> On Jan 20, 2022, at 12:37 PM, David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 20.01.22 21:09, David Hildenbrand wrote:
>> On 20.01.22 21:07, Matthew Wilcox wrote:
>>> On Thu, Jan 20, 2022 at 08:55:12PM +0100, David Hildenbrand wrote:
>>>>>>> David, does any of it regards the lru_cache_add() reference issue that I
>>>>>>> mentioned? [1]
>>>
>>>> +++ b/mm/memory.c
>>>> @@ -3291,19 +3291,28 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
>>>> if (PageAnon(vmf->page)) {
>>>> struct page *page = vmf->page;
>>>>
>>>> - /* PageKsm() doesn't necessarily raise the page refcount */
>>>> - if (PageKsm(page) || page_count(page) != 1)
>>>> + /*
>>>> + * PageKsm() doesn't necessarily raise the page refcount.
>>>> + *
>>>> + * These checks are racy as long as we haven't locked the page;
>>>> + * they are a pure optimization to avoid trying to lock the page
>>>> + * and trying to free the swap cache when there is little hope
>>>> + * it will actually result in a refcount of 1.
>>>> + */
>>>> + if (PageKsm(page) || page_count(page) > 1 + PageSwapCache(page))
>>>> goto copy;
>>>> if (!trylock_page(page))
>>>> goto copy;
>>>> - if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
>>>> + if (PageSwapCache(page))
>>>> + try_to_free_swap(page);
>>>> + if (PageKsm(page) || page_count(page) != 1) {
>>>> unlock_page(page);
>>>> goto copy;
>>>> }
>>>> /*
>>>> - * Ok, we've got the only map reference, and the only
>>>> - * page count reference, and the page is locked,
>>>> - * it's dark out, and we're wearing sunglasses. Hit it.
>>>> + * Ok, we've got the only page reference from our mapping
>>>> + * and the page is locked, it's dark out, and we're wearing
>>>> + * sunglasses. Hit it.
>>>> */
>>>> unlock_page(page);
>>>> wp_page_reuse(vmf);
>>>>
>>>>
>>>> I added some vmstats that monitor various paths. After one run of
>>>> ./forceswap 2 1000000 1
>>>> I'm left with a rough delta (including some noise) of
>>>> anon_wp_copy_count 1799
>>>> anon_wp_copy_count_early 1
>>>> anon_wp_copy_lock 983396
>>>> anon_wp_reuse 0
>>>>
>>>> The relevant part of your reproducer is
>>>>
>>>> for (i = 0; i < nops; i++) {
>>>> if (madvise((void *)p, PAGE_SIZE * npages, MADV_PAGEOUT)) {
>>>> perror("madvise");
>>>> exit(-1);
>>>> }
>>>>
>>>> for (j = 0; j < npages; j++) {
>>>> c = p[j * PAGE_SIZE];
>>>> c++;
>>>> time -= rdtscp();
>>>> p[j * PAGE_SIZE] = c;
>>>> time += rdtscp();
>>>> }
>>>> }
>>>>
>>>> For this specific reproducer at least, the page lock seems to be the thingy that prohibits
>>>> reuse if I interpret the numbers correctly. We pass the initial page_count() check.
>>>>
>>>> Haven't looked into the details, and I would be curious how that performs with actual
>>>> workloads, if we can reproduce similar behavior.
>>>
>>> I don't see how that patch addresses the lru issue. Wouldn't we need
>>> something like ...
>>>
>>> if (!PageLRU(page))
>>> lru_add_drain_all();
>>>
>
> lru_add_drain_all() takes a mutex ... best we can do I guess is drain
> the local CPU using lru_add_drain(). I'll go play with it and see what
> breaks :)
>
I did hack something similar and it solved the problem, but I felt it is
a hack. If the thread is scheduled on another core, or if the write fault
is triggered by another thread it wouldn’t work.
If you look for a real-world workload that behaves similarly, you can try
memcached with memory pressure and low latency device (I used
pmem-emulated). This is the workload in which I encountered the issue
first.
