Date: Mon, 7 Dec 2015 11:14:57 +0000 (UTC) From: Konstantin Belousov <kib@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-10@freebsd.org Subject: svn commit: r291934 - stable/10/sys/vm Message-ID: <201512071114.tB7BEvaT009352@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: kib Date: Mon Dec 7 11:14:57 2015 New Revision: 291934 URL: https://svnweb.freebsd.org/changeset/base/291934 Log: MFC r290917: Provide the OOM-specific vm_pageout_oom_pagecount() function which estimates the amount of reclamaible memory which could be stolen if the process is killed. Modified: stable/10/sys/vm/vm_pageout.c Directory Properties: stable/10/ (props changed) Modified: stable/10/sys/vm/vm_pageout.c ============================================================================== --- stable/10/sys/vm/vm_pageout.c Mon Dec 7 11:12:03 2015 (r291933) +++ stable/10/sys/vm/vm_pageout.c Mon Dec 7 11:14:57 2015 (r291934) @@ -1526,6 +1526,65 @@ vm_pageout_mightbe_oom(struct vm_domain atomic_subtract_int(&vm_pageout_oom_vote, 1); } +/* + * The OOM killer is the page daemon's action of last resort when + * memory allocation requests have been stalled for a prolonged period + * of time because it cannot reclaim memory. This function computes + * the approximate number of physical pages that could be reclaimed if + * the specified address space is destroyed. + * + * Private, anonymous memory owned by the address space is the + * principal resource that we expect to recover after an OOM kill. + * Since the physical pages mapped by the address space's COW entries + * are typically shared pages, they are unlikely to be released and so + * they are not counted. + * + * To get to the point where the page daemon runs the OOM killer, its + * efforts to write-back vnode-backed pages may have stalled. This + * could be caused by a memory allocation deadlock in the write path + * that might be resolved by an OOM kill. Therefore, physical pages + * belonging to vnode-backed objects are counted, because they might + * be freed without being written out first if the address space holds + * the last reference to an unlinked vnode. + * + * Similarly, physical pages belonging to OBJT_PHYS objects are + * counted because the address space might hold the last reference to + * the object. + */ +static long +vm_pageout_oom_pagecount(struct vmspace *vmspace) +{ + vm_map_t map; + vm_map_entry_t entry; + vm_object_t obj; + long res; + + map = &vmspace->vm_map; + KASSERT(!map->system_map, ("system map")); + sx_assert(&map->lock, SA_LOCKED); + res = 0; + for (entry = map->header.next; entry != &map->header; + entry = entry->next) { + if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) + continue; + obj = entry->object.vm_object; + if (obj == NULL) + continue; + if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && + obj->ref_count != 1) + continue; + switch (obj->type) { + case OBJT_DEFAULT: + case OBJT_SWAP: + case OBJT_PHYS: + case OBJT_VNODE: + res += obj->resident_page_count; + break; + } + } + return (res); +} + void vm_pageout_oom(int shortage) { @@ -1599,12 +1658,13 @@ vm_pageout_oom(int shortage) } PROC_UNLOCK(p); size = vmspace_swap_count(vm); - vm_map_unlock_read(&vm->vm_map); if (shortage == VM_OOM_MEM) - size += vmspace_resident_count(vm); + size += vm_pageout_oom_pagecount(vm); + vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); + /* - * if the this process is bigger than the biggest one + * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) {
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201512071114.tB7BEvaT009352>