Backport #36 to 6.18-lts: 9p/xen: reconnect to the backend on resume

[edera-rebase-app]

Backport of #36 from edera/mainline to edera/6.18-lts (750906adddc6), prepared by Claude. Run log.

Warning

• the checker found differences from the mainline commits (see below)
• Claude's report asks for a decision

Merge with Rebase and merge, so edera/6.18-lts stays linear.

Independent checks

Builds: success (x86_64 and arm64 vmlinux with the Edera features built in).

LTS tip before	750906adddc6
LTS tip after	bd6704b0b89f
Commits requested	4

• New branch fast-forwards the LTS tip.
• Backported commits are linear.
• Backport differs from the request: 3 of 4 unchanged; 1 adapted, 0 retitled, 0 left out, 0 added, 0 not requested or out of order.

	Backport	Mainline source
adapted	0a28a6fc9ca3 9p/xen: reconnect to the backend on resume	9a1668bd2e37 9p/xen: reconnect to the backend on resume

Adapted commits (source +/- lines vs. backport)

### 0a28a6fc9ca3 9p/xen: reconnect to the backend on resume
--- mainline
+++ backport
@@ -34,7 +34,7 @@
 +		goto lookup;
 +	}
 +
-+				      !READ_ONCE(priv->connected) ||
++				   !READ_ONCE(priv->connected) ||
 +	if (!READ_ONCE(priv->connected)) {
 +		up_read(&priv->rings_sem);
 +		goto lookup;
@@ -49,56 +49,42 @@
 +		struct xen_9pfs_dataring *ring = &priv->rings[i];
 -			cancel_work_sync(&ring->work);
 +		cancel_work_sync(&ring->work);
--			if (!ring->intf)
+-			if (!priv->rings[i].intf)
 -				break;
--			if (ring->irq >= 0) {
--				unbind_from_irqhandler(ring->irq, ring);
--				ring->irq = -1;
--			}
--			if (ring->data.in) {
--				for (j = 0; j < (1 << ring->intf->ring_order);
+-			if (priv->rings[i].irq > 0)
+-				unbind_from_irqhandler(priv->rings[i].irq, ring);
+-			if (priv->rings[i].data.in) {
+-				for (j = 0;
+-				     j < (1 << priv->rings[i].intf->ring_order);
 -				     j++) {
 -					grant_ref_t ref;
 -
--					ref = ring->intf->ref[j];
+-					ref = priv->rings[i].intf->ref[j];
 -					gnttab_end_foreign_access(ref, NULL);
--					ring->intf->ref[j] = INVALID_GRANT_REF;
 -				}
--				free_pages_exact(ring->data.in,
--						 1UL << (ring->intf->ring_order +
--							 XEN_PAGE_SHIFT));
--				ring->data.in = NULL;
--				ring->data.out = NULL;
--			}
--			if (ring->ref != INVALID_GRANT_REF) {
--				gnttab_end_foreign_access(ring->ref, NULL);
--				ring->ref = INVALID_GRANT_REF;
-+		if (!ring->intf)
+-				free_pages_exact(priv->rings[i].data.in,
+-				   1UL << (priv->rings[i].intf->ring_order +
+-					   XEN_PAGE_SHIFT));
++		if (!priv->rings[i].intf)
 +			break;
-+		if (ring->irq >= 0) {
-+			unbind_from_irqhandler(ring->irq, ring);
-+			ring->irq = -1;
-+		}
-+		if (ring->data.in) {
-+			for (j = 0; j < (1 << ring->intf->ring_order); j++) {
++		if (priv->rings[i].irq > 0)
++			unbind_from_irqhandler(priv->rings[i].irq, ring);
++		if (priv->rings[i].data.in) {
++			for (j = 0;
++			     j < (1 << priv->rings[i].intf->ring_order);
++			     j++) {
 +				grant_ref_t ref;
 +
-+				ref = ring->intf->ref[j];
++				ref = priv->rings[i].intf->ref[j];
 +				gnttab_end_foreign_access(ref, NULL);
-+				ring->intf->ref[j] = INVALID_GRANT_REF;
--			free_page((unsigned long)ring->intf);
--			ring->intf = NULL;
-+			free_pages_exact(ring->data.in,
-+					 1UL << (ring->intf->ring_order +
-+						 XEN_PAGE_SHIFT));
-+			ring->data.in = NULL;
-+			ring->data.out = NULL;
-+		}
-+		if (ring->ref != INVALID_GRANT_REF) {
-+			gnttab_end_foreign_access(ring->ref, NULL);
-+			ring->ref = INVALID_GRANT_REF;
-+		free_page((unsigned long)ring->intf);
-+		ring->intf = NULL;
+-			gnttab_end_foreign_access(priv->rings[i].ref, NULL);
+-			free_page((unsigned long)priv->rings[i].intf);
++			free_pages_exact(priv->rings[i].data.in,
++			   1UL << (priv->rings[i].intf->ring_order +
++				   XEN_PAGE_SHIFT));
++		gnttab_end_foreign_access(priv->rings[i].ref, NULL);
++		free_page((unsigned long)priv->rings[i].intf);
++		priv->rings[i].intf = NULL;
 +	}
 +}
 +

Claude's report

Summary

All 4 requested commits from #36 are backported in order with -x trailers and unchanged subjects. Three apply unchanged: "wake ring-space waiters with wake_up()", "carry the 9P session across a restore" and "advertise session transfer to the toolstack". One commit, "reconnect to the backend on resume", is adapted because 6.18 lacks three upstream changes the mainline diff touches. None are left out. verify-backport.sh reports 3 unchanged and 1 adapted, which is expected. The adaptation is explained below. Mainline's merge re-committed the PR's commits (d9dc03ab10e7, 6a7cd40129aa, 177888523fdf, 9aa35aecaab1 in PENDING). git range-diff shows them identical to the requested PR-branch commits, so nothing else in PENDING is needed.

Needs a decision

• UNSURE: "9p/xen: reconnect to the backend on resume" calls xen_9pfs_front_free_rings() more than once on the same rings: on resume, again on remove or on a failed reconnect, and the rings are allocated again in between. Mainline relies on upstream 72cb9ee4f6d8 ("9p/trans_xen: make cleanup idempotent after dataring alloc errors"), which 6.18 does not have. There are two ways to handle this:

  1. Adapt minimally on 6.18 without that upstream change. The new free_rings keeps 6.18's teardown and only sets priv->rings[i].intf = NULL after freeing each ring. A second call then stops at the first ring, as mainline's would. This is what I committed. What remains is the old 6.18 bug that 72cb9ee4f6d8 fixes: xen_9pfs_front_alloc_dataring() frees intf on failure but leaves the pointer set. If allocating a ring fails partway, the cleanup that follows will end the grants and free that ring's page a second time. This already happens on 6.18's first-connect (front_init) error path. With this backport the new front_reconnect error path can hit it too.
  2. Backport 72cb9ee4f6d8 to 6.18 first, as its own upstream-stable fix outside this PR. The reconnect commit's free_rings could then match mainline line for line.

  If option 2 is preferred, take 72cb9ee4f6d8 into 6.18 first, then re-run this backport.

Adapted commits

• 9p/xen: reconnect to the backend on resume (9a1668bd2e37)
  • Ring-space wait in p9_xen_request(): mainline uses io_wait_event_killable(), which comes from upstream 51ffeabe836b ("9p: Track 9P RPC waiting time as IO") and is not in 6.18. I kept 6.18's wait_event_killable() and added the commit's new !READ_ONCE(priv->connected) || condition, so the behaviour on a disconnect is the same.
  • xen_9pfs_front_free() split into xen_9pfs_front_free_rings(): mainline moves a teardown body that already has the idempotent changes from upstream 72cb9ee4f6d8. I moved 6.18's own body instead: the priv->rings[i] style, the irq > 0 check and the unconditional grant release. I added one line, priv->rings[i].intf = NULL, so that calling free_rings again after a resume is safe. See Needs a decision.
  • xen_9pfs_front_init(): kept 6.18's kcalloc() instead of mainline's kzalloc_objs(), which 6.18 lacks. Kept 6.18's versions parsing; mainline's version comes from upstream 8fc518e489c1. The commit's new ring_order and init_rwsem() lines are added unchanged.
• 9p/xen: carry the 9P session across a restore (e271536038f6): the checker shows it as unchanged, but two conflicts were in context lines only. The mainline-only #include <linux/fs_context.h> context line is dropped, and the new delay.h and unaligned.h includes are added. The new INIT_LIST_HEAD/init_completion lines are placed before 6.18's kcalloc(). Everything the commit uses exists on 6.18: xenbus_driver.suspend, P9_HDRSZ, P9_NOTAG, P9_RLERROR, p9_req_try_get and linux/unaligned.h.

Left out

None.

Builds

BUILDS_PENDING

[azenla]

Backport of #36; the same four commits.