[bug report] mm/mglru: fix underprotected page cache

[bug report] mm/mglru: fix underprotected page cache

[Dan Carpenter]

Hello Yu Zhao,

The patch 081488051d28: "mm/mglru: fix underprotected page cache"
from Dec 7, 2023 (linux-next), leads to the following Smatch static
checker warning:

	mm/workingset.c:318 lru_gen_refault()
	warn: always true condition '(refs >= ((((1))) << (__builtin_choose_expr((4 == 4), (((2) < (64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))) ?(2):(64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))), ))) - 1) => (s32min-s32max >= 0)'

mm/workingset.c
    308         atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
    309         mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
    310 
    311         /*
    312          * Count the following two cases as stalls:
    313          * 1. For pages accessed through page tables, hotter pages pushed out
    314          *    hot pages which refaulted immediately.
    315          * 2. For pages accessed multiple times through file descriptors,
    316          *    they would have been protected by sort_folio().
    317          */
--> 318         if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) {
                                                  ^^^^^^^^^^^^^^^^^^^^^^^
In my x86 allmodconfig this is zero and refs is always >= 0.  This could
be intentional, but I thought it was worth checking.

    319                 set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset));
    320                 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta);
    321         }
    322 unlock:
    323         rcu_read_unlock();
    324 }

regards,
dan carpenter

Re: [bug report] mm/mglru: fix underprotected page cache

[Yu Zhao]

On Fri, Jan 12, 2024 at 12:13 AM Dan Carpenter <dan.carpenter@linaro.org> wrote:
>
> Hello Yu Zhao,
>
> The patch 081488051d28: "mm/mglru: fix underprotected page cache"
> from Dec 7, 2023 (linux-next), leads to the following Smatch static
> checker warning:
>
>         mm/workingset.c:318 lru_gen_refault()
>         warn: always true condition '(refs >= ((((1))) << (__builtin_choose_expr((4 == 4), (((2) < (64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))) ?(2):(64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))), ))) - 1) => (s32min-s32max >= 0)'
>
> mm/workingset.c
>     308         atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
>     309         mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
>     310
>     311         /*
>     312          * Count the following two cases as stalls:
>     313          * 1. For pages accessed through page tables, hotter pages pushed out
>     314          *    hot pages which refaulted immediately.
>     315          * 2. For pages accessed multiple times through file descriptors,
>     316          *    they would have been protected by sort_folio().
>     317          */
> --> 318         if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) {
>                                                   ^^^^^^^^^^^^^^^^^^^^^^^
> In my x86 allmodconfig this is zero and refs is always >= 0.  This could
> be intentional, but I thought it was worth checking.

Appreciate the manual work!

This indeed is intentional if LRU_REFS_WIDTH is 0. What would be the
best way to suppress this warning?

Re: [bug report] mm/mglru: fix underprotected page cache

[Dan Carpenter]

On Fri, Jan 12, 2024 at 12:42:37AM -0700, Yu Zhao wrote:
> On Fri, Jan 12, 2024 at 12:13 AM Dan Carpenter <dan.carpenter@linaro.org> wrote:
> >
> > Hello Yu Zhao,
> >
> > The patch 081488051d28: "mm/mglru: fix underprotected page cache"
> > from Dec 7, 2023 (linux-next), leads to the following Smatch static
> > checker warning:
> >
> >         mm/workingset.c:318 lru_gen_refault()
> >         warn: always true condition '(refs >= ((((1))) << (__builtin_choose_expr((4 == 4), (((2) < (64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))) ?(2):(64 - 26 - 3 - 3 - 0 - 10 - 0 - (8 + 14))), ))) - 1) => (s32min-s32max >= 0)'
> >
> > mm/workingset.c
> >     308         atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
> >     309         mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
> >     310
> >     311         /*
> >     312          * Count the following two cases as stalls:
> >     313          * 1. For pages accessed through page tables, hotter pages pushed out
> >     314          *    hot pages which refaulted immediately.
> >     315          * 2. For pages accessed multiple times through file descriptors,
> >     316          *    they would have been protected by sort_folio().
> >     317          */
> > --> 318         if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) {
> >                                                   ^^^^^^^^^^^^^^^^^^^^^^^
> > In my x86 allmodconfig this is zero and refs is always >= 0.  This could
> > be intentional, but I thought it was worth checking.
> 
> Appreciate the manual work!
> 
> This indeed is intentional if LRU_REFS_WIDTH is 0. What would be the
> best way to suppress this warning?

No need.

These warnings are a one time thing.  Ignore the false positives.  If
people are curious in the future they can find this conversation on
lore.

regards,
dan carpenter