Pull xpc-disengage into release branch

4 files changed, 1130 insertions, 368 deletions

diff --git a/arch/ia64/sn/kernel/xpc.h b/arch/ia64/sn/kernel/xpc.h
index d0ee635daf2..33df1b3758b 100644
--- a/arch/ia64/sn/kernel/xpc.h
+++ b/arch/ia64/sn/kernel/xpc.h
@@ -57,7 +57,7 @@
 #define XPC_NASID_FROM_W_B(_w, _b) (((_w) * 64 + (_b)) * 2)
 
 #define XPC_HB_DEFAULT_INTERVAL		5	/* incr HB every x secs */
-#define XPC_HB_CHECK_DEFAULT_TIMEOUT	20	/* check HB every x secs */
+#define XPC_HB_CHECK_DEFAULT_INTERVAL	20	/* check HB every x secs */
 
 /* define the process name of HB checker and the CPU it is pinned to */
 #define XPC_HB_CHECK_THREAD_NAME	"xpc_hb"
@@ -67,34 +67,82 @@
 #define XPC_DISCOVERY_THREAD_NAME	"xpc_discovery"
 
 
-#define XPC_HB_ALLOWED(_p, _v)	((_v)->heartbeating_to_mask & (1UL << (_p)))
-#define XPC_ALLOW_HB(_p, _v)	(_v)->heartbeating_to_mask |= (1UL << (_p))
-#define XPC_DISALLOW_HB(_p, _v)	(_v)->heartbeating_to_mask &= (~(1UL << (_p)))
-
-
 /*
- * Reserved Page provided by SAL.
+ * the reserved page
+ *
+ *   SAL reserves one page of memory per partition for XPC. Though a full page
+ *   in length (16384 bytes), its starting address is not page aligned, but it
+ *   is cacheline aligned. The reserved page consists of the following:
+ *
+ *   reserved page header
+ *
+ *     The first cacheline of the reserved page contains the header
+ *     (struct xpc_rsvd_page). Before SAL initialization has completed,
+ *     SAL has set up the following fields of the reserved page header:
+ *     SAL_signature, SAL_version, partid, and nasids_size. The other
+ *     fields are set up by XPC. (xpc_rsvd_page points to the local
+ *     partition's reserved page.)
  *
- * SAL provides one page per partition of reserved memory.  When SAL
- * initialization is complete, SAL_signature, SAL_version, partid,
- * part_nasids, and mach_nasids are set.
+ *   part_nasids mask
+ *   mach_nasids mask
+ *
+ *     SAL also sets up two bitmaps (or masks), one that reflects the actual
+ *     nasids in this partition (part_nasids), and the other that reflects
+ *     the actual nasids in the entire machine (mach_nasids). We're only
+ *     interested in the even numbered nasids (which contain the processors
+ *     and/or memory), so we only need half as many bits to represent the
+ *     nasids. The part_nasids mask is located starting at the first cacheline
+ *     following the reserved page header. The mach_nasids mask follows right
+ *     after the part_nasids mask. The size in bytes of each mask is reflected
+ *     by the reserved page header field 'nasids_size'. (Local partition's
+ *     mask pointers are xpc_part_nasids and xpc_mach_nasids.)
+ *
+ *   vars
+ *   vars part
+ *
+ *     Immediately following the mach_nasids mask are the XPC variables
+ *     required by other partitions. First are those that are generic to all
+ *     partitions (vars), followed on the next available cacheline by those
+ *     which are partition specific (vars part). These are setup by XPC.
+ *     (Local partition's vars pointers are xpc_vars and xpc_vars_part.)
  *
  * Note: Until vars_pa is set, the partition XPC code has not been initialized.
  */
 struct xpc_rsvd_page {
-	u64 SAL_signature;	/* SAL unique signature */
-	u64 SAL_version;	/* SAL specified version */
-	u8 partid;		/* partition ID from SAL */
+	u64 SAL_signature;	/* SAL: unique signature */
+	u64 SAL_version;	/* SAL: version */
+	u8 partid;		/* SAL: partition ID */
 	u8 version;
-	u8 pad[6];		/* pad to u64 align */
+	u8 pad1[6];		/* align to next u64 in cacheline */
 	volatile u64 vars_pa;
-	u64 part_nasids[XP_NASID_MASK_WORDS] ____cacheline_aligned;
-	u64 mach_nasids[XP_NASID_MASK_WORDS] ____cacheline_aligned;
+	struct timespec stamp;	/* time when reserved page was setup by XPC */
+	u64 pad2[9];		/* align to last u64 in cacheline */
+	u64 nasids_size;	/* SAL: size of each nasid mask in bytes */
 };
-#define XPC_RP_VERSION _XPC_VERSION(1,0) /* version 1.0 of the reserved page */
 
-#define XPC_RSVD_PAGE_ALIGNED_SIZE \
-			(L1_CACHE_ALIGN(sizeof(struct xpc_rsvd_page)))
+#define XPC_RP_VERSION _XPC_VERSION(1,1) /* version 1.1 of the reserved page */
+
+#define XPC_SUPPORTS_RP_STAMP(_version) \
+			(_version >= _XPC_VERSION(1,1))
+
+/*
+ * compare stamps - the return value is:
+ *
+ *	< 0,	if stamp1 < stamp2
+ *	= 0,	if stamp1 == stamp2
+ *	> 0,	if stamp1 > stamp2
+ */
+static inline int
+xpc_compare_stamps(struct timespec *stamp1, struct timespec *stamp2)
+{
+	int ret;
+
+
+	if ((ret = stamp1->tv_sec - stamp2->tv_sec) == 0) {
+		ret = stamp1->tv_nsec - stamp2->tv_nsec;
+	}
+	return ret;
+}
 
 
 /*
@@ -121,11 +169,58 @@ struct xpc_vars {
 	u64 vars_part_pa;
 	u64 amos_page_pa;	/* paddr of page of AMOs from MSPEC driver */
 	AMO_t *amos_page;	/* vaddr of page of AMOs from MSPEC driver */
-	AMO_t *act_amos;	/* pointer to the first activation AMO */
 };
-#define XPC_V_VERSION _XPC_VERSION(3,0) /* version 3.0 of the cross vars */
 
-#define XPC_VARS_ALIGNED_SIZE  (L1_CACHE_ALIGN(sizeof(struct xpc_vars)))
+#define XPC_V_VERSION _XPC_VERSION(3,1) /* version 3.1 of the cross vars */
+
+#define XPC_SUPPORTS_DISENGAGE_REQUEST(_version) \
+			(_version >= _XPC_VERSION(3,1))
+
+
+static inline int
+xpc_hb_allowed(partid_t partid, struct xpc_vars *vars)
+{
+	return ((vars->heartbeating_to_mask & (1UL << partid)) != 0);
+}
+
+static inline void
+xpc_allow_hb(partid_t partid, struct xpc_vars *vars)
+{
+	u64 old_mask, new_mask;
+
+	do {
+		old_mask = vars->heartbeating_to_mask;
+		new_mask = (old_mask | (1UL << partid));
+	} while (cmpxchg(&vars->heartbeating_to_mask, old_mask, new_mask) !=
+							old_mask);
+}
+
+static inline void
+xpc_disallow_hb(partid_t partid, struct xpc_vars *vars)
+{
+	u64 old_mask, new_mask;
+
+	do {
+		old_mask = vars->heartbeating_to_mask;
+		new_mask = (old_mask & ~(1UL << partid));
+	} while (cmpxchg(&vars->heartbeating_to_mask, old_mask, new_mask) !=
+							old_mask);
+}
+
+
+/*
+ * The AMOs page consists of a number of AMO variables which are divided into
+ * four groups, The first two groups are used to identify an IRQ's sender.
+ * These two groups consist of 64 and 128 AMO variables respectively. The last
+ * two groups, consisting of just one AMO variable each, are used to identify
+ * the remote partitions that are currently engaged (from the viewpoint of
+ * the XPC running on the remote partition).
+ */
+#define XPC_NOTIFY_IRQ_AMOS	   0
+#define XPC_ACTIVATE_IRQ_AMOS	   (XPC_NOTIFY_IRQ_AMOS + XP_MAX_PARTITIONS)
+#define XPC_ENGAGED_PARTITIONS_AMO (XPC_ACTIVATE_IRQ_AMOS + XP_NASID_MASK_WORDS)
+#define XPC_DISENGAGE_REQUEST_AMO  (XPC_ENGAGED_PARTITIONS_AMO + 1)
+
 
 /*
  * The following structure describes the per partition specific variables.
@@ -165,6 +260,16 @@ struct xpc_vars_part {
 #define XPC_VP_MAGIC2	0x0073726176435058L  /* 'XPCvars\0'L (little endian) */
 
 
+/* the reserved page sizes and offsets */
+
+#define XPC_RP_HEADER_SIZE	L1_CACHE_ALIGN(sizeof(struct xpc_rsvd_page))
+#define XPC_RP_VARS_SIZE 	L1_CACHE_ALIGN(sizeof(struct xpc_vars))
+
+#define XPC_RP_PART_NASIDS(_rp) (u64 *) ((u8 *) _rp + XPC_RP_HEADER_SIZE)
+#define XPC_RP_MACH_NASIDS(_rp) (XPC_RP_PART_NASIDS(_rp) + xp_nasid_mask_words)
+#define XPC_RP_VARS(_rp)	((struct xpc_vars *) XPC_RP_MACH_NASIDS(_rp) + xp_nasid_mask_words)
+#define XPC_RP_VARS_PART(_rp)	(struct xpc_vars_part *) ((u8 *) XPC_RP_VARS(rp) + XPC_RP_VARS_SIZE)
+
 
 /*
  * Functions registered by add_timer() or called by kernel_thread() only
@@ -349,6 +454,9 @@ struct xpc_channel {
 	atomic_t n_on_msg_allocate_wq;   /* #on msg allocation wait queue */
 	wait_queue_head_t msg_allocate_wq; /* msg allocation wait queue */
 
+	u8 delayed_IPI_flags;		/* IPI flags received, but delayed */
+					/* action until channel disconnected */
+
 	/* queue of msg senders who want to be notified when msg received */
 
 	atomic_t n_to_notify;		/* #of msg senders to notify */
@@ -358,7 +466,7 @@ struct xpc_channel {
 	void *key;			/* pointer to user's key */
 
 	struct semaphore msg_to_pull_sema; /* next msg to pull serialization */
-	struct semaphore teardown_sema;    /* wait for teardown completion */
+	struct semaphore wdisconnect_sema; /* wait for channel disconnect */
 
 	struct xpc_openclose_args *local_openclose_args; /* args passed on */
 					/* opening or closing of channel */
@@ -410,6 +518,8 @@ struct xpc_channel {
 
 #define	XPC_C_DISCONNECTED	0x00002000 /* channel is disconnected */
 #define	XPC_C_DISCONNECTING	0x00004000 /* channel is being disconnected */
+#define	XPC_C_DISCONNECTCALLOUT	0x00008000 /* chan disconnected callout made */
+#define	XPC_C_WDISCONNECT	0x00010000 /* waiting for channel disconnect */
 
 
 
@@ -422,6 +532,8 @@ struct xpc_partition {
 
 	/* XPC HB infrastructure */
 
+	u8 remote_rp_version;		/* version# of partition's rsvd pg */
+	struct timespec remote_rp_stamp;/* time when rsvd pg was initialized */
 	u64 remote_rp_pa;		/* phys addr of partition's rsvd pg */
 	u64 remote_vars_pa;		/* phys addr of partition's vars */
 	u64 remote_vars_part_pa;	/* phys addr of partition's vars part */
@@ -432,14 +544,18 @@ struct xpc_partition {
 	u32 act_IRQ_rcvd;		/* IRQs since activation */
 	spinlock_t act_lock;		/* protect updating of act_state */
 	u8 act_state;			/* from XPC HB viewpoint */
+	u8 remote_vars_version;		/* version# of partition's vars */
 	enum xpc_retval reason;		/* reason partition is deactivating */
 	int reason_line;		/* line# deactivation initiated from */
 	int reactivate_nasid;		/* nasid in partition to reactivate */
 
+	unsigned long disengage_request_timeout; /* timeout in jiffies */
+	struct timer_list disengage_request_timer;
+
 
 	/* XPC infrastructure referencing and teardown control */
 
-	volatile u8 setup_state;			/* infrastructure setup state */
+	volatile u8 setup_state;	/* infrastructure setup state */
 	wait_queue_head_t teardown_wq;	/* kthread waiting to teardown infra */
 	atomic_t references;		/* #of references to infrastructure */
 
@@ -454,6 +570,7 @@ struct xpc_partition {
 
 	u8 nchannels;		   /* #of defined channels supported */
 	atomic_t nchannels_active; /* #of channels that are not DISCONNECTED */
+	atomic_t nchannels_engaged;/* #of channels engaged with remote part */
 	struct xpc_channel *channels;/* array of channel structures */
 
 	void *local_GPs_base;	  /* base address of kmalloc'd space */
@@ -518,6 +635,7 @@ struct xpc_partition {
 #define XPC_P_TORNDOWN		0x03	/* infrastructure is torndown */
 
 
+
 /*
  * struct xpc_partition IPI_timer #of seconds to wait before checking for
  * dropped IPIs. These occur whenever an IPI amo write doesn't complete until
@@ -526,6 +644,13 @@ struct xpc_partition {
 #define XPC_P_DROPPED_IPI_WAIT	(0.25 * HZ)
 
 
+/* number of seconds to wait for other partitions to disengage */
+#define XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT	90
+
+/* interval in seconds to print 'waiting disengagement' messages */
+#define XPC_DISENGAGE_PRINTMSG_INTERVAL		10
+
+
 #define XPC_PARTID(_p)	((partid_t) ((_p) - &xpc_partitions[0]))
 
 
@@ -534,24 +659,20 @@ struct xpc_partition {
 extern struct xpc_registration xpc_registrations[];
 
 
-/* >>> found in xpc_main.c only */
+/* found in xpc_main.c */
 extern struct device *xpc_part;
 extern struct device *xpc_chan;
+extern int xpc_disengage_request_timelimit;
 extern irqreturn_t xpc_notify_IRQ_handler(int, void *, struct pt_regs *);
 extern void xpc_dropped_IPI_check(struct xpc_partition *);
+extern void xpc_activate_partition(struct xpc_partition *);
 extern void xpc_activate_kthreads(struct xpc_channel *, int);
 extern void xpc_create_kthreads(struct xpc_channel *, int);
 extern void xpc_disconnect_wait(int);
 
 
-/* found in xpc_main.c and efi-xpc.c */
-extern void xpc_activate_partition(struct xpc_partition *);
-
-
 /* found in xpc_partition.c */
 extern int xpc_exiting;
-extern int xpc_hb_interval;
-extern int xpc_hb_check_interval;
 extern struct xpc_vars *xpc_vars;
 extern struct xpc_rsvd_page *xpc_rsvd_page;
 extern struct xpc_vars_part *xpc_vars_part;
@@ -561,6 +682,7 @@ extern struct xpc_rsvd_page *xpc_rsvd_page_init(void);
 extern void xpc_allow_IPI_ops(void);
 extern void xpc_restrict_IPI_ops(void);
 extern int xpc_identify_act_IRQ_sender(void);
+extern int xpc_partition_disengaged(struct xpc_partition *);
 extern enum xpc_retval xpc_mark_partition_active(struct xpc_partition *);
 extern void xpc_mark_partition_inactive(struct xpc_partition *);
 extern void xpc_discovery(void);
@@ -585,8 +707,8 @@ extern void xpc_connected_callout(struct xpc_channel *);
 extern void xpc_deliver_msg(struct xpc_channel *);
 extern void xpc_disconnect_channel(const int, struct xpc_channel *,
 					enum xpc_retval, unsigned long *);
-extern void xpc_disconnected_callout(struct xpc_channel *);
-extern void xpc_partition_down(struct xpc_partition *, enum xpc_retval);
+extern void xpc_disconnecting_callout(struct xpc_channel *);
+extern void xpc_partition_going_down(struct xpc_partition *, enum xpc_retval);
 extern void xpc_teardown_infrastructure(struct xpc_partition *);
 
 
@@ -674,6 +796,157 @@ xpc_part_ref(struct xpc_partition *part)
 
 
 /*
+ * This next set of inlines are used to keep track of when a partition is
+ * potentially engaged in accessing memory belonging to another partition.
+ */
+
+static inline void
+xpc_mark_partition_engaged(struct xpc_partition *part)
+{
+	unsigned long irq_flags;
+	AMO_t *amo = (AMO_t *) __va(part->remote_amos_page_pa +
+				(XPC_ENGAGED_PARTITIONS_AMO * sizeof(AMO_t)));
+
+
+	local_irq_save(irq_flags);
+
+	/* set bit corresponding to our partid in remote partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_OR,
+						(1UL << sn_partition_id));
+	/*
+	 * We must always use the nofault function regardless of whether we
+	 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
+	 * didn't, we'd never know that the other partition is down and would
+	 * keep sending IPIs and AMOs to it until the heartbeat times out.
+	 */
+	(void) xp_nofault_PIOR((u64 *) GLOBAL_MMR_ADDR(NASID_GET(&amo->
+				variable), xp_nofault_PIOR_target));
+
+	local_irq_restore(irq_flags);
+}
+
+static inline void
+xpc_mark_partition_disengaged(struct xpc_partition *part)
+{
+	unsigned long irq_flags;
+	AMO_t *amo = (AMO_t *) __va(part->remote_amos_page_pa +
+				(XPC_ENGAGED_PARTITIONS_AMO * sizeof(AMO_t)));
+
+
+	local_irq_save(irq_flags);
+
+	/* clear bit corresponding to our partid in remote partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_AND,
+						~(1UL << sn_partition_id));
+	/*
+	 * We must always use the nofault function regardless of whether we
+	 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
+	 * didn't, we'd never know that the other partition is down and would
+	 * keep sending IPIs and AMOs to it until the heartbeat times out.
+	 */
+	(void) xp_nofault_PIOR((u64 *) GLOBAL_MMR_ADDR(NASID_GET(&amo->
+				variable), xp_nofault_PIOR_target));
+
+	local_irq_restore(irq_flags);
+}
+
+static inline void
+xpc_request_partition_disengage(struct xpc_partition *part)
+{
+	unsigned long irq_flags;
+	AMO_t *amo = (AMO_t *) __va(part->remote_amos_page_pa +
+				(XPC_DISENGAGE_REQUEST_AMO * sizeof(AMO_t)));
+
+
+	local_irq_save(irq_flags);
+
+	/* set bit corresponding to our partid in remote partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_OR,
+						(1UL << sn_partition_id));
+	/*
+	 * We must always use the nofault function regardless of whether we
+	 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
+	 * didn't, we'd never know that the other partition is down and would
+	 * keep sending IPIs and AMOs to it until the heartbeat times out.
+	 */
+	(void) xp_nofault_PIOR((u64 *) GLOBAL_MMR_ADDR(NASID_GET(&amo->
+				variable), xp_nofault_PIOR_target));
+
+	local_irq_restore(irq_flags);
+}
+
+static inline void
+xpc_cancel_partition_disengage_request(struct xpc_partition *part)
+{
+	unsigned long irq_flags;
+	AMO_t *amo = (AMO_t *) __va(part->remote_amos_page_pa +
+				(XPC_DISENGAGE_REQUEST_AMO * sizeof(AMO_t)));
+
+
+	local_irq_save(irq_flags);
+
+	/* clear bit corresponding to our partid in remote partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_AND,
+						~(1UL << sn_partition_id));
+	/*
+	 * We must always use the nofault function regardless of whether we
+	 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
+	 * didn't, we'd never know that the other partition is down and would
+	 * keep sending IPIs and AMOs to it until the heartbeat times out.
+	 */
+	(void) xp_nofault_PIOR((u64 *) GLOBAL_MMR_ADDR(NASID_GET(&amo->
+				variable), xp_nofault_PIOR_target));
+
+	local_irq_restore(irq_flags);
+}
+
+static inline u64
+xpc_partition_engaged(u64 partid_mask)
+{
+	AMO_t *amo = xpc_vars->amos_page + XPC_ENGAGED_PARTITIONS_AMO;
+
+
+	/* return our partition's AMO variable ANDed with partid_mask */
+	return (FETCHOP_LOAD_OP(TO_AMO((u64) &amo->variable), FETCHOP_LOAD) &
+								partid_mask);
+}
+
+static inline u64
+xpc_partition_disengage_requested(u64 partid_mask)
+{
+	AMO_t *amo = xpc_vars->amos_page + XPC_DISENGAGE_REQUEST_AMO;
+
+
+	/* return our partition's AMO variable ANDed with partid_mask */
+	return (FETCHOP_LOAD_OP(TO_AMO((u64) &amo->variable), FETCHOP_LOAD) &
+								partid_mask);
+}
+
+static inline void
+xpc_clear_partition_engaged(u64 partid_mask)
+{
+	AMO_t *amo = xpc_vars->amos_page + XPC_ENGAGED_PARTITIONS_AMO;
+
+
+	/* clear bit(s) based on partid_mask in our partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_AND,
+								~partid_mask);
+}
+
+static inline void
+xpc_clear_partition_disengage_request(u64 partid_mask)
+{
+	AMO_t *amo = xpc_vars->amos_page + XPC_DISENGAGE_REQUEST_AMO;
+
+
+	/* clear bit(s) based on partid_mask in our partition's AMO */
+	FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_AND,
+								~partid_mask);
+}
+
+
+
+/*
  * The following set of macros and inlines are used for the sending and
  * receiving of IPIs (also known as IRQs). There are two flavors of IPIs,
  * one that is associated with partition activity (SGI_XPC_ACTIVATE) and
@@ -722,13 +995,13 @@ xpc_IPI_send(AMO_t *amo, u64 flag, int nasid, int phys_cpuid, int vector)
  * Flag the appropriate AMO variable and send an IPI to the specified node.
  */
 static inline void
-xpc_activate_IRQ_send(u64 amos_page, int from_nasid, int to_nasid,
+xpc_activate_IRQ_send(u64 amos_page_pa, int from_nasid, int to_nasid,
 			int to_phys_cpuid)
 {
 	int w_index = XPC_NASID_W_INDEX(from_nasid);
 	int b_index = XPC_NASID_B_INDEX(from_nasid);
-	AMO_t *amos = (AMO_t *) __va(amos_page +
-					(XP_MAX_PARTITIONS * sizeof(AMO_t)));
+	AMO_t *amos = (AMO_t *) __va(amos_page_pa +
+				(XPC_ACTIVATE_IRQ_AMOS * sizeof(AMO_t)));
 
 
 	(void) xpc_IPI_send(&amos[w_index], (1UL << b_index), to_nasid,
@@ -756,6 +1029,13 @@ xpc_IPI_send_reactivate(struct xpc_partition *part)
 				xpc_vars->act_nasid, xpc_vars->act_phys_cpuid);
 }
 
+static inline void
+xpc_IPI_send_disengage(struct xpc_partition *part)
+{
+	xpc_activate_IRQ_send(part->remote_amos_page_pa, cnodeid_to_nasid(0),
+			part->remote_act_nasid, part->remote_act_phys_cpuid);
+}
+
 
 /*
  * IPIs associated with SGI_XPC_NOTIFY IRQ.
@@ -836,6 +1116,7 @@ xpc_notify_IRQ_send_local(struct xpc_channel *ch, u8 ipi_flag,
 
 /* given an AMO variable and a channel#, get its associated IPI flags */
 #define XPC_GET_IPI_FLAGS(_amo, _c)	((u8) (((_amo) >> ((_c) * 8)) & 0xff))
+#define XPC_SET_IPI_FLAGS(_amo, _c, _f)	(_amo) |= ((u64) (_f) << ((_c) * 8))
 
 #define	XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(_amo) ((_amo) & 0x0f0f0f0f0f0f0f0f)
 #define XPC_ANY_MSG_IPI_FLAGS_SET(_amo)       ((_amo) & 0x1010101010101010)
@@ -903,17 +1184,18 @@ xpc_IPI_send_local_msgrequest(struct xpc_channel *ch)
  * cacheable mapping for the entire region. This will prevent speculative
  * reading of cached copies of our lines from being issued which will cause
  * a PI FSB Protocol error to be generated by the SHUB. For XPC, we need 64
- * (XP_MAX_PARTITIONS) AMO variables for message notification (xpc_main.c)
- * and an additional 16 AMO variables for partition activation (xpc_hb.c).
+ * AMO variables (based on XP_MAX_PARTITIONS) for message notification and an
+ * additional 128 AMO variables (based on XP_NASID_MASK_WORDS) for partition
+ * activation and 2 AMO variables for partition deactivation.
  */
 static inline AMO_t *
-xpc_IPI_init(partid_t partid)
+xpc_IPI_init(int index)
 {
-	AMO_t *part_amo = xpc_vars->amos_page + partid;
+	AMO_t *amo = xpc_vars->amos_page + index;
 
 
-	xpc_IPI_receive(part_amo);
-	return part_amo;
+	(void) xpc_IPI_receive(amo);	/* clear AMO variable */
+	return amo;
 }
 
 
diff --git a/arch/ia64/sn/kernel/xpc_channel.c b/arch/ia64/sn/kernel/xpc_channel.c
index 94698bea7be..abf4fc2a87b 100644
--- a/arch/ia64/sn/kernel/xpc_channel.c
+++ b/arch/ia64/sn/kernel/xpc_channel.c
@@ -57,6 +57,7 @@ xpc_initialize_channels(struct xpc_partition *part, partid_t partid)
 
 		spin_lock_init(&ch->lock);
 		sema_init(&ch->msg_to_pull_sema, 1);	/* mutex */
+		sema_init(&ch->wdisconnect_sema, 0);	/* event wait */
 
 		atomic_set(&ch->n_on_msg_allocate_wq, 0);
 		init_waitqueue_head(&ch->msg_allocate_wq);
@@ -166,6 +167,7 @@ xpc_setup_infrastructure(struct xpc_partition *part)
 	xpc_initialize_channels(part, partid);
 
 	atomic_set(&part->nchannels_active, 0);
+	atomic_set(&part->nchannels_engaged, 0);
 
 
 	/* local_IPI_amo were set to 0 by an earlier memset() */
@@ -555,8 +557,6 @@ xpc_allocate_msgqueues(struct xpc_channel *ch)
 		sema_init(&ch->notify_queue[i].sema, 0);
 	}
 
-	sema_init(&ch->teardown_sema, 0);	/* event wait */
-
 	spin_lock_irqsave(&ch->lock, irq_flags);
 	ch->flags |= XPC_C_SETUP;
 	spin_unlock_irqrestore(&ch->lock, irq_flags);
@@ -626,6 +626,55 @@ xpc_process_connect(struct xpc_channel *ch, unsigned long *irq_flags)
 
 
 /*
+ * Notify those who wanted to be notified upon delivery of their message.
+ */
+static void
+xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
+{
+	struct xpc_notify *notify;
+	u8 notify_type;
+	s64 get = ch->w_remote_GP.get - 1;
+
+
+	while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
+
+		notify = &ch->notify_queue[get % ch->local_nentries];
+
+		/*
+		 * See if the notify entry indicates it was associated with
+		 * a message who's sender wants to be notified. It is possible
+		 * that it is, but someone else is doing or has done the
+		 * notification.
+		 */
+		notify_type = notify->type;
+		if (notify_type == 0 ||
+				cmpxchg(&notify->type, notify_type, 0) !=
+								notify_type) {
+			continue;
+		}
+
+		DBUG_ON(notify_type != XPC_N_CALL);
+
+		atomic_dec(&ch->n_to_notify);
+
+		if (notify->func != NULL) {
+			dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
+				"msg_number=%ld, partid=%d, channel=%d\n",
+				(void *) notify, get, ch->partid, ch->number);
+
+			notify->func(reason, ch->partid, ch->number,
+								notify->key);
+
+			dev_dbg(xpc_chan, "notify->func() returned, "
+				"notify=0x%p, msg_number=%ld, partid=%d, "
+				"channel=%d\n", (void *) notify, get,
+				ch->partid, ch->number);
+		}
+	}
+}
+
+
+/*
  * Free up message queues and other stuff that were allocated for the specified
  * channel.
  *
@@ -669,9 +718,6 @@ xpc_free_msgqueues(struct xpc_channel *ch)
 		ch->remote_msgqueue = NULL;
 		kfree(ch->notify_queue);
 		ch->notify_queue = NULL;
-
-		/* in case someone is waiting for the teardown to complete */
-		up(&ch->teardown_sema);
 	}
 }
 
@@ -683,7 +729,7 @@ static void
 xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 {
 	struct xpc_partition *part = &xpc_partitions[ch->partid];
-	u32 ch_flags = ch->flags;
+	u32 channel_was_connected = (ch->flags & XPC_C_WASCONNECTED);
 
 
 	DBUG_ON(!spin_is_locked(&ch->lock));
@@ -701,12 +747,13 @@ xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 	}
 	DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
 
-	/* it's now safe to free the channel's message queues */
-
-	xpc_free_msgqueues(ch);
-	DBUG_ON(ch->flags & XPC_C_SETUP);
+	if (part->act_state == XPC_P_DEACTIVATING) {
+		/* can't proceed until the other side disengages from us */
+		if (xpc_partition_engaged(1UL << ch->partid)) {
+			return;
+		}
 
-	if (part->act_state != XPC_P_DEACTIVATING) {
+	} else {
 
 		/* as long as the other side is up do the full protocol */
 
@@ -724,16 +771,42 @@ xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
 		}
 	}
 
+	/* wake those waiting for notify completion */
+	if (atomic_read(&ch->n_to_notify) > 0) {
+		/* >>> we do callout while holding ch->lock */
+		xpc_notify_senders(ch, ch->reason, ch->w_local_GP.put);
+	}
+
 	/* both sides are disconnected now */
 
-	ch->flags = XPC_C_DISCONNECTED;	/* clear all flags, but this one */
+	/* it's now safe to free the channel's message queues */
+	xpc_free_msgqueues(ch);
+
+	/* mark disconnected, clear all other flags except XPC_C_WDISCONNECT */
+	ch->flags = (XPC_C_DISCONNECTED | (ch->flags & XPC_C_WDISCONNECT));
 
 	atomic_dec(&part->nchannels_active);
 
-	if (ch_flags & XPC_C_WASCONNECTED) {
+	if (channel_was_connected) {
 		dev_info(xpc_chan, "channel %d to partition %d disconnected, "
 			"reason=%d\n", ch->number, ch->partid, ch->reason);
 	}
+
+	if (ch->flags & XPC_C_WDISCONNECT) {
+		spin_unlock_irqrestore(&ch->lock, *irq_flags);
+		up(&ch->wdisconnect_sema);
+		spin_lock_irqsave(&ch->lock, *irq_flags);
+
+	} else if (ch->delayed_IPI_flags) {
+		if (part->act_state != XPC_P_DEACTIVATING) {
+			/* time to take action on any delayed IPI flags */
+			spin_lock(&part->IPI_lock);
+			XPC_SET_IPI_FLAGS(part->local_IPI_amo, ch->number,
+							ch->delayed_IPI_flags);
+			spin_unlock(&part->IPI_lock);
+		}
+		ch->delayed_IPI_flags = 0;
+	}
 }
 
 
@@ -754,6 +827,19 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 
 	spin_lock_irqsave(&ch->lock, irq_flags);
 
+again:
+
+	if ((ch->flags & XPC_C_DISCONNECTED) &&
+					(ch->flags & XPC_C_WDISCONNECT)) {
+		/*
+		 * Delay processing IPI flags until thread waiting disconnect
+		 * has had a chance to see that the channel is disconnected.
+		 */
+		ch->delayed_IPI_flags |= IPI_flags;
+		spin_unlock_irqrestore(&ch->lock, irq_flags);
+		return;
+	}
+
 
 	if (IPI_flags & XPC_IPI_CLOSEREQUEST) {
 
@@ -764,7 +850,7 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 		/*
 		 * If RCLOSEREQUEST is set, we're probably waiting for
 		 * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
-		 * with this RCLOSEQREUQEST in the IPI_flags.
+		 * with this RCLOSEREQUEST in the IPI_flags.
 		 */
 
 		if (ch->flags & XPC_C_RCLOSEREQUEST) {
@@ -779,14 +865,22 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 
 			/* both sides have finished disconnecting */
 			xpc_process_disconnect(ch, &irq_flags);
+			DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
+			goto again;
 		}
 
 		if (ch->flags & XPC_C_DISCONNECTED) {
-			// >>> explain this section
-
 			if (!(IPI_flags & XPC_IPI_OPENREQUEST)) {
-				DBUG_ON(part->act_state !=
-							XPC_P_DEACTIVATING);
+				if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo,
+					 ch_number) & XPC_IPI_OPENREQUEST)) {
+
+					DBUG_ON(ch->delayed_IPI_flags != 0);
+					spin_lock(&part->IPI_lock);
+					XPC_SET_IPI_FLAGS(part->local_IPI_amo,
+							ch_number,
+							XPC_IPI_CLOSEREQUEST);
+					spin_unlock(&part->IPI_lock);
+				}
 				spin_unlock_irqrestore(&ch->lock, irq_flags);
 				return;
 			}
@@ -816,9 +910,13 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 			}
 
 			XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
-		} else {
-			xpc_process_disconnect(ch, &irq_flags);
+
+			DBUG_ON(IPI_flags & XPC_IPI_CLOSEREPLY);
+			spin_unlock_irqrestore(&ch->lock, irq_flags);
+			return;
 		}
+
+		xpc_process_disconnect(ch, &irq_flags);
 	}
 
 
@@ -834,7 +932,20 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 		}
 
 		DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
-		DBUG_ON(!(ch->flags & XPC_C_RCLOSEREQUEST));
+
+		if (!(ch->flags & XPC_C_RCLOSEREQUEST)) {
+			if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo, ch_number)
+						& XPC_IPI_CLOSEREQUEST)) {
+
+				DBUG_ON(ch->delayed_IPI_flags != 0);
+				spin_lock(&part->IPI_lock);
+				XPC_SET_IPI_FLAGS(part->local_IPI_amo,
+						ch_number, XPC_IPI_CLOSEREPLY);
+				spin_unlock(&part->IPI_lock);
+			}
+			spin_unlock_irqrestore(&ch->lock, irq_flags);
+			return;
+		}
 
 		ch->flags |= XPC_C_RCLOSEREPLY;
 
@@ -852,8 +963,14 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 			"channel=%d\n", args->msg_size, args->local_nentries,
 			ch->partid, ch->number);
 
-		if ((ch->flags & XPC_C_DISCONNECTING) ||
-					part->act_state == XPC_P_DEACTIVATING) {
+		if (part->act_state == XPC_P_DEACTIVATING ||
+					(ch->flags & XPC_C_ROPENREQUEST)) {
+			spin_unlock_irqrestore(&ch->lock, irq_flags);
+			return;
+		}
+
+		if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) {
+			ch->delayed_IPI_flags |= XPC_IPI_OPENREQUEST;
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
 			return;
 		}
@@ -867,8 +984,11 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 		 *      msg_size = size of channel's messages in bytes
 		 *      local_nentries = remote partition's local_nentries
 		 */
-		DBUG_ON(args->msg_size == 0);
-		DBUG_ON(args->local_nentries == 0);
+		if (args->msg_size == 0 || args->local_nentries == 0) {
+			/* assume OPENREQUEST was delayed by mistake */
+			spin_unlock_irqrestore(&ch->lock, irq_flags);
+			return;
+		}
 
 		ch->flags |= (XPC_C_ROPENREQUEST | XPC_C_CONNECTING);
 		ch->remote_nentries = args->local_nentries;
@@ -906,7 +1026,13 @@ xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
 			return;
 		}
-		DBUG_ON(!(ch->flags & XPC_C_OPENREQUEST));
+		if (!(ch->flags & XPC_C_OPENREQUEST)) {
+			XPC_DISCONNECT_CHANNEL(ch, xpcOpenCloseError,
+								&irq_flags);
+			spin_unlock_irqrestore(&ch->lock, irq_flags);
+			return;
+		}
+
 		DBUG_ON(!(ch->flags & XPC_C_ROPENREQUEST));
 		DBUG_ON(ch->flags & XPC_C_CONNECTED);
 
@@ -960,8 +1086,8 @@ xpc_connect_channel(struct xpc_channel *ch)
 	struct xpc_registration *registration = &xpc_registrations[ch->number];
 
 
-	if (down_interruptible(&registration->sema) != 0) {
-		return xpcInterrupted;
+	if (down_trylock(&registration->sema) != 0) {
+		return xpcRetry;
 	}
 
 	if (!XPC_CHANNEL_REGISTERED(ch->number)) {
@@ -1040,55 +1166,6 @@ xpc_connect_channel(struct xpc_channel *ch)
 
 
 /*
- * Notify those who wanted to be notified upon delivery of their message.
- */
-static void
-xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
-{
-	struct xpc_notify *notify;
-	u8 notify_type;
-	s64 get = ch->w_remote_GP.get - 1;
-
-
-	while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
-
-		notify = &ch->notify_queue[get % ch->local_nentries];
-
-		/*
-		 * See if the notify entry indicates it was associated with
-		 * a message who's sender wants to be notified. It is possible
-		 * that it is, but someone else is doing or has done the
-		 * notification.
-		 */
-		notify_type = notify->type;
-		if (notify_type == 0 ||
-				cmpxchg(&notify->type, notify_type, 0) !=
-								notify_type) {
-			continue;
-		}
-
-		DBUG_ON(notify_type != XPC_N_CALL);
-
-		atomic_dec(&ch->n_to_notify);
-
-		if (notify->func != NULL) {
-			dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
-				"msg_number=%ld, partid=%d, channel=%d\n",
-				(void *) notify, get, ch->partid, ch->number);
-
-			notify->func(reason, ch->partid, ch->number,
-								notify->key);
-
-			dev_dbg(xpc_chan, "notify->func() returned, "
-				"notify=0x%p, msg_number=%ld, partid=%d, "
-				"channel=%d\n", (void *) notify, get,
-				ch->partid, ch->number);
-		}
-	}
-}
-
-
-/*
  * Clear some of the msg flags in the local message queue.
  */
 static inline void
@@ -1240,6 +1317,7 @@ xpc_process_channel_activity(struct xpc_partition *part)
 	u64 IPI_amo, IPI_flags;
 	struct xpc_channel *ch;
 	int ch_number;
+	u32 ch_flags;
 
 
 	IPI_amo = xpc_get_IPI_flags(part);
@@ -1266,8 +1344,9 @@ xpc_process_channel_activity(struct xpc_partition *part)
 			xpc_process_openclose_IPI(part, ch_number, IPI_flags);
 		}
 
+		ch_flags = ch->flags;	/* need an atomic snapshot of flags */
 
-		if (ch->flags & XPC_C_DISCONNECTING) {
+		if (ch_flags & XPC_C_DISCONNECTING) {
 			spin_lock_irqsave(&ch->lock, irq_flags);
 			xpc_process_disconnect(ch, &irq_flags);
 			spin_unlock_irqrestore(&ch->lock, irq_flags);
@@ -1278,9 +1357,9 @@ xpc_process_channel_activity(struct xpc_partition *part)
 			continue;
 		}
 
-		if (!(ch->flags & XPC_C_CONNECTED)) {
-			if (!(ch->flags & XPC_C_OPENREQUEST)) {
-				DBUG_ON(ch->flags & XPC_C_SETUP);
+		if (!(ch_flags & XPC_C_CONNECTED)) {
+			if (!(ch_flags & XPC_C_OPENREQUEST)) {
+				DBUG_ON(ch_flags & XPC_C_SETUP);
 				(void) xpc_connect_channel(ch);
 			} else {
 				spin_lock_irqsave(&ch->lock, irq_flags);
@@ -1305,8 +1384,8 @@ xpc_process_channel_activity(struct xpc_partition *part)
 
 
 /*
- * XPC's heartbeat code calls this function to inform XPC that a partition has
- * gone down.  XPC responds by tearing down the XPartition Communication
+ * XPC's heartbeat code calls this function to inform XPC that a partition is
+ * going down.  XPC responds by tearing down the XPartition Communication
  * infrastructure used for the just downed partition.
  *
  * XPC's heartbeat code will never call this function and xpc_partition_up()
@@ -1314,7 +1393,7 @@ xpc_process_channel_activity(struct xpc_partition *part)
  * at the same time.
  */
 void
-xpc_partition_down(struct xpc_partition *part, enum xpc_retval reason)
+xpc_partition_going_down(struct xpc_partition *part, enum xpc_retval reason)
 {
 	unsigned long irq_flags;
 	int ch_number;
@@ -1330,12 +1409,11 @@ xpc_partition_down(struct xpc_partition *part, enum xpc_retval reason)
 	}
 
 
-	/* disconnect all channels associated with the downed partition */
+	/* disconnect channels associated with the partition going down */
 
 	for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
 		ch = &part->channels[ch_number];
 
-
 		xpc_msgqueue_ref(ch);
 		spin_lock_irqsave(&ch->lock, irq_flags);
 
@