Index | Thread | Search

From:
Kirill A. Korinsky <kirill@korins.ky>
Subject:
Re: need help with test: sys/vfs: stop checkalias replacing vnodes
To:
Yamasaki Yuu <miagetegorann@gmail.com>
Cc:
tech@openbsd.org
Date:
Wed, 11 Nov 2026 12:22:18 +0100

Download raw body.

Thread
On Wed, 11 Nov 2026 03:46:40 +0100,
Yamasaki Yuu <miagetegorann@gmail.com> wrote:
> 
> Since "sys/vfs: stop checkalias replacing vnodes" (9274ed7885f in the
> git mirror, 2026-11-08), the root swap partition is configured twice
> at boot.  swapctl -A in rc now adds it a second time:
> 
>   swapctl: adding 0b85468de71a22af.b as swap device at priority 0
> 
>   $ pstat -s
>   Device      512-blocks     Used    Avail Capacity  Priority
>   /dev/sd0b       524256        0   524256     0%    0
>   /dev/sd0b       524256        0   524256     0%    0
>   Total          1048512        0  1048512     0%
> 
> GENERIC.MP built from the parent commit lists sd0b once, and
> swapctl -A gets EBUSY there as before.  The duplicate is still there
> at 6afb24b0786.  Seen on amd64 under QEMU with an AHCI disk.
> 
> swapmount() gets its vnode from bdevvp(), swapctl -A gets one through
> namei(), and swaplist_find() compares vnode pointers.  ffs_vinit()
> used to replace its vnode with the VT_NON alias returned by
> checkalias(), so both were the same vnode and SWAP_ON failed with
> EBUSY.  Now they are two vnodes for the same dev_t.
> 
> The same lookup is used by SWAP_OFF and SWAP_CTL, so the partition the
> kernel configured can't be removed by path any more.  Of two SWAP_OFF
> calls on /dev/sd0b, the first removes the entry rc added and the
> second fails with ENXIO; sd0b stays in the list.  swapctl -d makes
> the same call.
> 
> Both entries cover the same blocks, so paging corrupts memory.  Under
> QEMU with 512 MB of RAM, a pattern test that fills and checks 560 MB
> finished with no bad words on the parent commit.  On 9274ed7885f the
> test program and sshd crashed during the same test.
> 
> The diff below makes swaplist_find() match block devices on swd_dev as
> well.  With it on top of 6afb24b0786 (amd64 GENERIC.MP, QEMU, 512 MB of
> RAM): rc no longer adds sd0b and pstat -s lists it once; a second
> SWAP_ON fails with EBUSY; the pattern test finishes with no bad words;
> SWAP_OFF of /dev/sd0b, SWAP_ON and SWAP_ON again (EBUSY) work.
> 

Thansk for the report.

Yesterday I had send to miod@ this diff which addressed the problem.

Anyway, I had reverted my checkaliases diff which uncovered the issue.

The diff:

Index: sys/uvm/uvm_swap.c
===================================================================
RCS file: /home/cvs/src/sys/uvm/uvm_swap.c,v
diff -u -p -r1.185 uvm_swap.c
--- sys/uvm/uvm_swap.c	8 Nov 2026 08:42:55 -0000	1.185
+++ sys/uvm/uvm_swap.c	10 Nov 2026 22:19:13 -0000
@@ -58,6 +58,7 @@
 #include <nfs/nfsproto.h>
 #include <nfs/nfsdiskless.h>
 #endif
+#include <sys/lock.h>
 
 #include <uvm/uvm.h>
 #ifdef UVM_SWAP_ENCRYPT
@@ -471,7 +472,8 @@ swaplist_find(struct vnode *vp, boolean_
 	 */
 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
 		TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
-			if (sdp->swd_vp != vp)
+			if (sdp->swd_vp != vp &&
+			    (vp->v_type != VBLK || sdp->swd_dev != vp->v_rdev))
 				continue;
 			if (remove) {
 				TAILQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
@@ -773,8 +775,17 @@ sys_swapctl(struct proc *p, void *v, reg
 		}
 
 		/*
-		 * do the real work.
+		 * swap_off() drops the swap reference before closing the vnode.
+		 * Hold a reference to that vnode even when named by an alias.
 		 */
+		if (vp != sdp->swd_vp) {
+			mtx_leave(&uvm_swap_data_lock);
+			vput(vp);
+			vp = sdp->swd_vp;
+			if ((error = vget(vp, LK_EXCLUSIVE | LK_RETRY)) != 0)
+				goto out;
+			mtx_enter(&uvm_swap_data_lock);
+		}
 		error = swap_off(p, sdp);
 		break;
 	default:



-- 
wbr, Kirill