On Mon, Oct 25, 2010 at 2:32 AM, Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> Karel, Michael,
>
> In kernel 2.6.36 (and in stable kernel 2.6.35.5) I made a tiny change
> to the swapon(const char *path, int swapflags) system call interface:
> kernel commit 3399446632739fcd05fd8b272b476a69c6e6d14a
> swap: discard while swapping only if SWAP_FLAG_DISCARD
>
> I added this definition to the kernel's include/linux/swap.h:
> #define SWAP_FLAG_DISCARD 0x10000 /* discard swap cluster after use */
> (which I think needs to find its way into /usr/include/sys/swap.h somehow);
> and if that bit is set in swapflags (previously devoted to swap priority),
> then swapping will discard clusters of swap pages in between freeing them
> and re-writing to them, if the swap device supports that (discard or trim
> being notification from OS to device that stale data is no longer needed).
>
> (No error occurs if the device does not support discard, and the only sign
> that it is discarding is "D" at the end of the kernel's "Adding XXXk swap"
> message. And that flag only governs discarding while swapping: whether
> or not it is supplied, every swapon tries to discard the entire swap area,
> less its swap header page - that will normally be an efficient operation.
> Earlier kernels than 2.6.35.5 will just ignore the flag.)
>
> The story behind is this. In 2.6.29 I implemented this behaviour for any
> device that supports discard, before having any real SSD that supported
> it: in the hope that it would give a useful boost to swapping on SSD once
> device support arrived. About a year ago, OCZ implemented discard in their
> Vertex, then Intel in their X25: testing around 2.6.32, the OCZ showed
> measurable (though disappointing) benefit, the Intel arguable benefit.
>
> But since then, the kernel's block-layer support for discard has gone
> through a series of changes (with most recently a shift from reliance
> on I/O barriers to waiting on completions), and the SSD firmwares have
> been updated too. With the result that once I tested again, shortly
> after 2.6.35 came out (alerted by a bugreport from Nigel Cunningham),
> it proved to be a disadvantage on the Intel, and a disaster on the OCZ
> (even after a patch from hch removed the most obvious overhead).
>
> As things stand at present, we could just remove the swap discard
> support; but since several filesystems (including ext4 and btrfs and
> fat) are offering a "discard" mount option, I thought swap should take
> the same course, and offer a "--discard" or "-d" option to swapon(8).
>
> Below is the patch I tested with. I could sign it off, except for that
> #ifndef SWAP_FLAG_DISCARD business at the start (I'm rather hoping that
> you, Karel, will have a hotline to getting that addition to sys/swap.h).
> Something else questionable about it, is that final "if (all && discard)
> swapon_usage(stderr, 1)" hunk: I'm not sure what's best when you
> "swapon -a -d", given that /etc/fstab can itself say "discard" or not;
> I took the easiest way to code it up, and made "swapon -a -d" an error.
>
> I apologize to you both for taking so long to alert you (and sorry for
> being so verbose! but I think the background makes more sense of it).
> But as you can see from the above, there is no great urgency to adding
> and documenting this support.
Hugh, would you write a patch for the man-page? Or just supply me with
the raw text that should be added to the man page?
Thanks,
Michael
> --- util-linux-ng-2.18/mount/swapon.c.0 2010-06-14 02:19:19.000000000 -0700
> +++ util-linux-ng-2.18/mount/swapon.c 2010-09-05 22:07:48.000000000 -0700
> @@ -29,6 +29,9 @@
> #ifdef HAVE_SYS_SWAP_H
> # include <sys/swap.h>
> #endif
> +#ifndef SWAP_FLAG_DISCARD
> +#define SWAP_FLAG_DISCARD 0x10000 /* discard swap cluster after use */
> +#endif
>
> #ifndef SWAPON_HAS_TWO_ARGS
> /* libc is insane, let's call the kernel */
> @@ -54,6 +57,7 @@ enum {
>
> int all;
> int priority = -1; /* non-prioritized swap by default */
> +int discard;
>
> /* If true, don't complain if the device/file doesn't exist */
> int ifexists;
> @@ -65,6 +69,7 @@ char *progname;
> static struct option longswaponopts[] = {
> /* swapon only */
> { "priority", required_argument, 0, 'p' },
> + { "discard", 0, 0, 'd' },
> { "ifexists", 0, 0, 'e' },
> { "summary", 0, 0, 's' },
> { "fixpgsz", 0, 0, 'f' },
> @@ -92,7 +97,7 @@ static void
> swapon_usage(FILE *fp, int n) {
> fprintf(fp, _("\nUsage:\n"
> " %1$s -a [-e] [-v] [-f] enable all swaps from /etc/fstab\n"
> - " %1$s [-p priority] [-v] [-f] <special> enable given swap\n"
> + " %1$s [-p priority] [-d] [-v] [-f] <special> enable given swap\n"
> " %1$s -s display swap usage summary\n"
> " %1$s -h display help\n"
> " %1$s -V display version\n\n"), progname);
> @@ -493,6 +498,8 @@ do_swapon(const char *orig_special, int
> << SWAP_FLAG_PRIO_SHIFT);
> }
> #endif
> + if (discard)
> + flags |= SWAP_FLAG_DISCARD;
> status = swapon(special, flags);
> if (status < 0)
> warn(_("%s: swapon failed"), orig_special);
> @@ -582,6 +589,8 @@ swapon_all(void) {
> opt = strtok(NULL, ",")) {
> if (strncmp(opt, "pri=", 4) == 0)
> pri = atoi(opt+4);
> + if (strcmp(opt, "discard") == 0)
> + discard = 1;
> if (strcmp(opt, "noauto") == 0)
> skip = 1;
> }
> @@ -605,6 +614,7 @@ swapon_all(void) {
> }
> fclose(fp);
>
> + discard = 0;
> return status;
> }
>
> @@ -632,7 +642,7 @@ main_swapon(int argc, char *argv[]) {
> int status = 0;
> int c, i;
>
> - while ((c = getopt_long(argc, argv, "ahefp:svVL:U:",
> + while ((c = getopt_long(argc, argv, "ahdefp:svVL:U:",
> longswaponopts, NULL)) != -1) {
> switch (c) {
> case 'a': /* all */
> @@ -650,6 +660,9 @@ main_swapon(int argc, char *argv[]) {
> case 'U':
> addu(optarg);
> break;
> + case 'd':
> + discard = 1;
> + break;
> case 'e': /* ifexists */
> ifexists = 1;
> break;
> @@ -680,6 +693,9 @@ main_swapon(int argc, char *argv[]) {
> if (ifexists && (!all || strcmp(progname, "swapon")))
> swapon_usage(stderr, 1);
>
> + if (all && discard)
> + swapon_usage(stderr, 1);
> +
> if (all)
> status |= swapon_all();
>
>
--
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Author of "The Linux Programming Interface"; http://man7.org/tlpi/
--
To unsubscribe from this list: send the line "unsubscribe util-linux-ng" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
