swap: Fix swap size in case of block devices

During swapon, swap size is set to number of usable pages in the given swap file/block device minus 1 (for header page). In case of block devices, this size is incorrectly set as one page less than the actual due to an off-by-one error. For regular files, this size is set correctly.

Signed-off-by Nitin Gupta <[email protected]>

---

 mm/swapfile.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

mm/swapfile.c +0 -0

index 8ffdc0d..3d37b97 100644
@@ -1951,9 +1951,9 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
 	if (error)

Quoted 6 lines · Nitin Gupta

During swapon, swap size is set to number of usable pages in the given swap file/block device minus 1 (for header page). In case of block devices, this size is incorrectly set as one page less than the actual due to an off-by-one error. For regular files, this size is set correctly.

Signed-off-by Nitin Gupta<[email protected]>

I agree that there's an off-by-one disagreement between swapon and mkswap regarding last_page. The kernel seems to interpret it as the index of what I'd call the end page, the first page beyond the swap area.

I'd never noticed that until the beginning of this year, and out of caution I've done nothing about it. I believe that the kernel has been wrong since Linux 2.2.? or 2.3.?, ever since swap header version 1 was first introduced; and that mkswap has set it the same way all along.

But I've not spent time on the research to establish that for sure.

What if there used to be a version of mkswap which set last_page one greater as the kernel expects? Neither Karel nor I think that's the case, but we're not absolutely certain. And what if (I suppose I'm getting even more wildly cautious here!) someone has learnt that that page remains untouched and is now putting it to other use? or has compensated for the off-by-one and is setting it one greater, beyond the end of the partition, not using mkswap?

Since nobody has been hurt by it in all these years, I felt safer to go on leaving that discrepancy as is. Call me over cautious.

Regarding your patch comment: I'm puzzled by the remark "For regular files, this size is set correctly". Do you mean that mkswap is setting last_page one higher when dealing with a regular file rather than a block device (I was unaware of that, but never looked to see)? But your patch appears to be to code shared equally between block devices and regular files, so then you'd be introducing a bug on regular files? And shouldn't mkswap be fixed to be consistent with itself? Hopefully I've misunderstood: please explain further.

And regarding the patch itself: my understanding is that the problem is with the interpretation of last_page, so I don't think one change to nr_good_pages would be enough to fix it - you'd need to change the other places where last_page is referred to too.

I'm still disinclined to make any change here myself (beyond a comment noting the discrepancy); but tell me I'm a fool.

mkswap sets last_page correctly: 0-based index of last usable swap page. To explain why this bug affects only block swap devices, some code walkthrough is done below:
(BTW, I only checked mkswap which is part of util-linux-ng 2.14.2).

swapon()
{
...

        nr_good_pages = swap_header->info.last_page -
                        swap_header->info.nr_badpages -
                        1 /* header page */;

====
off-by-one error: for both regular and block device case, but...
====

        if (nr_good_pages) {
                swap_map[0] = SWAP_MAP_BAD;
                p->max = maxpages;
                p->pages = nr_good_pages;
                nr_extents = setup_swap_extents(p, &span);
====
For block devices, setup_swap_extents() leaves p->pages untouched.
For regular files, it sets p->pages
	== total usable swap pages (including header page) - 1;

====

                if (nr_extents < 0) {
                        error = nr_extents;
                        goto bad_swap;
                }
                nr_good_pages = p->pages;

====
So, for block device, nr_good_pages == last_page - nr_badpages - 1

				== (total pages - 1) - nr_badpages - 1 (error)
			For regular files, nr_good_pages == total pages - 1

(correct)
====

        }

...
}

With this fix, block device case is corrected to last_page - nr_badpages - 1 while regular file case remain correct since setup_swap_extents() still gives same correct value in p->pages (== total pages - 1).

Quoted 5 lines

And regarding the patch itself: my understanding is that the problem is with the interpretation of last_page, so I don't think one change to nr_good_pages would be enough to fix it - you'd need to change the other places where last_page is referred to too.

I looked at other instances of last_page in swapon() -- all these other instances looked correct to me.

Quoted 4 lines

I'm still disinclined to make any change here myself (beyond a comment noting the discrepancy); but tell me I'm a fool.

I agree that nobody would bother losing 1 swap slot, so it might not be desirable to have this fix. But IMHO, I don't see any reason to leave this discrepancy between regular files and swap devices -- its just so odd.

Quoted 3 lines · Nitin Gupta

For block devices, setup_swap_extents() leaves p->pages untouched. For regular files, it sets p->pages
      == total usable swap pages (including header page) - 1;

I think you're overlooking the "page < sis->max" condition in setup_swap_extents()'s loop.  So at the end of the loop, if no pages were lost to fragmentation, we have

               sis->max = page_no;             /* no change */ sis->pages = page_no - 1;       /* no change */

Yes, I'd dislike that discrepancy between regular files and block devices, if I could see it. Though I'd probably still be cautious about the disk partitions.

dd if=/dev/zero of=/swap bs=200k        # says 204800 bytes (205kB)
mkswap /swap                            # says size = 196 KiB
swapon /swap                            # dmesg says Adding 192k swap

which is what I've come to expect from the off-by-one, even on regular files.

Quoted 5 lines · Hugh Dickins

Quoted 3 lines · Nitin Gupta

For block devices, setup_swap_extents() leaves p->pages untouched. For regular files, it sets p->pages
      == total usable swap pages (including header page) - 1;

Quoted 7 lines

I think you're overlooking the "page < sis->max" condition in setup_swap_extents()'s loop.  So at the end of the loop, if no pages were lost to fragmentation, we have

               sis->max = page_no;             /* no change */ sis->pages = page_no - 1;       /* no change */

Oh, I missed this loop condition. The variable naming is so bad, I find it very hard to follow this part of code.

Still, if there is even a single page in swap file that is not usable (i.e. non-contiguous on disk) -- which is what usually happens for swap files of any practical size -- setup_swap_extents() gives correct value in sis->pages == total usable pages (including header) - 1;

However, if all the file pages are usable, it gives off-by-one error, as you noted.

Quoted 11 lines

Yes, I'd dislike that discrepancy between regular files and block devices, if I could see it. Though I'd probably still be cautious about the disk partitions.

dd if=/dev/zero of=/swap bs=200k        # says 204800 bytes (205kB)
mkswap /swap                            # says size = 196 KiB
swapon /swap                            # dmesg says Adding 192k swap

which is what I've come to expect from the off-by-one, even on regular files.

In general, its not correct to compare size repored by mkswap and swapon like this. The size reported by mkswap includes pages which are not contiguous on disk. While, kernel considers only PAGE_SIZE-length, PAGE_SIZE-aligned contiguous run of blocks. So, size reported by mkswap and swapon can vary wildly. For e.g.:

(on mtdram with ext2 fs)
dd if=/dev/zero of=swap.dd bs=1M count=10
mkswap swap.dd # says size = 10236 KiB
swapon swap.dd # says Adding 10112k swap

====

So, to summarize:

1. mkswap always behaves correctly: It sets number of pages in swap file minus one as 'last_page' in swap header (since this is a 0-based index). This same value (total pages - 1) is printed out as size since it knows that first page is swap header.

2. swapon() for block devices: off-by-one error causing last swap page to remain unused.

3. swapon() for regular files:
3.1 off-by-one error if every swap page in this file is usable i.e.
every PAGE_SIZE-length, PAGE_SIZE-aligned chunk is contiguous on disk.
3.2 correct size value if there is at least one swap page which is
unusable -- which is expected from swap file of any practical size.

I will go through swap code again to find other possible off-by-one errors. The revised patch will fix these inconsistencies.