kswapd logic improvement

kswapd logic improvement

[Rik van Riel]

Hi Linus,

here's another short patch for 2.1.88. Basically
it improves the kswapd logic. Currently kswapd
will give up when it fails three times in a row,
even when it hasn't any memory yet.

With my patch, it'll try (free_pages_high - nr_free_pages)
times, and four times as often when nr_free_pages < free_pages_low.

Furthermore, it doesn't do synchronous swapouts any more
when nr_free_pages < min_free_pages because that's just
too slow (disk seek time etc.), but instead it runs the
disk task queue whenever nr_async_pages >= SWAP_CLUSTER_MAX

I hope it passes your test for 2.1.89 (I've been testing
it for quite a while, and it seems to give a small improvement).

Rik.
+-----------------------------+------------------------------+
| For Linux mm-patches, go to | "I'm busy managing memory.." |
| my homepage (via LinuxHQ).  | H.H.vanRiel@fys.ruu.nl       |
| ...submissions welcome...   | http://www.fys.ruu.nl/~riel/ |
+-----------------------------+------------------------------+

--- linux2188orig/mm/vmscan.c	Wed Feb 25 16:39:55 1998
+++ linux-2.1.88/mm/vmscan.c	Thu Feb 26 17:27:39 1998
@@ -434,7 +434,6 @@
        printk ("Starting kswapd v%.*s\n", i, s);
 }
 
-#define MAX_SWAP_FAIL 3
 /*
  * The background pageout daemon.
  * Started as a kernel thread from the init process.
@@ -462,7 +461,7 @@
 	init_swap_timer();
 	
 	while (1) {
-		int fail;
+		int tries;
 
 		kswapd_awake = 0;
 		flush_signals(current);
@@ -471,13 +470,16 @@
 		kswapd_awake = 1;
 		swapstats.wakeups++;
 		/* Do the background pageout: 
-		 * We now only swap out as many pages as needed.
-		 * When we are truly low on memory, we swap out
-		 * synchronously (WAIT == 1).  -- Rik.
-		 * If we've had too many consecutive failures,
-		 * go back to sleep to let other tasks run.
+		 * We try free_pages_high - nr_free_pages times,
+		 * only when we're truly low on memory we'll try
+		 * more often. -- Rik.
 		 */
-		for (fail = 0; fail++ < MAX_SWAP_FAIL;) {
+		tries = (free_pages_high - nr_free_pages);
+		if (nr_free_pages < free_pages_low)
+			tries <<= 2;
+		if (tries < min_free_pages)
+			tries = min_free_pages;
+		while (tries--) {
 			int pages, gfp_mask;
 
 			pages = nr_free_pages;
@@ -486,10 +488,14 @@
 			if (pages >= free_pages_high)
 				break;
 			gfp_mask = __GFP_IO;
-			if (pages < free_pages_low)
-				gfp_mask |= __GFP_WAIT;
-			if (try_to_free_page(gfp_mask))
-				fail = 0;
+			try_to_free_page(gfp_mask);
+			/* We used to swap out syncronously when we
+			 * were low on memory, but this is simply
+			 * faster. And a fast recovery is all that
+			 * matters when nr_free_pages is too low.
+			 * -- Rik. */
+			if (atomic_read(&nr_async_pages) >= SWAP_CLUSTER_MAX)
+				run_task_queue(&tq_disk);
 		}
 		/*
 		 * Report failure if we couldn't reach the minimum goal.

Re: kswapd logic improvement

[Linus Torvalds]

On Thu, 26 Feb 1998, Rik van Riel wrote:
> 
> here's another short patch for 2.1.88. Basically
> it improves the kswapd logic. Currently kswapd
> will give up when it fails three times in a row,
> even when it hasn't any memory yet.

Look at the pre-89's on ftp.kernel.org - I changed kswapd around quite a
bit, because I want the machine to remain up when you get lots of atomic
get_free_page() calls that would otherwise not even have woken up kswapd
depending on how "min_pages_free" was set up and what the fragmentation
was. 

I'm not saying that the pre-89 is any better, but it is sufficiently
different that I'd like to hear comments about it. I suspect it really
needs tweaking kswapd, for example - it might result in bad latency right
now because kswapd is so high-priority and if it isn't able to unfragment
easily it might run for a while.. 

Could you send patches relative to that (including your other patch, I'd
be happier to know that you've tested that patch against my current kernel
too),

		Linus