Dan Magenheimer
2009-Dec-08 19:16 UTC
[Xen-devel] [PATCH] tmem: reclaim minimal memory proactively
When a single domain is using most/all of tmem memory for ephemeral pages belonging to the same object, e.g. when copying a single huge file larger than ephemeral memory, long lists are traversed looking for a page to evict that doesn''t belong to this object (as pages in the object for which a page is currently being inserted are locked and cannot be evicted). This is essentially a livelock. Avoid this by proactively ensuring there is a margin of available memory (1MB) before locks are taken on the object. Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> diff -r 7f611de6b93c xen/common/tmem.c --- a/xen/common/tmem.c Tue Dec 08 14:14:27 2009 +0000 +++ b/xen/common/tmem.c Tue Dec 08 12:09:05 2009 -0700 @@ -1091,6 +1091,24 @@ static unsigned long tmem_relinquish_npa if ( avail_pages ) tmh_release_avail_pages_to_host(); return avail_pages; +} + +/* Under certain conditions (e.g. if each client is putting pages for exactly + * one object), once locks are held, freeing up memory may + * result in livelocks and very long "put" times, so we try to ensure there + * is a minimum amount of memory (1MB) available BEFORE any data structure + * locks are held */ +static inline void tmem_ensure_avail_pages(void) +{ + int failed_evict = 10; + + while ( !tmh_free_mb() ) + { + if ( tmem_evict() ) + continue; + else if ( failed_evict-- <= 0 ) + break; + } } /************ TMEM CORE OPERATIONS ************************************/ @@ -2315,10 +2333,12 @@ EXPORT long do_tmem_op(tmem_cli_op_t uop op.u.new.uuid[0], op.u.new.uuid[1]); break; case TMEM_NEW_PAGE: + tmem_ensure_avail_pages(); rc = do_tmem_put(pool, op.u.gen.object, op.u.gen.index, op.u.gen.cmfn, 0, 0, 0, NULL); break; case TMEM_PUT_PAGE: + tmem_ensure_avail_pages(); rc = do_tmem_put(pool, op.u.gen.object, op.u.gen.index, op.u.gen.cmfn, 0, 0, PAGE_SIZE, NULL); if (rc == 1) succ_put = 1; diff -r 7f611de6b93c xen/include/xen/tmem_xen.h --- a/xen/include/xen/tmem_xen.h Tue Dec 08 14:14:27 2009 +0000 +++ b/xen/include/xen/tmem_xen.h Tue Dec 08 12:09:05 2009 -0700 @@ -252,6 +252,11 @@ static inline unsigned long tmh_freeable (20 - PAGE_SHIFT); } +static inline unsigned long tmh_free_mb(void) +{ + return (tmh_avail_pages() + total_free_pages()) >> (20 - PAGE_SHIFT); +} + /* * Memory allocation for "infrastructure" data */ _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel