device-mapper-multipath/0163-limpathpersist-Handle-changing-key-corner-case.patch
Benjamin Marzinski 5a8898b817 device-mapper-multipath-0.8.7-40
Add 0150-libmpathpersist-retry-commands-on-other-paths-in-mpa.patch
Add 0151-libmpathpersist-check-released-key-against-the-reser.patch
Add 0152-multipathd-remove-thread-from-mpath_pr_event_handle.patch
Add 0153-libmpathpersist-remove-uneeded-wrapper-function.patch
Add 0154-libmpathpersist-reduce-log-level-for-persistent-rese.patch
Add 0155-libmpathpersist-remove-pointless-update_map_pr-ret-v.patch
Add 0156-multipathd-use-update_map_pr-in-mpath_pr_event_handl.patch
Add 0157-libmpathpersist-limit-changing-prflag-in-update_map_.patch
Add 0158-multipathd-Don-t-call-update_map_pr-unnecessarily.patch
Add 0159-libmpathpersist-remove-useless-function-send_prout_a.patch
Add 0160-libmpathpersist-fix-memory-leak-in-mpath_prout_rel.patch
Add 0161-limpathpersist-redesign-failed-release-workaround.patch
Add 0162-libmpathpersist-fail-the-release-if-all-threads-fail.patch
Add 0163-limpathpersist-Handle-changing-key-corner-case.patch
Add 0164-libmpathpersist-fix-command-keyword-ordering.patch
Add 0165-libmpathpersist-use-conf-timeout-for-updating-persis.patch
Add 0166-libmapthpersist-Handle-REGISTER-AND-IGNORE-changing-.patch
Add 0167-libmultipath-rename-prflag_value-enums.patch
Add 0168-libmpathpersist-use-a-switch-statement-for-prout-com.patch
Add 0169-libmpathpersist-Add-safety-check-for-preempting-on-k.patch
Add 0170-limpathpersist-remove-update_map_pr-code-for-NULL-pp.patch
Add 0171-libmpathpersist-move-update_map_pr-to-multipathd.patch
Add 0172-multipathd-clean-up-update_map_pr-and-mpath_pr_event.patch
Add 0173-libmpathpersist-clean-up-duplicate-function-declarat.patch
Add 0174-multipathd-wrap-setting-and-unsetting-prflag.patch
Add 0175-multipathd-unregister-PR-key-when-path-is-restored-i.patch
Add 0176-libmpathpersist-Fix-up-reservation_key-checking.patch
Add 0177-libmpathpersist-change-how-reservation-conflicts-are.patch
Add 0178-libmpathpersist-Clear-prkey-in-multipathd-before-unr.patch
Add 0179-libmpathpersist-only-clear-the-key-if-we-are-using-t.patch
Add 0180-libmpathpersist-Restore-old-reservation-key-on-failu.patch
Add 0181-libmpatpersist-update-reservation-key-before-checkin.patch
Add 0182-libmpathpersist-retry-on-conflicts-in-mpath_prout_co.patch
Add 0183-libmpathpersist-Don-t-always-fail-registrations-for-.patch
Add 0184-libmpathpersist-Don-t-try-release-workaround-for-inv.patch
Add 0185-libmpathpersist-Don-t-fail-RESERVE-commands-unnecess.patch
Add 0186-libmpathpersist-reregister-keys-when-self-preempting.patch
Add 0187-libmpathpersist-handle-updating-key-race-condition.patch
Add 0188-libmpathpersist-handle-preempting-all-registrants-re.patch
Add 0189-libmpathpersist-Fix-REGISTER-AND-IGNORE-while-holdin.patch
Add 0190-libmpathpersist-Handle-RESERVE-with-reservation-held.patch
Add 0191-libmpathpersist-use-check_holding_reservation-in-mpa.patch
Add 0192-libmpathpersist-Fix-unregistering-while-holding-the-.patch
Add 0193-libmpathpersist-Fix-race-between-restoring-a-path-an.patch
Add 0194-multipathd-Fix-tracking-of-old-PR-key.patch
  * Fixes RHEL-118515 ("There are many bugs in multipath's persistent
    reservation handling")
Resolves: RHEL-118515
2025-10-01 15:39:06 -04:00

125 lines
5.3 KiB
Diff

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Benjamin Marzinski <bmarzins@redhat.com>
Date: Fri, 27 Jun 2025 18:17:40 -0400
Subject: [PATCH] limpathpersist: Handle changing key corner case
When you change the reservation key of a registered multipath device,
some of paths might be down or even deleted since you originally
registered the key. libmpathpersist won't be able to change these
registrations. This can be a problem if the path that is down is holding
the reservation. Other nodes will assume that the reservation is now
under your new key, when it is still under the old one. If they try to
preempt the reservation, they will succeed. But they will just
unregister all the paths with the new key. They won't take over the
reservation, and if the path holding it comes back up, it will still be
able to write to the device.
To avoid this, after libmpathpersist changes the key of a registered
device, it now checks to see if its old key is still holding the
reservation. If it is, limpathpersist preempts the reservation.
Currently this only works on REGISTER commands, not REGISTER AND IGNORE
ones. A future patch will add that capability.
I should note that this relies on the fact that libmpathpersist has
never worked correctly if two nodes use the same reservation key for the
same device. If another node used the same key, it could be the one
holding the reservation, and preempting the reservation would deregister
it. However, multipathd has always relied on the fact that two nodes
will not use the same registration key to know if it is safe to register
a key on a path that was just added or just came back up. If there is
already a registered key matching the configured key, multipathd assumes
that the device has not been preempted, and registers the key on the
path. If there are no keys matching the configured key, multipathd
assumes that the device has been preempted, and won't register a key.
Again, if another node shared the same key, multipathd could registered
paths on a node that had been preempted.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
---
libmpathpersist/mpath_persist.c | 58 ++++++++++++++++++++++++++++++++-
1 file changed, 57 insertions(+), 1 deletion(-)
diff --git a/libmpathpersist/mpath_persist.c b/libmpathpersist/mpath_persist.c
index 04a08b76..c09fbbe7 100644
--- a/libmpathpersist/mpath_persist.c
+++ b/libmpathpersist/mpath_persist.c
@@ -425,6 +425,60 @@ get_mpvec (vector curmp, vector pathvec, char * refwwid)
return MPATH_PR_SUCCESS ;
}
+/*
+ * If you are changing the key registered to a device, and that device is
+ * holding the reservation on a path that couldn't get its key updated,
+ * either because it is down or no longer part of the multipath device,
+ * you need to preempt the reservation to a usable path with the new key
+ */
+void preempt_missing_path(struct multipath *mpp, uint8_t *key, uint8_t *sa_key,
+ int noisy)
+{
+ uint8_t zero[8] = {0};
+ struct prin_resp resp = {0};
+ int rq_scope;
+ unsigned int rq_type;
+ struct prout_param_descriptor paramp = {0};
+ int status;
+
+ /*
+ * If you previously didn't have a key registered or you didn't
+ * switch to a different key, there's no need to preempt. Also, you
+ * can't preempt if you no longer have a registered key
+ */
+ if (memcmp(key, zero, 8) == 0 || memcmp(sa_key, zero, 8) == 0 ||
+ memcmp(key, sa_key, 8) == 0)
+ return;
+
+ status = mpath_prin_activepath (mpp, MPATH_PRIN_RRES_SA, &resp, noisy);
+ if (status != MPATH_PR_SUCCESS) {
+ condlog(0, "%s: register: pr in read reservation command failed.", mpp->wwid);
+ return;
+ }
+ /* If there is no reservation, there's nothing to preempt */
+ if (!resp.prin_descriptor.prin_readresv.additional_length)
+ return;
+ /*
+ * If there reservation is not held by the old key, you don't
+ * want to preempt it
+ */
+ if (memcmp(key, resp.prin_descriptor.prin_readresv.key, 8) != 0)
+ return;
+ /* Assume this key is being held by an inaccessable path on this
+ * node. libmpathpersist has never worked if multiple nodes share
+ * the same reservation key for a device
+ */
+ rq_type = resp.prin_descriptor.prin_readresv.scope_type & MPATH_PR_TYPE_MASK;
+ rq_scope = (resp.prin_descriptor.prin_readresv.scope_type & MPATH_PR_SCOPE_MASK) >> 4;
+ memcpy(paramp.key, sa_key, 8);
+ memcpy(paramp.sa_key, key, 8);
+ status = mpath_prout_common(mpp, MPATH_PROUT_PREE_SA, rq_scope, rq_type,
+ &paramp, noisy, NULL);
+ if (status != MPATH_PR_SUCCESS)
+ condlog(0, "%s: register: pr preempt command failed.",
+ mpp->wwid);
+}
+
int mpath_prout_reg(struct multipath *mpp,int rq_servact, int rq_scope,
unsigned int rq_type, struct prout_param_descriptor * paramp, int noisy)
{
@@ -564,6 +618,8 @@ int mpath_prout_reg(struct multipath *mpp,int rq_servact, int rq_scope,
}
pthread_attr_destroy(&attr);
+ if (status == MPATH_PR_SUCCESS)
+ preempt_missing_path(mpp, paramp->key, paramp->sa_key, noisy);
return (status == MPATH_PR_RETRYABLE_ERROR) ? MPATH_PR_OTHER : status;
}
@@ -624,7 +680,7 @@ int mpath_prout_rel(struct multipath *mpp,int rq_servact, int rq_scope,
int rc;
int count = 0;
int status = MPATH_PR_SUCCESS;
- struct prin_resp resp;
+ struct prin_resp resp = {0};
uint16_t udev_flags = (mpp->skip_kpartx)? MPATH_UDEV_NO_KPARTX_FLAG : 0;
bool did_resume = false;
bool all_threads_failed;