On 4/12/22 5:30 PM, Oscar Salvador wrote:
On Tue, Apr 12, 2022 at 05:25:52PM +0800, Miaohe Lin wrote:
On 2022/4/12 16:31, Oscar Salvador wrote:
On Sun, Apr 10, 2022 at 11:22:34PM +0800, Xu Yu wrote:
Kernel panic when injecting memory_failure for the global huge_zero_page,
when CONFIG_DEBUG_VM is enabled, as follows.
...
In fact, huge_zero_page is unhandlable currently in either soft offline
or memory failure injection. With CONFIG_DEBUG_VM disabled,
huge_zero_page is bailed out when checking HWPoisonHandlable() in
get_any_page(), or checking page mapping in split_huge_page_to_list().
This makes huge_zero_page bail out early in madvise_inject_error(), and
panic above won't happen again.
I would not special case this in madvise_inject_error() but rather
handle it in memory-failure code.
We do already have HWPoisonHandlable(), which tells us whether the page
is of a type we can really do something about, so why not add another
check in HWPoisonHandlable() for huge_zero_page(), and have that checked
in memory_failure().
IIUC, this does not work. Because HWPoisonHandlable is only called in !MF_COUNT_INCREASED case.
But MF_COUNT_INCREASED is always set when called from madvise_inject_error, so HWPoisonHandlable
is not even called in this scene. Or am I miss something?
But nothing stops you from calling it in memory_failure(), right?
if (MF_COUNT_INCREASED not set) {
....
...
} else if(!HWPoisonHandable(p)) {
action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
res = -EBUSY;
goto unlock_mutex;
}
BTW: IIRC, LRU isn't set on huge_zero_page. So the origin HWPoisonHandlable can already filter out this page.
I would rather have it as a explicit check than buried in that kind of
assumption.
And this is also why I decided to bail out huge zero page early for both
soft offline and memory failure.
But after all, Naoya's suggestion might just be better and more focused.
Agree, thanks!
--
Thanks,
Yu
