[LU-12534] Add delete_pagevec_from_page_cache

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Assignee: WC Triage
• Reporter: Patrick Farrell (Inactive)

Description

When doing shared file reads, we back up in page discard (this can also happen when writing a file, but is less of an issue).

a
-   32.79%     0.10%  ior              [kernel.kallsyms]        [k] osc_page_init           a
   - 32.69% osc_page_init                                                                   a
      - 32.22% osc_lru_alloc                                                                a
         - 31.58% osc_lru_reclaim                                                           a
            - 31.57% osc_lru_shrink                                                         a
               - 30.24% discard_pagevec                                                     a
                  - 27.27% cl_page_discard                                                  a
                     - 27.22% vvp_page_discard                                              a
                        - 26.55% delete_from_page_cache                                     a
                           + 23.77% _raw_qspin_lock_irq                                     a
                           + 1.58% mem_cgroup_uncharge_cache_page                           a
                             1.03% __delete_from_page_cache                                 a

This is seen when doing the i/o 500 hard read with more data than will fit in cache on a node (and many threads).

This will be very similar to LU-9920, for dirtying pages, where we can take the lock once per pagevec, rather than once per page.  The operation in question is trivial and also quite similar, and will be easy to convert to a pagevec:

 void delete_from_page_cache(struct page *page)
{
        struct address_space *mapping = page->mapping;
        void (*freepage)(struct page *);        BUG_ON(!PageLocked(page));        freepage = mapping->a_ops->freepage;
        spin_lock_irq(&mapping->tree_lock);
        __delete_from_page_cache(page, NULL);
        spin_unlock_irq(&mapping->tree_lock);
        mem_cgroup_uncharge_cache_page(page);        if (freepage)
                freepage(page);
        page_cache_release(page);
}
EXPORT_SYMBOL(delete_from_page_cache);

And this should offer a decent speed up.