panic in zfs arc_release during zfs send of encrypted dataset

Authored by: Paul Zuchowski <pzuchowski@racktopsystems.com>
Reviewed by: Jerry Jelinek <gjelinek@racktopsystems.com>
Reviewed by: Gordon Ross <gordon.w.ross@gmail.com>
Reviewed by: Toomas Soome <tsoome@me.com>
Approved by: Robert Mustacchi <rm@fingolfin.org>

Illumos-issue: https://www.illumos.org/issues/14003
Illumos-commit: illumos/illumos-gate@b541cf35

Porting Notes:

The b_cv was removed in #15340 preventing us from adopting this
change without modification.  Instead, we adopt the approach taken
in #15340 to wait for any in progress IO.

Additionally, update trace_arc.h on Linux to include the new dtrace
probe.  While not required I opted to keep it.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #13479

Author: behlendorf

include/os/linux/zfs/sys/trace_arc.h

@@ -105,6 +105,7 @@ DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__hit);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__iohit);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__evict);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_arc__delete);
+DEFINE_ARC_BUF_HDR_EVENT(zfs_arc_release__io);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_new_state__mru);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_new_state__mfu);
 DEFINE_ARC_BUF_HDR_EVENT(zfs_new_state__uncached);
@@ -388,6 +389,7 @@ DEFINE_DTRACE_PROBE1(arc__hit);
 DEFINE_DTRACE_PROBE1(arc__iohit);
 DEFINE_DTRACE_PROBE1(arc__evict);
 DEFINE_DTRACE_PROBE1(arc__delete);
+DEFINE_DTRACE_PROBE1(arc_release__io);
 DEFINE_DTRACE_PROBE1(new_state__mru);
 DEFINE_DTRACE_PROBE1(new_state__mfu);
 DEFINE_DTRACE_PROBE1(new_state__uncached);

module/zfs/arc.c

@@ -6623,6 +6623,39 @@ arc_release(arc_buf_t *buf, const void *tag)
 	kmutex_t *hash_lock = HDR_LOCK(hdr);
 	mutex_enter(hash_lock);
 
+	/*
+	 * Wait for any other IO for this hdr, as additional
+	 * buf(s) could be about to appear, in which case
+	 * we would not want to transition hdr to arc_anon.
+	 */
+	while (HDR_IO_IN_PROGRESS(hdr)) {
+		arc_callback_t *acb;
+
+		DTRACE_PROBE1(arc_release__io, arc_buf_hdr_t *, hdr);
+
+		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
+		acb->acb_wait = B_TRUE;
+		mutex_init(&acb->acb_wait_lock, NULL, MUTEX_DEFAULT, NULL);
+		cv_init(&acb->acb_wait_cv, NULL, CV_DEFAULT, NULL);
+
+		acb->acb_zio_head = hdr->b_l1hdr.b_acb->acb_zio_head;
+		acb->acb_next = hdr->b_l1hdr.b_acb;
+		hdr->b_l1hdr.b_acb->acb_prev = acb;
+		hdr->b_l1hdr.b_acb = acb;
+
+		mutex_exit(hash_lock);
+		mutex_enter(&acb->acb_wait_lock);
+		while (acb->acb_wait)
+			cv_wait(&acb->acb_wait_cv, &acb->acb_wait_lock);
+
+		mutex_exit(&acb->acb_wait_lock);
+		mutex_destroy(&acb->acb_wait_lock);
+		cv_destroy(&acb->acb_wait_cv);
+		kmem_free(acb, sizeof (arc_callback_t));
+
+		mutex_enter(hash_lock);
+	}
+
 	/*
 	 * This assignment is only valid as long as the hash_lock is
 	 * held, we must be careful not to reference state or the