE) {
> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> index 004a02f44271..c97fb2b7ab4a 100644
> --- a/mm/memory-failure.c
> +++ b/mm/memory-failure.c
> @@ -791,8 +791,10 @@ static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
> goto out;
> }
>
> - if (pmd_trans_unstable(pmdp))
> + if (pmd_trans_unstable(pmdp)) {
> + walk->action = ACTION_AGAIN;
> goto out;
> + }

This may be worth retrying IMHO.

>
> mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
> addr, &ptl);
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> index f06ca8c18e62..af8907b4aad1 100644
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -514,8 +514,10 @@ static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
> if (ptl)
> return queue_folios_pmd(pmd, ptl, addr, end, walk);
>
> - if (pmd_trans_unstable(pmd))
> + if (pmd_trans_unstable(pmd)) {
> + walk->action = ACTION_AGAIN;
> return 0;
> + }

Either retry or keep it as is is fine.

>
> mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
> for (; addr != end; pte++, addr += PAGE_SIZE) {
> --
> 2.40.1
>
>