[PATCH 0/1] checkers: add alua path checker

[PATCH 0/1] checkers: add alua path checker

[Brian Bunker]

For ALUA-capable storage arrays, multipath-tools currently uses TUR
(Test Unit Ready) as the default path checker while a separate
prioritizer determines ALUA state. When sysfs provides ALUA state
(the common case), no RTPG command is needed for priority. However,
if sysfs is unavailable or detect_prio is off, the prioritizer issues
RTPG, resulting in two SCSI commands per path check cycle: TUR + RTPG.

This patch introduces an 'alua' path checker that uses RTPG as the path
check command, providing two benefits:

1. For configurations where RTPG is needed for priority, this eliminates
   duplicate I/O by combining path checking and ALUA state retrieval
   into a single RTPG command.

2. RTPG provides richer path state information than TUR:
   * TRANSITIONING state detected and mapped to PATH_PENDING,
     preventing false all paths down during controller failover
   * STANDBY state detected and mapped to PATH_GHOST
   * UNAVAILABLE/OFFLINE states properly distinguished

The richer state information is valuable even when sysfs is available,
as TUR cannot distinguish between these ALUA-specific conditions.

This design intentionally couples the checker and prioritizer: the
alua_cached prioritizer consumes state cached by the alua checker.
While this blurs the traditional separation between these components,
for ALUA the path state and priority fundamentally derive from the
same underlying data, making the coupling a natural fit.

This patch sets the alua checker as the auto-detected default for ALUA
devices to demonstrate its benefits. We recognize that adopting a new
default checker requires careful consideration; the auto-detection can
be adjusted or removed based on review feedback.

Brian Bunker (1):
  checkers: add alua path checker

 libmultipath/Makefile                   |   5 +
 libmultipath/checkers.c                 |   1 +
 libmultipath/checkers.h                 |   1 +
 libmultipath/checkers/Makefile          |   3 +-
 libmultipath/checkers/alua.c            | 426 ++++++++++++++++++++++++
 libmultipath/checkers/alua.h            |  15 +
 libmultipath/prio.c                     |   1 +
 libmultipath/prio.h                     |   1 +
 libmultipath/prioritizers/Makefile      |   1 +
 libmultipath/prioritizers/alua_cached.c | 225 +++++++++++++
 libmultipath/prioritizers/alua_rtpg.c   |  12 +-
 libmultipath/prioritizers/alua_rtpg.h   |   1 +
 libmultipath/prioritizers/sysfs.c       |   7 +-
 libmultipath/propsel.c                  |  25 +-
 libmultipath/structs.c                  |   2 +
 libmultipath/structs.h                  |   3 +
 16 files changed, 721 insertions(+), 8 deletions(-)
 create mode 100644 libmultipath/checkers/alua.c
 create mode 100644 libmultipath/checkers/alua.h
 create mode 100644 libmultipath/prioritizers/alua_cached.c

-- 
2.50.1 (Apple Git-155)

[PATCH 1/1] checkers: add alua path checker

[Brian Bunker]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset=y, Size: 27648 bytes --]

Add an 'alua' path checker that uses RTPG to check path state, and an
'alua_cached' prioritizer that uses the checker's cached ALUA state.

The checker maps ALUA Asymmetric Access States to path states:
  - AAS_OPTIMIZED, AAS_NON_OPTIMIZED -> PATH_UP
  - AAS_STANDBY -> PATH_GHOST
  - AAS_TRANSITIONING -> PATH_PENDING
  - AAS_UNAVAILABLE, AAS_OFFLINE -> PATH_DOWN

The alua_cached prioritizer first checks for cached state from the alua
checker, then falls back to sysfs, then to issuing RTPG directly. This
maintains backward compatibility with other checker configurations.

When detect_checker is enabled, the alua checker is auto-selected for
devices with TPGS support. When the alua checker is in use, detect_prio
auto-selects the alua_cached prioritizer.

Signed-off-by: Brian Bunker <brian@purestorage.com>
---
 libmultipath/Makefile                   |   5 +
 libmultipath/checkers.c                 |   1 +
 libmultipath/checkers.h                 |   1 +
 libmultipath/checkers/Makefile          |   3 +-
 libmultipath/checkers/alua.c            | 426 ++++++++++++++++++++++++
 libmultipath/checkers/alua.h            |  15 +
 libmultipath/prio.c                     |   1 +
 libmultipath/prio.h                     |   1 +
 libmultipath/prioritizers/Makefile      |   1 +
 libmultipath/prioritizers/alua_cached.c | 224 +++++++++++++
 libmultipath/prioritizers/alua_rtpg.c   |  12 +-
 libmultipath/prioritizers/alua_rtpg.h   |   1 +
 libmultipath/prioritizers/sysfs.c       |   7 +-
 libmultipath/propsel.c                  |  20 +-
 libmultipath/structs.c                  |   2 +
 libmultipath/structs.h                  |   3 +
 16 files changed, 716 insertions(+), 7 deletions(-)
 create mode 100644 libmultipath/checkers/alua.c
 create mode 100644 libmultipath/checkers/alua.h
 create mode 100644 libmultipath/prioritizers/alua_cached.c

diff --git a/libmultipath/Makefile b/libmultipath/Makefile
index 85767ab4..379f4d02 100644
--- a/libmultipath/Makefile
+++ b/libmultipath/Makefile
@@ -38,6 +38,11 @@ nvme-lib.o: nvme-lib.c nvme-ioctl.c nvme-ioctl.h
 dict.o:	dict.c
 	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -Wno-unused-parameter -c -o $@ $<
 
+# checkers/alua.o needs to find headers in parent directory
+checkers/alua.o: checkers/alua.c
+	@echo building $@ because of $?
+	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -I. -c -o $@ $<
+
 make_static = $(shell sed '/^static/!s/^\([a-z]\{1,\} \)/static \1/' <$1 >$2)
 
 nvme-ioctl.c: nvme/nvme-ioctl.c
diff --git a/libmultipath/checkers.c b/libmultipath/checkers.c
index bb6ad1ee..21aa52d4 100644
--- a/libmultipath/checkers.c
+++ b/libmultipath/checkers.c
@@ -461,6 +461,7 @@ int init_checkers(void)
 		EMC_CLARIION,
 		READSECTOR0,
 		CCISS_TUR,
+		ALUA_RTPG,
 	};
 	unsigned int i;
 
diff --git a/libmultipath/checkers.h b/libmultipath/checkers.h
index a969e7d1..2905f3fc 100644
--- a/libmultipath/checkers.h
+++ b/libmultipath/checkers.h
@@ -98,6 +98,7 @@ enum path_check_state {
 #define EMC_CLARIION "emc_clariion"
 #define READSECTOR0  "readsector0"
 #define CCISS_TUR    "cciss_tur"
+#define ALUA         "alua"
 #define NONE         "none"
 #define INVALID      "invalid"
 
diff --git a/libmultipath/checkers/Makefile b/libmultipath/checkers/Makefile
index 6f7cfb95..dfdd13a1 100644
--- a/libmultipath/checkers/Makefile
+++ b/libmultipath/checkers/Makefile
@@ -17,7 +17,8 @@ LIBS= \
 	libcheckdirectio.so \
 	libcheckemc_clariion.so \
 	libcheckhp_sw.so \
-	libcheckrdac.so
+	libcheckrdac.so \
+	libcheckalua.so
 
 all: $(LIBS)
 
diff --git a/libmultipath/checkers/alua.c b/libmultipath/checkers/alua.c
new file mode 100644
index 00000000..cb3d7000
--- /dev/null
+++ b/libmultipath/checkers/alua.c
@@ -0,0 +1,426 @@
+/*
+ * ALUA Path Checker
+ *
+ * Copyright (c) 2024
+ * This file is released under the GPL.
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <pthread.h>
+#include <urcu.h>
+#include <urcu/uatomic.h>
+
+#include "checkers.h"
+#include "debug.h"
+#include "structs.h"
+#include "prio.h"
+#include "util.h"
+#include "time-util.h"
+#include "../prioritizers/alua.h"
+#include "../prioritizers/alua_rtpg.h"
+#include "alua.h"
+
+#define MAX_NR_TIMEOUTS 1
+
+enum {
+	MSG_ALUA_RTPG_NOT_SUPPORTED = CHECKER_FIRST_MSGID,
+	MSG_ALUA_RTPG_TRANSITIONING,
+	MSG_ALUA_RUNNING,
+	MSG_ALUA_TIMEOUT,
+};
+
+#define IDX_(x) (MSG_ ## x - CHECKER_FIRST_MSGID)
+const char *libcheck_msgtable[] = {
+	[IDX_(ALUA_RTPG_NOT_SUPPORTED)] = " ALUA not supported",
+	[IDX_(ALUA_RTPG_TRANSITIONING)] = " path is transitioning",
+	[IDX_(ALUA_RUNNING)] = " still running",
+	[IDX_(ALUA_TIMEOUT)] = " timed out",
+	NULL,
+};
+
+struct alua_checker_context {
+	dev_t devt;
+	int state;
+	int running; /* uatomic access only */
+	int fd;
+	unsigned int timeout;
+	time_t time;
+	pthread_t thread;
+	pthread_mutex_t lock;
+	pthread_cond_t active;
+	int holders; /* uatomic access only */
+	int msgid;
+	struct checker_context ctx;
+	unsigned int nr_timeouts;
+	bool checked_state;
+	/* ALUA-specific cached data for prioritizer */
+	int aas;           /* Asymmetric Access State */
+	int tpg;           /* Target Port Group */
+	time_t timestamp;  /* When this data was collected */
+	/* Data needed for RTPG in thread */
+	struct path *pp;   /* Path pointer - only valid during check */
+};
+
+int libcheck_init(struct checker *c)
+{
+	struct alua_checker_context *ct;
+	struct stat sb;
+
+	ct = malloc(sizeof(struct alua_checker_context));
+	if (!ct)
+		return 1;
+	memset(ct, 0, sizeof(struct alua_checker_context));
+
+	ct->state = PATH_UNCHECKED;
+	ct->fd = -1;
+	ct->aas = -1;
+	ct->tpg = -1;
+	uatomic_set(&ct->holders, 1);
+	pthread_cond_init_mono(&ct->active);
+	pthread_mutex_init(&ct->lock, NULL);
+	if (fstat(c->fd, &sb) == 0)
+		ct->devt = sb.st_rdev;
+	ct->ctx.cls = c->cls;
+	c->context = ct;
+
+	return 0;
+}
+
+static void cleanup_context(struct alua_checker_context *ct)
+{
+	pthread_mutex_destroy(&ct->lock);
+	pthread_cond_destroy(&ct->active);
+	free(ct);
+}
+
+void libcheck_free(struct checker *c)
+{
+	if (c->context) {
+		struct alua_checker_context *ct = c->context;
+		int holders;
+		int running;
+
+		running = uatomic_xchg(&ct->running, 0);
+		if (running)
+			pthread_cancel(ct->thread);
+		ct->thread = 0;
+		holders = uatomic_sub_return(&ct->holders, 1);
+		if (!holders)
+			cleanup_context(ct);
+		c->context = NULL;
+	}
+	return;
+}
+
+/*
+ * Map ALUA Asymmetric Access State to path state
+ */
+static int alua_state_to_path_state(int aas, short *msgid)
+{
+	switch (aas & 0x0f) {
+	case AAS_OPTIMIZED:
+	case AAS_NON_OPTIMIZED:
+		*msgid = CHECKER_MSGID_UP;
+		return PATH_UP;
+	case AAS_STANDBY:
+		*msgid = CHECKER_MSGID_GHOST;
+		return PATH_GHOST;
+	case AAS_TRANSITIONING:
+		*msgid = MSG_ALUA_RTPG_TRANSITIONING;
+		return PATH_PENDING;
+	case AAS_UNAVAILABLE:
+	case AAS_OFFLINE:
+	default:
+		*msgid = CHECKER_MSGID_DOWN;
+		return PATH_DOWN;
+	}
+}
+
+/*
+ * Main ALUA check function - uses alua_rtpg.c library.
+ * This is called either in sync mode or from the async thread.
+ */
+static int
+alua_check(struct path *pp, int *aas_out, short *msgid)
+{
+	int aas;
+
+	if (pp->tpg_id == GROUP_ID_UNDEF) {
+		*msgid = CHECKER_MSGID_DOWN;
+		return PATH_DOWN;
+	}
+
+	aas = get_asymmetric_access_state(pp, pp->tpg_id);
+	if (aas < 0) {
+		if (aas == -RTPG_LUN_DISCONNECTED) {
+			*msgid = CHECKER_MSGID_DISCONNECTED;
+			return PATH_DISCONNECTED;
+		}
+		*msgid = CHECKER_MSGID_DOWN;
+		return PATH_DOWN;
+	}
+
+	*aas_out = aas;
+	return alua_state_to_path_state(aas, msgid);
+}
+
+#define alua_thread_cleanup_push(ct) pthread_cleanup_push(cleanup_func, ct)
+#define alua_thread_cleanup_pop(ct) pthread_cleanup_pop(1)
+
+static void cleanup_func(void *data)
+{
+	int holders;
+	struct alua_checker_context *ct = data;
+
+	holders = uatomic_sub_return(&ct->holders, 1);
+	if (!holders)
+		cleanup_context(ct);
+}
+
+void *libcheck_thread(struct checker_context *ctx)
+{
+	struct alua_checker_context *ct =
+		container_of(ctx, struct alua_checker_context, ctx);
+	int state, running;
+	short msgid;
+	int aas;
+
+	/* This thread can be canceled, so setup clean up */
+	alua_thread_cleanup_push(ct);
+
+	condlog(4, "%d:%d : alua checker starting up", major(ct->devt),
+		minor(ct->devt));
+
+	state = alua_check(ct->pp, &aas, &msgid);
+	pthread_testcancel();
+
+	/* ALUA checker done */
+	pthread_mutex_lock(&ct->lock);
+	ct->state = state;
+	ct->msgid = msgid;
+	ct->aas = aas;
+	ct->tpg = ct->pp->tpg_id;
+	ct->timestamp = time(NULL);
+	/* Update path-level cache for prioritizer use */
+	ct->pp->alua_state = aas;
+	ct->pp->alua_timestamp = ct->timestamp;
+	pthread_cond_signal(&ct->active);
+	pthread_mutex_unlock(&ct->lock);
+
+	condlog(4, "%d:%d : alua checker finished, state %s", major(ct->devt),
+		minor(ct->devt), checker_state_name(state));
+
+	running = uatomic_xchg(&ct->running, 0);
+	if (!running)
+		pause();
+
+	alua_thread_cleanup_pop(ct);
+
+	return ((void *)0);
+}
+
+static void alua_set_async_timeout(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+	struct timespec now;
+
+	get_monotonic_time(&now);
+	ct->time = now.tv_sec + c->timeout;
+}
+
+static int alua_check_async_timeout(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+	struct timespec now;
+
+	get_monotonic_time(&now);
+	return (now.tv_sec > ct->time);
+}
+
+static int check_pending(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+	int alua_status = PATH_PENDING;
+
+	pthread_mutex_lock(&ct->lock);
+	if (ct->state != PATH_PENDING || ct->msgid != MSG_ALUA_RUNNING) {
+		alua_status = ct->state;
+		c->msgid = ct->msgid;
+	}
+	pthread_mutex_unlock(&ct->lock);
+	if (alua_status == PATH_PENDING && c->msgid == MSG_ALUA_RUNNING) {
+		condlog(4, "%d:%d : alua checker still running",
+			major(ct->devt), minor(ct->devt));
+	} else {
+		int running = uatomic_xchg(&ct->running, 0);
+		if (running)
+			pthread_cancel(ct->thread);
+		ct->thread = 0;
+	}
+
+	ct->checked_state = true;
+	return alua_status;
+}
+
+bool libcheck_need_wait(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+	return (ct && ct->thread && uatomic_read(&ct->running) != 0 &&
+		!ct->checked_state);
+}
+
+int libcheck_pending(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+
+	/* The if path checker isn't running, just return the exiting value. */
+	if (!ct || !ct->thread)
+		return c->path_state;
+
+	return check_pending(c);
+}
+
+int libcheck_check(struct checker *c)
+{
+	struct alua_checker_context *ct = c->context;
+	struct path *pp = container_of(c, struct path, checker);
+	pthread_attr_t attr;
+	int alua_status = PATH_PENDING, r;
+	int aas;
+
+	if (!ct)
+		return PATH_UNCHECKED;
+
+	if (checker_is_sync(c)) {
+		alua_status = alua_check(pp, &aas, &c->msgid);
+		ct->tpg = pp->tpg_id;
+		ct->aas = aas;
+		ct->timestamp = time(NULL);
+		/* Update path-level cache for prioritizer use */
+		pp->alua_state = aas;
+		pp->alua_timestamp = ct->timestamp;
+		return alua_status;
+	}
+
+	/* Async mode */
+	if (ct->thread) {
+		ct->checked_state = true;
+		if (alua_check_async_timeout(c)) {
+			int running = uatomic_xchg(&ct->running, 0);
+			if (running) {
+				pthread_cancel(ct->thread);
+				condlog(3, "%d:%d : alua checker timeout",
+					major(ct->devt), minor(ct->devt));
+				c->msgid = MSG_ALUA_TIMEOUT;
+				alua_status = PATH_TIMEOUT;
+			} else {
+				pthread_mutex_lock(&ct->lock);
+				alua_status = ct->state;
+				c->msgid = ct->msgid;
+				pthread_mutex_unlock(&ct->lock);
+			}
+			ct->thread = 0;
+		} else if (uatomic_read(&ct->running) != 0) {
+			/*
+			 * ct->running may be stale - the thread sets
+			 * msgid/state under lock, then clears running
+			 * after releasing it. Check msgid under lock
+			 * to get the truth. Only MSG_ALUA_RUNNING means
+			 * no result yet.
+			 */
+			pthread_mutex_lock(&ct->lock);
+			if (ct->msgid != MSG_ALUA_RUNNING) {
+				alua_status = ct->state;
+				c->msgid = ct->msgid;
+				ct->thread = 0;
+			}
+			pthread_mutex_unlock(&ct->lock);
+		} else {
+			/* ALUA checker done */
+			ct->thread = 0;
+			pthread_mutex_lock(&ct->lock);
+			alua_status = ct->state;
+			c->msgid = ct->msgid;
+			pthread_mutex_unlock(&ct->lock);
+		}
+	} else {
+		if (uatomic_read(&ct->holders) > 1) {
+			/* The thread has been cancelled but hasn't quit. */
+			if (ct->nr_timeouts <= MAX_NR_TIMEOUTS) {
+				if (ct->nr_timeouts == MAX_NR_TIMEOUTS) {
+					condlog(2, "%d:%d : waiting for stalled alua thread to finish",
+						major(ct->devt), minor(ct->devt));
+				}
+				ct->nr_timeouts++;
+				c->msgid = MSG_ALUA_TIMEOUT;
+				return PATH_TIMEOUT;
+			}
+			ct->nr_timeouts++;
+			/*
+			 * Start a new thread while the old one is stalled.
+			 * We have to prevent it from interfering with the new
+			 * thread. We create a new context and leave the old
+			 * one with the stale thread, hoping it will clean up
+			 * eventually.
+			 */
+			condlog(3, "%d:%d : alua thread not responding",
+				major(ct->devt), minor(ct->devt));
+
+			if (libcheck_init(c) != 0) {
+				c->msgid = CHECKER_MSGID_DOWN;
+				return PATH_UNCHECKED;
+			}
+			((struct alua_checker_context *)c->context)->nr_timeouts = ct->nr_timeouts;
+
+			if (!uatomic_sub_return(&ct->holders, 1)) {
+				/* It did terminate, eventually */
+				cleanup_context(ct);
+				((struct alua_checker_context *)c->context)->nr_timeouts = 0;
+			}
+
+			ct = c->context;
+		} else
+			ct->nr_timeouts = 0;
+
+		/* Start new ALUA checker */
+		pthread_mutex_lock(&ct->lock);
+		alua_status = ct->state = PATH_PENDING;
+		c->msgid = ct->msgid = MSG_ALUA_RUNNING;
+		pthread_mutex_unlock(&ct->lock);
+		ct->fd = c->fd;
+		ct->timeout = c->timeout;
+		ct->pp = pp;
+		ct->checked_state = false;
+		uatomic_add(&ct->holders, 1);
+		uatomic_set(&ct->running, 1);
+		alua_set_async_timeout(c);
+		setup_thread_attr(&attr, 32 * 1024, 1);
+		r = start_checker_thread(&ct->thread, &attr, &ct->ctx);
+		pthread_attr_destroy(&attr);
+		if (r) {
+			uatomic_sub(&ct->holders, 1);
+			uatomic_set(&ct->running, 0);
+			ct->thread = 0;
+			condlog(3, "%d:%d : failed to start alua thread, using"
+				" sync mode", major(ct->devt), minor(ct->devt));
+			alua_status = alua_check(pp, &aas, &c->msgid);
+			ct->tpg = pp->tpg_id;
+			ct->aas = aas;
+			ct->timestamp = time(NULL);
+			/* Update path-level cache for prioritizer use */
+			pp->alua_state = aas;
+			pp->alua_timestamp = ct->timestamp;
+			return alua_status;
+		}
+	}
+
+	return alua_status;
+}
+
diff --git a/libmultipath/checkers/alua.h b/libmultipath/checkers/alua.h
new file mode 100644
index 00000000..b53cef66
--- /dev/null
+++ b/libmultipath/checkers/alua.h
@@ -0,0 +1,15 @@
+/*
+ * ALUA Path Checker Header
+ *
+ * This file is released under the GPL.
+ */
+#ifndef _ALUA_CHECKER_H
+#define _ALUA_CHECKER_H
+
+/* ALUA checker caches state in struct path for prioritizer use:
+ *   pp->alua_state     - cached AAS value (-1 if not cached)
+ *   pp->alua_timestamp - when the state was last updated
+ */
+
+#endif /* _ALUA_CHECKER_H */
+
diff --git a/libmultipath/prio.c b/libmultipath/prio.c
index 24f825bd..2d59d04e 100644
--- a/libmultipath/prio.c
+++ b/libmultipath/prio.c
@@ -29,6 +29,7 @@ int init_prio(void)
 #ifdef LOAD_ALL_SHARED_LIBS
 	static const char *const all_prios[] = {
 		PRIO_ALUA,
+		PRIO_ALUA_CACHED,
 		PRIO_CONST,
 		PRIO_DATACORE,
 		PRIO_EMC,
diff --git a/libmultipath/prio.h b/libmultipath/prio.h
index 119b75f2..b54ece0d 100644
--- a/libmultipath/prio.h
+++ b/libmultipath/prio.h
@@ -17,6 +17,7 @@ struct path;
  * Known prioritizers for use in hwtable.c
  */
 #define PRIO_ALUA		"alua"
+#define PRIO_ALUA_CACHED	"alua_cached"
 #define PRIO_CONST		"const"
 #define PRIO_DATACORE		"datacore"
 #define PRIO_EMC		"emc"
diff --git a/libmultipath/prioritizers/Makefile b/libmultipath/prioritizers/Makefile
index ff2524c2..201c52c0 100644
--- a/libmultipath/prioritizers/Makefile
+++ b/libmultipath/prioritizers/Makefile
@@ -13,6 +13,7 @@ LIBDEPS = -lmultipath -lmpathutil -lm -lpthread -lrt
 # If you add or remove a prioritizer also update multipath/multipath.conf.5
 LIBS = \
 	libprioalua.so \
+	libprioalua_cached.so \
 	libprioconst.so \
 	libpriodatacore.so \
 	libprioemc.so \
diff --git a/libmultipath/prioritizers/alua_cached.c b/libmultipath/prioritizers/alua_cached.c
new file mode 100644
index 00000000..2ba5b2e7
--- /dev/null
+++ b/libmultipath/prioritizers/alua_cached.c
@@ -0,0 +1,224 @@
+/*
+ * (C) Copyright IBM Corp. 2004, 2005   All Rights Reserved.
+ * Modified 2024 to support cached ALUA state from checker
+ *
+ * ALUA prioritizer that can use cached state from alua checker
+ * to avoid duplicate RTPG commands.
+ *
+ * This file is released under the GPL.
+ */
+#include <stdio.h>
+#include <string.h>
+
+#include <time.h>
+
+#include "debug.h"
+#include "prio.h"
+#include "structs.h"
+#include "checkers.h"
+#include "discovery.h"
+
+#include "alua.h"
+
+#define ALUA_PRIO_NOT_SUPPORTED			1
+#define ALUA_PRIO_RTPG_FAILED			2
+#define ALUA_PRIO_GETAAS_FAILED			3
+#define ALUA_PRIO_TPGS_FAILED			4
+#define ALUA_PRIO_NO_INFORMATION		5
+#define ALUA_PRIO_CACHED_STALE			6
+
+/* Maximum age of cached ALUA state in seconds */
+#define ALUA_CACHE_MAX_AGE_SECS			5
+
+static const char * aas_string[] = {
+	[AAS_OPTIMIZED]		= "active/optimized",
+	[AAS_NON_OPTIMIZED]	= "active/non-optimized",
+	[AAS_STANDBY]		= "standby",
+	[AAS_UNAVAILABLE]	= "unavailable",
+	[AAS_LBA_DEPENDENT]	= "logical block dependent",
+	[AAS_RESERVED]		= "ARRAY BUG: invalid TPGs state!",
+	[AAS_OFFLINE]		= "offline",
+	[AAS_TRANSITIONING]	= "transitioning between states",
+};
+
+static const char *aas_print_string(int rc)
+{
+	rc &= 0x7f;
+
+	if (rc & 0x70)
+		return aas_string[AAS_RESERVED];
+	rc &= 0x0f;
+	if (rc > AAS_RESERVED && rc < AAS_OFFLINE)
+		return aas_string[AAS_RESERVED];
+	else
+		return aas_string[rc];
+}
+
+/*
+ * Try to get ALUA info with optimization:
+ * 1. If using alua checker: use its cached state (it just polled!)
+ * 2. Try sysfs (kernel maintains this, no I/O needed)
+ * 3. Fall back to issuing RTPG directly
+ *
+ * Rationale: The checker is our active poller. If we're using alua
+ * and its cache is fresh, we trust that data first. If the cache misses
+ * (stale or empty), we try sysfs before issuing RTPG. This provides a
+ * three-tier optimization: checker cache → sysfs → RTPG.
+ */
+/* Maximum length of sysfs access_state string */
+#define ALUA_ACCESS_STATE_SIZE 32
+
+int
+get_alua_info_cached(struct path * pp, bool *used_cache)
+{
+	int	rc;
+	int	tpg;
+	bool	diff_tpg;
+	char	buff[ALUA_ACCESS_STATE_SIZE];
+
+	*used_cache = false;
+
+	/* First: if using alua checker, use its fresh data from path structure */
+	if (checker_selected(&pp->checker)) {
+		const char *chk_name = checker_name(&pp->checker);
+
+		if (chk_name && strcmp(chk_name, "alua") == 0 &&
+		    pp->alua_state >= 0) {
+			time_t now = time(NULL);
+			time_t age = now - pp->alua_timestamp;
+
+			if (pp->alua_timestamp > 0 && age <= ALUA_CACHE_MAX_AGE_SECS) {
+				condlog(4, "%s: using cached ALUA state from checker (fresh poll)", pp->dev);
+				*used_cache = true;
+				return pp->alua_state;
+			}
+			condlog(4, "%s: cached ALUA state not available or stale, trying sysfs",
+				pp->dev);
+			/* Cache miss - try sysfs before issuing RTPG */
+		}
+	}
+
+	/* Second: try sysfs (if cache missed or not using alua checker) */
+	rc = sysfs_get_asymmetric_access_state(pp, buff, sizeof(buff));
+	if (rc >= 0) {
+		/* Parse sysfs state string to ALUA state value */
+		int aas = -1;
+		int priopath = rc; /* rc contains preferred bit */
+
+		if (!strncmp(buff, "active/optimized", 16))
+			aas = AAS_OPTIMIZED;
+		else if (!strncmp(buff, "active/non-optimized", 20))
+			aas = AAS_NON_OPTIMIZED;
+		else if (!strncmp(buff, "lba-dependent", 13))
+			aas = AAS_LBA_DEPENDENT;
+		else if (!strncmp(buff, "standby", 7))
+			aas = AAS_STANDBY;
+
+		if (aas >= 0) {
+			condlog(4, "%s: using sysfs ALUA state (no I/O)", pp->dev);
+			*used_cache = true;
+			return (priopath ? 0x80 : 0) | aas;
+		}
+	}
+
+	/* Last resort: issue RTPG directly */
+	tpg = get_target_port_group(pp);
+	if (tpg < 0) {
+		rc = get_target_port_group_support(pp);
+		if (rc < 0)
+			return -ALUA_PRIO_TPGS_FAILED;
+		if (rc == TPGS_NONE)
+			return -ALUA_PRIO_NOT_SUPPORTED;
+		return -ALUA_PRIO_RTPG_FAILED;
+	}
+	diff_tpg = (pp->tpg_id != GROUP_ID_UNDEF && pp->tpg_id != tpg);
+	pp->tpg_id = tpg;
+	condlog((diff_tpg) ? 2 : 4, "%s: reported target port group is %i",
+		pp->dev, tpg);
+	rc = get_asymmetric_access_state(pp, tpg);
+	if (rc < 0) {
+		condlog(2, "%s: get_asymmetric_access_state returned %d",
+			__func__, rc);
+		return -ALUA_PRIO_GETAAS_FAILED;
+	}
+
+	condlog(3, "%s: aas = %02x [%s]%s", pp->dev, rc, aas_print_string(rc),
+		(rc & 0x80) ? " [preferred]" : "");
+	return rc;
+}
+
+int get_exclusive_pref_arg(char *args)
+{
+	char *ptr;
+
+	if (args == NULL)
+		return 0;
+	ptr = strstr(args, "exclusive_pref_bit");
+	if (!ptr)
+		return 0;
+	if (ptr[18] != '\0' && ptr[18] != ' ' && ptr[18] != '\t')
+		return 0;
+	if (ptr != args && ptr[-1] != ' ' && ptr[-1] != '\t')
+		return 0;
+	return 1;
+}
+
+int getprio (struct path * pp, char * args)
+{
+	int rc;
+	int aas;
+	int priopath;
+	int exclusive_pref;
+	bool used_cache = false;
+
+	if (pp->fd < 0)
+		return -ALUA_PRIO_NO_INFORMATION;
+
+	exclusive_pref = get_exclusive_pref_arg(args);
+	rc = get_alua_info_cached(pp, &used_cache);
+
+	if (used_cache) {
+		condlog(4, "%s: priority calculated from cached ALUA state (no RTPG issued)",
+			pp->dev);
+	}
+
+	if (rc >= 0) {
+		aas = (rc & 0x0f);
+		priopath = (rc & 0x80);
+		switch(aas) {
+		case AAS_OPTIMIZED:
+			rc = 50;
+			break;
+		case AAS_NON_OPTIMIZED:
+			rc = 10;
+			break;
+		case AAS_LBA_DEPENDENT:
+			rc = 5;
+			break;
+		case AAS_STANDBY:
+			rc = 1;
+			break;
+		default:
+			rc = 0;
+		}
+		if (priopath && (aas != AAS_OPTIMIZED || exclusive_pref))
+			rc += 80;
+	} else {
+		switch(-rc) {
+		case ALUA_PRIO_NOT_SUPPORTED:
+			condlog(0, "%s: alua not supported", pp->dev);
+			break;
+		case ALUA_PRIO_RTPG_FAILED:
+			condlog(0, "%s: couldn't get target port group", pp->dev);
+			break;
+		case ALUA_PRIO_GETAAS_FAILED:
+			condlog(0, "%s: couldn't get asymmetric access state", pp->dev);
+			break;
+		case ALUA_PRIO_TPGS_FAILED:
+			condlog(3, "%s: couldn't get supported alua states", pp->dev);
+			break;
+		}
+	}
+	return rc;
+}
+
diff --git a/libmultipath/prioritizers/alua_rtpg.c b/libmultipath/prioritizers/alua_rtpg.c
index 053cccb7..5da06b50 100644
--- a/libmultipath/prioritizers/alua_rtpg.c
+++ b/libmultipath/prioritizers/alua_rtpg.c
@@ -75,6 +75,7 @@ enum scsi_disposition {
 	SCSI_GOOD = 0,
 	SCSI_ERROR,
 	SCSI_RETRY,
+	SCSI_DISCONNECTED,
 };
 
 static int
@@ -124,6 +125,12 @@ scsi_error(struct sg_io_hdr *hdr, int opcode)
 	PRINT_DEBUG("alua: SCSI error for command %02x: status %02x, sense %02x/%02x/%02x",
 		    opcode, hdr->status, sense_key, asc, ascq);
 
+	/* Check for disconnected LUN (unmapped at target) */
+	if (sense_key == 0x5 && asc == 0x25 && ascq == 0x00) {
+		/* ILLEGAL REQUEST / LOGICAL UNIT NOT SUPPORTED */
+		return SCSI_DISCONNECTED;
+	}
+
 	if (sense_key == UNIT_ATTENTION || sense_key == NOT_READY)
 		return SCSI_RETRY;
 	else
@@ -325,7 +332,10 @@ retry:
 	}
 
 	rc = scsi_error(&hdr, OPERATION_CODE_RTPG);
-	if (rc == SCSI_ERROR) {
+	if (rc == SCSI_DISCONNECTED) {
+		PRINT_DEBUG("do_rtpg: LUN disconnected (unmapped at target)!");
+		return -RTPG_LUN_DISCONNECTED;
+	} else if (rc == SCSI_ERROR) {
 		PRINT_DEBUG("do_rtpg: SCSI error!");
 		return -RTPG_RTPG_FAILED;
 	} else if (rc == SCSI_RETRY) {
diff --git a/libmultipath/prioritizers/alua_rtpg.h b/libmultipath/prioritizers/alua_rtpg.h
index ff6ec0ef..030ca58d 100644
--- a/libmultipath/prioritizers/alua_rtpg.h
+++ b/libmultipath/prioritizers/alua_rtpg.h
@@ -21,6 +21,7 @@
 #define RTPG_NO_TPG_IDENTIFIER			2
 #define RTPG_RTPG_FAILED			3
 #define RTPG_TPG_NOT_FOUND			4
+#define RTPG_LUN_DISCONNECTED			5
 
 int get_target_port_group_support(const struct path *pp);
 int get_target_port_group(const struct path *pp);
diff --git a/libmultipath/prioritizers/sysfs.c b/libmultipath/prioritizers/sysfs.c
index 5e8adc05..ab058ea9 100644
--- a/libmultipath/prioritizers/sysfs.c
+++ b/libmultipath/prioritizers/sysfs.c
@@ -10,6 +10,9 @@
 #include "discovery.h"
 #include "prio.h"
 
+/* Maximum length of sysfs access_state string */
+#define ALUA_ACCESS_STATE_SIZE 32
+
 static const struct {
 	unsigned char value;
 	char *name;
@@ -39,11 +42,11 @@ int get_exclusive_pref_arg(char *args)
 int getprio (struct path * pp, char *args)
 {
 	int prio = 0, rc, i;
-	char buff[512];
+	char buff[ALUA_ACCESS_STATE_SIZE];
 	int exclusive_pref;
 
 	exclusive_pref = get_exclusive_pref_arg(args);
-	rc = sysfs_get_asymmetric_access_state(pp, buff, 512);
+	rc = sysfs_get_asymmetric_access_state(pp, buff, sizeof(buff));
 	if (rc < 0)
 		return PRIO_UNDEF;
 	prio = 0;
diff --git a/libmultipath/propsel.c b/libmultipath/propsel.c
index 56895c4b..e421abba 100644
--- a/libmultipath/propsel.c
+++ b/libmultipath/propsel.c
@@ -262,6 +262,7 @@ verify_alua_prio(struct multipath *mp)
 		if (!prio_selected(&pp->prio))
 			continue;
 		if (strncmp(name, PRIO_ALUA, PRIO_NAME_LEN) &&
+		    strncmp(name, PRIO_ALUA_CACHED, PRIO_NAME_LEN) &&
 		    strncmp(name, PRIO_SYSFS, PRIO_NAME_LEN))
 			 return false;
 		assume_alua = true;
@@ -690,7 +691,12 @@ int select_checker(struct config *conf, struct path *pp)
 		}
 		(void)path_get_tpgs(pp);
 		if (pp->tpgs != TPGS_NONE && pp->tpgs != TPGS_UNDEF) {
-			ckr_name = TUR;
+			/*
+			 * For ALUA devices (any TPGS support), use alua
+			 * checker which combines path checking with ALUA state
+			 * caching for optimal performance.
+			 */
+			ckr_name = ALUA;
 			goto out;
 		}
 	}
@@ -767,8 +773,15 @@ void detect_prio(struct path *pp)
 		if ((tpgs == TPGS_EXPLICIT || !check_rdac(pp)) &&
 		    sysfs_get_asymmetric_access_state(pp, buff, 512) >= 0)
 			default_prio = PRIO_SYSFS;
-		else
+		else {
 			default_prio = PRIO_ALUA;
+			/* If using alua checker, use cached prioritizer */
+			if (checker_selected(&pp->checker)) {
+				const char *chk_name = checker_name(&pp->checker);
+				if (chk_name && strcmp(chk_name, ALUA) == 0)
+					default_prio = PRIO_ALUA_CACHED;
+			}
+		}
 		break;
 	default:
 		return;
@@ -831,7 +844,8 @@ out:
 	/*
 	 * fetch tpgs mode for alua, if its not already obtained
 	 */
-	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN)) {
+	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN) ||
+	    !strncmp(prio_name(p), PRIO_ALUA_CACHED, PRIO_NAME_LEN)) {
 		int tpgs = path_get_tpgs(pp);
 
 		if (tpgs == TPGS_NONE) {
diff --git a/libmultipath/structs.c b/libmultipath/structs.c
index f4413127..6483a568 100644
--- a/libmultipath/structs.c
+++ b/libmultipath/structs.c
@@ -126,6 +126,8 @@ alloc_path (void)
 		pp->fd = -1;
 		pp->tpgs = TPGS_UNDEF;
 		pp->tpg_id = GROUP_ID_UNDEF;
+		pp->alua_state = -1;
+		pp->alua_timestamp = 0;
 		pp->priority = PRIO_UNDEF;
 		pp->checkint = CHECKINT_UNDEF;
 		checker_clear(&pp->checker);
diff --git a/libmultipath/structs.h b/libmultipath/structs.h
index 9adedde2..c4bddee9 100644
--- a/libmultipath/structs.h
+++ b/libmultipath/structs.h
@@ -444,6 +444,9 @@ struct path {
 	vector hwe;
 	struct gen_path generic_path;
 	int tpg_id;
+	/* Cached ALUA state from checker for prioritizer use */
+	int alua_state;		/* AAS value, -1 if not cached */
+	time_t alua_timestamp;	/* When alua_state was last updated */
 	enum ioctl_info_states ioctl_info;
 };
 
-- 
2.50.1 (Apple Git-155)

Re: [PATCH 1/1] checkers: add alua path checker

[Martin Wilck]

Hi Brian,

On Wed, 2026-03-11 at 17:16 -0700, Brian Bunker wrote:
> Add an 'alua' path checker that uses RTPG to check path state, and an
> 'alua_cached' prioritizer that uses the checker's cached ALUA state.
> 
> The checker maps ALUA Asymmetric Access States to path states:
>   - AAS_OPTIMIZED, AAS_NON_OPTIMIZED -> PATH_UP
>   - AAS_STANDBY -> PATH_GHOST
>   - AAS_TRANSITIONING -> PATH_PENDING
>   - AAS_UNAVAILABLE, AAS_OFFLINE -> PATH_DOWN
> 
> The alua_cached prioritizer first checks for cached state from the
> alua
> checker, then falls back to sysfs, then to issuing RTPG directly.
> This
> maintains backward compatibility with other checker configurations.
> 
> When detect_checker is enabled, the alua checker is auto-selected for
> devices with TPGS support. When the alua checker is in use,
> detect_prio
> auto-selects the alua_cached prioritizer.
> 
> Signed-off-by: Brian Bunker <brian@purestorage.com>

Thanks! Just a few remarks for now.

Martin

> ---
>  libmultipath/Makefile                   |   5 +
>  libmultipath/checkers.c                 |   1 +
>  libmultipath/checkers.h                 |   1 +
>  libmultipath/checkers/Makefile          |   3 +-
>  libmultipath/checkers/alua.c            | 426
> ++++++++++++++++++++++++
>  libmultipath/checkers/alua.h            |  15 +
>  libmultipath/prio.c                     |   1 +
>  libmultipath/prio.h                     |   1 +
>  libmultipath/prioritizers/Makefile      |   1 +
>  libmultipath/prioritizers/alua_cached.c | 224 +++++++++++++
>  libmultipath/prioritizers/alua_rtpg.c   |  12 +-
>  libmultipath/prioritizers/alua_rtpg.h   |   1 +
>  libmultipath/prioritizers/sysfs.c       |   7 +-
>  libmultipath/propsel.c                  |  20 +-
>  libmultipath/structs.c                  |   2 +
>  libmultipath/structs.h                  |   3 +
>  16 files changed, 716 insertions(+), 7 deletions(-)
>  create mode 100644 libmultipath/checkers/alua.c
>  create mode 100644 libmultipath/checkers/alua.h
>  create mode 100644 libmultipath/prioritizers/alua_cached.c
> 
> diff --git a/libmultipath/Makefile b/libmultipath/Makefile
> index 85767ab4..379f4d02 100644
> --- a/libmultipath/Makefile
> +++ b/libmultipath/Makefile
> @@ -38,6 +38,11 @@ nvme-lib.o: nvme-lib.c nvme-ioctl.c nvme-ioctl.h
>  dict.o:	dict.c
>  	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -Wno-unused-parameter -c -o
> $@ $<
>  
> +# checkers/alua.o needs to find headers in parent directory
> +checkers/alua.o: checkers/alua.c
> +	@echo building $@ because of $?
> +	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -I. -c -o $@ $<
> +
>  make_static = $(shell sed '/^static/!s/^\([a-z]\{1,\} \)/static \1/'
> <$1 >$2)
>  
>  nvme-ioctl.c: nvme/nvme-ioctl.c
> diff --git a/libmultipath/checkers.c b/libmultipath/checkers.c
> index bb6ad1ee..21aa52d4 100644
> --- a/libmultipath/checkers.c
> +++ b/libmultipath/checkers.c
> @@ -461,6 +461,7 @@ int init_checkers(void)
>  		EMC_CLARIION,
>  		READSECTOR0,
>  		CCISS_TUR,
> +		ALUA_RTPG,
>  	};
>  	unsigned int i;
>  
> diff --git a/libmultipath/checkers.h b/libmultipath/checkers.h
> index a969e7d1..2905f3fc 100644
> --- a/libmultipath/checkers.h
> +++ b/libmultipath/checkers.h
> @@ -98,6 +98,7 @@ enum path_check_state {
>  #define EMC_CLARIION "emc_clariion"
>  #define READSECTOR0  "readsector0"
>  #define CCISS_TUR    "cciss_tur"
> +#define ALUA         "alua"
>  #define NONE         "none"
>  #define INVALID      "invalid"
>  
> diff --git a/libmultipath/checkers/Makefile
> b/libmultipath/checkers/Makefile
> index 6f7cfb95..dfdd13a1 100644
> --- a/libmultipath/checkers/Makefile
> +++ b/libmultipath/checkers/Makefile
> @@ -17,7 +17,8 @@ LIBS= \
>  	libcheckdirectio.so \
>  	libcheckemc_clariion.so \
>  	libcheckhp_sw.so \
> -	libcheckrdac.so
> +	libcheckrdac.so \
> +	libcheckalua.so
>  
>  all: $(LIBS)
>  
> diff --git a/libmultipath/checkers/alua.c
> b/libmultipath/checkers/alua.c
> new file mode 100644
> index 00000000..cb3d7000
> --- /dev/null
> +++ b/libmultipath/checkers/alua.c
> @@ -0,0 +1,426 @@
> +/*
> + * ALUA Path Checker
> + *
> + * Copyright (c) 2024
> + * This file is released under the GPL.

Please add a proper license header.

> + */
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <string.h>
> +#include <sys/types.h>
> +#include <sys/stat.h>
> +#include <sys/sysmacros.h>
> +#include <unistd.h>
> +#include <fcntl.h>
> +#include <errno.h>
> +#include <pthread.h>
> +#include <urcu.h>
> +#include <urcu/uatomic.h>
> +
> +#include "checkers.h"
> +#include "debug.h"
> +#include "structs.h"
> +#include "prio.h"
> +#include "util.h"
> +#include "time-util.h"
> +#include "../prioritizers/alua.h"
> +#include "../prioritizers/alua_rtpg.h"
> +#include "alua.h"
> +
> +#define MAX_NR_TIMEOUTS 1
> +
> +enum {
> +	MSG_ALUA_RTPG_NOT_SUPPORTED = CHECKER_FIRST_MSGID,
> +	MSG_ALUA_RTPG_TRANSITIONING,
> +	MSG_ALUA_RUNNING,
> +	MSG_ALUA_TIMEOUT,
> +};
> +
> +#define IDX_(x) (MSG_ ## x - CHECKER_FIRST_MSGID)
> +const char *libcheck_msgtable[] = {
> +	[IDX_(ALUA_RTPG_NOT_SUPPORTED)] = " ALUA not supported",
> +	[IDX_(ALUA_RTPG_TRANSITIONING)] = " path is transitioning",
> +	[IDX_(ALUA_RUNNING)] = " still running",
> +	[IDX_(ALUA_TIMEOUT)] = " timed out",
> +	NULL,
> +};
> +
> +struct alua_checker_context {
> +	dev_t devt;
> +	int state;
> +	int running; /* uatomic access only */
> +	int fd;
> +	unsigned int timeout;
> +	time_t time;
> +	pthread_t thread;
> +	pthread_mutex_t lock;
> +	pthread_cond_t active;
> +	int holders; /* uatomic access only */
> +	int msgid;
> +	struct checker_context ctx;
> +	unsigned int nr_timeouts;
> +	bool checked_state;
> +	/* ALUA-specific cached data for prioritizer */
> +	int aas;           /* Asymmetric Access State */
> +	int tpg;           /* Target Port Group */
> +	time_t timestamp;  /* When this data was collected */
> +	/* Data needed for RTPG in thread */
> +	struct path *pp;   /* Path pointer - only valid during check
> */
> +};

Instead of duplicating the threading implementation of the TUR checker
here we should abstract it out and use the common framework by both
checkers (and others, too, maybe). This is a thing I've wanted to do
for a long time. Maybe it's time to start doing it now.




> +
> +int libcheck_init(struct checker *c)
> +{
> +	struct alua_checker_context *ct;
> +	struct stat sb;
> +
> +	ct = malloc(sizeof(struct alua_checker_context));
> +	if (!ct)
> +		return 1;
> +	memset(ct, 0, sizeof(struct alua_checker_context));
> +
> +	ct->state = PATH_UNCHECKED;
> +	ct->fd = -1;
> +	ct->aas = -1;
> +	ct->tpg = -1;
> +	uatomic_set(&ct->holders, 1);
> +	pthread_cond_init_mono(&ct->active);
> +	pthread_mutex_init(&ct->lock, NULL);
> +	if (fstat(c->fd, &sb) == 0)
> +		ct->devt = sb.st_rdev;
> +	ct->ctx.cls = c->cls;
> +	c->context = ct;
> +
> +	return 0;
> +}
> +
> +static void cleanup_context(struct alua_checker_context *ct)
> +{
> +	pthread_mutex_destroy(&ct->lock);
> +	pthread_cond_destroy(&ct->active);
> +	free(ct);
> +}
> +
> +void libcheck_free(struct checker *c)
> +{
> +	if (c->context) {
> +		struct alua_checker_context *ct = c->context;
> +		int holders;
> +		int running;
> +
> +		running = uatomic_xchg(&ct->running, 0);
> +		if (running)
> +			pthread_cancel(ct->thread);
> +		ct->thread = 0;
> +		holders = uatomic_sub_return(&ct->holders, 1);
> +		if (!holders)
> +			cleanup_context(ct);
> +		c->context = NULL;
> +	}
> +	return;
> +}
> +
> +/*
> + * Map ALUA Asymmetric Access State to path state
> + */
> +static int alua_state_to_path_state(int aas, short *msgid)
> +{
> +	switch (aas & 0x0f) {
> +	case AAS_OPTIMIZED:
> +	case AAS_NON_OPTIMIZED:
> +		*msgid = CHECKER_MSGID_UP;
> +		return PATH_UP;
> +	case AAS_STANDBY:
> +		*msgid = CHECKER_MSGID_GHOST;
> +		return PATH_GHOST;
> +	case AAS_TRANSITIONING:
> +		*msgid = MSG_ALUA_RTPG_TRANSITIONING;
> +		return PATH_PENDING;
> +	case AAS_UNAVAILABLE:
> +	case AAS_OFFLINE:
> +	default:
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +}
> +
> +/*
> + * Main ALUA check function - uses alua_rtpg.c library.
> + * This is called either in sync mode or from the async thread.
> + */
> +static int
> +alua_check(struct path *pp, int *aas_out, short *msgid)
> +{
> +	int aas;
> +
> +	if (pp->tpg_id == GROUP_ID_UNDEF) {
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +
> +	aas = get_asymmetric_access_state(pp, pp->tpg_id);
> +	if (aas < 0) {
> +		if (aas == -RTPG_LUN_DISCONNECTED) {
> +			*msgid = CHECKER_MSGID_DISCONNECTED;
> +			return PATH_DISCONNECTED;
> +		}
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +
> +	*aas_out = aas;
> +	return alua_state_to_path_state(aas, msgid);
> +}
> +
> +#define alua_thread_cleanup_push(ct)
> pthread_cleanup_push(cleanup_func, ct)
> +#define alua_thread_cleanup_pop(ct) pthread_cleanup_pop(1)
> +
> +static void cleanup_func(void *data)
> +{
> +	int holders;
> +	struct alua_checker_context *ct = data;
> +
> +	holders = uatomic_sub_return(&ct->holders, 1);
> +	if (!holders)
> +		cleanup_context(ct);
> +}
> +
> +void *libcheck_thread(struct checker_context *ctx)
> +{
> +	struct alua_checker_context *ct =
> +		container_of(ctx, struct alua_checker_context, ctx);
> +	int state, running;
> +	short msgid;
> +	int aas;
> +
> +	/* This thread can be canceled, so setup clean up */
> +	alua_thread_cleanup_push(ct);
> +
> +	condlog(4, "%d:%d : alua checker starting up", major(ct-
> >devt),
> +		minor(ct->devt));
> +
> +	state = alua_check(ct->pp, &aas, &msgid);
> +	pthread_testcancel();
> +
> +	/* ALUA checker done */
> +	pthread_mutex_lock(&ct->lock);
> +	ct->state = state;
> +	ct->msgid = msgid;
> +	ct->aas = aas;
> +	ct->tpg = ct->pp->tpg_id;
> +	ct->timestamp = time(NULL);
> +	/* Update path-level cache for prioritizer use */
> +	ct->pp->alua_state = aas;
> +	ct->pp->alua_timestamp = ct->timestamp;
> +	pthread_cond_signal(&ct->active);
> +	pthread_mutex_unlock(&ct->lock);
> +
> +	condlog(4, "%d:%d : alua checker finished, state %s",
> major(ct->devt),
> +		minor(ct->devt), checker_state_name(state));
> +
> +	running = uatomic_xchg(&ct->running, 0);
> +	if (!running)
> +		pause();
> +
> +	alua_thread_cleanup_pop(ct);
> +
> +	return ((void *)0);
> +}
> +
> +static void alua_set_async_timeout(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct timespec now;
> +
> +	get_monotonic_time(&now);
> +	ct->time = now.tv_sec + c->timeout;
> +}
> +
> +static int alua_check_async_timeout(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct timespec now;
> +
> +	get_monotonic_time(&now);
> +	return (now.tv_sec > ct->time);
> +}
> +
> +static int check_pending(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	int alua_status = PATH_PENDING;
> +
> +	pthread_mutex_lock(&ct->lock);
> +	if (ct->state != PATH_PENDING || ct->msgid !=
> MSG_ALUA_RUNNING) {
> +		alua_status = ct->state;
> +		c->msgid = ct->msgid;
> +	}
> +	pthread_mutex_unlock(&ct->lock);
> +	if (alua_status == PATH_PENDING && c->msgid ==
> MSG_ALUA_RUNNING) {
> +		condlog(4, "%d:%d : alua checker still running",
> +			major(ct->devt), minor(ct->devt));
> +	} else {
> +		int running = uatomic_xchg(&ct->running, 0);
> +		if (running)
> +			pthread_cancel(ct->thread);
> +		ct->thread = 0;
> +	}
> +
> +	ct->checked_state = true;
> +	return alua_status;
> +}
> +
> +bool libcheck_need_wait(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	return (ct && ct->thread && uatomic_read(&ct->running) != 0
> &&
> +		!ct->checked_state);
> +}
> +
> +int libcheck_pending(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +
> +	/* The if path checker isn't running, just return the
> exiting value. */
> +	if (!ct || !ct->thread)
> +		return c->path_state;
> +
> +	return check_pending(c);
> +}
> +
> +int libcheck_check(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct path *pp = container_of(c, struct path, checker);
> +	pthread_attr_t attr;
> +	int alua_status = PATH_PENDING, r;
> +	int aas;
> +
> +	if (!ct)
> +		return PATH_UNCHECKED;
> +
> +	if (checker_is_sync(c)) {
> +		alua_status = alua_check(pp, &aas, &c->msgid);
> +		ct->tpg = pp->tpg_id;
> +		ct->aas = aas;
> +		ct->timestamp = time(NULL);
> +		/* Update path-level cache for prioritizer use */
> +		pp->alua_state = aas;
> +		pp->alua_timestamp = ct->timestamp;
> +		return alua_status;
> +	}
> +
> +	/* Async mode */
> +	if (ct->thread) {
> +		ct->checked_state = true;
> +		if (alua_check_async_timeout(c)) {
> +			int running = uatomic_xchg(&ct->running, 0);
> +			if (running) {
> +				pthread_cancel(ct->thread);
> +				condlog(3, "%d:%d : alua checker
> timeout",
> +					major(ct->devt), minor(ct-
> >devt));
> +				c->msgid = MSG_ALUA_TIMEOUT;
> +				alua_status = PATH_TIMEOUT;
> +			} else {
> +				pthread_mutex_lock(&ct->lock);
> +				alua_status = ct->state;
> +				c->msgid = ct->msgid;
> +				pthread_mutex_unlock(&ct->lock);
> +			}
> +			ct->thread = 0;
> +		} else if (uatomic_read(&ct->running) != 0) {
> +			/*
> +			 * ct->running may be stale - the thread
> sets
> +			 * msgid/state under lock, then clears
> running
> +			 * after releasing it. Check msgid under
> lock
> +			 * to get the truth. Only MSG_ALUA_RUNNING
> means
> +			 * no result yet.
> +			 */
> +			pthread_mutex_lock(&ct->lock);
> +			if (ct->msgid != MSG_ALUA_RUNNING) {
> +				alua_status = ct->state;
> +				c->msgid = ct->msgid;
> +				ct->thread = 0;
> +			}
> +			pthread_mutex_unlock(&ct->lock);
> +		} else {
> +			/* ALUA checker done */
> +			ct->thread = 0;
> +			pthread_mutex_lock(&ct->lock);
> +			alua_status = ct->state;
> +			c->msgid = ct->msgid;
> +			pthread_mutex_unlock(&ct->lock);
> +		}
> +	} else {
> +		if (uatomic_read(&ct->holders) > 1) {
> +			/* The thread has been cancelled but hasn't
> quit. */
> +			if (ct->nr_timeouts <= MAX_NR_TIMEOUTS) {
> +				if (ct->nr_timeouts ==
> MAX_NR_TIMEOUTS) {
> +					condlog(2, "%d:%d : waiting
> for stalled alua thread to finish",
> +						major(ct->devt),
> minor(ct->devt));
> +				}
> +				ct->nr_timeouts++;
> +				c->msgid = MSG_ALUA_TIMEOUT;
> +				return PATH_TIMEOUT;
> +			}
> +			ct->nr_timeouts++;
> +			/*
> +			 * Start a new thread while the old one is
> stalled.
> +			 * We have to prevent it from interfering
> with the new
> +			 * thread. We create a new context and leave
> the old
> +			 * one with the stale thread, hoping it will
> clean up
> +			 * eventually.
> +			 */
> +			condlog(3, "%d:%d : alua thread not
> responding",
> +				major(ct->devt), minor(ct->devt));
> +
> +			if (libcheck_init(c) != 0) {
> +				c->msgid = CHECKER_MSGID_DOWN;
> +				return PATH_UNCHECKED;
> +			}
> +			((struct alua_checker_context *)c->context)-
> >nr_timeouts = ct->nr_timeouts;
> +
> +			if (!uatomic_sub_return(&ct->holders, 1)) {
> +				/* It did terminate, eventually */
> +				cleanup_context(ct);
> +				((struct alua_checker_context *)c-
> >context)->nr_timeouts = 0;
> +			}
> +
> +			ct = c->context;
> +		} else
> +			ct->nr_timeouts = 0;
> +
> +		/* Start new ALUA checker */
> +		pthread_mutex_lock(&ct->lock);
> +		alua_status = ct->state = PATH_PENDING;
> +		c->msgid = ct->msgid = MSG_ALUA_RUNNING;
> +		pthread_mutex_unlock(&ct->lock);
> +		ct->fd = c->fd;
> +		ct->timeout = c->timeout;
> +		ct->pp = pp;
> +		ct->checked_state = false;
> +		uatomic_add(&ct->holders, 1);
> +		uatomic_set(&ct->running, 1);
> +		alua_set_async_timeout(c);
> +		setup_thread_attr(&attr, 32 * 1024, 1);
> +		r = start_checker_thread(&ct->thread, &attr, &ct-
> >ctx);
> +		pthread_attr_destroy(&attr);
> +		if (r) {
> +			uatomic_sub(&ct->holders, 1);
> +			uatomic_set(&ct->running, 0);
> +			ct->thread = 0;
> +			condlog(3, "%d:%d : failed to start alua
> thread, using"
> +				" sync mode", major(ct->devt),
> minor(ct->devt));
> +			alua_status = alua_check(pp, &aas, &c-
> >msgid);
> +			ct->tpg = pp->tpg_id;
> +			ct->aas = aas;
> +			ct->timestamp = time(NULL);
> +			/* Update path-level cache for prioritizer
> use */
> +			pp->alua_state = aas;
> +			pp->alua_timestamp = ct->timestamp;
> +			return alua_status;
> +		}
> +	}
> +
> +	return alua_status;
> +}
> +
> diff --git a/libmultipath/checkers/alua.h
> b/libmultipath/checkers/alua.h
> new file mode 100644
> index 00000000..b53cef66
> --- /dev/null
> +++ b/libmultipath/checkers/alua.h
> @@ -0,0 +1,15 @@
> +/*
> + * ALUA Path Checker Header
> + *
> + * This file is released under the GPL.

Please add a proper license header.

> + */
> +#ifndef _ALUA_CHECKER_H
> +#define _ALUA_CHECKER_H
> +
> +/* ALUA checker caches state in struct path for prioritizer use:
> + *   pp->alua_state     - cached AAS value (-1 if not cached)
> + *   pp->alua_timestamp - when the state was last updated
> + */
> +
> +#endif /* _ALUA_CHECKER_H */
> +

Only comments in this file?


> diff --git a/libmultipath/prio.c b/libmultipath/prio.c
> index 24f825bd..2d59d04e 100644
> --- a/libmultipath/prio.c
> +++ b/libmultipath/prio.c
> @@ -29,6 +29,7 @@ int init_prio(void)
>  #ifdef LOAD_ALL_SHARED_LIBS
>  	static const char *const all_prios[] = {
>  		PRIO_ALUA,
> +		PRIO_ALUA_CACHED,
>  		PRIO_CONST,
>  		PRIO_DATACORE,
>  		PRIO_EMC,
> diff --git a/libmultipath/prio.h b/libmultipath/prio.h
> index 119b75f2..b54ece0d 100644
> --- a/libmultipath/prio.h
> +++ b/libmultipath/prio.h
> @@ -17,6 +17,7 @@ struct path;
>   * Known prioritizers for use in hwtable.c
>   */
>  #define PRIO_ALUA		"alua"
> +#define PRIO_ALUA_CACHED	"alua_cached"
>  #define PRIO_CONST		"const"
>  #define PRIO_DATACORE		"datacore"
>  #define PRIO_EMC		"emc"
> diff --git a/libmultipath/prioritizers/Makefile
> b/libmultipath/prioritizers/Makefile
> index ff2524c2..201c52c0 100644
> --- a/libmultipath/prioritizers/Makefile
> +++ b/libmultipath/prioritizers/Makefile
> @@ -13,6 +13,7 @@ LIBDEPS = -lmultipath -lmpathutil -lm -lpthread -
> lrt
>  # If you add or remove a prioritizer also update
> multipath/multipath.conf.5
>  LIBS = \
>  	libprioalua.so \
> +	libprioalua_cached.so \
>  	libprioconst.so \
>  	libpriodatacore.so \
>  	libprioemc.so \

I think the below should go into a separate patch.

> diff --git a/libmultipath/prioritizers/alua_cached.c
> b/libmultipath/prioritizers/alua_cached.c
> new file mode 100644
> index 00000000..2ba5b2e7
> --- /dev/null
> +++ b/libmultipath/prioritizers/alua_cached.c
> @@ -0,0 +1,224 @@
> +/*
> + * (C) Copyright IBM Corp. 2004, 2005   All Rights Reserved.
> + * Modified 2024 to support cached ALUA state from checker
> + *
> + * ALUA prioritizer that can use cached state from alua checker
> + * to avoid duplicate RTPG commands.
> + *
> + * This file is released under the GPL.

Please add a proper license header.

> + */
> +#include <stdio.h>
> +#include <string.h>
> +
> +#include <time.h>
> +
> +#include "debug.h"
> +#include "prio.h"
> +#include "structs.h"
> +#include "checkers.h"
> +#include "discovery.h"
> +
> +#include "alua.h"
> +
> +#define ALUA_PRIO_NOT_SUPPORTED			1
> +#define ALUA_PRIO_RTPG_FAILED			2
> +#define ALUA_PRIO_GETAAS_FAILED			3
> +#define ALUA_PRIO_TPGS_FAILED			4
> +#define ALUA_PRIO_NO_INFORMATION		5
> +#define ALUA_PRIO_CACHED_STALE			6
> +
> +/* Maximum age of cached ALUA state in seconds */
> +#define ALUA_CACHE_MAX_AGE_SECS			5
> +
> +static const char * aas_string[] = {
> +	[AAS_OPTIMIZED]		= "active/optimized",
> +	[AAS_NON_OPTIMIZED]	= "active/non-optimized",
> +	[AAS_STANDBY]		= "standby",
> +	[AAS_UNAVAILABLE]	= "unavailable",
> +	[AAS_LBA_DEPENDENT]	= "logical block dependent",
> +	[AAS_RESERVED]		= "ARRAY BUG: invalid TPGs state!",
> +	[AAS_OFFLINE]		= "offline",
> +	[AAS_TRANSITIONING]	= "transitioning between states",
> +};

We shouldn't duplicate the definition of these constants.

> +
> +static const char *aas_print_string(int rc)
> +{
> +	rc &= 0x7f;
> +
> +	if (rc & 0x70)
> +		return aas_string[AAS_RESERVED];
> +	rc &= 0x0f;
> +	if (rc > AAS_RESERVED && rc < AAS_OFFLINE)
> +		return aas_string[AAS_RESERVED];
> +	else
> +		return aas_string[rc];
> +}
> +
> +/*
> + * Try to get ALUA info with optimization:
> + * 1. If using alua checker: use its cached state (it just polled!)
> + * 2. Try sysfs (kernel maintains this, no I/O needed)
> + * 3. Fall back to issuing RTPG directly
> + *
> + * Rationale: The checker is our active poller. If we're using alua
> + * and its cache is fresh, we trust that data first. If the cache
> misses
> + * (stale or empty), we try sysfs before issuing RTPG. This provides
> a
> + * three-tier optimization: checker cache → sysfs → RTPG.
> + */
> +/* Maximum length of sysfs access_state string */
> +#define ALUA_ACCESS_STATE_SIZE 32
> +
> +int
> +get_alua_info_cached(struct path * pp, bool *used_cache)
> +{
> +	int	rc;
> +	int	tpg;
> +	bool	diff_tpg;
> +	char	buff[ALUA_ACCESS_STATE_SIZE];
> +
> +	*used_cache = false;
> +
> +	/* First: if using alua checker, use its fresh data from
> path structure */
> +	if (checker_selected(&pp->checker)) {
> +		const char *chk_name = checker_name(&pp->checker);
> +
> +		if (chk_name && strcmp(chk_name, "alua") == 0 &&
> +		    pp->alua_state >= 0) {
> +			time_t now = time(NULL);
> +			time_t age = now - pp->alua_timestamp;
> +
> +			if (pp->alua_timestamp > 0 && age <=
> ALUA_CACHE_MAX_AGE_SECS) {
> +				condlog(4, "%s: using cached ALUA
> state from checker (fresh poll)", pp->dev);
> +				*used_cache = true;
> +				return pp->alua_state;
> +			}
> +			condlog(4, "%s: cached ALUA state not
> available or stale, trying sysfs",
> +				pp->dev);
> +			/* Cache miss - try sysfs before issuing
> RTPG */
> +		}
> +	}
> +
> +	/* Second: try sysfs (if cache missed or not using alua
> checker) */
> +	rc = sysfs_get_asymmetric_access_state(pp, buff,
> sizeof(buff));
> +	if (rc >= 0) {
> +		/* Parse sysfs state string to ALUA state value */
> +		int aas = -1;
> +		int priopath = rc; /* rc contains preferred bit */
> +
> +		if (!strncmp(buff, "active/optimized", 16))
> +			aas = AAS_OPTIMIZED;
> +		else if (!strncmp(buff, "active/non-optimized", 20))
> +			aas = AAS_NON_OPTIMIZED;
> +		else if (!strncmp(buff, "lba-dependent", 13))
> +			aas = AAS_LBA_DEPENDENT;
> +		else if (!strncmp(buff, "standby", 7))
> +			aas = AAS_STANDBY;
> +
> +		if (aas >= 0) {
> +			condlog(4, "%s: using sysfs ALUA state (no
> I/O)", pp->dev);
> +			*used_cache = true;
> +			return (priopath ? 0x80 : 0) | aas;
> +		}
> +	}
> +
> +	/* Last resort: issue RTPG directly */
> +	tpg = get_target_port_group(pp);
> +	if (tpg < 0) {
> +		rc = get_target_port_group_support(pp);
> +		if (rc < 0)
> +			return -ALUA_PRIO_TPGS_FAILED;
> +		if (rc == TPGS_NONE)
> +			return -ALUA_PRIO_NOT_SUPPORTED;
> +		return -ALUA_PRIO_RTPG_FAILED;
> +	}
> +	diff_tpg = (pp->tpg_id != GROUP_ID_UNDEF && pp->tpg_id !=
> tpg);
> +	pp->tpg_id = tpg;
> +	condlog((diff_tpg) ? 2 : 4, "%s: reported target port group
> is %i",
> +		pp->dev, tpg);
> +	rc = get_asymmetric_access_state(pp, tpg);
> +	if (rc < 0) {
> +		condlog(2, "%s: get_asymmetric_access_state returned
> %d",
> +			__func__, rc);
> +		return -ALUA_PRIO_GETAAS_FAILED;
> +	}
> +
> +	condlog(3, "%s: aas = %02x [%s]%s", pp->dev, rc,
> aas_print_string(rc),
> +		(rc & 0x80) ? " [preferred]" : "");
> +	return rc;
> +}
> +
> +int get_exclusive_pref_arg(char *args)
> +{
> +	char *ptr;
> +
> +	if (args == NULL)
> +		return 0;
> +	ptr = strstr(args, "exclusive_pref_bit");
> +	if (!ptr)
> +		return 0;
> +	if (ptr[18] != '\0' && ptr[18] != ' ' && ptr[18] != '\t')
> +		return 0;
> +	if (ptr != args && ptr[-1] != ' ' && ptr[-1] != '\t')
> +		return 0;
> +	return 1;
> +}
> +
> +int getprio (struct path * pp, char * args)
> +{
> +	int rc;
> +	int aas;
> +	int priopath;
> +	int exclusive_pref;
> +	bool used_cache = false;
> +
> +	if (pp->fd < 0)
> +		return -ALUA_PRIO_NO_INFORMATION;
> +
> +	exclusive_pref = get_exclusive_pref_arg(args);
> +	rc = get_alua_info_cached(pp, &used_cache);
> +
> +	if (used_cache) {
> +		condlog(4, "%s: priority calculated from cached ALUA
> state (no RTPG issued)",
> +			pp->dev);
> +	}
> +
> +	if (rc >= 0) {
> +		aas = (rc & 0x0f);
> +		priopath = (rc & 0x80);
> +		switch(aas) {
> +		case AAS_OPTIMIZED:
> +			rc = 50;
> +			break;
> +		case AAS_NON_OPTIMIZED:
> +			rc = 10;
> +			break;
> +		case AAS_LBA_DEPENDENT:
> +			rc = 5;
> +			break;
> +		case AAS_STANDBY:
> +			rc = 1;
> +			break;
> +		default:
> +			rc = 0;
> +		}
> +		if (priopath && (aas != AAS_OPTIMIZED ||
> exclusive_pref))
> +			rc += 80;
> +	} else {
> +		switch(-rc) {
> +		case ALUA_PRIO_NOT_SUPPORTED:
> +			condlog(0, "%s: alua not supported", pp-
> >dev);
> +			break;
> +		case ALUA_PRIO_RTPG_FAILED:
> +			condlog(0, "%s: couldn't get target port
> group", pp->dev);
> +			break;
> +		case ALUA_PRIO_GETAAS_FAILED:
> +			condlog(0, "%s: couldn't get asymmetric
> access state", pp->dev);
> +			break;
> +		case ALUA_PRIO_TPGS_FAILED:
> +			condlog(3, "%s: couldn't get supported alua
> states", pp->dev);
> +			break;
> +		}
> +	}
> +	return rc;
> +}
> +
> diff --git a/libmultipath/prioritizers/alua_rtpg.c
> b/libmultipath/prioritizers/alua_rtpg.c
> index 053cccb7..5da06b50 100644
> --- a/libmultipath/prioritizers/alua_rtpg.c
> +++ b/libmultipath/prioritizers/alua_rtpg.c
> @@ -75,6 +75,7 @@ enum scsi_disposition {
>  	SCSI_GOOD = 0,
>  	SCSI_ERROR,
>  	SCSI_RETRY,
> +	SCSI_DISCONNECTED,
>  };
>  
>  static int
> @@ -124,6 +125,12 @@ scsi_error(struct sg_io_hdr *hdr, int opcode)
>  	PRINT_DEBUG("alua: SCSI error for command %02x: status %02x,
> sense %02x/%02x/%02x",
>  		    opcode, hdr->status, sense_key, asc, ascq);
>  
> +	/* Check for disconnected LUN (unmapped at target) */
> +	if (sense_key == 0x5 && asc == 0x25 && ascq == 0x00) {
> +		/* ILLEGAL REQUEST / LOGICAL UNIT NOT SUPPORTED */
> +		return SCSI_DISCONNECTED;
> +	}
> +
>  	if (sense_key == UNIT_ATTENTION || sense_key == NOT_READY)
>  		return SCSI_RETRY;
>  	else
> @@ -325,7 +332,10 @@ retry:
>  	}
>  
>  	rc = scsi_error(&hdr, OPERATION_CODE_RTPG);
> -	if (rc == SCSI_ERROR) {
> +	if (rc == SCSI_DISCONNECTED) {
> +		PRINT_DEBUG("do_rtpg: LUN disconnected (unmapped at
> target)!");
> +		return -RTPG_LUN_DISCONNECTED;
> +	} else if (rc == SCSI_ERROR) {
>  		PRINT_DEBUG("do_rtpg: SCSI error!");
>  		return -RTPG_RTPG_FAILED;
>  	} else if (rc == SCSI_RETRY) {
> diff --git a/libmultipath/prioritizers/alua_rtpg.h
> b/libmultipath/prioritizers/alua_rtpg.h
> index ff6ec0ef..030ca58d 100644
> --- a/libmultipath/prioritizers/alua_rtpg.h
> +++ b/libmultipath/prioritizers/alua_rtpg.h
> @@ -21,6 +21,7 @@
>  #define RTPG_NO_TPG_IDENTIFIER			2
>  #define RTPG_RTPG_FAILED			3
>  #define RTPG_TPG_NOT_FOUND			4
> +#define RTPG_LUN_DISCONNECTED			5
>  
>  int get_target_port_group_support(const struct path *pp);
>  int get_target_port_group(const struct path *pp);
> diff --git a/libmultipath/prioritizers/sysfs.c
> b/libmultipath/prioritizers/sysfs.c
> index 5e8adc05..ab058ea9 100644
> --- a/libmultipath/prioritizers/sysfs.c
> +++ b/libmultipath/prioritizers/sysfs.c
> @@ -10,6 +10,9 @@
>  #include "discovery.h"
>  #include "prio.h"
>  
> +/* Maximum length of sysfs access_state string */
> +#define ALUA_ACCESS_STATE_SIZE 32
> +
>  static const struct {
>  	unsigned char value;
>  	char *name;
> @@ -39,11 +42,11 @@ int get_exclusive_pref_arg(char *args)
>  int getprio (struct path * pp, char *args)
>  {
>  	int prio = 0, rc, i;
> -	char buff[512];
> +	char buff[ALUA_ACCESS_STATE_SIZE];
>  	int exclusive_pref;
>  
>  	exclusive_pref = get_exclusive_pref_arg(args);
> -	rc = sysfs_get_asymmetric_access_state(pp, buff, 512);
> +	rc = sysfs_get_asymmetric_access_state(pp, buff,
> sizeof(buff));
>  	if (rc < 0)
>  		return PRIO_UNDEF;
>  	prio = 0;
> diff --git a/libmultipath/propsel.c b/libmultipath/propsel.c
> index 56895c4b..e421abba 100644
> --- a/libmultipath/propsel.c
> +++ b/libmultipath/propsel.c
> @@ -262,6 +262,7 @@ verify_alua_prio(struct multipath *mp)
>  		if (!prio_selected(&pp->prio))
>  			continue;
>  		if (strncmp(name, PRIO_ALUA, PRIO_NAME_LEN) &&
> +		    strncmp(name, PRIO_ALUA_CACHED, PRIO_NAME_LEN)
> &&
>  		    strncmp(name, PRIO_SYSFS, PRIO_NAME_LEN))
>  			 return false;
>  		assume_alua = true;
> @@ -690,7 +691,12 @@ int select_checker(struct config *conf, struct
> path *pp)
>  		}
>  		(void)path_get_tpgs(pp);
>  		if (pp->tpgs != TPGS_NONE && pp->tpgs != TPGS_UNDEF)
> {
> -			ckr_name = TUR;
> +			/*
> +			 * For ALUA devices (any TPGS support), use
> alua
> +			 * checker which combines path checking with
> ALUA state
> +			 * caching for optimal performance.
> +			 */
> +			ckr_name = ALUA;
>  			goto out;
>  		}
>  	}
> @@ -767,8 +773,15 @@ void detect_prio(struct path *pp)
>  		if ((tpgs == TPGS_EXPLICIT || !check_rdac(pp)) &&
>  		    sysfs_get_asymmetric_access_state(pp, buff, 512)
> >= 0)
>  			default_prio = PRIO_SYSFS;
> -		else
> +		else {
>  			default_prio = PRIO_ALUA;
> +			/* If using alua checker, use cached
> prioritizer */
> +			if (checker_selected(&pp->checker)) {
> +				const char *chk_name =
> checker_name(&pp->checker);
> +				if (chk_name && strcmp(chk_name,
> ALUA) == 0)
> +					default_prio =
> PRIO_ALUA_CACHED;
> +			}
> +		}
>  		break;
>  	default:
>  		return;
> @@ -831,7 +844,8 @@ out:
>  	/*
>  	 * fetch tpgs mode for alua, if its not already obtained
>  	 */
> -	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN)) {
> +	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN) ||
> +	    !strncmp(prio_name(p), PRIO_ALUA_CACHED, PRIO_NAME_LEN))
> {
>  		int tpgs = path_get_tpgs(pp);
>  
>  		if (tpgs == TPGS_NONE) {
> diff --git a/libmultipath/structs.c b/libmultipath/structs.c
> index f4413127..6483a568 100644
> --- a/libmultipath/structs.c
> +++ b/libmultipath/structs.c
> @@ -126,6 +126,8 @@ alloc_path (void)
>  		pp->fd = -1;
>  		pp->tpgs = TPGS_UNDEF;
>  		pp->tpg_id = GROUP_ID_UNDEF;
> +		pp->alua_state = -1;
> +		pp->alua_timestamp = 0;
>  		pp->priority = PRIO_UNDEF;
>  		pp->checkint = CHECKINT_UNDEF;
>  		checker_clear(&pp->checker);
> diff --git a/libmultipath/structs.h b/libmultipath/structs.h
> index 9adedde2..c4bddee9 100644
> --- a/libmultipath/structs.h
> +++ b/libmultipath/structs.h
> @@ -444,6 +444,9 @@ struct path {
>  	vector hwe;
>  	struct gen_path generic_path;
>  	int tpg_id;
> +	/* Cached ALUA state from checker for prioritizer use */
> +	int alua_state;		/* AAS value, -1 if not
> cached */
> +	time_t alua_timestamp;	/* When alua_state was last updated
> */
>  	enum ioctl_info_states ioctl_info;
>  };
>  

Re: [PATCH 1/1] checkers: add alua path checker

[Martin Wilck]

On Thu, 2026-03-12 at 11:38 -0700, Brian Bunker wrote:
> On Thu, Mar 12, 2026 at 9:32 AM Martin Wilck <mwilck@suse.com> wrote:
> 
> > Hi Brian,
> > 
> > On Wed, 2026-03-11 at 17:16 -0700, Brian Bunker wrote:
> > > Add an 'alua' path checker that uses RTPG to check path state,
> > > and an
> > > 'alua_cached' prioritizer that uses the checker's cached ALUA
> > > state.
> > > 
> > > The checker maps ALUA Asymmetric Access States to path states:
> > >   - AAS_OPTIMIZED, AAS_NON_OPTIMIZED -> PATH_UP
> > >   - AAS_STANDBY -> PATH_GHOST
> > >   - AAS_TRANSITIONING -> PATH_PENDING
> > >   - AAS_UNAVAILABLE, AAS_OFFLINE -> PATH_DOWN
> > > 
> > > The alua_cached prioritizer first checks for cached state from
> > > the
> > > alua
> > > checker, then falls back to sysfs, then to issuing RTPG directly.
> > > This
> > > maintains backward compatibility with other checker
> > > configurations.
> > > 
> > > When detect_checker is enabled, the alua checker is auto-selected
> > > for
> > > devices with TPGS support. When the alua checker is in use,
> > > detect_prio
> > > auto-selects the alua_cached prioritizer.
> > > 
> > > Signed-off-by: Brian Bunker <brian@purestorage.com>
> > 
> > Thanks! Just a few remarks for now.
> > 
> > Martin
> > 
> > > ---
> > >  libmultipath/Makefile                   |   5 +
> > >  libmultipath/checkers.c                 |   1 +
> > >  libmultipath/checkers.h                 |   1 +
> > >  libmultipath/checkers/Makefile          |   3 +-
> > >  libmultipath/checkers/alua.c            | 426
> > > ++++++++++++++++++++++++
> > >  libmultipath/checkers/alua.h            |  15 +
> > >  libmultipath/prio.c                     |   1 +
> > >  libmultipath/prio.h                     |   1 +
> > >  libmultipath/prioritizers/Makefile      |   1 +
> > >  libmultipath/prioritizers/alua_cached.c | 224 +++++++++++++
> > >  libmultipath/prioritizers/alua_rtpg.c   |  12 +-
> > >  libmultipath/prioritizers/alua_rtpg.h   |   1 +
> > >  libmultipath/prioritizers/sysfs.c       |   7 +-
> > >  libmultipath/propsel.c                  |  20 +-
> > >  libmultipath/structs.c                  |   2 +
> > >  libmultipath/structs.h                  |   3 +
> > >  16 files changed, 716 insertions(+), 7 deletions(-)
> > >  create mode 100644 libmultipath/checkers/alua.c
> > >  create mode 100644 libmultipath/checkers/alua.h
> > >  create mode 100644 libmultipath/prioritizers/alua_cached.c
> > 

[...]

> > Instead of duplicating the threading implementation of the TUR
> > checker
> > here we should abstract it out and use the common framework by both
> > checkers (and others, too, maybe). This is a thing I've wanted to
> > do
> > for a long time. Maybe it's time to start doing it now.
> > 
> I considered that sort of re-work, but I was concerned that changes
> adding to the existing checker might muddy the waters. There are
> certainly
> changes where helper functions could be extracted for both checkers
> to leverage them.
> 

I didn't mean to say that _you_ have to do it. I see it as my job,
actually. The question is whether you are willing to wait until I get
to it...

[...]

> > > --- a/libmultipath/prioritizers/Makefile
> > > +++ b/libmultipath/prioritizers/Makefile
> > > @@ -13,6 +13,7 @@ LIBDEPS = -lmultipath -lmpathutil -lm -lpthread
> > > -
> > > lrt
> > >  # If you add or remove a prioritizer also update
> > > multipath/multipath.conf.5
> > >  LIBS = \
> > >       libprioalua.so \
> > > +     libprioalua_cached.so \
> > >       libprioconst.so \
> > >       libpriodatacore.so \
> > >       libprioemc.so \
> > 
> > I think the below should go into a separate patch.
> > 
> Can they be separated? The prioritizer's logic relies
> on the checker populating values for its decision.
> 

First the checker, then the prioritizer?

> > 
> > > diff --git a/libmultipath/prioritizers/alua_cached.c
> > > b/libmultipath/prioritizers/alua_cached.c
> > > new file mode 100644
> > > index 00000000..2ba5b2e7
> > > --- /dev/null
> > > +++ b/libmultipath/prioritizers/alua_cached.c
> > > @@ -0,0 +1,224 @@
> > > +/*
> > > + * (C) Copyright IBM Corp. 2004, 2005   All Rights Reserved.
> > > + * Modified 2024 to support cached ALUA state from checker
> > > + *
> > > + * ALUA prioritizer that can use cached state from alua checker
> > > + * to avoid duplicate RTPG commands.
> > > + *
> > > + * This file is released under the GPL.
> > 
> > Please add a proper license header.
> > 
> > > + */
> > > +#include <stdio.h>
> > > +#include <string.h>
> > > +
> > > +#include <time.h>
> > > +
> > > +#include "debug.h"
> > > +#include "prio.h"
> > > +#include "structs.h"
> > > +#include "checkers.h"
> > > +#include "discovery.h"
> > > +
> > > +#include "alua.h"
> > > +
> > > +#define ALUA_PRIO_NOT_SUPPORTED                      1
> > > +#define ALUA_PRIO_RTPG_FAILED                        2
> > > +#define ALUA_PRIO_GETAAS_FAILED                      3
> > > +#define ALUA_PRIO_TPGS_FAILED                        4
> > > +#define ALUA_PRIO_NO_INFORMATION             5
> > > +#define ALUA_PRIO_CACHED_STALE                       6
> > > +
> > > +/* Maximum age of cached ALUA state in seconds */
> > > +#define ALUA_CACHE_MAX_AGE_SECS                      5
> > > +
> > > +static const char * aas_string[] = {
> > > +     [AAS_OPTIMIZED]         = "active/optimized",
> > > +     [AAS_NON_OPTIMIZED]     = "active/non-optimized",
> > > +     [AAS_STANDBY]           = "standby",
> > > +     [AAS_UNAVAILABLE]       = "unavailable",
> > > +     [AAS_LBA_DEPENDENT]     = "logical block dependent",
> > > +     [AAS_RESERVED]          = "ARRAY BUG: invalid TPGs state!",
> > > +     [AAS_OFFLINE]           = "offline",
> > > +     [AAS_TRANSITIONING]     = "transitioning between states",
> > > +};
> > 
> > We shouldn't duplicate the definition of these constants.
> 
> This was somewhat intentional though not efficient. I will
> do better with leveraging in the next patch.
> 
> 

OK.

Thanks
Martin

Re: [PATCH 1/1] checkers: add alua path checker

[Brian Bunker]

On Thu, Mar 12, 2026 at 9:32 AM Martin Wilck <mwilck@suse.com> wrote:
>
> Hi Brian,
>
> On Wed, 2026-03-11 at 17:16 -0700, Brian Bunker wrote:
> > Add an 'alua' path checker that uses RTPG to check path state, and an
> > 'alua_cached' prioritizer that uses the checker's cached ALUA state.
> >
> > The checker maps ALUA Asymmetric Access States to path states:
> >   - AAS_OPTIMIZED, AAS_NON_OPTIMIZED -> PATH_UP
> >   - AAS_STANDBY -> PATH_GHOST
> >   - AAS_TRANSITIONING -> PATH_PENDING
> >   - AAS_UNAVAILABLE, AAS_OFFLINE -> PATH_DOWN
> >
> > The alua_cached prioritizer first checks for cached state from the
> > alua
> > checker, then falls back to sysfs, then to issuing RTPG directly.
> > This
> > maintains backward compatibility with other checker configurations.
> >
> > When detect_checker is enabled, the alua checker is auto-selected for
> > devices with TPGS support. When the alua checker is in use,
> > detect_prio
> > auto-selects the alua_cached prioritizer.
> >
> > Signed-off-by: Brian Bunker <brian@purestorage.com>
>
> Thanks! Just a few remarks for now.
>
> Martin
>
> > ---
> >  libmultipath/Makefile                   |   5 +
> >  libmultipath/checkers.c                 |   1 +
> >  libmultipath/checkers.h                 |   1 +
> >  libmultipath/checkers/Makefile          |   3 +-
> >  libmultipath/checkers/alua.c            | 426
> > ++++++++++++++++++++++++
> >  libmultipath/checkers/alua.h            |  15 +
> >  libmultipath/prio.c                     |   1 +
> >  libmultipath/prio.h                     |   1 +
> >  libmultipath/prioritizers/Makefile      |   1 +
> >  libmultipath/prioritizers/alua_cached.c | 224 +++++++++++++
> >  libmultipath/prioritizers/alua_rtpg.c   |  12 +-
> >  libmultipath/prioritizers/alua_rtpg.h   |   1 +
> >  libmultipath/prioritizers/sysfs.c       |   7 +-
> >  libmultipath/propsel.c                  |  20 +-
> >  libmultipath/structs.c                  |   2 +
> >  libmultipath/structs.h                  |   3 +
> >  16 files changed, 716 insertions(+), 7 deletions(-)
> >  create mode 100644 libmultipath/checkers/alua.c
> >  create mode 100644 libmultipath/checkers/alua.h
> >  create mode 100644 libmultipath/prioritizers/alua_cached.c
> >
> > diff --git a/libmultipath/Makefile b/libmultipath/Makefile
> > index 85767ab4..379f4d02 100644
> > --- a/libmultipath/Makefile
> > +++ b/libmultipath/Makefile
> > @@ -38,6 +38,11 @@ nvme-lib.o: nvme-lib.c nvme-ioctl.c nvme-ioctl.h
> >  dict.o:      dict.c
> >       $(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -Wno-unused-parameter -c -o
> > $@ $<
> >
> > +# checkers/alua.o needs to find headers in parent directory
> > +checkers/alua.o: checkers/alua.c
> > +     @echo building $@ because of $?
> > +     $(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -I. -c -o $@ $<
> > +
> >  make_static = $(shell sed '/^static/!s/^\([a-z]\{1,\} \)/static \1/'
> > <$1 >$2)
> >
> >  nvme-ioctl.c: nvme/nvme-ioctl.c
> > diff --git a/libmultipath/checkers.c b/libmultipath/checkers.c
> > index bb6ad1ee..21aa52d4 100644
> > --- a/libmultipath/checkers.c
> > +++ b/libmultipath/checkers.c
> > @@ -461,6 +461,7 @@ int init_checkers(void)
> >               EMC_CLARIION,
> >               READSECTOR0,
> >               CCISS_TUR,
> > +             ALUA_RTPG,
> >       };
> >       unsigned int i;
> >
> > diff --git a/libmultipath/checkers.h b/libmultipath/checkers.h
> > index a969e7d1..2905f3fc 100644
> > --- a/libmultipath/checkers.h
> > +++ b/libmultipath/checkers.h
> > @@ -98,6 +98,7 @@ enum path_check_state {
> >  #define EMC_CLARIION "emc_clariion"
> >  #define READSECTOR0  "readsector0"
> >  #define CCISS_TUR    "cciss_tur"
> > +#define ALUA         "alua"
> >  #define NONE         "none"
> >  #define INVALID      "invalid"
> >
> > diff --git a/libmultipath/checkers/Makefile
> > b/libmultipath/checkers/Makefile
> > index 6f7cfb95..dfdd13a1 100644
> > --- a/libmultipath/checkers/Makefile
> > +++ b/libmultipath/checkers/Makefile
> > @@ -17,7 +17,8 @@ LIBS= \
> >       libcheckdirectio.so \
> >       libcheckemc_clariion.so \
> >       libcheckhp_sw.so \
> > -     libcheckrdac.so
> > +     libcheckrdac.so \
> > +     libcheckalua.so
> >
> >  all: $(LIBS)
> >
> > diff --git a/libmultipath/checkers/alua.c
> > b/libmultipath/checkers/alua.c
> > new file mode 100644
> > index 00000000..cb3d7000
> > --- /dev/null
> > +++ b/libmultipath/checkers/alua.c
> > @@ -0,0 +1,426 @@
> > +/*
> > + * ALUA Path Checker
> > + *
> > + * Copyright (c) 2024
> > + * This file is released under the GPL.
>
> Please add a proper license header.
I will do that.
>
> > + */
> > +#include <stdio.h>
> > +#include <stdlib.h>
> > +#include <string.h>
> > +#include <sys/types.h>
> > +#include <sys/stat.h>
> > +#include <sys/sysmacros.h>
> > +#include <unistd.h>
> > +#include <fcntl.h>
> > +#include <errno.h>
> > +#include <pthread.h>
> > +#include <urcu.h>
> > +#include <urcu/uatomic.h>
> > +
> > +#include "checkers.h"
> > +#include "debug.h"
> > +#include "structs.h"
> > +#include "prio.h"
> > +#include "util.h"
> > +#include "time-util.h"
> > +#include "../prioritizers/alua.h"
> > +#include "../prioritizers/alua_rtpg.h"
> > +#include "alua.h"
> > +
> > +#define MAX_NR_TIMEOUTS 1
> > +
> > +enum {
> > +     MSG_ALUA_RTPG_NOT_SUPPORTED = CHECKER_FIRST_MSGID,
> > +     MSG_ALUA_RTPG_TRANSITIONING,
> > +     MSG_ALUA_RUNNING,
> > +     MSG_ALUA_TIMEOUT,
> > +};
> > +
> > +#define IDX_(x) (MSG_ ## x - CHECKER_FIRST_MSGID)
> > +const char *libcheck_msgtable[] = {
> > +     [IDX_(ALUA_RTPG_NOT_SUPPORTED)] = " ALUA not supported",
> > +     [IDX_(ALUA_RTPG_TRANSITIONING)] = " path is transitioning",
> > +     [IDX_(ALUA_RUNNING)] = " still running",
> > +     [IDX_(ALUA_TIMEOUT)] = " timed out",
> > +     NULL,
> > +};
> > +
> > +struct alua_checker_context {
> > +     dev_t devt;
> > +     int state;
> > +     int running; /* uatomic access only */
> > +     int fd;
> > +     unsigned int timeout;
> > +     time_t time;
> > +     pthread_t thread;
> > +     pthread_mutex_t lock;
> > +     pthread_cond_t active;
> > +     int holders; /* uatomic access only */
> > +     int msgid;
> > +     struct checker_context ctx;
> > +     unsigned int nr_timeouts;
> > +     bool checked_state;
> > +     /* ALUA-specific cached data for prioritizer */
> > +     int aas;           /* Asymmetric Access State */
> > +     int tpg;           /* Target Port Group */
> > +     time_t timestamp;  /* When this data was collected */
> > +     /* Data needed for RTPG in thread */
> > +     struct path *pp;   /* Path pointer - only valid during check
> > */
> > +};
>
> Instead of duplicating the threading implementation of the TUR checker
> here we should abstract it out and use the common framework by both
> checkers (and others, too, maybe). This is a thing I've wanted to do
> for a long time. Maybe it's time to start doing it now.
I considered that sort of re-work, but I was concerned that changes
adding to the existing checker might muddy the waters. There are certainly
changes where helper functions could be extracted for both checkers
to leverage them.
>
>
>
>
> > +
> > +int libcheck_init(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct;
> > +     struct stat sb;
> > +
> > +     ct = malloc(sizeof(struct alua_checker_context));
> > +     if (!ct)
> > +             return 1;
> > +     memset(ct, 0, sizeof(struct alua_checker_context));
> > +
> > +     ct->state = PATH_UNCHECKED;
> > +     ct->fd = -1;
> > +     ct->aas = -1;
> > +     ct->tpg = -1;
> > +     uatomic_set(&ct->holders, 1);
> > +     pthread_cond_init_mono(&ct->active);
> > +     pthread_mutex_init(&ct->lock, NULL);
> > +     if (fstat(c->fd, &sb) == 0)
> > +             ct->devt = sb.st_rdev;
> > +     ct->ctx.cls = c->cls;
> > +     c->context = ct;
> > +
> > +     return 0;
> > +}
> > +
> > +static void cleanup_context(struct alua_checker_context *ct)
> > +{
> > +     pthread_mutex_destroy(&ct->lock);
> > +     pthread_cond_destroy(&ct->active);
> > +     free(ct);
> > +}
> > +
> > +void libcheck_free(struct checker *c)
> > +{
> > +     if (c->context) {
> > +             struct alua_checker_context *ct = c->context;
> > +             int holders;
> > +             int running;
> > +
> > +             running = uatomic_xchg(&ct->running, 0);
> > +             if (running)
> > +                     pthread_cancel(ct->thread);
> > +             ct->thread = 0;
> > +             holders = uatomic_sub_return(&ct->holders, 1);
> > +             if (!holders)
> > +                     cleanup_context(ct);
> > +             c->context = NULL;
> > +     }
> > +     return;
> > +}
> > +
> > +/*
> > + * Map ALUA Asymmetric Access State to path state
> > + */
> > +static int alua_state_to_path_state(int aas, short *msgid)
> > +{
> > +     switch (aas & 0x0f) {
> > +     case AAS_OPTIMIZED:
> > +     case AAS_NON_OPTIMIZED:
> > +             *msgid = CHECKER_MSGID_UP;
> > +             return PATH_UP;
> > +     case AAS_STANDBY:
> > +             *msgid = CHECKER_MSGID_GHOST;
> > +             return PATH_GHOST;
> > +     case AAS_TRANSITIONING:
> > +             *msgid = MSG_ALUA_RTPG_TRANSITIONING;
> > +             return PATH_PENDING;
> > +     case AAS_UNAVAILABLE:
> > +     case AAS_OFFLINE:
> > +     default:
> > +             *msgid = CHECKER_MSGID_DOWN;
> > +             return PATH_DOWN;
> > +     }
> > +}
> > +
> > +/*
> > + * Main ALUA check function - uses alua_rtpg.c library.
> > + * This is called either in sync mode or from the async thread.
> > + */
> > +static int
> > +alua_check(struct path *pp, int *aas_out, short *msgid)
> > +{
> > +     int aas;
> > +
> > +     if (pp->tpg_id == GROUP_ID_UNDEF) {
> > +             *msgid = CHECKER_MSGID_DOWN;
> > +             return PATH_DOWN;
> > +     }
> > +
> > +     aas = get_asymmetric_access_state(pp, pp->tpg_id);
> > +     if (aas < 0) {
> > +             if (aas == -RTPG_LUN_DISCONNECTED) {
> > +                     *msgid = CHECKER_MSGID_DISCONNECTED;
> > +                     return PATH_DISCONNECTED;
> > +             }
> > +             *msgid = CHECKER_MSGID_DOWN;
> > +             return PATH_DOWN;
> > +     }
> > +
> > +     *aas_out = aas;
> > +     return alua_state_to_path_state(aas, msgid);
> > +}
> > +
> > +#define alua_thread_cleanup_push(ct)
> > pthread_cleanup_push(cleanup_func, ct)
> > +#define alua_thread_cleanup_pop(ct) pthread_cleanup_pop(1)
> > +
> > +static void cleanup_func(void *data)
> > +{
> > +     int holders;
> > +     struct alua_checker_context *ct = data;
> > +
> > +     holders = uatomic_sub_return(&ct->holders, 1);
> > +     if (!holders)
> > +             cleanup_context(ct);
> > +}
> > +
> > +void *libcheck_thread(struct checker_context *ctx)
> > +{
> > +     struct alua_checker_context *ct =
> > +             container_of(ctx, struct alua_checker_context, ctx);
> > +     int state, running;
> > +     short msgid;
> > +     int aas;
> > +
> > +     /* This thread can be canceled, so setup clean up */
> > +     alua_thread_cleanup_push(ct);
> > +
> > +     condlog(4, "%d:%d : alua checker starting up", major(ct-
> > >devt),
> > +             minor(ct->devt));
> > +
> > +     state = alua_check(ct->pp, &aas, &msgid);
> > +     pthread_testcancel();
> > +
> > +     /* ALUA checker done */
> > +     pthread_mutex_lock(&ct->lock);
> > +     ct->state = state;
> > +     ct->msgid = msgid;
> > +     ct->aas = aas;
> > +     ct->tpg = ct->pp->tpg_id;
> > +     ct->timestamp = time(NULL);
> > +     /* Update path-level cache for prioritizer use */
> > +     ct->pp->alua_state = aas;
> > +     ct->pp->alua_timestamp = ct->timestamp;
> > +     pthread_cond_signal(&ct->active);
> > +     pthread_mutex_unlock(&ct->lock);
> > +
> > +     condlog(4, "%d:%d : alua checker finished, state %s",
> > major(ct->devt),
> > +             minor(ct->devt), checker_state_name(state));
> > +
> > +     running = uatomic_xchg(&ct->running, 0);
> > +     if (!running)
> > +             pause();
> > +
> > +     alua_thread_cleanup_pop(ct);
> > +
> > +     return ((void *)0);
> > +}
> > +
> > +static void alua_set_async_timeout(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +     struct timespec now;
> > +
> > +     get_monotonic_time(&now);
> > +     ct->time = now.tv_sec + c->timeout;
> > +}
> > +
> > +static int alua_check_async_timeout(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +     struct timespec now;
> > +
> > +     get_monotonic_time(&now);
> > +     return (now.tv_sec > ct->time);
> > +}
> > +
> > +static int check_pending(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +     int alua_status = PATH_PENDING;
> > +
> > +     pthread_mutex_lock(&ct->lock);
> > +     if (ct->state != PATH_PENDING || ct->msgid !=
> > MSG_ALUA_RUNNING) {
> > +             alua_status = ct->state;
> > +             c->msgid = ct->msgid;
> > +     }
> > +     pthread_mutex_unlock(&ct->lock);
> > +     if (alua_status == PATH_PENDING && c->msgid ==
> > MSG_ALUA_RUNNING) {
> > +             condlog(4, "%d:%d : alua checker still running",
> > +                     major(ct->devt), minor(ct->devt));
> > +     } else {
> > +             int running = uatomic_xchg(&ct->running, 0);
> > +             if (running)
> > +                     pthread_cancel(ct->thread);
> > +             ct->thread = 0;
> > +     }
> > +
> > +     ct->checked_state = true;
> > +     return alua_status;
> > +}
> > +
> > +bool libcheck_need_wait(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +     return (ct && ct->thread && uatomic_read(&ct->running) != 0
> > &&
> > +             !ct->checked_state);
> > +}
> > +
> > +int libcheck_pending(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +
> > +     /* The if path checker isn't running, just return the
> > exiting value. */
> > +     if (!ct || !ct->thread)
> > +             return c->path_state;
> > +
> > +     return check_pending(c);
> > +}
> > +
> > +int libcheck_check(struct checker *c)
> > +{
> > +     struct alua_checker_context *ct = c->context;
> > +     struct path *pp = container_of(c, struct path, checker);
> > +     pthread_attr_t attr;
> > +     int alua_status = PATH_PENDING, r;
> > +     int aas;
> > +
> > +     if (!ct)
> > +             return PATH_UNCHECKED;
> > +
> > +     if (checker_is_sync(c)) {
> > +             alua_status = alua_check(pp, &aas, &c->msgid);
> > +             ct->tpg = pp->tpg_id;
> > +             ct->aas = aas;
> > +             ct->timestamp = time(NULL);
> > +             /* Update path-level cache for prioritizer use */
> > +             pp->alua_state = aas;
> > +             pp->alua_timestamp = ct->timestamp;
> > +             return alua_status;
> > +     }
> > +
> > +     /* Async mode */
> > +     if (ct->thread) {
> > +             ct->checked_state = true;
> > +             if (alua_check_async_timeout(c)) {
> > +                     int running = uatomic_xchg(&ct->running, 0);
> > +                     if (running) {
> > +                             pthread_cancel(ct->thread);
> > +                             condlog(3, "%d:%d : alua checker
> > timeout",
> > +                                     major(ct->devt), minor(ct-
> > >devt));
> > +                             c->msgid = MSG_ALUA_TIMEOUT;
> > +                             alua_status = PATH_TIMEOUT;
> > +                     } else {
> > +                             pthread_mutex_lock(&ct->lock);
> > +                             alua_status = ct->state;
> > +                             c->msgid = ct->msgid;
> > +                             pthread_mutex_unlock(&ct->lock);
> > +                     }
> > +                     ct->thread = 0;
> > +             } else if (uatomic_read(&ct->running) != 0) {
> > +                     /*
> > +                      * ct->running may be stale - the thread
> > sets
> > +                      * msgid/state under lock, then clears
> > running
> > +                      * after releasing it. Check msgid under
> > lock
> > +                      * to get the truth. Only MSG_ALUA_RUNNING
> > means
> > +                      * no result yet.
> > +                      */
> > +                     pthread_mutex_lock(&ct->lock);
> > +                     if (ct->msgid != MSG_ALUA_RUNNING) {
> > +                             alua_status = ct->state;
> > +                             c->msgid = ct->msgid;
> > +                             ct->thread = 0;
> > +                     }
> > +                     pthread_mutex_unlock(&ct->lock);
> > +             } else {
> > +                     /* ALUA checker done */
> > +                     ct->thread = 0;
> > +                     pthread_mutex_lock(&ct->lock);
> > +                     alua_status = ct->state;
> > +                     c->msgid = ct->msgid;
> > +                     pthread_mutex_unlock(&ct->lock);
> > +             }
> > +     } else {
> > +             if (uatomic_read(&ct->holders) > 1) {
> > +                     /* The thread has been cancelled but hasn't
> > quit. */
> > +                     if (ct->nr_timeouts <= MAX_NR_TIMEOUTS) {
> > +                             if (ct->nr_timeouts ==
> > MAX_NR_TIMEOUTS) {
> > +                                     condlog(2, "%d:%d : waiting
> > for stalled alua thread to finish",
> > +                                             major(ct->devt),
> > minor(ct->devt));
> > +                             }
> > +                             ct->nr_timeouts++;
> > +                             c->msgid = MSG_ALUA_TIMEOUT;
> > +                             return PATH_TIMEOUT;
> > +                     }
> > +                     ct->nr_timeouts++;
> > +                     /*
> > +                      * Start a new thread while the old one is
> > stalled.
> > +                      * We have to prevent it from interfering
> > with the new
> > +                      * thread. We create a new context and leave
> > the old
> > +                      * one with the stale thread, hoping it will
> > clean up
> > +                      * eventually.
> > +                      */
> > +                     condlog(3, "%d:%d : alua thread not
> > responding",
> > +                             major(ct->devt), minor(ct->devt));
> > +
> > +                     if (libcheck_init(c) != 0) {
> > +                             c->msgid = CHECKER_MSGID_DOWN;
> > +                             return PATH_UNCHECKED;
> > +                     }
> > +                     ((struct alua_checker_context *)c->context)-
> > >nr_timeouts = ct->nr_timeouts;
> > +
> > +                     if (!uatomic_sub_return(&ct->holders, 1)) {
> > +                             /* It did terminate, eventually */
> > +                             cleanup_context(ct);
> > +                             ((struct alua_checker_context *)c-
> > >context)->nr_timeouts = 0;
> > +                     }
> > +
> > +                     ct = c->context;
> > +             } else
> > +                     ct->nr_timeouts = 0;
> > +
> > +             /* Start new ALUA checker */
> > +             pthread_mutex_lock(&ct->lock);
> > +             alua_status = ct->state = PATH_PENDING;
> > +             c->msgid = ct->msgid = MSG_ALUA_RUNNING;
> > +             pthread_mutex_unlock(&ct->lock);
> > +             ct->fd = c->fd;
> > +             ct->timeout = c->timeout;
> > +             ct->pp = pp;
> > +             ct->checked_state = false;
> > +             uatomic_add(&ct->holders, 1);
> > +             uatomic_set(&ct->running, 1);
> > +             alua_set_async_timeout(c);
> > +             setup_thread_attr(&attr, 32 * 1024, 1);
> > +             r = start_checker_thread(&ct->thread, &attr, &ct-
> > >ctx);
> > +             pthread_attr_destroy(&attr);
> > +             if (r) {
> > +                     uatomic_sub(&ct->holders, 1);
> > +                     uatomic_set(&ct->running, 0);
> > +                     ct->thread = 0;
> > +                     condlog(3, "%d:%d : failed to start alua
> > thread, using"
> > +                             " sync mode", major(ct->devt),
> > minor(ct->devt));
> > +                     alua_status = alua_check(pp, &aas, &c-
> > >msgid);
> > +                     ct->tpg = pp->tpg_id;
> > +                     ct->aas = aas;
> > +                     ct->timestamp = time(NULL);
> > +                     /* Update path-level cache for prioritizer
> > use */
> > +                     pp->alua_state = aas;
> > +                     pp->alua_timestamp = ct->timestamp;
> > +                     return alua_status;
> > +             }
> > +     }
> > +
> > +     return alua_status;
> > +}
> > +
> > diff --git a/libmultipath/checkers/alua.h
> > b/libmultipath/checkers/alua.h
> > new file mode 100644
> > index 00000000..b53cef66
> > --- /dev/null
> > +++ b/libmultipath/checkers/alua.h
> > @@ -0,0 +1,15 @@
> > +/*
> > + * ALUA Path Checker Header
> > + *
> > + * This file is released under the GPL.
>
> Please add a proper license header.
>
> > + */
> > +#ifndef _ALUA_CHECKER_H
> > +#define _ALUA_CHECKER_H
> > +
> > +/* ALUA checker caches state in struct path for prioritizer use:
> > + *   pp->alua_state     - cached AAS value (-1 if not cached)
> > + *   pp->alua_timestamp - when the state was last updated
> > + */
> > +
> > +#endif /* _ALUA_CHECKER_H */
> > +
>
> Only comments in this file?
I will look at this. It was likely leftover.
>
>
> > diff --git a/libmultipath/prio.c b/libmultipath/prio.c
> > index 24f825bd..2d59d04e 100644
> > --- a/libmultipath/prio.c
> > +++ b/libmultipath/prio.c
> > @@ -29,6 +29,7 @@ int init_prio(void)
> >  #ifdef LOAD_ALL_SHARED_LIBS
> >       static const char *const all_prios[] = {
> >               PRIO_ALUA,
> > +             PRIO_ALUA_CACHED,
> >               PRIO_CONST,
> >               PRIO_DATACORE,
> >               PRIO_EMC,
> > diff --git a/libmultipath/prio.h b/libmultipath/prio.h
> > index 119b75f2..b54ece0d 100644
> > --- a/libmultipath/prio.h
> > +++ b/libmultipath/prio.h
> > @@ -17,6 +17,7 @@ struct path;
> >   * Known prioritizers for use in hwtable.c
> >   */
> >  #define PRIO_ALUA            "alua"
> > +#define PRIO_ALUA_CACHED     "alua_cached"
> >  #define PRIO_CONST           "const"
> >  #define PRIO_DATACORE                "datacore"
> >  #define PRIO_EMC             "emc"
> > diff --git a/libmultipath/prioritizers/Makefile
> > b/libmultipath/prioritizers/Makefile
> > index ff2524c2..201c52c0 100644
> > --- a/libmultipath/prioritizers/Makefile
> > +++ b/libmultipath/prioritizers/Makefile
> > @@ -13,6 +13,7 @@ LIBDEPS = -lmultipath -lmpathutil -lm -lpthread -
> > lrt
> >  # If you add or remove a prioritizer also update
> > multipath/multipath.conf.5
> >  LIBS = \
> >       libprioalua.so \
> > +     libprioalua_cached.so \
> >       libprioconst.so \
> >       libpriodatacore.so \
> >       libprioemc.so \
>
> I think the below should go into a separate patch.
Can they be separated? The prioritizer's logic relies
on the checker populating values for its decision.
>
> > diff --git a/libmultipath/prioritizers/alua_cached.c
> > b/libmultipath/prioritizers/alua_cached.c
> > new file mode 100644
> > index 00000000..2ba5b2e7
> > --- /dev/null
> > +++ b/libmultipath/prioritizers/alua_cached.c
> > @@ -0,0 +1,224 @@
> > +/*
> > + * (C) Copyright IBM Corp. 2004, 2005   All Rights Reserved.
> > + * Modified 2024 to support cached ALUA state from checker
> > + *
> > + * ALUA prioritizer that can use cached state from alua checker
> > + * to avoid duplicate RTPG commands.
> > + *
> > + * This file is released under the GPL.
>
> Please add a proper license header.
>
> > + */
> > +#include <stdio.h>
> > +#include <string.h>
> > +
> > +#include <time.h>
> > +
> > +#include "debug.h"
> > +#include "prio.h"
> > +#include "structs.h"
> > +#include "checkers.h"
> > +#include "discovery.h"
> > +
> > +#include "alua.h"
> > +
> > +#define ALUA_PRIO_NOT_SUPPORTED                      1
> > +#define ALUA_PRIO_RTPG_FAILED                        2
> > +#define ALUA_PRIO_GETAAS_FAILED                      3
> > +#define ALUA_PRIO_TPGS_FAILED                        4
> > +#define ALUA_PRIO_NO_INFORMATION             5
> > +#define ALUA_PRIO_CACHED_STALE                       6
> > +
> > +/* Maximum age of cached ALUA state in seconds */
> > +#define ALUA_CACHE_MAX_AGE_SECS                      5
> > +
> > +static const char * aas_string[] = {
> > +     [AAS_OPTIMIZED]         = "active/optimized",
> > +     [AAS_NON_OPTIMIZED]     = "active/non-optimized",
> > +     [AAS_STANDBY]           = "standby",
> > +     [AAS_UNAVAILABLE]       = "unavailable",
> > +     [AAS_LBA_DEPENDENT]     = "logical block dependent",
> > +     [AAS_RESERVED]          = "ARRAY BUG: invalid TPGs state!",
> > +     [AAS_OFFLINE]           = "offline",
> > +     [AAS_TRANSITIONING]     = "transitioning between states",
> > +};
>
> We shouldn't duplicate the definition of these constants.
This was somewhat intentional though not efficient. I will
do better with leveraging in the next patch.
>
> > +
> > +static const char *aas_print_string(int rc)
> > +{
> > +     rc &= 0x7f;
> > +
> > +     if (rc & 0x70)
> > +             return aas_string[AAS_RESERVED];
> > +     rc &= 0x0f;
> > +     if (rc > AAS_RESERVED && rc < AAS_OFFLINE)
> > +             return aas_string[AAS_RESERVED];
> > +     else
> > +             return aas_string[rc];
> > +}
> > +
> > +/*
> > + * Try to get ALUA info with optimization:
> > + * 1. If using alua checker: use its cached state (it just polled!)
> > + * 2. Try sysfs (kernel maintains this, no I/O needed)
> > + * 3. Fall back to issuing RTPG directly
> > + *
> > + * Rationale: The checker is our active poller. If we're using alua
> > + * and its cache is fresh, we trust that data first. If the cache
> > misses
> > + * (stale or empty), we try sysfs before issuing RTPG. This provides
> > a
> > + * three-tier optimization: checker cache → sysfs → RTPG.
> > + */
> > +/* Maximum length of sysfs access_state string */
> > +#define ALUA_ACCESS_STATE_SIZE 32
> > +
> > +int
> > +get_alua_info_cached(struct path * pp, bool *used_cache)
> > +{
> > +     int     rc;
> > +     int     tpg;
> > +     bool    diff_tpg;
> > +     char    buff[ALUA_ACCESS_STATE_SIZE];
> > +
> > +     *used_cache = false;
> > +
> > +     /* First: if using alua checker, use its fresh data from
> > path structure */
> > +     if (checker_selected(&pp->checker)) {
> > +             const char *chk_name = checker_name(&pp->checker);
> > +
> > +             if (chk_name && strcmp(chk_name, "alua") == 0 &&
> > +                 pp->alua_state >= 0) {
> > +                     time_t now = time(NULL);
> > +                     time_t age = now - pp->alua_timestamp;
> > +
> > +                     if (pp->alua_timestamp > 0 && age <=
> > ALUA_CACHE_MAX_AGE_SECS) {
> > +                             condlog(4, "%s: using cached ALUA
> > state from checker (fresh poll)", pp->dev);
> > +                             *used_cache = true;
> > +                             return pp->alua_state;
> > +                     }
> > +                     condlog(4, "%s: cached ALUA state not
> > available or stale, trying sysfs",
> > +                             pp->dev);
> > +                     /* Cache miss - try sysfs before issuing
> > RTPG */
> > +             }
> > +     }
> > +
> > +     /* Second: try sysfs (if cache missed or not using alua
> > checker) */
> > +     rc = sysfs_get_asymmetric_access_state(pp, buff,
> > sizeof(buff));
> > +     if (rc >= 0) {
> > +             /* Parse sysfs state string to ALUA state value */
> > +             int aas = -1;
> > +             int priopath = rc; /* rc contains preferred bit */
> > +
> > +             if (!strncmp(buff, "active/optimized", 16))
> > +                     aas = AAS_OPTIMIZED;
> > +             else if (!strncmp(buff, "active/non-optimized", 20))
> > +                     aas = AAS_NON_OPTIMIZED;
> > +             else if (!strncmp(buff, "lba-dependent", 13))
> > +                     aas = AAS_LBA_DEPENDENT;
> > +             else if (!strncmp(buff, "standby", 7))
> > +                     aas = AAS_STANDBY;
> > +
> > +             if (aas >= 0) {
> > +                     condlog(4, "%s: using sysfs ALUA state (no
> > I/O)", pp->dev);
> > +                     *used_cache = true;
> > +                     return (priopath ? 0x80 : 0) | aas;
> > +             }
> > +     }
> > +
> > +     /* Last resort: issue RTPG directly */
> > +     tpg = get_target_port_group(pp);
> > +     if (tpg < 0) {
> > +             rc = get_target_port_group_support(pp);
> > +             if (rc < 0)
> > +                     return -ALUA_PRIO_TPGS_FAILED;
> > +             if (rc == TPGS_NONE)
> > +                     return -ALUA_PRIO_NOT_SUPPORTED;
> > +             return -ALUA_PRIO_RTPG_FAILED;
> > +     }
> > +     diff_tpg = (pp->tpg_id != GROUP_ID_UNDEF && pp->tpg_id !=
> > tpg);
> > +     pp->tpg_id = tpg;
> > +     condlog((diff_tpg) ? 2 : 4, "%s: reported target port group
> > is %i",
> > +             pp->dev, tpg);
> > +     rc = get_asymmetric_access_state(pp, tpg);
> > +     if (rc < 0) {
> > +             condlog(2, "%s: get_asymmetric_access_state returned
> > %d",
> > +                     __func__, rc);
> > +             return -ALUA_PRIO_GETAAS_FAILED;
> > +     }
> > +
> > +     condlog(3, "%s: aas = %02x [%s]%s", pp->dev, rc,
> > aas_print_string(rc),
> > +             (rc & 0x80) ? " [preferred]" : "");
> > +     return rc;
> > +}
> > +
> > +int get_exclusive_pref_arg(char *args)
> > +{
> > +     char *ptr;
> > +
> > +     if (args == NULL)
> > +             return 0;
> > +     ptr = strstr(args, "exclusive_pref_bit");
> > +     if (!ptr)
> > +             return 0;
> > +     if (ptr[18] != '\0' && ptr[18] != ' ' && ptr[18] != '\t')
> > +             return 0;
> > +     if (ptr != args && ptr[-1] != ' ' && ptr[-1] != '\t')
> > +             return 0;
> > +     return 1;
> > +}
> > +
> > +int getprio (struct path * pp, char * args)
> > +{
> > +     int rc;
> > +     int aas;
> > +     int priopath;
> > +     int exclusive_pref;
> > +     bool used_cache = false;
> > +
> > +     if (pp->fd < 0)
> > +             return -ALUA_PRIO_NO_INFORMATION;
> > +
> > +     exclusive_pref = get_exclusive_pref_arg(args);
> > +     rc = get_alua_info_cached(pp, &used_cache);
> > +
> > +     if (used_cache) {
> > +             condlog(4, "%s: priority calculated from cached ALUA
> > state (no RTPG issued)",
> > +                     pp->dev);
> > +     }
> > +
> > +     if (rc >= 0) {
> > +             aas = (rc & 0x0f);
> > +             priopath = (rc & 0x80);
> > +             switch(aas) {
> > +             case AAS_OPTIMIZED:
> > +                     rc = 50;
> > +                     break;
> > +             case AAS_NON_OPTIMIZED:
> > +                     rc = 10;
> > +                     break;
> > +             case AAS_LBA_DEPENDENT:
> > +                     rc = 5;
> > +                     break;
> > +             case AAS_STANDBY:
> > +                     rc = 1;
> > +                     break;
> > +             default:
> > +                     rc = 0;
> > +             }
> > +             if (priopath && (aas != AAS_OPTIMIZED ||
> > exclusive_pref))
> > +                     rc += 80;
> > +     } else {
> > +             switch(-rc) {
> > +             case ALUA_PRIO_NOT_SUPPORTED:
> > +                     condlog(0, "%s: alua not supported", pp-
> > >dev);
> > +                     break;
> > +             case ALUA_PRIO_RTPG_FAILED:
> > +                     condlog(0, "%s: couldn't get target port
> > group", pp->dev);
> > +                     break;
> > +             case ALUA_PRIO_GETAAS_FAILED:
> > +                     condlog(0, "%s: couldn't get asymmetric
> > access state", pp->dev);
> > +                     break;
> > +             case ALUA_PRIO_TPGS_FAILED:
> > +                     condlog(3, "%s: couldn't get supported alua
> > states", pp->dev);
> > +                     break;
> > +             }
> > +     }
> > +     return rc;
> > +}
> > +
> > diff --git a/libmultipath/prioritizers/alua_rtpg.c
> > b/libmultipath/prioritizers/alua_rtpg.c
> > index 053cccb7..5da06b50 100644
> > --- a/libmultipath/prioritizers/alua_rtpg.c
> > +++ b/libmultipath/prioritizers/alua_rtpg.c
> > @@ -75,6 +75,7 @@ enum scsi_disposition {
> >       SCSI_GOOD = 0,
> >       SCSI_ERROR,
> >       SCSI_RETRY,
> > +     SCSI_DISCONNECTED,
> >  };
> >
> >  static int
> > @@ -124,6 +125,12 @@ scsi_error(struct sg_io_hdr *hdr, int opcode)
> >       PRINT_DEBUG("alua: SCSI error for command %02x: status %02x,
> > sense %02x/%02x/%02x",
> >                   opcode, hdr->status, sense_key, asc, ascq);
> >
> > +     /* Check for disconnected LUN (unmapped at target) */
> > +     if (sense_key == 0x5 && asc == 0x25 && ascq == 0x00) {
> > +             /* ILLEGAL REQUEST / LOGICAL UNIT NOT SUPPORTED */
> > +             return SCSI_DISCONNECTED;
> > +     }
> > +
> >       if (sense_key == UNIT_ATTENTION || sense_key == NOT_READY)
> >               return SCSI_RETRY;
> >       else
> > @@ -325,7 +332,10 @@ retry:
> >       }
> >
> >       rc = scsi_error(&hdr, OPERATION_CODE_RTPG);
> > -     if (rc == SCSI_ERROR) {
> > +     if (rc == SCSI_DISCONNECTED) {
> > +             PRINT_DEBUG("do_rtpg: LUN disconnected (unmapped at
> > target)!");
> > +             return -RTPG_LUN_DISCONNECTED;
> > +     } else if (rc == SCSI_ERROR) {
> >               PRINT_DEBUG("do_rtpg: SCSI error!");
> >               return -RTPG_RTPG_FAILED;
> >       } else if (rc == SCSI_RETRY) {
> > diff --git a/libmultipath/prioritizers/alua_rtpg.h
> > b/libmultipath/prioritizers/alua_rtpg.h
> > index ff6ec0ef..030ca58d 100644
> > --- a/libmultipath/prioritizers/alua_rtpg.h
> > +++ b/libmultipath/prioritizers/alua_rtpg.h
> > @@ -21,6 +21,7 @@
> >  #define RTPG_NO_TPG_IDENTIFIER                       2
> >  #define RTPG_RTPG_FAILED                     3
> >  #define RTPG_TPG_NOT_FOUND                   4
> > +#define RTPG_LUN_DISCONNECTED                        5
> >
> >  int get_target_port_group_support(const struct path *pp);
> >  int get_target_port_group(const struct path *pp);
> > diff --git a/libmultipath/prioritizers/sysfs.c
> > b/libmultipath/prioritizers/sysfs.c
> > index 5e8adc05..ab058ea9 100644
> > --- a/libmultipath/prioritizers/sysfs.c
> > +++ b/libmultipath/prioritizers/sysfs.c
> > @@ -10,6 +10,9 @@
> >  #include "discovery.h"
> >  #include "prio.h"
> >
> > +/* Maximum length of sysfs access_state string */
> > +#define ALUA_ACCESS_STATE_SIZE 32
> > +
> >  static const struct {
> >       unsigned char value;
> >       char *name;
> > @@ -39,11 +42,11 @@ int get_exclusive_pref_arg(char *args)
> >  int getprio (struct path * pp, char *args)
> >  {
> >       int prio = 0, rc, i;
> > -     char buff[512];
> > +     char buff[ALUA_ACCESS_STATE_SIZE];
> >       int exclusive_pref;
> >
> >       exclusive_pref = get_exclusive_pref_arg(args);
> > -     rc = sysfs_get_asymmetric_access_state(pp, buff, 512);
> > +     rc = sysfs_get_asymmetric_access_state(pp, buff,
> > sizeof(buff));
> >       if (rc < 0)
> >               return PRIO_UNDEF;
> >       prio = 0;
> > diff --git a/libmultipath/propsel.c b/libmultipath/propsel.c
> > index 56895c4b..e421abba 100644
> > --- a/libmultipath/propsel.c
> > +++ b/libmultipath/propsel.c
> > @@ -262,6 +262,7 @@ verify_alua_prio(struct multipath *mp)
> >               if (!prio_selected(&pp->prio))
> >                       continue;
> >               if (strncmp(name, PRIO_ALUA, PRIO_NAME_LEN) &&
> > +                 strncmp(name, PRIO_ALUA_CACHED, PRIO_NAME_LEN)
> > &&
> >                   strncmp(name, PRIO_SYSFS, PRIO_NAME_LEN))
> >                        return false;
> >               assume_alua = true;
> > @@ -690,7 +691,12 @@ int select_checker(struct config *conf, struct
> > path *pp)
> >               }
> >               (void)path_get_tpgs(pp);
> >               if (pp->tpgs != TPGS_NONE && pp->tpgs != TPGS_UNDEF)
> > {
> > -                     ckr_name = TUR;
> > +                     /*
> > +                      * For ALUA devices (any TPGS support), use
> > alua
> > +                      * checker which combines path checking with
> > ALUA state
> > +                      * caching for optimal performance.
> > +                      */
> > +                     ckr_name = ALUA;
> >                       goto out;
> >               }
> >       }
> > @@ -767,8 +773,15 @@ void detect_prio(struct path *pp)
> >               if ((tpgs == TPGS_EXPLICIT || !check_rdac(pp)) &&
> >                   sysfs_get_asymmetric_access_state(pp, buff, 512)
> > >= 0)
> >                       default_prio = PRIO_SYSFS;
> > -             else
> > +             else {
> >                       default_prio = PRIO_ALUA;
> > +                     /* If using alua checker, use cached
> > prioritizer */
> > +                     if (checker_selected(&pp->checker)) {
> > +                             const char *chk_name =
> > checker_name(&pp->checker);
> > +                             if (chk_name && strcmp(chk_name,
> > ALUA) == 0)
> > +                                     default_prio =
> > PRIO_ALUA_CACHED;
> > +                     }
> > +             }
> >               break;
> >       default:
> >               return;
> > @@ -831,7 +844,8 @@ out:
> >       /*
> >        * fetch tpgs mode for alua, if its not already obtained
> >        */
> > -     if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN)) {
> > +     if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN) ||
> > +         !strncmp(prio_name(p), PRIO_ALUA_CACHED, PRIO_NAME_LEN))
> > {
> >               int tpgs = path_get_tpgs(pp);
> >
> >               if (tpgs == TPGS_NONE) {
> > diff --git a/libmultipath/structs.c b/libmultipath/structs.c
> > index f4413127..6483a568 100644
> > --- a/libmultipath/structs.c
> > +++ b/libmultipath/structs.c
> > @@ -126,6 +126,8 @@ alloc_path (void)
> >               pp->fd = -1;
> >               pp->tpgs = TPGS_UNDEF;
> >               pp->tpg_id = GROUP_ID_UNDEF;
> > +             pp->alua_state = -1;
> > +             pp->alua_timestamp = 0;
> >               pp->priority = PRIO_UNDEF;
> >               pp->checkint = CHECKINT_UNDEF;
> >               checker_clear(&pp->checker);
> > diff --git a/libmultipath/structs.h b/libmultipath/structs.h
> > index 9adedde2..c4bddee9 100644
> > --- a/libmultipath/structs.h
> > +++ b/libmultipath/structs.h
> > @@ -444,6 +444,9 @@ struct path {
> >       vector hwe;
> >       struct gen_path generic_path;
> >       int tpg_id;
> > +     /* Cached ALUA state from checker for prioritizer use */
> > +     int alua_state;         /* AAS value, -1 if not
> > cached */
> > +     time_t alua_timestamp;  /* When alua_state was last updated
> > */
> >       enum ioctl_info_states ioctl_info;
> >  };
> >
I will put together a 2nd version which addresses these comments.
Thanks,
Brian
-- 
Brian Bunker
PURE Storage, Inc.
brian@purestorage.com

Re: [PATCH 1/1] checkers: add alua path checker

[Xose Vazquez Perez]

On 3/12/26 5:32 PM, Martin Wilck wrote:

> On Wed, 2026-03-11 at 17:16 -0700, Brian Bunker wrote:

>> diff --git a/libmultipath/checkers/alua.c
>> b/libmultipath/checkers/alua.c
>> new file mode 100644
>> index 00000000..cb3d7000
>> --- /dev/null
>> +++ b/libmultipath/checkers/alua.c
>> @@ -0,0 +1,426 @@
>> +/*
>> + * ALUA Path Checker
>> + *
>> + * Copyright (c) 2024

There is no Copyright holder.

>> + * This file is released under the GPL.
> 
> Please add a proper license header.
If the code comes from libmultipath/prioritizers/alua*

For consistency with:
libmultipath/prioritizers/alua.c: * This file is released under the GPL.
libmultipath/prioritizers/alua_rtpg.c: * This file is released under the GPL.
libmultipath/prioritizers/alua_rtpg.h: * This file is released under the GPL.
libmultipath/prioritizers/alua_spc3.h: * This file is released under the GPL.

Otherwise, the license should be explicitly specified.

Re: [PATCH 1/1] checkers: add alua path checker

[Benjamin Marzinski]

On Wed, Mar 11, 2026 at 05:16:37PM -0700, Brian Bunker wrote:
> Add an 'alua' path checker that uses RTPG to check path state, and an
> 'alua_cached' prioritizer that uses the checker's cached ALUA state.
> 
> The checker maps ALUA Asymmetric Access States to path states:
>   - AAS_OPTIMIZED, AAS_NON_OPTIMIZED -> PATH_UP
>   - AAS_STANDBY -> PATH_GHOST
>   - AAS_TRANSITIONING -> PATH_PENDING
>   - AAS_UNAVAILABLE, AAS_OFFLINE -> PATH_DOWN
> 
> The alua_cached prioritizer first checks for cached state from the alua
> checker, then falls back to sysfs, then to issuing RTPG directly. This
> maintains backward compatibility with other checker configurations.
> 
> When detect_checker is enabled, the alua checker is auto-selected for
> devices with TPGS support. When the alua checker is in use, detect_prio
> auto-selects the alua_cached prioritizer.
> 
> Signed-off-by: Brian Bunker <brian@purestorage.com>
> ---
>  libmultipath/Makefile                   |   5 +
>  libmultipath/checkers.c                 |   1 +
>  libmultipath/checkers.h                 |   1 +
>  libmultipath/checkers/Makefile          |   3 +-
>  libmultipath/checkers/alua.c            | 426 ++++++++++++++++++++++++
>  libmultipath/checkers/alua.h            |  15 +
>  libmultipath/prio.c                     |   1 +
>  libmultipath/prio.h                     |   1 +
>  libmultipath/prioritizers/Makefile      |   1 +
>  libmultipath/prioritizers/alua_cached.c | 224 +++++++++++++
>  libmultipath/prioritizers/alua_rtpg.c   |  12 +-
>  libmultipath/prioritizers/alua_rtpg.h   |   1 +
>  libmultipath/prioritizers/sysfs.c       |   7 +-
>  libmultipath/propsel.c                  |  20 +-
>  libmultipath/structs.c                  |   2 +
>  libmultipath/structs.h                  |   3 +
>  16 files changed, 716 insertions(+), 7 deletions(-)
>  create mode 100644 libmultipath/checkers/alua.c
>  create mode 100644 libmultipath/checkers/alua.h
>  create mode 100644 libmultipath/prioritizers/alua_cached.c
> 
> diff --git a/libmultipath/Makefile b/libmultipath/Makefile
> index 85767ab4..379f4d02 100644
> --- a/libmultipath/Makefile
> +++ b/libmultipath/Makefile
> @@ -38,6 +38,11 @@ nvme-lib.o: nvme-lib.c nvme-ioctl.c nvme-ioctl.h
>  dict.o:	dict.c
>  	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -Wno-unused-parameter -c -o $@ $<
>  
> +# checkers/alua.o needs to find headers in parent directory
> +checkers/alua.o: checkers/alua.c
> +	@echo building $@ because of $?
> +	$(Q)$(CC) $(CPPFLAGS) $(CFLAGS) -I. -c -o $@ $<
> +

At the risk of sounding stupid, why is this necessary?
libmultipath/checkers/Makefile includes libmultipath in CPPFLAGS.

>  make_static = $(shell sed '/^static/!s/^\([a-z]\{1,\} \)/static \1/' <$1 >$2)
>  
>  nvme-ioctl.c: nvme/nvme-ioctl.c
> diff --git a/libmultipath/checkers.c b/libmultipath/checkers.c
> index bb6ad1ee..21aa52d4 100644
> --- a/libmultipath/checkers.c
> +++ b/libmultipath/checkers.c
> @@ -461,6 +461,7 @@ int init_checkers(void)
>  		EMC_CLARIION,
>  		READSECTOR0,
>  		CCISS_TUR,
> +		ALUA_RTPG,
>  	};
>  	unsigned int i;
>  
> diff --git a/libmultipath/checkers.h b/libmultipath/checkers.h
> index a969e7d1..2905f3fc 100644
> --- a/libmultipath/checkers.h
> +++ b/libmultipath/checkers.h
> @@ -98,6 +98,7 @@ enum path_check_state {
>  #define EMC_CLARIION "emc_clariion"
>  #define READSECTOR0  "readsector0"
>  #define CCISS_TUR    "cciss_tur"
> +#define ALUA         "alua"
>  #define NONE         "none"
>  #define INVALID      "invalid"
>  
> diff --git a/libmultipath/checkers/Makefile b/libmultipath/checkers/Makefile
> index 6f7cfb95..dfdd13a1 100644
> --- a/libmultipath/checkers/Makefile
> +++ b/libmultipath/checkers/Makefile
> @@ -17,7 +17,8 @@ LIBS= \
>  	libcheckdirectio.so \
>  	libcheckemc_clariion.so \
>  	libcheckhp_sw.so \
> -	libcheckrdac.so
> +	libcheckrdac.so \
> +	libcheckalua.so
>  
>  all: $(LIBS)
>  
> diff --git a/libmultipath/checkers/alua.c b/libmultipath/checkers/alua.c
> new file mode 100644
> index 00000000..cb3d7000
> --- /dev/null
> +++ b/libmultipath/checkers/alua.c
> @@ -0,0 +1,426 @@
> +/*
> + * ALUA Path Checker
> + *
> + * Copyright (c) 2024
> + * This file is released under the GPL.
> + */
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <string.h>
> +#include <sys/types.h>
> +#include <sys/stat.h>
> +#include <sys/sysmacros.h>
> +#include <unistd.h>
> +#include <fcntl.h>
> +#include <errno.h>
> +#include <pthread.h>
> +#include <urcu.h>
> +#include <urcu/uatomic.h>
> +
> +#include "checkers.h"
> +#include "debug.h"
> +#include "structs.h"
> +#include "prio.h"
> +#include "util.h"
> +#include "time-util.h"
> +#include "../prioritizers/alua.h"
> +#include "../prioritizers/alua_rtpg.h"
> +#include "alua.h"
> +
> +#define MAX_NR_TIMEOUTS 1
> +
> +enum {
> +	MSG_ALUA_RTPG_NOT_SUPPORTED = CHECKER_FIRST_MSGID,
> +	MSG_ALUA_RTPG_TRANSITIONING,
> +	MSG_ALUA_RUNNING,
> +	MSG_ALUA_TIMEOUT,
> +};
> +
> +#define IDX_(x) (MSG_ ## x - CHECKER_FIRST_MSGID)
> +const char *libcheck_msgtable[] = {
> +	[IDX_(ALUA_RTPG_NOT_SUPPORTED)] = " ALUA not supported",
> +	[IDX_(ALUA_RTPG_TRANSITIONING)] = " path is transitioning",
> +	[IDX_(ALUA_RUNNING)] = " still running",
> +	[IDX_(ALUA_TIMEOUT)] = " timed out",
> +	NULL,
> +};
> +
> +struct alua_checker_context {
> +	dev_t devt;
> +	int state;
> +	int running; /* uatomic access only */
> +	int fd;
> +	unsigned int timeout;
> +	time_t time;
> +	pthread_t thread;
> +	pthread_mutex_t lock;
> +	pthread_cond_t active;
> +	int holders; /* uatomic access only */
> +	int msgid;
> +	struct checker_context ctx;
> +	unsigned int nr_timeouts;
> +	bool checked_state;
> +	/* ALUA-specific cached data for prioritizer */
> +	int aas;           /* Asymmetric Access State */
> +	int tpg;           /* Target Port Group */
> +	time_t timestamp;  /* When this data was collected */
> +	/* Data needed for RTPG in thread */
> +	struct path *pp;   /* Path pointer - only valid during check */
> +};
> +
> +int libcheck_init(struct checker *c)
> +{
> +	struct alua_checker_context *ct;
> +	struct stat sb;
> +
> +	ct = malloc(sizeof(struct alua_checker_context));
> +	if (!ct)
> +		return 1;
> +	memset(ct, 0, sizeof(struct alua_checker_context));
> +
> +	ct->state = PATH_UNCHECKED;
> +	ct->fd = -1;
> +	ct->aas = -1;
> +	ct->tpg = -1;
> +	uatomic_set(&ct->holders, 1);
> +	pthread_cond_init_mono(&ct->active);
> +	pthread_mutex_init(&ct->lock, NULL);
> +	if (fstat(c->fd, &sb) == 0)
> +		ct->devt = sb.st_rdev;
> +	ct->ctx.cls = c->cls;
> +	c->context = ct;
> +
> +	return 0;
> +}
> +
> +static void cleanup_context(struct alua_checker_context *ct)
> +{
> +	pthread_mutex_destroy(&ct->lock);
> +	pthread_cond_destroy(&ct->active);
> +	free(ct);
> +}
> +
> +void libcheck_free(struct checker *c)
> +{
> +	if (c->context) {
> +		struct alua_checker_context *ct = c->context;
> +		int holders;
> +		int running;
> +
> +		running = uatomic_xchg(&ct->running, 0);
> +		if (running)
> +			pthread_cancel(ct->thread);
> +		ct->thread = 0;
> +		holders = uatomic_sub_return(&ct->holders, 1);
> +		if (!holders)
> +			cleanup_context(ct);
> +		c->context = NULL;
> +	}
> +	return;
> +}
> +
> +/*
> + * Map ALUA Asymmetric Access State to path state
> + */
> +static int alua_state_to_path_state(int aas, short *msgid)
> +{
> +	switch (aas & 0x0f) {
> +	case AAS_OPTIMIZED:
> +	case AAS_NON_OPTIMIZED:
> +		*msgid = CHECKER_MSGID_UP;
> +		return PATH_UP;
> +	case AAS_STANDBY:
> +		*msgid = CHECKER_MSGID_GHOST;
> +		return PATH_GHOST;
> +	case AAS_TRANSITIONING:
> +		*msgid = MSG_ALUA_RTPG_TRANSITIONING;
> +		return PATH_PENDING;
> +	case AAS_UNAVAILABLE:
> +	case AAS_OFFLINE:
> +	default:
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +}
> +
> +/*
> + * Main ALUA check function - uses alua_rtpg.c library.
> + * This is called either in sync mode or from the async thread.
> + */
> +static int
> +alua_check(struct path *pp, int *aas_out, short *msgid)
> +{
> +	int aas;
> +
> +	if (pp->tpg_id == GROUP_ID_UNDEF) {
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +
> +	aas = get_asymmetric_access_state(pp, pp->tpg_id);
> +	if (aas < 0) {
> +		if (aas == -RTPG_LUN_DISCONNECTED) {
> +			*msgid = CHECKER_MSGID_DISCONNECTED;
> +			return PATH_DISCONNECTED;
> +		}
> +		*msgid = CHECKER_MSGID_DOWN;
> +		return PATH_DOWN;
> +	}
> +
> +	*aas_out = aas;
> +	return alua_state_to_path_state(aas, msgid);
> +}
> +
> +#define alua_thread_cleanup_push(ct) pthread_cleanup_push(cleanup_func, ct)
> +#define alua_thread_cleanup_pop(ct) pthread_cleanup_pop(1)
> +
> +static void cleanup_func(void *data)
> +{
> +	int holders;
> +	struct alua_checker_context *ct = data;
> +
> +	holders = uatomic_sub_return(&ct->holders, 1);
> +	if (!holders)
> +		cleanup_context(ct);
> +}
> +
> +void *libcheck_thread(struct checker_context *ctx)
> +{
> +	struct alua_checker_context *ct =
> +		container_of(ctx, struct alua_checker_context, ctx);
> +	int state, running;
> +	short msgid;
> +	int aas;
> +
> +	/* This thread can be canceled, so setup clean up */
> +	alua_thread_cleanup_push(ct);
> +
> +	condlog(4, "%d:%d : alua checker starting up", major(ct->devt),
> +		minor(ct->devt));
> +
> +	state = alua_check(ct->pp, &aas, &msgid);
> +	pthread_testcancel();
> +
> +	/* ALUA checker done */
> +	pthread_mutex_lock(&ct->lock);
> +	ct->state = state;
> +	ct->msgid = msgid;
> +	ct->aas = aas;
> +	ct->tpg = ct->pp->tpg_id;
> +	ct->timestamp = time(NULL);

The multipath code pretty much exclusively uses get_monotonic_time() to
track the time. That can get used here as well.

> +	/* Update path-level cache for prioritizer use */
> +	ct->pp->alua_state = aas;
> +	ct->pp->alua_timestamp = ct->timestamp;

This is not safe. The checker_context was specifically designed so that
it was possible to free a path while an async checker was still running.
The path holds a reference to the context (ct->holders), as does the
thread.  When the last reference is dropped, the context is freed. So
there's no way you can know that the path still exists while the thread
is running.  The only place where it is safe to access the path is in
libcheck_check().

> +	pthread_cond_signal(&ct->active);
> +	pthread_mutex_unlock(&ct->lock);
> +
> +	condlog(4, "%d:%d : alua checker finished, state %s", major(ct->devt),
> +		minor(ct->devt), checker_state_name(state));
> +
> +	running = uatomic_xchg(&ct->running, 0);
> +	if (!running)
> +		pause();
> +
> +	alua_thread_cleanup_pop(ct);
> +
> +	return ((void *)0);
> +}
> +
> +static void alua_set_async_timeout(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct timespec now;
> +
> +	get_monotonic_time(&now);
> +	ct->time = now.tv_sec + c->timeout;
> +}
> +
> +static int alua_check_async_timeout(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct timespec now;
> +
> +	get_monotonic_time(&now);
> +	return (now.tv_sec > ct->time);
> +}
> +
> +static int check_pending(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	int alua_status = PATH_PENDING;
> +
> +	pthread_mutex_lock(&ct->lock);
> +	if (ct->state != PATH_PENDING || ct->msgid != MSG_ALUA_RUNNING) {
> +		alua_status = ct->state;
> +		c->msgid = ct->msgid;
> +	}
> +	pthread_mutex_unlock(&ct->lock);
> +	if (alua_status == PATH_PENDING && c->msgid == MSG_ALUA_RUNNING) {
> +		condlog(4, "%d:%d : alua checker still running",
> +			major(ct->devt), minor(ct->devt));
> +	} else {
> +		int running = uatomic_xchg(&ct->running, 0);
> +		if (running)
> +			pthread_cancel(ct->thread);
> +		ct->thread = 0;
> +	}
> +
> +	ct->checked_state = true;
> +	return alua_status;
> +}
> +
> +bool libcheck_need_wait(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	return (ct && ct->thread && uatomic_read(&ct->running) != 0 &&
> +		!ct->checked_state);
> +}
> +
> +int libcheck_pending(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +
> +	/* The if path checker isn't running, just return the exiting value. */
> +	if (!ct || !ct->thread)
> +		return c->path_state;
> +
> +	return check_pending(c);
> +}
> +
> +int libcheck_check(struct checker *c)
> +{
> +	struct alua_checker_context *ct = c->context;
> +	struct path *pp = container_of(c, struct path, checker);
> +	pthread_attr_t attr;
> +	int alua_status = PATH_PENDING, r;
> +	int aas;
> +
> +	if (!ct)
> +		return PATH_UNCHECKED;
> +
> +	if (checker_is_sync(c)) {
> +		alua_status = alua_check(pp, &aas, &c->msgid);
> +		ct->tpg = pp->tpg_id;
> +		ct->aas = aas;
> +		ct->timestamp = time(NULL);

The purpose of the context is to be able to copy data from the thread
back to the path checker. I don't see the need to update it when
you aren't in the thread. And again, we can use get_monotonic_time()
to get the time here.

> +		/* Update path-level cache for prioritizer use */
> +		pp->alua_state = aas;
> +		pp->alua_timestamp = ct->timestamp;
> +		return alua_status;
> +	}
> +
> +	/* Async mode */
> +	if (ct->thread) {
> +		ct->checked_state = true;
> +		if (alua_check_async_timeout(c)) {
> +			int running = uatomic_xchg(&ct->running, 0);
> +			if (running) {
> +				pthread_cancel(ct->thread);
> +				condlog(3, "%d:%d : alua checker timeout",
> +					major(ct->devt), minor(ct->devt));
> +				c->msgid = MSG_ALUA_TIMEOUT;
> +				alua_status = PATH_TIMEOUT;
> +			} else {
> +				pthread_mutex_lock(&ct->lock);
> +				alua_status = ct->state;
> +				c->msgid = ct->msgid;
> +				pthread_mutex_unlock(&ct->lock);
> +			}
> +			ct->thread = 0;
> +		} else if (uatomic_read(&ct->running) != 0) {
> +			/*
> +			 * ct->running may be stale - the thread sets
> +			 * msgid/state under lock, then clears running
> +			 * after releasing it. Check msgid under lock
> +			 * to get the truth. Only MSG_ALUA_RUNNING means
> +			 * no result yet.
> +			 */

How hard is it to trigger the checker hang you saw with this checker.
I'd really like to understand why it happens because I still can't see
the necessity of this.

> +			pthread_mutex_lock(&ct->lock);
> +			if (ct->msgid != MSG_ALUA_RUNNING) {
> +				alua_status = ct->state;
> +				c->msgid = ct->msgid;
> +				ct->thread = 0;
> +			}
> +			pthread_mutex_unlock(&ct->lock);
> +		} else {
> +			/* ALUA checker done */
> +			ct->thread = 0;
> +			pthread_mutex_lock(&ct->lock);
> +			alua_status = ct->state;
> +			c->msgid = ct->msgid;
> +			pthread_mutex_unlock(&ct->lock);
> +		}
> +	} else {
> +		if (uatomic_read(&ct->holders) > 1) {
> +			/* The thread has been cancelled but hasn't quit. */
> +			if (ct->nr_timeouts <= MAX_NR_TIMEOUTS) {

I think you meant ">=" here, instead of "<=".

> +				if (ct->nr_timeouts == MAX_NR_TIMEOUTS) {
> +					condlog(2, "%d:%d : waiting for stalled alua thread to finish",
> +						major(ct->devt), minor(ct->devt));
> +				}
> +				ct->nr_timeouts++;
> +				c->msgid = MSG_ALUA_TIMEOUT;
> +				return PATH_TIMEOUT;
> +			}
> +			ct->nr_timeouts++;
> +			/*
> +			 * Start a new thread while the old one is stalled.
> +			 * We have to prevent it from interfering with the new
> +			 * thread. We create a new context and leave the old
> +			 * one with the stale thread, hoping it will clean up
> +			 * eventually.
> +			 */
> +			condlog(3, "%d:%d : alua thread not responding",
> +				major(ct->devt), minor(ct->devt));
> +
> +			if (libcheck_init(c) != 0) {
> +				c->msgid = CHECKER_MSGID_DOWN;
> +				return PATH_UNCHECKED;
> +			}
> +			((struct alua_checker_context *)c->context)->nr_timeouts = ct->nr_timeouts;
> +
> +			if (!uatomic_sub_return(&ct->holders, 1)) {
> +				/* It did terminate, eventually */
> +				cleanup_context(ct);
> +				((struct alua_checker_context *)c->context)->nr_timeouts = 0;
> +			}
> +
> +			ct = c->context;
> +		} else
> +			ct->nr_timeouts = 0;
> +
> +		/* Start new ALUA checker */
> +		pthread_mutex_lock(&ct->lock);
> +		alua_status = ct->state = PATH_PENDING;
> +		c->msgid = ct->msgid = MSG_ALUA_RUNNING;
> +		pthread_mutex_unlock(&ct->lock);
> +		ct->fd = c->fd;
> +		ct->timeout = c->timeout;
> +		ct->pp = pp;
> +		ct->checked_state = false;
> +		uatomic_add(&ct->holders, 1);
> +		uatomic_set(&ct->running, 1);
> +		alua_set_async_timeout(c);
> +		setup_thread_attr(&attr, 32 * 1024, 1);
> +		r = start_checker_thread(&ct->thread, &attr, &ct->ctx);
> +		pthread_attr_destroy(&attr);
> +		if (r) {
> +			uatomic_sub(&ct->holders, 1);
> +			uatomic_set(&ct->running, 0);
> +			ct->thread = 0;
> +			condlog(3, "%d:%d : failed to start alua thread, using"
> +				" sync mode", major(ct->devt), minor(ct->devt));
> +			alua_status = alua_check(pp, &aas, &c->msgid);
> +			ct->tpg = pp->tpg_id;
> +			ct->aas = aas;
> +			ct->timestamp = time(NULL);

Again with not needing to update ct here and using get_monotonic_time.

> +			/* Update path-level cache for prioritizer use */
> +			pp->alua_state = aas;
> +			pp->alua_timestamp = ct->timestamp;
> +			return alua_status;
> +		}
> +	}
> +
> +	return alua_status;
> +}
> +
> diff --git a/libmultipath/checkers/alua.h b/libmultipath/checkers/alua.h
> new file mode 100644
> index 00000000..b53cef66
> --- /dev/null
> +++ b/libmultipath/checkers/alua.h
> @@ -0,0 +1,15 @@
> +/*
> + * ALUA Path Checker Header
> + *
> + * This file is released under the GPL.
> + */
> +#ifndef _ALUA_CHECKER_H
> +#define _ALUA_CHECKER_H
> +
> +/* ALUA checker caches state in struct path for prioritizer use:
> + *   pp->alua_state     - cached AAS value (-1 if not cached)
> + *   pp->alua_timestamp - when the state was last updated
> + */
> +
> +#endif /* _ALUA_CHECKER_H */
> +
> diff --git a/libmultipath/prio.c b/libmultipath/prio.c
> index 24f825bd..2d59d04e 100644
> --- a/libmultipath/prio.c
> +++ b/libmultipath/prio.c
> @@ -29,6 +29,7 @@ int init_prio(void)
>  #ifdef LOAD_ALL_SHARED_LIBS
>  	static const char *const all_prios[] = {
>  		PRIO_ALUA,
> +		PRIO_ALUA_CACHED,
>  		PRIO_CONST,
>  		PRIO_DATACORE,
>  		PRIO_EMC,
> diff --git a/libmultipath/prio.h b/libmultipath/prio.h
> index 119b75f2..b54ece0d 100644
> --- a/libmultipath/prio.h
> +++ b/libmultipath/prio.h
> @@ -17,6 +17,7 @@ struct path;
>   * Known prioritizers for use in hwtable.c
>   */
>  #define PRIO_ALUA		"alua"
> +#define PRIO_ALUA_CACHED	"alua_cached"
>  #define PRIO_CONST		"const"
>  #define PRIO_DATACORE		"datacore"
>  #define PRIO_EMC		"emc"
> diff --git a/libmultipath/prioritizers/Makefile b/libmultipath/prioritizers/Makefile
> index ff2524c2..201c52c0 100644
> --- a/libmultipath/prioritizers/Makefile
> +++ b/libmultipath/prioritizers/Makefile
> @@ -13,6 +13,7 @@ LIBDEPS = -lmultipath -lmpathutil -lm -lpthread -lrt
>  # If you add or remove a prioritizer also update multipath/multipath.conf.5
>  LIBS = \
>  	libprioalua.so \
> +	libprioalua_cached.so \
>  	libprioconst.so \
>  	libpriodatacore.so \
>  	libprioemc.so \
> diff --git a/libmultipath/prioritizers/alua_cached.c b/libmultipath/prioritizers/alua_cached.c
> new file mode 100644
> index 00000000..2ba5b2e7
> --- /dev/null
> +++ b/libmultipath/prioritizers/alua_cached.c

There is so much duplication in these ALUA checkers, I am a little
perplexed about what to do with this. It makes me wonder if we should
just have one alua prioritzer that can use all the sources like this,
and the different config options would just control which methods the
prioritizer would use.

> @@ -0,0 +1,224 @@
> +/*
> + * (C) Copyright IBM Corp. 2004, 2005   All Rights Reserved.
> + * Modified 2024 to support cached ALUA state from checker
> + *
> + * ALUA prioritizer that can use cached state from alua checker
> + * to avoid duplicate RTPG commands.
> + *
> + * This file is released under the GPL.
> + */
> +#include <stdio.h>
> +#include <string.h>
> +
> +#include <time.h>
> +
> +#include "debug.h"
> +#include "prio.h"
> +#include "structs.h"
> +#include "checkers.h"
> +#include "discovery.h"
> +
> +#include "alua.h"
> +
> +#define ALUA_PRIO_NOT_SUPPORTED			1
> +#define ALUA_PRIO_RTPG_FAILED			2
> +#define ALUA_PRIO_GETAAS_FAILED			3
> +#define ALUA_PRIO_TPGS_FAILED			4
> +#define ALUA_PRIO_NO_INFORMATION		5
> +#define ALUA_PRIO_CACHED_STALE			6
> +
> +/* Maximum age of cached ALUA state in seconds */
> +#define ALUA_CACHE_MAX_AGE_SECS			5
> +
> +static const char * aas_string[] = {
> +	[AAS_OPTIMIZED]		= "active/optimized",
> +	[AAS_NON_OPTIMIZED]	= "active/non-optimized",
> +	[AAS_STANDBY]		= "standby",
> +	[AAS_UNAVAILABLE]	= "unavailable",
> +	[AAS_LBA_DEPENDENT]	= "logical block dependent",
> +	[AAS_RESERVED]		= "ARRAY BUG: invalid TPGs state!",
> +	[AAS_OFFLINE]		= "offline",
> +	[AAS_TRANSITIONING]	= "transitioning between states",
> +};
> +
> +static const char *aas_print_string(int rc)
> +{
> +	rc &= 0x7f;
> +
> +	if (rc & 0x70)
> +		return aas_string[AAS_RESERVED];
> +	rc &= 0x0f;
> +	if (rc > AAS_RESERVED && rc < AAS_OFFLINE)
> +		return aas_string[AAS_RESERVED];
> +	else
> +		return aas_string[rc];
> +}
> +
> +/*
> + * Try to get ALUA info with optimization:
> + * 1. If using alua checker: use its cached state (it just polled!)
> + * 2. Try sysfs (kernel maintains this, no I/O needed)
> + * 3. Fall back to issuing RTPG directly
> + *
> + * Rationale: The checker is our active poller. If we're using alua
> + * and its cache is fresh, we trust that data first. If the cache misses
> + * (stale or empty), we try sysfs before issuing RTPG. This provides a
> + * three-tier optimization: checker cache → sysfs → RTPG.
> + */
> +/* Maximum length of sysfs access_state string */
> +#define ALUA_ACCESS_STATE_SIZE 32
> +
> +int
> +get_alua_info_cached(struct path * pp, bool *used_cache)
> +{
> +	int	rc;
> +	int	tpg;
> +	bool	diff_tpg;
> +	char	buff[ALUA_ACCESS_STATE_SIZE];
> +
> +	*used_cache = false;
> +
> +	/* First: if using alua checker, use its fresh data from path structure */
> +	if (checker_selected(&pp->checker)) {
> +		const char *chk_name = checker_name(&pp->checker);
> +
> +		if (chk_name && strcmp(chk_name, "alua") == 0 &&
> +		    pp->alua_state >= 0) {
> +			time_t now = time(NULL);

Again, you can use get_monotonic_time() & timespecsub() to check the
difference.

> +			time_t age = now - pp->alua_timestamp;
> +
> +			if (pp->alua_timestamp > 0 && age <= ALUA_CACHE_MAX_AGE_SECS) {
> +				condlog(4, "%s: using cached ALUA state from checker (fresh poll)", pp->dev);
> +				*used_cache = true;
> +				return pp->alua_state;
> +			}
> +			condlog(4, "%s: cached ALUA state not available or stale, trying sysfs",
> +				pp->dev);
> +			/* Cache miss - try sysfs before issuing RTPG */
> +		}
> +	}
> +
> +	/* Second: try sysfs (if cache missed or not using alua checker) */
> +	rc = sysfs_get_asymmetric_access_state(pp, buff, sizeof(buff));
> +	if (rc >= 0) {
> +		/* Parse sysfs state string to ALUA state value */
> +		int aas = -1;
> +		int priopath = rc; /* rc contains preferred bit */
> +
> +		if (!strncmp(buff, "active/optimized", 16))
> +			aas = AAS_OPTIMIZED;
> +		else if (!strncmp(buff, "active/non-optimized", 20))
> +			aas = AAS_NON_OPTIMIZED;
> +		else if (!strncmp(buff, "lba-dependent", 13))
> +			aas = AAS_LBA_DEPENDENT;
> +		else if (!strncmp(buff, "standby", 7))
> +			aas = AAS_STANDBY;
> +
> +		if (aas >= 0) {
> +			condlog(4, "%s: using sysfs ALUA state (no I/O)", pp->dev);
> +			*used_cache = true;
> +			return (priopath ? 0x80 : 0) | aas;
> +		}
> +	}
> +
> +	/* Last resort: issue RTPG directly */
> +	tpg = get_target_port_group(pp);
> +	if (tpg < 0) {
> +		rc = get_target_port_group_support(pp);
> +		if (rc < 0)
> +			return -ALUA_PRIO_TPGS_FAILED;
> +		if (rc == TPGS_NONE)
> +			return -ALUA_PRIO_NOT_SUPPORTED;
> +		return -ALUA_PRIO_RTPG_FAILED;
> +	}
> +	diff_tpg = (pp->tpg_id != GROUP_ID_UNDEF && pp->tpg_id != tpg);
> +	pp->tpg_id = tpg;
> +	condlog((diff_tpg) ? 2 : 4, "%s: reported target port group is %i",
> +		pp->dev, tpg);
> +	rc = get_asymmetric_access_state(pp, tpg);
> +	if (rc < 0) {
> +		condlog(2, "%s: get_asymmetric_access_state returned %d",
> +			__func__, rc);
> +		return -ALUA_PRIO_GETAAS_FAILED;
> +	}
> +
> +	condlog(3, "%s: aas = %02x [%s]%s", pp->dev, rc, aas_print_string(rc),
> +		(rc & 0x80) ? " [preferred]" : "");
> +	return rc;
> +}
> +
> +int get_exclusive_pref_arg(char *args)
> +{
> +	char *ptr;
> +
> +	if (args == NULL)
> +		return 0;
> +	ptr = strstr(args, "exclusive_pref_bit");
> +	if (!ptr)
> +		return 0;
> +	if (ptr[18] != '\0' && ptr[18] != ' ' && ptr[18] != '\t')
> +		return 0;
> +	if (ptr != args && ptr[-1] != ' ' && ptr[-1] != '\t')
> +		return 0;
> +	return 1;
> +}
> +
> +int getprio (struct path * pp, char * args)
> +{
> +	int rc;
> +	int aas;
> +	int priopath;
> +	int exclusive_pref;
> +	bool used_cache = false;
> +
> +	if (pp->fd < 0)
> +		return -ALUA_PRIO_NO_INFORMATION;
> +
> +	exclusive_pref = get_exclusive_pref_arg(args);
> +	rc = get_alua_info_cached(pp, &used_cache);
> +
> +	if (used_cache) {
> +		condlog(4, "%s: priority calculated from cached ALUA state (no RTPG issued)",
> +			pp->dev);
> +	}
> +
> +	if (rc >= 0) {
> +		aas = (rc & 0x0f);
> +		priopath = (rc & 0x80);
> +		switch(aas) {
> +		case AAS_OPTIMIZED:
> +			rc = 50;
> +			break;
> +		case AAS_NON_OPTIMIZED:
> +			rc = 10;
> +			break;
> +		case AAS_LBA_DEPENDENT:
> +			rc = 5;
> +			break;
> +		case AAS_STANDBY:
> +			rc = 1;
> +			break;
> +		default:
> +			rc = 0;
> +		}
> +		if (priopath && (aas != AAS_OPTIMIZED || exclusive_pref))
> +			rc += 80;
> +	} else {
> +		switch(-rc) {
> +		case ALUA_PRIO_NOT_SUPPORTED:
> +			condlog(0, "%s: alua not supported", pp->dev);
> +			break;
> +		case ALUA_PRIO_RTPG_FAILED:
> +			condlog(0, "%s: couldn't get target port group", pp->dev);
> +			break;
> +		case ALUA_PRIO_GETAAS_FAILED:
> +			condlog(0, "%s: couldn't get asymmetric access state", pp->dev);
> +			break;
> +		case ALUA_PRIO_TPGS_FAILED:
> +			condlog(3, "%s: couldn't get supported alua states", pp->dev);
> +			break;
> +		}
> +	}
> +	return rc;
> +}
> +
> diff --git a/libmultipath/prioritizers/alua_rtpg.c b/libmultipath/prioritizers/alua_rtpg.c
> index 053cccb7..5da06b50 100644
> --- a/libmultipath/prioritizers/alua_rtpg.c
> +++ b/libmultipath/prioritizers/alua_rtpg.c
> @@ -75,6 +75,7 @@ enum scsi_disposition {
>  	SCSI_GOOD = 0,
>  	SCSI_ERROR,
>  	SCSI_RETRY,
> +	SCSI_DISCONNECTED,
>  };
>  
>  static int
> @@ -124,6 +125,12 @@ scsi_error(struct sg_io_hdr *hdr, int opcode)
>  	PRINT_DEBUG("alua: SCSI error for command %02x: status %02x, sense %02x/%02x/%02x",
>  		    opcode, hdr->status, sense_key, asc, ascq);
>  
> +	/* Check for disconnected LUN (unmapped at target) */
> +	if (sense_key == 0x5 && asc == 0x25 && ascq == 0x00) {
> +		/* ILLEGAL REQUEST / LOGICAL UNIT NOT SUPPORTED */
> +		return SCSI_DISCONNECTED;
> +	}
> +

You need to make sure to handle SCSI_DISCONNECTED in do_inquiry_sg(),
which also calls scsi_error().

>  	if (sense_key == UNIT_ATTENTION || sense_key == NOT_READY)
>  		return SCSI_RETRY;
>  	else
> @@ -325,7 +332,10 @@ retry:
>  	}
>  
>  	rc = scsi_error(&hdr, OPERATION_CODE_RTPG);
> -	if (rc == SCSI_ERROR) {
> +	if (rc == SCSI_DISCONNECTED) {
> +		PRINT_DEBUG("do_rtpg: LUN disconnected (unmapped at target)!");
> +		return -RTPG_LUN_DISCONNECTED;
> +	} else if (rc == SCSI_ERROR) {
>  		PRINT_DEBUG("do_rtpg: SCSI error!");
>  		return -RTPG_RTPG_FAILED;
>  	} else if (rc == SCSI_RETRY) {
> diff --git a/libmultipath/prioritizers/alua_rtpg.h b/libmultipath/prioritizers/alua_rtpg.h
> index ff6ec0ef..030ca58d 100644
> --- a/libmultipath/prioritizers/alua_rtpg.h
> +++ b/libmultipath/prioritizers/alua_rtpg.h
> @@ -21,6 +21,7 @@
>  #define RTPG_NO_TPG_IDENTIFIER			2
>  #define RTPG_RTPG_FAILED			3
>  #define RTPG_TPG_NOT_FOUND			4
> +#define RTPG_LUN_DISCONNECTED			5
>  
>  int get_target_port_group_support(const struct path *pp);
>  int get_target_port_group(const struct path *pp);
> diff --git a/libmultipath/prioritizers/sysfs.c b/libmultipath/prioritizers/sysfs.c
> index 5e8adc05..ab058ea9 100644
> --- a/libmultipath/prioritizers/sysfs.c
> +++ b/libmultipath/prioritizers/sysfs.c
> @@ -10,6 +10,9 @@
>  #include "discovery.h"
>  #include "prio.h"
>  
> +/* Maximum length of sysfs access_state string */
> +#define ALUA_ACCESS_STATE_SIZE 32
> +
>  static const struct {
>  	unsigned char value;
>  	char *name;
> @@ -39,11 +42,11 @@ int get_exclusive_pref_arg(char *args)
>  int getprio (struct path * pp, char *args)
>  {
>  	int prio = 0, rc, i;
> -	char buff[512];
> +	char buff[ALUA_ACCESS_STATE_SIZE];
>  	int exclusive_pref;
>  
>  	exclusive_pref = get_exclusive_pref_arg(args);
> -	rc = sysfs_get_asymmetric_access_state(pp, buff, 512);
> +	rc = sysfs_get_asymmetric_access_state(pp, buff, sizeof(buff));
>  	if (rc < 0)
>  		return PRIO_UNDEF;
>  	prio = 0;
> diff --git a/libmultipath/propsel.c b/libmultipath/propsel.c
> index 56895c4b..e421abba 100644
> --- a/libmultipath/propsel.c
> +++ b/libmultipath/propsel.c
> @@ -262,6 +262,7 @@ verify_alua_prio(struct multipath *mp)
>  		if (!prio_selected(&pp->prio))
>  			continue;
>  		if (strncmp(name, PRIO_ALUA, PRIO_NAME_LEN) &&
> +		    strncmp(name, PRIO_ALUA_CACHED, PRIO_NAME_LEN) &&
>  		    strncmp(name, PRIO_SYSFS, PRIO_NAME_LEN))
>  			 return false;
>  		assume_alua = true;
> @@ -690,7 +691,12 @@ int select_checker(struct config *conf, struct path *pp)
>  		}
>  		(void)path_get_tpgs(pp);
>  		if (pp->tpgs != TPGS_NONE && pp->tpgs != TPGS_UNDEF) {
> -			ckr_name = TUR;
> +			/*
> +			 * For ALUA devices (any TPGS support), use alua
> +			 * checker which combines path checking with ALUA state
> +			 * caching for optimal performance.
> +			 */
> +			ckr_name = ALUA;
>  			goto out;
>  		}
>  	}

I think we should actually have a separate option to distinguish between
ALUA and TUR as the default autodetected checker for ALUA devices.
Something like detect_pgpolicy_use_tpg does. That way there is just one
config option that needs to get changed to switch between the new and
old behavior.

-Ben

> @@ -767,8 +773,15 @@ void detect_prio(struct path *pp)
>  		if ((tpgs == TPGS_EXPLICIT || !check_rdac(pp)) &&
>  		    sysfs_get_asymmetric_access_state(pp, buff, 512) >= 0)
>  			default_prio = PRIO_SYSFS;
> -		else
> +		else {
>  			default_prio = PRIO_ALUA;
> +			/* If using alua checker, use cached prioritizer */
> +			if (checker_selected(&pp->checker)) {
> +				const char *chk_name = checker_name(&pp->checker);
> +				if (chk_name && strcmp(chk_name, ALUA) == 0)
> +					default_prio = PRIO_ALUA_CACHED;
> +			}
> +		}
>  		break;
>  	default:
>  		return;
> @@ -831,7 +844,8 @@ out:
>  	/*
>  	 * fetch tpgs mode for alua, if its not already obtained
>  	 */
> -	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN)) {
> +	if (!strncmp(prio_name(p), PRIO_ALUA, PRIO_NAME_LEN) ||
> +	    !strncmp(prio_name(p), PRIO_ALUA_CACHED, PRIO_NAME_LEN)) {
>  		int tpgs = path_get_tpgs(pp);
>  
>  		if (tpgs == TPGS_NONE) {
> diff --git a/libmultipath/structs.c b/libmultipath/structs.c
> index f4413127..6483a568 100644
> --- a/libmultipath/structs.c
> +++ b/libmultipath/structs.c
> @@ -126,6 +126,8 @@ alloc_path (void)
>  		pp->fd = -1;
>  		pp->tpgs = TPGS_UNDEF;
>  		pp->tpg_id = GROUP_ID_UNDEF;
> +		pp->alua_state = -1;
> +		pp->alua_timestamp = 0;
>  		pp->priority = PRIO_UNDEF;
>  		pp->checkint = CHECKINT_UNDEF;
>  		checker_clear(&pp->checker);
> diff --git a/libmultipath/structs.h b/libmultipath/structs.h
> index 9adedde2..c4bddee9 100644
> --- a/libmultipath/structs.h
> +++ b/libmultipath/structs.h
> @@ -444,6 +444,9 @@ struct path {
>  	vector hwe;
>  	struct gen_path generic_path;
>  	int tpg_id;
> +	/* Cached ALUA state from checker for prioritizer use */
> +	int alua_state;		/* AAS value, -1 if not cached */
> +	time_t alua_timestamp;	/* When alua_state was last updated */
>  	enum ioctl_info_states ioctl_info;
>  };
>  
> -- 
> 2.50.1 (Apple Git-155)

Re: [PATCH 0/1] checkers: add alua path checker

[Hannes Reinecke]

On 3/12/26 01:16, Brian Bunker wrote:
> For ALUA-capable storage arrays, multipath-tools currently uses TUR
> (Test Unit Ready) as the default path checker while a separate
> prioritizer determines ALUA state. When sysfs provides ALUA state
> (the common case), no RTPG command is needed for priority. However,
> if sysfs is unavailable or detect_prio is off, the prioritizer issues
> RTPG, resulting in two SCSI commands per path check cycle: TUR + RTPG.
> 
> This patch introduces an 'alua' path checker that uses RTPG as the path
> check command, providing two benefits:
> 
> 1. For configurations where RTPG is needed for priority, this eliminates
>     duplicate I/O by combining path checking and ALUA state retrieval
>     into a single RTPG command.
> 
> 2. RTPG provides richer path state information than TUR:
>     * TRANSITIONING state detected and mapped to PATH_PENDING,
>       preventing false all paths down during controller failover
>     * STANDBY state detected and mapped to PATH_GHOST
>     * UNAVAILABLE/OFFLINE states properly distinguished
> 
> The richer state information is valuable even when sysfs is available,
> as TUR cannot distinguish between these ALUA-specific conditions.
> 
> This design intentionally couples the checker and prioritizer: the
> alua_cached prioritizer consumes state cached by the alua checker.
> While this blurs the traditional separation between these components,
> for ALUA the path state and priority fundamentally derive from the
> same underlying data, making the coupling a natural fit.
> 
> This patch sets the alua checker as the auto-detected default for ALUA
> devices to demonstrate its benefits. We recognize that adopting a new
> default checker requires careful consideration; the auto-detection can
> be adjusted or removed based on review feedback.
> 
> Brian Bunker (1):
>    checkers: add alua path checker
> 
>   libmultipath/Makefile                   |   5 +
>   libmultipath/checkers.c                 |   1 +
>   libmultipath/checkers.h                 |   1 +
>   libmultipath/checkers/Makefile          |   3 +-
>   libmultipath/checkers/alua.c            | 426 ++++++++++++++++++++++++
>   libmultipath/checkers/alua.h            |  15 +
>   libmultipath/prio.c                     |   1 +
>   libmultipath/prio.h                     |   1 +
>   libmultipath/prioritizers/Makefile      |   1 +
>   libmultipath/prioritizers/alua_cached.c | 225 +++++++++++++
>   libmultipath/prioritizers/alua_rtpg.c   |  12 +-
>   libmultipath/prioritizers/alua_rtpg.h   |   1 +
>   libmultipath/prioritizers/sysfs.c       |   7 +-
>   libmultipath/propsel.c                  |  25 +-
>   libmultipath/structs.c                  |   2 +
>   libmultipath/structs.h                  |   3 +
>   16 files changed, 721 insertions(+), 8 deletions(-)
>   create mode 100644 libmultipath/checkers/alua.c
>   create mode 100644 libmultipath/checkers/alua.h
>   create mode 100644 libmultipath/prioritizers/alua_cached.c
> 

Hmm. Some array vendor have been less than happy when I
tried a similar thing during development of the ALUA
device handler.
Problem is that RTPG requires a cluster callout on the
array side, as it needs to contain information about
the _entire_ target system. Which, for distributed systems,
is a costly operation. So using this as a path
checker is not making certain array vendors happy.

However: as you might have followed the discussion
about scsi multipathing, maybe we can turn this around
to profit both sides:
Can't we move ALUA state handling into the generic
code, leaving the device handler only to deal with
explicit ALUA?
That way we can intercept sense codes and send RTPG
only when required, such that the ALUA state of
the scsi device is correct.
And the the ALUA path checker just has to evaluate
the information in sysfs.
_And_ scsi multipathing can use that information, too.

Hmm?

Cheers,

Hannes
-- 
Dr. Hannes Reinecke                  Kernel Storage Architect
hare@suse.de                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich

Re: [PATCH 0/1] checkers: add alua path checker

[Martin Wilck]

On Thu, 2026-03-12 at 08:10 +0100, Hannes Reinecke wrote:
> > 
> 
> Hmm. Some array vendor have been less than happy when I
> tried a similar thing during development of the ALUA
> device handler.
> Problem is that RTPG requires a cluster callout on the
> array side, as it needs to contain information about
> the _entire_ target system. Which, for distributed systems,
> is a costly operation. So using this as a path
> checker is not making certain array vendors happy.
> 

I wonder if we could avoid this problem by listening to 
SDEV_EVT_ALUA_STATE_CHANGE_REPORTED uevents in multipathd.

Actually, multipathd could use TUR for checking unless we receive an
event of this type. multipathd could listen to those events and then
retrieve the new device state(s) from sysfs, without sending an RTPG
command itself.

We wouldn't switch to the alua checker by default anyway, so the
vendors that prefer the sysfs prioritizer won't be hurt even
if that doesn't work.

> However: as you might have followed the discussion
> about scsi multipathing, maybe we can turn this around
> to profit both sides:
> Can't we move ALUA state handling into the generic
> code, leaving the device handler only to deal with
> explicit ALUA?

When you say "generic code", I suppose you are talking about
the SCSI midlayer. It isn't clear to me how multipathd would benefit
from handling ALUA there, other than being able to read state from
sysfs, as it already does.

Regards
Martin

> That way we can intercept sense codes and send RTPG
> only when required, such that the ALUA state of
> the scsi device is correct.
> And the the ALUA path checker just has to evaluate
> the information in sysfs.
> _And_ scsi multipathing can use that information, too.
> 
> Hmm?
> 
> Cheers,
> 
> Hannes

Re: [PATCH 0/1] checkers: add alua path checker

[Xose Vazquez Perez]

On 3/12/26 5:48 PM, Martin Wilck wrote:

> Actually, multipathd could use TUR for checking unless we receive an
> event of this type. multipathd could listen to those events and then
> retrieve the new device state(s) from sysfs, without sending an RTPG
> command itself.
> 
> We wouldn't switch to the alua checker by default anyway, so the
> vendors that prefer the sysfs prioritizer won't be hurt even
> if that doesn't work.

Just one observation: currently there are ALUA arrays (NetApp E/EF, Dell Unity)
where their own checker is preferred.
As well as software-defined storage (Hitachi Vantara VSP One SDS Block, Linux-IO
(LIO) Target) where directio must be used.

Re: [PATCH 0/1] checkers: add alua path checker

[Brian Bunker]

On Thu, Mar 12, 2026 at 12:10 AM Hannes Reinecke <hare@suse.de> wrote:
>
> On 3/12/26 01:16, Brian Bunker wrote:
> > For ALUA-capable storage arrays, multipath-tools currently uses TUR
> > (Test Unit Ready) as the default path checker while a separate
> > prioritizer determines ALUA state. When sysfs provides ALUA state
> > (the common case), no RTPG command is needed for priority. However,
> > if sysfs is unavailable or detect_prio is off, the prioritizer issues
> > RTPG, resulting in two SCSI commands per path check cycle: TUR + RTPG.
> >
> > This patch introduces an 'alua' path checker that uses RTPG as the path
> > check command, providing two benefits:
> >
> > 1. For configurations where RTPG is needed for priority, this eliminates
> >     duplicate I/O by combining path checking and ALUA state retrieval
> >     into a single RTPG command.
> >
> > 2. RTPG provides richer path state information than TUR:
> >     * TRANSITIONING state detected and mapped to PATH_PENDING,
> >       preventing false all paths down during controller failover
> >     * STANDBY state detected and mapped to PATH_GHOST
> >     * UNAVAILABLE/OFFLINE states properly distinguished
> >
> > The richer state information is valuable even when sysfs is available,
> > as TUR cannot distinguish between these ALUA-specific conditions.
> >
> > This design intentionally couples the checker and prioritizer: the
> > alua_cached prioritizer consumes state cached by the alua checker.
> > While this blurs the traditional separation between these components,
> > for ALUA the path state and priority fundamentally derive from the
> > same underlying data, making the coupling a natural fit.
> >
> > This patch sets the alua checker as the auto-detected default for ALUA
> > devices to demonstrate its benefits. We recognize that adopting a new
> > default checker requires careful consideration; the auto-detection can
> > be adjusted or removed based on review feedback.
> >
> > Brian Bunker (1):
> >    checkers: add alua path checker
> >
> >   libmultipath/Makefile                   |   5 +
> >   libmultipath/checkers.c                 |   1 +
> >   libmultipath/checkers.h                 |   1 +
> >   libmultipath/checkers/Makefile          |   3 +-
> >   libmultipath/checkers/alua.c            | 426 ++++++++++++++++++++++++
> >   libmultipath/checkers/alua.h            |  15 +
> >   libmultipath/prio.c                     |   1 +
> >   libmultipath/prio.h                     |   1 +
> >   libmultipath/prioritizers/Makefile      |   1 +
> >   libmultipath/prioritizers/alua_cached.c | 225 +++++++++++++
> >   libmultipath/prioritizers/alua_rtpg.c   |  12 +-
> >   libmultipath/prioritizers/alua_rtpg.h   |   1 +
> >   libmultipath/prioritizers/sysfs.c       |   7 +-
> >   libmultipath/propsel.c                  |  25 +-
> >   libmultipath/structs.c                  |   2 +
> >   libmultipath/structs.h                  |   3 +
> >   16 files changed, 721 insertions(+), 8 deletions(-)
> >   create mode 100644 libmultipath/checkers/alua.c
> >   create mode 100644 libmultipath/checkers/alua.h
> >   create mode 100644 libmultipath/prioritizers/alua_cached.c
> >
>
> Hmm. Some array vendor have been less than happy when I
> tried a similar thing during development of the ALUA
> device handler.
> Problem is that RTPG requires a cluster callout on the
> array side, as it needs to contain information about
> the _entire_ target system. Which, for distributed systems,
> is a costly operation. So using this as a path
> checker is not making certain array vendors happy.
From working on one of these arrays, many of the path
state decisions often require the same set of information
to respond correctly to test unit ready.  Other OS's including
both Windows and ESX, once they determine that the array
supports ALUA, use RTPG rather than TUR as the periodic
path checker.
>
> However: as you might have followed the discussion
> about scsi multipathing, maybe we can turn this around
> to profit both sides:
> Can't we move ALUA state handling into the generic
> code, leaving the device handler only to deal with
> explicit ALUA?
> That way we can intercept sense codes and send RTPG
> only when required, such that the ALUA state of
> the scsi device is correct.
> And the the ALUA path checker just has to evaluate
> the information in sysfs.
> _And_ scsi multipathing can use that information, too.
>
> Hmm?
I think that there are likely ways to re-imagine the checker
and prioritizer separation that would have benefit. This
patch wasn't really about going in "as is" but to start that
sort of thought.

In general multipath-tools is "good enough" to support the
ALUA story from my perspective as a target vendor. There
are some improvements that would likely simplify things.

The Linux kernel, in my view, should take these two
changes to complete its ALUA story:
https://marc.info/?l=linux-scsi&m=177189801604075&w=2
https://marc.info/?l=linux-scsi&m=176530648031820&w=2

It would be great if someone could review those two changes.
>
> Cheers,
>
> Hannes
> --
> Dr. Hannes Reinecke                  Kernel Storage Architect
> hare@suse.de                                +49 911 74053 688
> SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
> HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich

Thanks,
Brian
-- 
Brian Bunker
PURE Storage, Inc.
brian@purestorage.com

Re: [PATCH 0/1] pr: add alua path checker

[Xose Vazquez Perez]

On 3/12/26 1:16 AM, Brian Bunker wrote:

> For ALUA-capable storage arrays, multipath-tools currently uses TUR
> (Test Unit Ready) as the default path checker while a separate
> prioritizer determines ALUA state. When sysfs provides ALUA state
> (the common case), no RTPG command is needed for priority. However,
> if sysfs is unavailable or detect_prio is off, the prioritizer issues
> RTPG, resulting in two SCSI commands per path check cycle: TUR + RTPG.

prio/checker changes should be documented in multipath/multipath.conf.5.in

Thanks.