[patch][rfc] delayacct: fix swapin delay accounting (maybe)

[Nick Piggin <npiggin@suse.de>]

Hi,

I can't convince myself that delay accounting for swapin is quite right
at the moment (not having a test setup handy to run and check for myself).
Maybe I'm not reading the swapin code very well...

lookup_swap_cache, and read_swap_cache_async should be non-blocking
operations for the most part.

read_swap_cache_async might, when allocating the new page, go into reclaim
and take a long time to come back. However is that any more a "swapin" delay
than eg. when we sleep on mmap_sem when first taking the fault, or any other
types of fault which require allocations? None of which we account for as
swapin delay.

But the most obvious delay, where we actually lock the page waiting for
the swap IO to finish, does not seem to be accounted at all!

My proposed fix is to just move the swaping delay accounting to the
point where the VM does actually wait, for the swapin.

I have no idea what uses swapin delay accounting, but it would be good to
see if this makes a positive (or at least not negative) impact on those
users...

Thanks,
Nick

--
Index: linux-2.6/mm/memory.c
===================================================================
--- linux-2.6.orig/mm/memory.c
+++ linux-2.6/mm/memory.c
@@ -2158,7 +2158,6 @@ static int do_swap_page(struct mm_struct
 		migration_entry_wait(mm, pmd, address);
 		goto out;
 	}
-	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
 	page = lookup_swap_cache(entry);
 	if (!page) {
 		grab_swap_token(); /* Contend for token _before_ read-in */
@@ -2172,7 +2171,6 @@ static int do_swap_page(struct mm_struct
 			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
 			if (likely(pte_same(*page_table, orig_pte)))
 				ret = VM_FAULT_OOM;
-			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
 			goto unlock;
 		}
 
@@ -2181,9 +2179,10 @@ static int do_swap_page(struct mm_struct
 		count_vm_event(PGMAJFAULT);
 	}
 
-	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
 	mark_page_accessed(page);
+	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
 	lock_page(page);
+	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
 
 	/*
 	 * Back out if somebody else already faulted in this pte.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>