Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1765330AbYFFQvX (ORCPT ); Fri, 6 Jun 2008 12:51:23 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1753699AbYFFQvO (ORCPT ); Fri, 6 Jun 2008 12:51:14 -0400 Received: from netops-testserver-3-out.sgi.com ([192.48.171.28]:60899 "EHLO relay.sgi.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1753241AbYFFQvK (ORCPT ); Fri, 6 Jun 2008 12:51:10 -0400 Date: Fri, 6 Jun 2008 11:51:09 -0500 From: Dean Nelson To: akpm@linux-foundation.org Cc: linux-kernel@vger.kernel.org Subject: [Patch 08/18] isolate xpc_vars structure to sn2 only Message-ID: <20080606165109.GI13695@sgi.com> References: <20080606164034.GA13695@sgi.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20080606164034.GA13695@sgi.com> User-Agent: Mutt/1.5.13 (2006-08-11) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 98151 Lines: 3163 Isolate the xpc_vars structure of XPC's reserved page to sn2 only. Signed-off-by: Dean Nelson --- drivers/misc/sgi-xp/xpc.h | 529 ++------------ drivers/misc/sgi-xp/xpc_channel.c | 315 -------- drivers/misc/sgi-xp/xpc_main.c | 152 ++-- drivers/misc/sgi-xp/xpc_partition.c | 454 ------------ drivers/misc/sgi-xp/xpc_sn2.c | 1181 +++++++++++++++++++++++++++++++- drivers/misc/sgi-xp/xpc_uv.c | 57 + 6 files changed, 1433 insertions(+), 1255 deletions(-) Index: linux-2.6/drivers/misc/sgi-xp/xpc.h =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc.h 2008-05-28 07:03:09.360488680 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc.h 2008-05-28 07:03:15.269223054 -0500 @@ -159,10 +159,10 @@ xpc_compare_stamps(struct timespec *stam * reflected by incrementing either the major or minor version numbers * of struct xpc_vars. */ -struct xpc_vars { +struct xpc_vars_sn2 { u8 version; u64 heartbeat; - u64 heartbeating_to_mask; + DECLARE_BITMAP(heartbeating_to_mask, XP_MAX_NPARTITIONS_SN2); u64 heartbeat_offline; /* if 0, heartbeat should be changing */ int act_nasid; int act_phys_cpuid; @@ -176,46 +176,23 @@ struct xpc_vars { #define XPC_SUPPORTS_DISENGAGE_REQUEST(_version) \ (_version >= _XPC_VERSION(3, 1)) -static inline int -xpc_hb_allowed(short partid, struct xpc_vars *vars) -{ - return ((vars->heartbeating_to_mask & (1UL << partid)) != 0); -} - -static inline void -xpc_allow_hb(short 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(short 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). + * The following pertains to ia64-sn2 only. + * + * Memory for XPC's AMO variables is allocated by the MSPEC driver. These + * pages are located in the lowest granule. The lowest granule uses 4k pages + * for cached references and an alternate TLB handler to never provide a + * 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 + * AMO variables (based on XP_MAX_NPARTITIONS_SN2) to identify the senders of + * NOTIFY IRQs, 128 AMO variables (based on XP_NASID_MASK_WORDS) to identify + * the senders of ACTIVATE IRQs, and 2 AMO variables to identify which remote + * partitions (i.e., XPCs) consider themselves currently engaged with the + * local XPC. */ -#define XPC_NOTIFY_IRQ_AMOS 0 -#define XPC_ACTIVATE_IRQ_AMOS (XPC_NOTIFY_IRQ_AMOS + XP_MAX_NPARTITIONS_SN2) +#define XPC_NOTIFY_IRQ_AMOS 0 +#define XPC_ACTIVATE_IRQ_AMOS (XPC_NOTIFY_IRQ_AMOS + XP_MAX_NPARTITIONS_SN2) #define XPC_ENGAGED_PARTITIONS_AMO (XPC_ACTIVATE_IRQ_AMOS + XP_NASID_MASK_WORDS) #define XPC_DISENGAGE_REQUEST_AMO (XPC_ENGAGED_PARTITIONS_AMO + 1) @@ -259,11 +236,11 @@ struct xpc_vars_part_sn2 { /* 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_VARS_SIZE L1_CACHE_ALIGN(sizeof(struct xpc_vars_sn2)) #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) + \ +#define XPC_RP_VARS(_rp) ((struct xpc_vars_sn2 *)(XPC_RP_MACH_NASIDS(_rp) + \ xp_nasid_mask_words)) /* @@ -344,6 +321,7 @@ struct xpc_notify { * allocated at the time a partition becomes active. The array contains one * of these structures for each potential channel connection to that partition. * +>>> sn2 only!!! * Each of these structures manages two message queues (circular buffers). * They are allocated at the time a channel connection is made. One of * these message queues (local_msgqueue) holds the locally created messages @@ -622,6 +600,9 @@ extern struct device *xpc_part; extern struct device *xpc_chan; extern int xpc_disengage_request_timelimit; extern int xpc_disengage_request_timedout; +extern atomic_t xpc_act_IRQ_rcvd; +extern wait_queue_head_t xpc_act_IRQ_wq; +extern void *xpc_heartbeating_to_mask; extern irqreturn_t xpc_notify_IRQ_handler(int, void *); extern void xpc_dropped_IPI_check(struct xpc_partition *); extern void xpc_activate_partition(struct xpc_partition *); @@ -629,15 +610,48 @@ extern void xpc_activate_kthreads(struct extern void xpc_create_kthreads(struct xpc_channel *, int, int); extern void xpc_disconnect_wait(int); extern enum xp_retval (*xpc_rsvd_page_init) (struct xpc_rsvd_page *); +extern void (*xpc_heartbeat_init) (void); +extern void (*xpc_heartbeat_exit) (void); +extern void (*xpc_increment_heartbeat) (void); +extern void (*xpc_offline_heartbeat) (void); +extern void (*xpc_online_heartbeat) (void); +extern void (*xpc_check_remote_hb) (void); extern enum xp_retval (*xpc_make_first_contact) (struct xpc_partition *); extern u64 (*xpc_get_IPI_flags) (struct xpc_partition *); extern struct xpc_msg *(*xpc_get_deliverable_msg) (struct xpc_channel *); +extern void (*xpc_initiate_partition_activation) (struct xpc_rsvd_page *, u64, + int); +extern void (*xpc_process_act_IRQ_rcvd) (int); extern enum xp_retval (*xpc_setup_infrastructure) (struct xpc_partition *); extern void (*xpc_teardown_infrastructure) (struct xpc_partition *); +extern void (*xpc_mark_partition_engaged) (struct xpc_partition *); +extern void (*xpc_mark_partition_disengaged) (struct xpc_partition *); +extern void (*xpc_request_partition_disengage) (struct xpc_partition *); +extern void (*xpc_cancel_partition_disengage_request) (struct xpc_partition *); +extern u64 (*xpc_partition_engaged) (u64); +extern u64 (*xpc_partition_disengage_requested) (u64);; +extern void (*xpc_clear_partition_engaged) (u64); +extern void (*xpc_clear_partition_disengage_request) (u64); + +extern void (*xpc_IPI_send_local_activate) (int); +extern void (*xpc_IPI_send_activated) (struct xpc_partition *); +extern void (*xpc_IPI_send_local_reactivate) (int); +extern void (*xpc_IPI_send_disengage) (struct xpc_partition *); + +extern void (*xpc_IPI_send_closerequest) (struct xpc_channel *, + unsigned long *); +extern void (*xpc_IPI_send_closereply) (struct xpc_channel *, unsigned long *); +extern void (*xpc_IPI_send_openrequest) (struct xpc_channel *, unsigned long *); +extern void (*xpc_IPI_send_openreply) (struct xpc_channel *, unsigned long *); + +extern enum xp_retval (*xpc_allocate_msg) (struct xpc_channel *, u32, + struct xpc_msg **); +extern enum xp_retval (*xpc_send_msg) (struct xpc_channel *, struct xpc_msg *, + u8, xpc_notify_func, void *); +extern void (*xpc_received_msg) (struct xpc_channel *, struct xpc_msg *); /* found in xpc_sn2.c */ extern void xpc_init_sn2(void); -extern struct xpc_vars *xpc_vars; /*>>> eliminate from here */ /* found in xpc_uv.c */ extern void xpc_init_uv(void); @@ -646,6 +660,7 @@ extern void xpc_init_uv(void); extern int xpc_exiting; extern int xp_nasid_mask_words; extern struct xpc_rsvd_page *xpc_rsvd_page; +extern u64 *xpc_mach_nasids; extern struct xpc_partition *xpc_partitions; extern char *xpc_remote_copy_buffer; extern void *xpc_remote_copy_buffer_base; @@ -658,7 +673,8 @@ extern int xpc_partition_disengaged(stru extern enum xp_retval xpc_mark_partition_active(struct xpc_partition *); extern void xpc_mark_partition_inactive(struct xpc_partition *); extern void xpc_discovery(void); -extern void xpc_check_remote_hb(void); +extern enum xp_retval xpc_get_remote_rp(int, u64 *, struct xpc_rsvd_page *, + u64 *); extern void xpc_deactivate_partition(const int, struct xpc_partition *, enum xp_retval); extern enum xp_retval xpc_initiate_partid_to_nasids(short, void *); @@ -667,6 +683,7 @@ extern enum xp_retval xpc_initiate_parti extern void *xpc_kzalloc_cacheline_aligned(size_t, gfp_t, void **); extern void xpc_initiate_connect(int); extern void xpc_initiate_disconnect(int); +extern enum xp_retval xpc_allocate_msg_wait(struct xpc_channel *); extern enum xp_retval xpc_initiate_allocate(short, int, u32, void **); extern enum xp_retval xpc_initiate_send(short, int, void *); extern enum xp_retval xpc_initiate_send_notify(short, int, void *, @@ -680,6 +697,40 @@ extern void xpc_disconnect_channel(const extern void xpc_disconnect_callout(struct xpc_channel *, enum xp_retval); extern void xpc_partition_going_down(struct xpc_partition *, enum xp_retval); +static inline int +xpc_hb_allowed(short partid, void *heartbeating_to_mask) +{ + return test_bit(partid, heartbeating_to_mask); +} + +static inline int +xpc_any_hbs_allowed(void) +{ + DBUG_ON(xpc_heartbeating_to_mask == NULL); + return !bitmap_empty(xpc_heartbeating_to_mask, xp_max_npartitions); +} + +static inline void +xpc_allow_hb(short partid) +{ + DBUG_ON(xpc_heartbeating_to_mask == NULL); + set_bit(partid, xpc_heartbeating_to_mask); +} + +static inline void +xpc_disallow_hb(short partid) +{ + DBUG_ON(xpc_heartbeating_to_mask == NULL); + clear_bit(partid, xpc_heartbeating_to_mask); +} + +static inline void +xpc_disallow_all_hbs(void) +{ + DBUG_ON(xpc_heartbeating_to_mask == NULL); + bitmap_zero(xpc_heartbeating_to_mask, xp_max_npartitions); +} + static inline void xpc_wakeup_channel_mgr(struct xpc_partition *part) { @@ -749,297 +800,7 @@ 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 - * the other that is associated with channel activity (SGI_XPC_NOTIFY). - */ - -static inline u64 -xpc_IPI_receive(AMO_t *amo) -{ - return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_CLEAR); -} - -static inline enum xp_retval -xpc_IPI_send(AMO_t *amo, u64 flag, int nasid, int phys_cpuid, int vector) -{ - int ret = 0; - unsigned long irq_flags; - - local_irq_save(irq_flags); - - FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR, flag); - sn_send_IPI_phys(nasid, phys_cpuid, vector, 0); - - /* - * 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. - */ - ret = xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->variable), - xp_nofault_PIOR_target)); - - local_irq_restore(irq_flags); - - return ((ret == 0) ? xpSuccess : xpPioReadError); -} - -/* - * IPIs associated with SGI_XPC_ACTIVATE IRQ. - */ - -/* - * Flag the appropriate AMO variable and send an IPI to the specified node. - */ -static inline void -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_pa + - (XPC_ACTIVATE_IRQ_AMOS * sizeof(AMO_t))); - - (void)xpc_IPI_send(&amos[w_index], (1UL << b_index), to_nasid, - to_phys_cpuid, SGI_XPC_ACTIVATE); -} - -static inline void -xpc_IPI_send_activate(struct xpc_vars *vars) -{ - xpc_activate_IRQ_send(vars->amos_page_pa, cnodeid_to_nasid(0), - vars->act_nasid, vars->act_phys_cpuid); -} - -static inline void -xpc_IPI_send_activated(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); -} - -static inline void -xpc_IPI_send_reactivate(struct xpc_partition *part) -{ - xpc_activate_IRQ_send(xpc_vars->amos_page_pa, part->reactivate_nasid, - 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. - */ - -/* - * Send an IPI to the remote partition that is associated with the - * specified channel. - */ -#define XPC_NOTIFY_IRQ_SEND(_ch, _ipi_f, _irq_f) \ - xpc_notify_IRQ_send(_ch, _ipi_f, #_ipi_f, _irq_f) - -static inline void -xpc_notify_IRQ_send(struct xpc_channel *ch, u8 ipi_flag, char *ipi_flag_string, - unsigned long *irq_flags) -{ - struct xpc_partition *part = &xpc_partitions[ch->partid]; - enum xp_retval ret; - - if (likely(part->act_state != XPC_P_DEACTIVATING)) { - ret = xpc_IPI_send(part->remote_IPI_amo_va, - (u64)ipi_flag << (ch->number * 8), - part->remote_IPI_nasid, - part->remote_IPI_phys_cpuid, SGI_XPC_NOTIFY); - dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n", - ipi_flag_string, ch->partid, ch->number, ret); - if (unlikely(ret != xpSuccess)) { - if (irq_flags != NULL) - spin_unlock_irqrestore(&ch->lock, *irq_flags); - XPC_DEACTIVATE_PARTITION(part, ret); - if (irq_flags != NULL) - spin_lock_irqsave(&ch->lock, *irq_flags); - } - } -} - -/* - * Make it look like the remote partition, which is associated with the - * specified channel, sent us an IPI. This faked IPI will be handled - * by xpc_dropped_IPI_check(). - */ -#define XPC_NOTIFY_IRQ_SEND_LOCAL(_ch, _ipi_f) \ - xpc_notify_IRQ_send_local(_ch, _ipi_f, #_ipi_f) - -static inline void -xpc_notify_IRQ_send_local(struct xpc_channel *ch, u8 ipi_flag, - char *ipi_flag_string) -{ - struct xpc_partition *part = &xpc_partitions[ch->partid]; - - FETCHOP_STORE_OP(TO_AMO((u64)&part->local_IPI_amo_va->variable), - FETCHOP_OR, ((u64)ipi_flag << (ch->number * 8))); - dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n", - ipi_flag_string, ch->partid, ch->number); -} - -/* - * The sending and receiving of IPIs includes the setting of an AMO variable + * The sending and receiving of IPIs includes the setting of an >>>AMO variable * to indicate the reason the IPI was sent. The 64-bit variable is divided * up into eight bytes, ordered from right to left. Byte zero pertains to * channel 0, byte one to channel 1, and so on. Each byte is described by @@ -1052,107 +813,11 @@ xpc_notify_IRQ_send_local(struct xpc_cha #define XPC_IPI_OPENREPLY 0x08 #define XPC_IPI_MSGREQUEST 0x10 -/* given an AMO variable and a channel#, get its associated IPI flags */ +/* 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) & 0x0f0f0f0f0f0f0f0fUL) #define XPC_ANY_MSG_IPI_FLAGS_SET(_amo) ((_amo) & 0x1010101010101010UL) -static inline void -xpc_IPI_send_closerequest(struct xpc_channel *ch, unsigned long *irq_flags) -{ - struct xpc_openclose_args *args = ch->local_openclose_args; - - args->reason = ch->reason; - - XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_CLOSEREQUEST, irq_flags); -} - -static inline void -xpc_IPI_send_closereply(struct xpc_channel *ch, unsigned long *irq_flags) -{ - XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_CLOSEREPLY, irq_flags); -} - -static inline void -xpc_IPI_send_openrequest(struct xpc_channel *ch, unsigned long *irq_flags) -{ - struct xpc_openclose_args *args = ch->local_openclose_args; - - args->msg_size = ch->msg_size; - args->local_nentries = ch->local_nentries; - - XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_OPENREQUEST, irq_flags); -} - -static inline void -xpc_IPI_send_openreply(struct xpc_channel *ch, unsigned long *irq_flags) -{ - struct xpc_openclose_args *args = ch->local_openclose_args; - - args->remote_nentries = ch->remote_nentries; - args->local_nentries = ch->local_nentries; - args->local_msgqueue_pa = __pa(ch->local_msgqueue); - - XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_OPENREPLY, irq_flags); -} - -static inline void -xpc_IPI_send_msgrequest(struct xpc_channel *ch) -{ - XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_MSGREQUEST, NULL); -} - -static inline void -xpc_IPI_send_local_msgrequest(struct xpc_channel *ch) -{ - XPC_NOTIFY_IRQ_SEND_LOCAL(ch, XPC_IPI_MSGREQUEST); -} - -/* ->>> this block comment needs to be moved and re-written. - * Memory for XPC's AMO variables is allocated by the MSPEC driver. These - * pages are located in the lowest granule. The lowest granule uses 4k pages - * for cached references and an alternate TLB handler to never provide a - * 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 - * AMO variables (based on xp_max_npartitions) 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(int index) -{ - AMO_t *amo = xpc_vars->amos_page + index; - - (void)xpc_IPI_receive(amo); /* clear AMO variable */ - return amo; -} - -/* - * Check to see if there is any channel activity to/from the specified - * partition. - */ -static inline void -xpc_check_for_channel_activity(struct xpc_partition *part) -{ - u64 IPI_amo; - unsigned long irq_flags; - - IPI_amo = xpc_IPI_receive(part->local_IPI_amo_va); - if (IPI_amo == 0) - return; - - spin_lock_irqsave(&part->IPI_lock, irq_flags); - part->local_IPI_amo |= IPI_amo; - spin_unlock_irqrestore(&part->IPI_lock, irq_flags); - - dev_dbg(xpc_chan, "received IPI from partid=%d, IPI_amo=0x%lx\n", - XPC_PARTID(part), IPI_amo); - - xpc_wakeup_channel_mgr(part); -} - #endif /* _DRIVERS_MISC_SGIXP_XPC_H */ Index: linux-2.6/drivers/misc/sgi-xp/xpc_sn2.c =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc_sn2.c 2008-05-28 07:03:09.420496138 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc_sn2.c 2008-05-28 07:03:15.285225043 -0500 @@ -19,9 +19,370 @@ #include #include "xpc.h" -struct xpc_vars *xpc_vars; +static struct xpc_vars_sn2 *xpc_vars; /* >>> Add _sn2 suffix? */ static struct xpc_vars_part_sn2 *xpc_vars_part; /* >>> Add _sn2 suffix? */ +/* + * The following set of macros and functions 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 + * the other that is associated with channel activity (SGI_XPC_NOTIFY). + */ + +static u64 +xpc_IPI_receive_sn2(AMO_t *amo) +{ + return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_CLEAR); +} + +static enum xp_retval +xpc_IPI_send_sn2(AMO_t *amo, u64 flag, int nasid, int phys_cpuid, int vector) +{ + int ret = 0; + unsigned long irq_flags; + + local_irq_save(irq_flags); + + FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR, flag); + sn_send_IPI_phys(nasid, phys_cpuid, vector, 0); + + /* + * 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. + */ + ret = xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->variable), + xp_nofault_PIOR_target)); + + local_irq_restore(irq_flags); + + return ((ret == 0) ? xpSuccess : xpPioReadError); +} + +static AMO_t * +xpc_IPI_init_sn2(int index) +{ + AMO_t *amo = xpc_vars->amos_page + index; + + (void)xpc_IPI_receive_sn2(amo); /* clear AMO variable */ + return amo; +} + +/* + * IPIs associated with SGI_XPC_ACTIVATE IRQ. + */ + +/* + * Flag the appropriate AMO variable and send an IPI to the specified node. + */ +static void +xpc_activate_IRQ_send_sn2(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_pa + + (XPC_ACTIVATE_IRQ_AMOS * sizeof(AMO_t))); + + (void)xpc_IPI_send_sn2(&amos[w_index], (1UL << b_index), to_nasid, + to_phys_cpuid, SGI_XPC_ACTIVATE); +} + +static void +xpc_activate_IRQ_send_local_sn2(int from_nasid) +{ + int w_index = XPC_NASID_W_INDEX(from_nasid); + int b_index = XPC_NASID_B_INDEX(from_nasid); + AMO_t *amos = (AMO_t *)__va(xpc_vars->amos_page_pa + + (XPC_ACTIVATE_IRQ_AMOS * sizeof(AMO_t))); + + /* fake the sending and receipt of an activate IRQ from remote nasid */ + FETCHOP_STORE_OP(TO_AMO((u64)&amos[w_index].variable), FETCHOP_OR, + (1UL << b_index)); + atomic_inc(&xpc_act_IRQ_rcvd); + wake_up_interruptible(&xpc_act_IRQ_wq); +} + +static void +xpc_IPI_send_local_activate_sn2(int from_nasid) +{ + xpc_activate_IRQ_send_local_sn2(from_nasid); +} + +static void +xpc_IPI_send_activated_sn2(struct xpc_partition *part) +{ + xpc_activate_IRQ_send_sn2(part->remote_amos_page_pa, + cnodeid_to_nasid(0), part->remote_act_nasid, + part->remote_act_phys_cpuid); +} + +static void +xpc_IPI_send_local_reactivate_sn2(int from_nasid) +{ + xpc_activate_IRQ_send_local_sn2(from_nasid); +} + +static void +xpc_IPI_send_disengage_sn2(struct xpc_partition *part) +{ + xpc_activate_IRQ_send_sn2(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. + */ + +/* + * Send an IPI to the remote partition that is associated with the + * specified channel. + */ +static void +xpc_notify_IRQ_send_sn2(struct xpc_channel *ch, u8 ipi_flag, + char *ipi_flag_string, unsigned long *irq_flags) +{ + struct xpc_partition *part = &xpc_partitions[ch->partid]; + enum xp_retval ret; + + if (likely(part->act_state != XPC_P_DEACTIVATING)) { + ret = xpc_IPI_send_sn2(part->remote_IPI_amo_va, + (u64)ipi_flag << (ch->number * 8), + part->remote_IPI_nasid, + part->remote_IPI_phys_cpuid, + SGI_XPC_NOTIFY); + dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n", + ipi_flag_string, ch->partid, ch->number, ret); + if (unlikely(ret != xpSuccess)) { + if (irq_flags != NULL) + spin_unlock_irqrestore(&ch->lock, *irq_flags); + XPC_DEACTIVATE_PARTITION(part, ret); + if (irq_flags != NULL) + spin_lock_irqsave(&ch->lock, *irq_flags); + } + } +} + +#define XPC_NOTIFY_IRQ_SEND_SN2(_ch, _ipi_f, _irq_f) \ + xpc_notify_IRQ_send_sn2(_ch, _ipi_f, #_ipi_f, _irq_f) + +/* + * Make it look like the remote partition, which is associated with the + * specified channel, sent us an IPI. This faked IPI will be handled + * by xpc_dropped_IPI_check(). + */ +static void +xpc_notify_IRQ_send_local_sn2(struct xpc_channel *ch, u8 ipi_flag, + char *ipi_flag_string) +{ + struct xpc_partition *part = &xpc_partitions[ch->partid]; + + FETCHOP_STORE_OP(TO_AMO((u64)&part->local_IPI_amo_va->variable), + FETCHOP_OR, ((u64)ipi_flag << (ch->number * 8))); + dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n", + ipi_flag_string, ch->partid, ch->number); +} + +#define XPC_NOTIFY_IRQ_SEND_LOCAL_SN2(_ch, _ipi_f) \ + xpc_notify_IRQ_send_local_sn2(_ch, _ipi_f, #_ipi_f) + +static void +xpc_IPI_send_closerequest_sn2(struct xpc_channel *ch, unsigned long *irq_flags) +{ + struct xpc_openclose_args *args = ch->local_openclose_args; + + args->reason = ch->reason; + XPC_NOTIFY_IRQ_SEND_SN2(ch, XPC_IPI_CLOSEREQUEST, irq_flags); +} + +static void +xpc_IPI_send_closereply_sn2(struct xpc_channel *ch, unsigned long *irq_flags) +{ + XPC_NOTIFY_IRQ_SEND_SN2(ch, XPC_IPI_CLOSEREPLY, irq_flags); +} + +static void +xpc_IPI_send_openrequest_sn2(struct xpc_channel *ch, unsigned long *irq_flags) +{ + struct xpc_openclose_args *args = ch->local_openclose_args; + + args->msg_size = ch->msg_size; + args->local_nentries = ch->local_nentries; + XPC_NOTIFY_IRQ_SEND_SN2(ch, XPC_IPI_OPENREQUEST, irq_flags); +} + +static void +xpc_IPI_send_openreply_sn2(struct xpc_channel *ch, unsigned long *irq_flags) +{ + struct xpc_openclose_args *args = ch->local_openclose_args; + + args->remote_nentries = ch->remote_nentries; + args->local_nentries = ch->local_nentries; + args->local_msgqueue_pa = __pa(ch->local_msgqueue); + XPC_NOTIFY_IRQ_SEND_SN2(ch, XPC_IPI_OPENREPLY, irq_flags); +} + +static void +xpc_IPI_send_msgrequest_sn2(struct xpc_channel *ch) +{ + XPC_NOTIFY_IRQ_SEND_SN2(ch, XPC_IPI_MSGREQUEST, NULL); +} + +static void +xpc_IPI_send_local_msgrequest_sn2(struct xpc_channel *ch) +{ + XPC_NOTIFY_IRQ_SEND_LOCAL_SN2(ch, XPC_IPI_MSGREQUEST); +} + +/* + * This next set of functions are used to keep track of when a partition is + * potentially engaged in accessing memory belonging to another partition. + */ + +static void +xpc_mark_partition_engaged_sn2(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 void +xpc_mark_partition_disengaged_sn2(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 void +xpc_request_partition_disengage_sn2(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 void +xpc_cancel_partition_disengage_request_sn2(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 u64 +xpc_partition_engaged_sn2(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 u64 +xpc_partition_disengage_requested_sn2(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 void +xpc_clear_partition_engaged_sn2(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 void +xpc_clear_partition_disengage_request_sn2(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); +} + static enum xp_retval xpc_rsvd_page_init_sn2(struct xpc_rsvd_page *rp) { @@ -79,7 +440,7 @@ xpc_rsvd_page_init_sn2(struct xpc_rsvd_p } /* clear xpc_vars */ - memset(xpc_vars, 0, sizeof(struct xpc_vars)); + memset(xpc_vars, 0, sizeof(struct xpc_vars_sn2)); xpc_vars->version = XPC_V_VERSION; xpc_vars->act_nasid = cpuid_to_nasid(0); @@ -94,15 +455,446 @@ xpc_rsvd_page_init_sn2(struct xpc_rsvd_p /* initialize the activate IRQ related AMO variables */ for (i = 0; i < xp_nasid_mask_words; i++) - (void)xpc_IPI_init(XPC_ACTIVATE_IRQ_AMOS + i); + (void)xpc_IPI_init_sn2(XPC_ACTIVATE_IRQ_AMOS + i); /* initialize the engaged remote partitions related AMO variables */ - (void)xpc_IPI_init(XPC_ENGAGED_PARTITIONS_AMO); - (void)xpc_IPI_init(XPC_DISENGAGE_REQUEST_AMO); + (void)xpc_IPI_init_sn2(XPC_ENGAGED_PARTITIONS_AMO); + (void)xpc_IPI_init_sn2(XPC_DISENGAGE_REQUEST_AMO); + + return xpSuccess; +} + +static void +xpc_increment_heartbeat_sn2(void) +{ + xpc_vars->heartbeat++; +} + +static void +xpc_offline_heartbeat_sn2(void) +{ + xpc_increment_heartbeat_sn2(); + xpc_vars->heartbeat_offline = 1; +} + +static void +xpc_online_heartbeat_sn2(void) +{ + xpc_increment_heartbeat_sn2(); + xpc_vars->heartbeat_offline = 0; +} + +static void +xpc_heartbeat_init_sn2(void) +{ + DBUG_ON(xpc_vars == NULL); + + bitmap_zero(xpc_vars->heartbeating_to_mask, XP_MAX_NPARTITIONS_SN2); + xpc_heartbeating_to_mask = &xpc_vars->heartbeating_to_mask[0]; + xpc_online_heartbeat_sn2(); +} + +static void +xpc_heartbeat_exit_sn2(void) +{ + xpc_offline_heartbeat_sn2(); +} + +/* + * At periodic intervals, scan through all active partitions and ensure + * their heartbeat is still active. If not, the partition is deactivated. + */ +static void +xpc_check_remote_hb_sn2(void) +{ + struct xpc_vars_sn2 *remote_vars; + struct xpc_partition *part; + short partid; + enum xp_retval ret; + + remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer; + + for (partid = 0; partid < xp_max_npartitions; partid++) { + + if (xpc_exiting) + break; + + if (partid == sn_partition_id) + continue; + + part = &xpc_partitions[partid]; + + if (part->act_state == XPC_P_INACTIVE || + part->act_state == XPC_P_DEACTIVATING) { + continue; + } + + /* pull the remote_hb cache line */ + ret = xp_remote_memcpy(remote_vars, + (void *)part->remote_vars_pa, + XPC_RP_VARS_SIZE); + if (ret != xpSuccess) { + XPC_DEACTIVATE_PARTITION(part, ret); + continue; + } + + dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat" + " = %ld, heartbeat_offline = %ld, HB_mask[0] = 0x%lx\n", + partid, remote_vars->heartbeat, part->last_heartbeat, + remote_vars->heartbeat_offline, + remote_vars->heartbeating_to_mask[0]); + + if (((remote_vars->heartbeat == part->last_heartbeat) && + (remote_vars->heartbeat_offline == 0)) || + !xpc_hb_allowed(sn_partition_id, + &remote_vars->heartbeating_to_mask)) { + + XPC_DEACTIVATE_PARTITION(part, xpNoHeartbeat); + continue; + } + + part->last_heartbeat = remote_vars->heartbeat; + } +} + +/* + * Get a copy of the remote partition's XPC variables from the reserved page. + * + * remote_vars points to a buffer that is cacheline aligned for BTE copies and + * assumed to be of size XPC_RP_VARS_SIZE. + */ +static enum xp_retval +xpc_get_remote_vars_sn2(u64 remote_vars_pa, struct xpc_vars_sn2 *remote_vars) +{ + enum xp_retval ret; + + if (remote_vars_pa == 0) + return xpVarsNotSet; + + /* pull over the cross partition variables */ + ret = xp_remote_memcpy(remote_vars, (void *)remote_vars_pa, + XPC_RP_VARS_SIZE); + if (ret != xpSuccess) + return ret; + + if (XPC_VERSION_MAJOR(remote_vars->version) != + XPC_VERSION_MAJOR(XPC_V_VERSION)) { + return xpBadVersion; + } return xpSuccess; } +static void +xpc_initiate_partition_activation_sn2(struct xpc_rsvd_page *remote_rp, + u64 remote_rp_pa, int nasid) +{ + xpc_IPI_send_local_activate(nasid); +} + +/* + * Update the remote partition's info. + */ +static void +xpc_update_partition_info_sn2(struct xpc_partition *part, u8 remote_rp_version, + struct timespec *remote_rp_stamp, + u64 remote_rp_pa, u64 remote_vars_pa, + struct xpc_vars_sn2 *remote_vars) +{ + part->remote_rp_version = remote_rp_version; + dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n", + part->remote_rp_version); + + part->remote_rp_stamp = *remote_rp_stamp; + dev_dbg(xpc_part, " remote_rp_stamp (tv_sec = 0x%lx tv_nsec = 0x%lx\n", + part->remote_rp_stamp.tv_sec, part->remote_rp_stamp.tv_nsec); + + part->remote_rp_pa = remote_rp_pa; + dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa); + + part->remote_vars_pa = remote_vars_pa; + dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n", + part->remote_vars_pa); + + part->last_heartbeat = remote_vars->heartbeat; + dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n", + part->last_heartbeat); + + part->remote_vars_part_pa = remote_vars->vars_part_pa; + dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n", + part->remote_vars_part_pa); + + part->remote_act_nasid = remote_vars->act_nasid; + dev_dbg(xpc_part, " remote_act_nasid = 0x%x\n", + part->remote_act_nasid); + + part->remote_act_phys_cpuid = remote_vars->act_phys_cpuid; + dev_dbg(xpc_part, " remote_act_phys_cpuid = 0x%x\n", + part->remote_act_phys_cpuid); + + part->remote_amos_page_pa = remote_vars->amos_page_pa; + dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n", + part->remote_amos_page_pa); + + part->remote_vars_version = remote_vars->version; + dev_dbg(xpc_part, " remote_vars_version = 0x%x\n", + part->remote_vars_version); +} + +/* + * Prior code has determined the nasid which generated an IPI. Inspect + * that nasid to determine if its partition needs to be activated or + * deactivated. + * + * A partition is consider "awaiting activation" if our partition + * flags indicate it is not active and it has a heartbeat. A + * partition is considered "awaiting deactivation" if our partition + * flags indicate it is active but it has no heartbeat or it is not + * sending its heartbeat to us. + * + * To determine the heartbeat, the remote nasid must have a properly + * initialized reserved page. + */ +static void +xpc_identify_act_IRQ_req_sn2(int nasid) +{ + struct xpc_rsvd_page *remote_rp; + struct xpc_vars_sn2 *remote_vars; + u64 remote_rp_pa; + u64 remote_vars_pa; + int remote_rp_version; + int reactivate = 0; + int stamp_diff; + struct timespec remote_rp_stamp = { 0, 0 }; /*>>> ZERO_STAMP */ + short partid; + struct xpc_partition *part; + enum xp_retval ret; + + /* pull over the reserved page structure */ + + remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer; + + ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa); + if (ret != xpSuccess) { + dev_warn(xpc_part, "unable to get reserved page from nasid %d, " + "which sent interrupt, reason=%d\n", nasid, ret); + return; + } + + remote_vars_pa = remote_rp->sn.vars_pa; + remote_rp_version = remote_rp->version; + if (XPC_SUPPORTS_RP_STAMP(remote_rp_version)) + remote_rp_stamp = remote_rp->stamp; + + partid = remote_rp->SAL_partid; + part = &xpc_partitions[partid]; + + /* pull over the cross partition variables */ + + remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer; + + ret = xpc_get_remote_vars_sn2(remote_vars_pa, remote_vars); + if (ret != xpSuccess) { + + dev_warn(xpc_part, "unable to get XPC variables from nasid %d, " + "which sent interrupt, reason=%d\n", nasid, ret); + + XPC_DEACTIVATE_PARTITION(part, ret); + return; + } + + part->act_IRQ_rcvd++; + + dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = " + "%ld:0x%lx\n", (int)nasid, (int)partid, part->act_IRQ_rcvd, + remote_vars->heartbeat, remote_vars->heartbeating_to_mask[0]); + + if (xpc_partition_disengaged(part) && + part->act_state == XPC_P_INACTIVE) { + + xpc_update_partition_info_sn2(part, remote_rp_version, + &remote_rp_stamp, remote_rp_pa, + remote_vars_pa, remote_vars); + + if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) { + if (xpc_partition_disengage_requested(1UL << partid)) { + /* + * Other side is waiting on us to disengage, + * even though we already have. + */ + return; + } + + } else { + /* other side doesn't support disengage requests */ + xpc_clear_partition_disengage_request(1UL << partid); + } + + xpc_activate_partition(part); + return; + } + + DBUG_ON(part->remote_rp_version == 0); + DBUG_ON(part->remote_vars_version == 0); + + if (!XPC_SUPPORTS_RP_STAMP(part->remote_rp_version)) { + DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(part-> + remote_vars_version)); + + if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) { + DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars-> + version)); + /* see if the other side rebooted */ + if (part->remote_amos_page_pa == + remote_vars->amos_page_pa && + xpc_hb_allowed(sn_partition_id, + &remote_vars->heartbeating_to_mask)) { + /* doesn't look that way, so ignore the IPI */ + return; + } + } + + /* + * Other side rebooted and previous XPC didn't support the + * disengage request, so we don't need to do anything special. + */ + + xpc_update_partition_info_sn2(part, remote_rp_version, + &remote_rp_stamp, remote_rp_pa, + remote_vars_pa, remote_vars); + part->reactivate_nasid = nasid; + XPC_DEACTIVATE_PARTITION(part, xpReactivating); + return; + } + + DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)); + + if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) { + DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version)); + + /* + * Other side rebooted and previous XPC did support the + * disengage request, but the new one doesn't. + */ + + xpc_clear_partition_engaged(1UL << partid); + xpc_clear_partition_disengage_request(1UL << partid); + + xpc_update_partition_info_sn2(part, remote_rp_version, + &remote_rp_stamp, remote_rp_pa, + remote_vars_pa, remote_vars); + reactivate = 1; + + } else { + DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version)); + + stamp_diff = xpc_compare_stamps(&part->remote_rp_stamp, + &remote_rp_stamp); + if (stamp_diff != 0) { + DBUG_ON(stamp_diff >= 0); + + /* + * Other side rebooted and the previous XPC did support + * the disengage request, as does the new one. + */ + + DBUG_ON(xpc_partition_engaged(1UL << partid)); + DBUG_ON(xpc_partition_disengage_requested(1UL << + partid)); + + xpc_update_partition_info_sn2(part, remote_rp_version, + &remote_rp_stamp, + remote_rp_pa, + remote_vars_pa, + remote_vars); + reactivate = 1; + } + } + + if (part->disengage_request_timeout > 0 && + !xpc_partition_disengaged(part)) { + /* still waiting on other side to disengage from us */ + return; + } + + if (reactivate) { + part->reactivate_nasid = nasid; + XPC_DEACTIVATE_PARTITION(part, xpReactivating); + + } else if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version) && + xpc_partition_disengage_requested(1UL << partid)) { + XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown); + } +} + +/* + * Loop through the activation AMO variables and process any bits + * which are set. Each bit indicates a nasid sending a partition + * activation or deactivation request. + * + * Return #of IRQs detected. + */ +int +xpc_identify_act_IRQ_sender_sn2(void) +{ + int word, bit; + u64 nasid_mask; + u64 nasid; /* remote nasid */ + int n_IRQs_detected = 0; + AMO_t *act_amos; + + act_amos = xpc_vars->amos_page + XPC_ACTIVATE_IRQ_AMOS; + + /* scan through act AMO variable looking for non-zero entries */ + for (word = 0; word < xp_nasid_mask_words; word++) { + + if (xpc_exiting) + break; + + nasid_mask = xpc_IPI_receive_sn2(&act_amos[word]); + if (nasid_mask == 0) { + /* no IRQs from nasids in this variable */ + continue; + } + + dev_dbg(xpc_part, "AMO[%d] gave back 0x%lx\n", word, + nasid_mask); + + /* + * If this nasid has been added to the machine since + * our partition was reset, this will retain the + * remote nasid in our reserved pages machine mask. + * This is used in the event of module reload. + */ + xpc_mach_nasids[word] |= nasid_mask; + + /* locate the nasid(s) which sent interrupts */ + + for (bit = 0; bit < (8 * sizeof(u64)); bit++) { + if (nasid_mask & (1UL << bit)) { + n_IRQs_detected++; + nasid = XPC_NASID_FROM_W_B(word, bit); + dev_dbg(xpc_part, "interrupt from nasid %ld\n", + nasid); + xpc_identify_act_IRQ_req_sn2(nasid); + } + } + } + return n_IRQs_detected; +} + +static void +xpc_process_act_IRQ_rcvd_sn2(int n_IRQs_expected) +{ + int n_IRQs_detected; + + n_IRQs_detected = xpc_identify_act_IRQ_sender_sn2(); + if (n_IRQs_detected < n_IRQs_expected) { + /* retry once to help avoid missing AMO */ + (void)xpc_identify_act_IRQ_sender_sn2(); + } +} + /* * Setup the infrastructure necessary to support XPartition Communication * between the specified remote partition and the local one. @@ -177,7 +969,7 @@ xpc_setup_infrastructure_sn2(struct xpc_ part->remote_openclose_args_pa = 0; - part->local_IPI_amo_va = xpc_IPI_init(partid); + part->local_IPI_amo_va = xpc_IPI_init_sn2(partid); part->local_IPI_amo = 0; spin_lock_init(&part->IPI_lock); @@ -468,6 +1260,28 @@ xpc_make_first_contact_sn2(struct xpc_pa { enum xp_retval ret; + /* + * Register the remote partition's AMOs with SAL so it can handle + * and cleanup errors within that address range should the remote + * partition go down. We don't unregister this range because it is + * difficult to tell when outstanding writes to the remote partition + * are finished and thus when it is safe to unregister. This should + * not result in wasted space in the SAL xp_addr_region table because + * we should get the same page for remote_amos_page_pa after module + * reloads and system reboots. + */ + if (sn_register_xp_addr_region(part->remote_amos_page_pa, + PAGE_SIZE, 1) < 0) { + dev_warn(xpc_part, "xpc_activating(%d) failed to register " + "xp_addr region\n", XPC_PARTID(part)); + + ret = xpPhysAddrRegFailed; + XPC_DEACTIVATE_PARTITION(part, ret); + return ret; + } + + xpc_IPI_send_activated(part); + while ((ret = xpc_pull_remote_vars_part_sn2(part)) != xpSuccess) { if (ret != xpRetry) { XPC_DEACTIVATE_PARTITION(part, ret); @@ -651,15 +1465,370 @@ xpc_get_deliverable_msg_sn2(struct xpc_c return msg; } +/* + * Now we actually send the messages that are ready to be sent by advancing + * the local message queue's Put value and then send an IPI to the recipient + * partition. + */ +static void +xpc_send_msgs_sn2(struct xpc_channel *ch, s64 initial_put) +{ + struct xpc_msg *msg; + s64 put = initial_put + 1; + int send_IPI = 0; + + while (1) { + + while (1) { + if (put == ch->w_local_GP.put) + break; + + msg = (struct xpc_msg *)((u64)ch->local_msgqueue + + (put % ch->local_nentries) * + ch->msg_size); + + if (!(msg->flags & XPC_M_READY)) + break; + + put++; + } + + if (put == initial_put) { + /* nothing's changed */ + break; + } + + if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) != + initial_put) { + /* someone else beat us to it */ + DBUG_ON(ch->local_GP->put < initial_put); + break; + } + + /* we just set the new value of local_GP->put */ + + dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, " + "channel=%d\n", put, ch->partid, ch->number); + + send_IPI = 1; + + /* + * We need to ensure that the message referenced by + * local_GP->put is not XPC_M_READY or that local_GP->put + * equals w_local_GP.put, so we'll go have a look. + */ + initial_put = put; + } + + if (send_IPI) + xpc_IPI_send_msgrequest_sn2(ch); +} + +/* + * Allocate an entry for a message from the message queue associated with the + * specified channel. + */ +static enum xp_retval +xpc_allocate_msg_sn2(struct xpc_channel *ch, u32 flags, + struct xpc_msg **address_of_msg) +{ + struct xpc_msg *msg; + enum xp_retval ret; + s64 put; + + /* this reference will be dropped in xpc_send_msg_sn2() */ + xpc_msgqueue_ref(ch); + + if (ch->flags & XPC_C_DISCONNECTING) { + xpc_msgqueue_deref(ch); + return ch->reason; + } + if (!(ch->flags & XPC_C_CONNECTED)) { + xpc_msgqueue_deref(ch); + return xpNotConnected; + } + + /* + * Get the next available message entry from the local message queue. + * If none are available, we'll make sure that we grab the latest + * GP values. + */ + ret = xpTimeout; + + while (1) { + + put = ch->w_local_GP.put; + rmb(); /* guarantee that .put loads before .get */ + if (put - ch->w_remote_GP.get < ch->local_nentries) { + + /* There are available message entries. We need to try + * to secure one for ourselves. We'll do this by trying + * to increment w_local_GP.put as long as someone else + * doesn't beat us to it. If they do, we'll have to + * try again. + */ + if (cmpxchg(&ch->w_local_GP.put, put, put + 1) == put) { + /* we got the entry referenced by put */ + break; + } + continue; /* try again */ + } + + /* + * There aren't any available msg entries at this time. + * + * In waiting for a message entry to become available, + * we set a timeout in case the other side is not + * sending completion IPIs. This lets us fake an IPI + * that will cause the IPI handler to fetch the latest + * GP values as if an IPI was sent by the other side. + */ + if (ret == xpTimeout) + xpc_IPI_send_local_msgrequest_sn2(ch); + + if (flags & XPC_NOWAIT) { + xpc_msgqueue_deref(ch); + return xpNoWait; + } + + ret = xpc_allocate_msg_wait(ch); + if (ret != xpInterrupted && ret != xpTimeout) { + xpc_msgqueue_deref(ch); + return ret; + } + } + + /* get the message's address and initialize it */ + msg = (struct xpc_msg *)((u64)ch->local_msgqueue + + (put % ch->local_nentries) * ch->msg_size); + + DBUG_ON(msg->flags != 0); + msg->number = put; + + dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, " + "msg_number=%ld, partid=%d, channel=%d\n", put + 1, + (void *)msg, msg->number, ch->partid, ch->number); + + *address_of_msg = msg; + + return xpSuccess; +} + +/* + * Common code that does the actual sending of the message by advancing the + * local message queue's Put value and sends an IPI to the partition the + * message is being sent to. + */ +static enum xp_retval +xpc_send_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type, + xpc_notify_func func, void *key) +{ + enum xp_retval ret = xpSuccess; + struct xpc_notify *notify = notify; + s64 put, msg_number = msg->number; + + DBUG_ON(notify_type == XPC_N_CALL && func == NULL); + DBUG_ON((((u64)msg - (u64)ch->local_msgqueue) / ch->msg_size) != + msg_number % ch->local_nentries); + DBUG_ON(msg->flags & XPC_M_READY); + + if (ch->flags & XPC_C_DISCONNECTING) { + /* drop the reference grabbed in xpc_allocate_msg_sn2() */ + xpc_msgqueue_deref(ch); + return ch->reason; + } + + if (notify_type != 0) { + /* + * Tell the remote side to send an ACK interrupt when the + * message has been delivered. + */ + msg->flags |= XPC_M_INTERRUPT; + + atomic_inc(&ch->n_to_notify); + + notify = &ch->notify_queue[msg_number % ch->local_nentries]; + notify->func = func; + notify->key = key; + notify->type = notify_type; + + /* >>> is a mb() needed here? */ + + if (ch->flags & XPC_C_DISCONNECTING) { + /* + * An error occurred between our last error check and + * this one. We will try to clear the type field from + * the notify entry. If we succeed then + * xpc_disconnect_channel() didn't already process + * the notify entry. + */ + if (cmpxchg(¬ify->type, notify_type, 0) == + notify_type) { + atomic_dec(&ch->n_to_notify); + ret = ch->reason; + } + + /* drop reference grabbed in xpc_allocate_msg_sn2() */ + xpc_msgqueue_deref(ch); + return ret; + } + } + + msg->flags |= XPC_M_READY; + + /* + * The preceding store of msg->flags must occur before the following + * load of ch->local_GP->put. + */ + mb(); + + /* see if the message is next in line to be sent, if so send it */ + + put = ch->local_GP->put; + if (put == msg_number) + xpc_send_msgs_sn2(ch, put); + + /* drop the reference grabbed in xpc_allocate_msg_sn2() */ + xpc_msgqueue_deref(ch); + return ret; +} + +/* + * Now we actually acknowledge the messages that have been delivered and ack'd + * by advancing the cached remote message queue's Get value and if requested + * send an IPI to the message sender's partition. + */ +static void +xpc_acknowledge_msgs_sn2(struct xpc_channel *ch, s64 initial_get, u8 msg_flags) +{ + struct xpc_msg *msg; + s64 get = initial_get + 1; + int send_IPI = 0; + + while (1) { + + while (1) { + if (get == ch->w_local_GP.get) + break; + + msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + + (get % ch->remote_nentries) * + ch->msg_size); + + if (!(msg->flags & XPC_M_DONE)) + break; + + msg_flags |= msg->flags; + get++; + } + + if (get == initial_get) { + /* nothing's changed */ + break; + } + + if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) != + initial_get) { + /* someone else beat us to it */ + DBUG_ON(ch->local_GP->get <= initial_get); + break; + } + + /* we just set the new value of local_GP->get */ + + dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, " + "channel=%d\n", get, ch->partid, ch->number); + + send_IPI = (msg_flags & XPC_M_INTERRUPT); + + /* + * We need to ensure that the message referenced by + * local_GP->get is not XPC_M_DONE or that local_GP->get + * equals w_local_GP.get, so we'll go have a look. + */ + initial_get = get; + } + + if (send_IPI) + xpc_IPI_send_msgrequest_sn2(ch); +} + +static void +xpc_received_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg) +{ + s64 get; + s64 msg_number = msg->number; + + dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n", + (void *)msg, msg_number, ch->partid, ch->number); + + DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) != + msg_number % ch->remote_nentries); + DBUG_ON(msg->flags & XPC_M_DONE); + + msg->flags |= XPC_M_DONE; + + /* + * The preceding store of msg->flags must occur before the following + * load of ch->local_GP->get. + */ + mb(); + + /* + * See if this message is next in line to be acknowledged as having + * been delivered. + */ + get = ch->local_GP->get; + if (get == msg_number) + xpc_acknowledge_msgs_sn2(ch, get, msg->flags); +} + void xpc_init_sn2(void) { xpc_rsvd_page_init = xpc_rsvd_page_init_sn2; + xpc_increment_heartbeat = xpc_increment_heartbeat_sn2; + xpc_offline_heartbeat = xpc_offline_heartbeat_sn2; + xpc_online_heartbeat = xpc_online_heartbeat_sn2; + xpc_heartbeat_init = xpc_heartbeat_init_sn2; + xpc_heartbeat_exit = xpc_heartbeat_exit_sn2; + xpc_check_remote_hb = xpc_check_remote_hb_sn2; + + xpc_initiate_partition_activation = + xpc_initiate_partition_activation_sn2; + xpc_process_act_IRQ_rcvd = xpc_process_act_IRQ_rcvd_sn2; xpc_setup_infrastructure = xpc_setup_infrastructure_sn2; xpc_teardown_infrastructure = xpc_teardown_infrastructure_sn2; xpc_make_first_contact = xpc_make_first_contact_sn2; xpc_get_IPI_flags = xpc_get_IPI_flags_sn2; xpc_get_deliverable_msg = xpc_get_deliverable_msg_sn2; + + xpc_mark_partition_engaged = xpc_mark_partition_engaged_sn2; + xpc_mark_partition_disengaged = xpc_mark_partition_disengaged_sn2; + xpc_request_partition_disengage = xpc_request_partition_disengage_sn2; + xpc_cancel_partition_disengage_request = + xpc_cancel_partition_disengage_request_sn2; + xpc_partition_engaged = xpc_partition_engaged_sn2; + xpc_partition_disengage_requested = + xpc_partition_disengage_requested_sn2; + xpc_clear_partition_engaged = xpc_clear_partition_engaged_sn2; + xpc_clear_partition_disengage_request = + xpc_clear_partition_disengage_request_sn2; + + xpc_IPI_send_local_activate = xpc_IPI_send_local_activate_sn2; + xpc_IPI_send_activated = xpc_IPI_send_activated_sn2; + xpc_IPI_send_local_reactivate = xpc_IPI_send_local_reactivate_sn2; + xpc_IPI_send_disengage = xpc_IPI_send_disengage_sn2; + + xpc_IPI_send_closerequest = xpc_IPI_send_closerequest_sn2; + xpc_IPI_send_closereply = xpc_IPI_send_closereply_sn2; + xpc_IPI_send_openrequest = xpc_IPI_send_openrequest_sn2; + xpc_IPI_send_openreply = xpc_IPI_send_openreply_sn2; + + xpc_allocate_msg = xpc_allocate_msg_sn2; + + xpc_send_msg = xpc_send_msg_sn2; + xpc_received_msg = xpc_received_msg_sn2; } void Index: linux-2.6/drivers/misc/sgi-xp/xpc_partition.c =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc_partition.c 2008-05-28 07:03:09.332485199 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc_partition.c 2008-05-28 07:03:15.309228026 -0500 @@ -42,7 +42,7 @@ u64 xpc_prot_vec[MAX_NUMNODES]; /* this partition's reserved page pointers */ struct xpc_rsvd_page *xpc_rsvd_page; static u64 *xpc_part_nasids; -static u64 *xpc_mach_nasids; +u64 *xpc_mach_nasids; /* >>> next two variables should be 'xpc_' if they remain here */ static int xp_sizeof_nasid_mask; /* actual size in bytes of nasid mask */ @@ -318,69 +318,13 @@ xpc_restrict_IPI_ops(void) } /* - * At periodic intervals, scan through all active partitions and ensure - * their heartbeat is still active. If not, the partition is deactivated. - */ -void -xpc_check_remote_hb(void) -{ - struct xpc_vars *remote_vars; - struct xpc_partition *part; - short partid; - enum xp_retval ret; - - remote_vars = (struct xpc_vars *)xpc_remote_copy_buffer; - - for (partid = 0; partid < xp_max_npartitions; partid++) { - - if (xpc_exiting) - break; - - if (partid == sn_partition_id) - continue; - - part = &xpc_partitions[partid]; - - if (part->act_state == XPC_P_INACTIVE || - part->act_state == XPC_P_DEACTIVATING) { - continue; - } - - /* pull the remote_hb cache line */ - ret = xp_remote_memcpy(remote_vars, - (void *)part->remote_vars_pa, - XPC_RP_VARS_SIZE); - if (ret != xpSuccess) { - XPC_DEACTIVATE_PARTITION(part, ret); - continue; - } - - dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat" - " = %ld, heartbeat_offline = %ld, HB_mask = 0x%lx\n", - partid, remote_vars->heartbeat, part->last_heartbeat, - remote_vars->heartbeat_offline, - remote_vars->heartbeating_to_mask); - - if (((remote_vars->heartbeat == part->last_heartbeat) && - (remote_vars->heartbeat_offline == 0)) || - !xpc_hb_allowed(sn_partition_id, remote_vars)) { - - XPC_DEACTIVATE_PARTITION(part, xpNoHeartbeat); - continue; - } - - part->last_heartbeat = remote_vars->heartbeat; - } -} - -/* * Get a copy of a portion of the remote partition's rsvd page. * * remote_rp points to a buffer that is cacheline aligned for BTE copies and * is large enough to contain a copy of their reserved page header and * part_nasids mask. */ -static enum xp_retval +enum xp_retval xpc_get_remote_rp(int nasid, u64 *discovered_nasids, struct xpc_rsvd_page *remote_rp, u64 *remote_rp_pa) { @@ -432,322 +376,6 @@ xpc_get_remote_rp(int nasid, u64 *discov } /* - * Get a copy of the remote partition's XPC variables from the reserved page. - * - * remote_vars points to a buffer that is cacheline aligned for BTE copies and - * assumed to be of size XPC_RP_VARS_SIZE. - */ -static enum xp_retval -xpc_get_remote_vars(u64 remote_vars_pa, struct xpc_vars *remote_vars) -{ - enum xp_retval ret; - - if (remote_vars_pa == 0) - return xpVarsNotSet; - - /* pull over the cross partition variables */ - ret = xp_remote_memcpy(remote_vars, (void *)remote_vars_pa, - XPC_RP_VARS_SIZE); - if (ret != xpSuccess) - return ret; - - if (XPC_VERSION_MAJOR(remote_vars->version) != - XPC_VERSION_MAJOR(XPC_V_VERSION)) { - return xpBadVersion; - } - - return xpSuccess; -} - -/* - * Update the remote partition's info. - */ -static void -xpc_update_partition_info(struct xpc_partition *part, u8 remote_rp_version, - struct timespec *remote_rp_stamp, u64 remote_rp_pa, - u64 remote_vars_pa, struct xpc_vars *remote_vars) -{ - part->remote_rp_version = remote_rp_version; - dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n", - part->remote_rp_version); - - part->remote_rp_stamp = *remote_rp_stamp; - dev_dbg(xpc_part, " remote_rp_stamp (tv_sec = 0x%lx tv_nsec = 0x%lx\n", - part->remote_rp_stamp.tv_sec, part->remote_rp_stamp.tv_nsec); - - part->remote_rp_pa = remote_rp_pa; - dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa); - - part->remote_vars_pa = remote_vars_pa; - dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n", - part->remote_vars_pa); - - part->last_heartbeat = remote_vars->heartbeat; - dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n", - part->last_heartbeat); - -/* >>> remote_vars_part_pa and vars_part_pa are sn2 only!!! */ - part->remote_vars_part_pa = remote_vars->vars_part_pa; - dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n", - part->remote_vars_part_pa); - - part->remote_act_nasid = remote_vars->act_nasid; - dev_dbg(xpc_part, " remote_act_nasid = 0x%x\n", - part->remote_act_nasid); - - part->remote_act_phys_cpuid = remote_vars->act_phys_cpuid; - dev_dbg(xpc_part, " remote_act_phys_cpuid = 0x%x\n", - part->remote_act_phys_cpuid); - - part->remote_amos_page_pa = remote_vars->amos_page_pa; - dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n", - part->remote_amos_page_pa); - - part->remote_vars_version = remote_vars->version; - dev_dbg(xpc_part, " remote_vars_version = 0x%x\n", - part->remote_vars_version); -} - -/* - * Prior code has determined the nasid which generated an IPI. Inspect - * that nasid to determine if its partition needs to be activated or - * deactivated. - * - * A partition is consider "awaiting activation" if our partition - * flags indicate it is not active and it has a heartbeat. A - * partition is considered "awaiting deactivation" if our partition - * flags indicate it is active but it has no heartbeat or it is not - * sending its heartbeat to us. - * - * To determine the heartbeat, the remote nasid must have a properly - * initialized reserved page. - */ -static void -xpc_identify_act_IRQ_req(int nasid) -{ - struct xpc_rsvd_page *remote_rp; - struct xpc_vars *remote_vars; - u64 remote_rp_pa; - u64 remote_vars_pa; - int remote_rp_version; - int reactivate = 0; - int stamp_diff; - struct timespec remote_rp_stamp = { 0, 0 }; /*>>> ZERO_STAMP */ - short partid; - struct xpc_partition *part; - enum xp_retval ret; - - /* pull over the reserved page structure */ - - remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer; - - ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa); - if (ret != xpSuccess) { - dev_warn(xpc_part, "unable to get reserved page from nasid %d, " - "which sent interrupt, reason=%d\n", nasid, ret); - return; - } - - remote_vars_pa = remote_rp->sn.vars_pa; - remote_rp_version = remote_rp->version; - if (XPC_SUPPORTS_RP_STAMP(remote_rp_version)) - remote_rp_stamp = remote_rp->stamp; - - partid = remote_rp->SAL_partid; - part = &xpc_partitions[partid]; - - /* pull over the cross partition variables */ - - remote_vars = (struct xpc_vars *)xpc_remote_copy_buffer; - - ret = xpc_get_remote_vars(remote_vars_pa, remote_vars); - if (ret != xpSuccess) { - - dev_warn(xpc_part, "unable to get XPC variables from nasid %d, " - "which sent interrupt, reason=%d\n", nasid, ret); - - XPC_DEACTIVATE_PARTITION(part, ret); - return; - } - - part->act_IRQ_rcvd++; - - dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = " - "%ld:0x%lx\n", (int)nasid, (int)partid, part->act_IRQ_rcvd, - remote_vars->heartbeat, remote_vars->heartbeating_to_mask); - - if (xpc_partition_disengaged(part) && - part->act_state == XPC_P_INACTIVE) { - - xpc_update_partition_info(part, remote_rp_version, - &remote_rp_stamp, remote_rp_pa, - remote_vars_pa, remote_vars); - - if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) { - if (xpc_partition_disengage_requested(1UL << partid)) { - /* - * Other side is waiting on us to disengage, - * even though we already have. - */ - return; - } - } else { - /* other side doesn't support disengage requests */ - xpc_clear_partition_disengage_request(1UL << partid); - } - - xpc_activate_partition(part); - return; - } - - DBUG_ON(part->remote_rp_version == 0); - DBUG_ON(part->remote_vars_version == 0); - - if (!XPC_SUPPORTS_RP_STAMP(part->remote_rp_version)) { - DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(part-> - remote_vars_version)); - - if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) { - DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars-> - version)); - /* see if the other side rebooted */ - if (part->remote_amos_page_pa == - remote_vars->amos_page_pa && - xpc_hb_allowed(sn_partition_id, remote_vars)) { - /* doesn't look that way, so ignore the IPI */ - return; - } - } - - /* - * Other side rebooted and previous XPC didn't support the - * disengage request, so we don't need to do anything special. - */ - - xpc_update_partition_info(part, remote_rp_version, - &remote_rp_stamp, remote_rp_pa, - remote_vars_pa, remote_vars); - part->reactivate_nasid = nasid; - XPC_DEACTIVATE_PARTITION(part, xpReactivating); - return; - } - - DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)); - - if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) { - DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version)); - - /* - * Other side rebooted and previous XPC did support the - * disengage request, but the new one doesn't. - */ - - xpc_clear_partition_engaged(1UL << partid); - xpc_clear_partition_disengage_request(1UL << partid); - - xpc_update_partition_info(part, remote_rp_version, - &remote_rp_stamp, remote_rp_pa, - remote_vars_pa, remote_vars); - reactivate = 1; - - } else { - DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version)); - - stamp_diff = xpc_compare_stamps(&part->remote_rp_stamp, - &remote_rp_stamp); - if (stamp_diff != 0) { - DBUG_ON(stamp_diff >= 0); - - /* - * Other side rebooted and the previous XPC did support - * the disengage request, as does the new one. - */ - - DBUG_ON(xpc_partition_engaged(1UL << partid)); - DBUG_ON(xpc_partition_disengage_requested(1UL << - partid)); - - xpc_update_partition_info(part, remote_rp_version, - &remote_rp_stamp, - remote_rp_pa, remote_vars_pa, - remote_vars); - reactivate = 1; - } - } - - if (part->disengage_request_timeout > 0 && - !xpc_partition_disengaged(part)) { - /* still waiting on other side to disengage from us */ - return; - } - - if (reactivate) { - part->reactivate_nasid = nasid; - XPC_DEACTIVATE_PARTITION(part, xpReactivating); - - } else if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version) && - xpc_partition_disengage_requested(1UL << partid)) { - XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown); - } -} - -/* - * Loop through the activation AMO variables and process any bits - * which are set. Each bit indicates a nasid sending a partition - * activation or deactivation request. - * - * Return #of IRQs detected. - */ -int -xpc_identify_act_IRQ_sender(void) -{ - int word, bit; - u64 nasid_mask; - u64 nasid; /* remote nasid */ - int n_IRQs_detected = 0; - AMO_t *act_amos; - - act_amos = xpc_vars->amos_page + XPC_ACTIVATE_IRQ_AMOS; - - /* scan through act AMO variable looking for non-zero entries */ - for (word = 0; word < xp_nasid_mask_words; word++) { - - if (xpc_exiting) - break; - - nasid_mask = xpc_IPI_receive(&act_amos[word]); - if (nasid_mask == 0) { - /* no IRQs from nasids in this variable */ - continue; - } - - dev_dbg(xpc_part, "AMO[%d] gave back 0x%lx\n", word, - nasid_mask); - - /* - * If this nasid has been added to the machine since - * our partition was reset, this will retain the - * remote nasid in our reserved pages machine mask. - * This is used in the event of module reload. - */ - xpc_mach_nasids[word] |= nasid_mask; - - /* locate the nasid(s) which sent interrupts */ - - for (bit = 0; bit < (8 * sizeof(u64)); bit++) { - if (nasid_mask & (1UL << bit)) { - n_IRQs_detected++; - nasid = XPC_NASID_FROM_W_B(word, bit); - dev_dbg(xpc_part, "interrupt from nasid %ld\n", - nasid); - xpc_identify_act_IRQ_req(nasid); - } - } - } - return n_IRQs_detected; -} - -/* * See if the other side has responded to a partition disengage request * from us. */ @@ -836,7 +464,7 @@ xpc_deactivate_partition(const int line, spin_unlock_irqrestore(&part->act_lock, irq_flags); if (reason == xpReactivating) { /* we interrupt ourselves to reactivate partition */ - xpc_IPI_send_reactivate(part); + xpc_IPI_send_local_reactivate(part->reactivate_nasid); } return; } @@ -903,16 +531,12 @@ xpc_discovery(void) { void *remote_rp_base; struct xpc_rsvd_page *remote_rp; - struct xpc_vars *remote_vars; u64 remote_rp_pa; - u64 remote_vars_pa; int region; int region_size; int max_regions; int nasid; struct xpc_rsvd_page *rp; - short partid; - struct xpc_partition *part; u64 *discovered_nasids; enum xp_retval ret; @@ -922,8 +546,6 @@ xpc_discovery(void) if (remote_rp == NULL) return; - remote_vars = (struct xpc_vars *)remote_rp; - discovered_nasids = kzalloc(sizeof(u64) * xp_nasid_mask_words, GFP_KERNEL); if (discovered_nasids == NULL) { @@ -988,7 +610,7 @@ xpc_discovery(void) continue; } - /* pull over the reserved page structure */ + /* pull over the rsvd page header & part_nasids mask */ ret = xpc_get_remote_rp(nasid, discovered_nasids, remote_rp, &remote_rp_pa); @@ -1003,72 +625,8 @@ xpc_discovery(void) continue; } - remote_vars_pa = remote_rp->sn.vars_pa; - - partid = remote_rp->SAL_partid; - part = &xpc_partitions[partid]; - - /* pull over the cross partition variables */ - - ret = xpc_get_remote_vars(remote_vars_pa, remote_vars); - if (ret != xpSuccess) { - dev_dbg(xpc_part, "unable to get XPC variables " - "from nasid %d, reason=%d\n", nasid, - ret); - - XPC_DEACTIVATE_PARTITION(part, ret); - continue; - } - - if (part->act_state != XPC_P_INACTIVE) { - dev_dbg(xpc_part, "partition %d on nasid %d is " - "already activating\n", partid, nasid); - break; - } - - /* - * Register the remote partition's AMOs with SAL so it - * can handle and cleanup errors within that address - * range should the remote partition go down. We don't - * unregister this range because it is difficult to - * tell when outstanding writes to the remote partition - * are finished and thus when it is thus safe to - * unregister. This should not result in wasted space - * in the SAL xp_addr_region table because we should - * get the same page for remote_act_amos_pa after - * module reloads and system reboots. - */ - if (sn_register_xp_addr_region - (remote_vars->amos_page_pa, PAGE_SIZE, 1) < 0) { - dev_dbg(xpc_part, - "partition %d failed to " - "register xp_addr region 0x%016lx\n", - partid, remote_vars->amos_page_pa); - - XPC_SET_REASON(part, xpPhysAddrRegFailed, - __LINE__); - break; - } - - /* - * The remote nasid is valid and available. - * Send an interrupt to that nasid to notify - * it that we are ready to begin activation. - */ - dev_dbg(xpc_part, "sending an interrupt to AMO 0x%lx, " - "nasid %d, phys_cpuid 0x%x\n", - remote_vars->amos_page_pa, - remote_vars->act_nasid, - remote_vars->act_phys_cpuid); - - if (XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars-> - version)) { - part->remote_amos_page_pa = - remote_vars->amos_page_pa; - xpc_mark_partition_disengaged(part); - xpc_cancel_partition_disengage_request(part); - } - xpc_IPI_send_activate(remote_vars); + xpc_initiate_partition_activation(remote_rp, + remote_rp_pa, nasid); } } Index: linux-2.6/drivers/misc/sgi-xp/xpc_main.c =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc_main.c 2008-05-28 07:03:09.400493652 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc_main.c 2008-05-28 07:03:15.329230512 -0500 @@ -148,12 +148,14 @@ static struct ctl_table_header *xpc_sysc int xpc_disengage_request_timedout; /* #of IRQs received */ -static atomic_t xpc_act_IRQ_rcvd; +atomic_t xpc_act_IRQ_rcvd; /* IRQ handler notifies this wait queue on receipt of an IRQ */ -static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq); +DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq); static unsigned long xpc_hb_check_timeout; +static struct timer_list xpc_hb_timer; +void *xpc_heartbeating_to_mask; /* notification that the xpc_hb_checker thread has exited */ static DECLARE_COMPLETION(xpc_hb_checker_exited); @@ -161,8 +163,6 @@ static DECLARE_COMPLETION(xpc_hb_checker /* notification that the xpc_discovery thread has exited */ static DECLARE_COMPLETION(xpc_discovery_exited); -static struct timer_list xpc_hb_timer; - static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *); static int xpc_system_reboot(struct notifier_block *, unsigned long, void *); @@ -176,12 +176,54 @@ static struct notifier_block xpc_die_not }; enum xp_retval (*xpc_rsvd_page_init) (struct xpc_rsvd_page *rp); +void (*xpc_heartbeat_init) (void); +void (*xpc_heartbeat_exit) (void); +void (*xpc_increment_heartbeat) (void); +void (*xpc_offline_heartbeat) (void); +void (*xpc_online_heartbeat) (void); +void (*xpc_check_remote_hb) (void); + enum xp_retval (*xpc_make_first_contact) (struct xpc_partition *part); u64 (*xpc_get_IPI_flags) (struct xpc_partition *part); struct xpc_msg *(*xpc_get_deliverable_msg) (struct xpc_channel *ch); + +void (*xpc_initiate_partition_activation) (struct xpc_rsvd_page *remote_rp, + u64 remote_rp_pa, int nasid); + +void (*xpc_process_act_IRQ_rcvd) (int n_IRQs_expected); enum xp_retval (*xpc_setup_infrastructure) (struct xpc_partition *part); void (*xpc_teardown_infrastructure) (struct xpc_partition *part); +void (*xpc_mark_partition_engaged) (struct xpc_partition *part); +void (*xpc_mark_partition_disengaged) (struct xpc_partition *part); +void (*xpc_request_partition_disengage) (struct xpc_partition *part); +void (*xpc_cancel_partition_disengage_request) (struct xpc_partition *part); +u64 (*xpc_partition_engaged) (u64 partid_mask); +u64 (*xpc_partition_disengage_requested) (u64 partid_mask); +void (*xpc_clear_partition_engaged) (u64 partid_mask); +void (*xpc_clear_partition_disengage_request) (u64 partid_mask); + +void (*xpc_IPI_send_local_activate) (int from_nasid); +void (*xpc_IPI_send_activated) (struct xpc_partition *part); +void (*xpc_IPI_send_local_reactivate) (int from_nasid); +void (*xpc_IPI_send_disengage) (struct xpc_partition *part); + +void (*xpc_IPI_send_closerequest) (struct xpc_channel *ch, + unsigned long *irq_flags); +void (*xpc_IPI_send_closereply) (struct xpc_channel *ch, + unsigned long *irq_flags); +void (*xpc_IPI_send_openrequest) (struct xpc_channel *ch, + unsigned long *irq_flags); +void (*xpc_IPI_send_openreply) (struct xpc_channel *ch, + unsigned long *irq_flags); + +enum xp_retval (*xpc_allocate_msg) (struct xpc_channel *ch, u32 flags, + struct xpc_msg **address_of_msg); + +enum xp_retval (*xpc_send_msg) (struct xpc_channel *ch, struct xpc_msg *msg, + u8 notify_type, xpc_notify_func func, + void *key); +void (*xpc_received_msg) (struct xpc_channel *ch, struct xpc_msg *msg); /* * Timer function to enforce the timelimit on the partition disengage request. @@ -218,7 +260,7 @@ xpc_act_IRQ_handler(int irq, void *dev_i static void xpc_hb_beater(unsigned long dummy) { - xpc_vars->heartbeat++; + xpc_increment_heartbeat(); if (time_after_eq(jiffies, xpc_hb_check_timeout)) wake_up_interruptible(&xpc_act_IRQ_wq); @@ -227,6 +269,22 @@ xpc_hb_beater(unsigned long dummy) add_timer(&xpc_hb_timer); } +static void +xpc_start_hb_beater(void) +{ + xpc_heartbeat_init(); + init_timer(&xpc_hb_timer); + xpc_hb_timer.function = xpc_hb_beater; + xpc_hb_beater(0); +} + +static void +xpc_stop_hb_beater(void) +{ + del_timer_sync(&xpc_hb_timer); + xpc_heartbeat_exit(); +} + /* * This thread is responsible for nearly all of the partition * activation/deactivation. @@ -244,7 +302,7 @@ xpc_hb_checker(void *ignore) /* set our heartbeating to other partitions into motion */ xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ); - xpc_hb_beater(0); + xpc_start_hb_beater(); while (!xpc_exiting) { @@ -274,11 +332,8 @@ xpc_hb_checker(void *ignore) dev_dbg(xpc_part, "found an IRQ to process; will be " "resetting xpc_hb_check_timeout\n"); - last_IRQ_count += xpc_identify_act_IRQ_sender(); - if (last_IRQ_count < new_IRQ_count) { - /* retry once to help avoid missing AMO */ - (void)xpc_identify_act_IRQ_sender(); - } + xpc_process_act_IRQ_rcvd(new_IRQ_count - + last_IRQ_count); last_IRQ_count = new_IRQ_count; xpc_hb_check_timeout = jiffies + @@ -294,6 +349,8 @@ xpc_hb_checker(void *ignore) xpc_exiting)); } + xpc_stop_hb_beater(); + dev_dbg(xpc_part, "heartbeat checker is exiting\n"); /* mark this thread as having exited */ @@ -401,31 +458,7 @@ xpc_activating(void *__partid) dev_dbg(xpc_part, "activating partition %d\n", partid); - /* - * Register the remote partition's AMOs with SAL so it can handle - * and cleanup errors within that address range should the remote - * partition go down. We don't unregister this range because it is - * difficult to tell when outstanding writes to the remote partition - * are finished and thus when it is safe to unregister. This should - * not result in wasted space in the SAL xp_addr_region table because - * we should get the same page for remote_amos_page_pa after module - * reloads and system reboots. - */ - if (sn_register_xp_addr_region(part->remote_amos_page_pa, - PAGE_SIZE, 1) < 0) { - dev_warn(xpc_part, "xpc_activating(%d) failed to register " - "xp_addr region\n", partid); - - spin_lock_irqsave(&part->act_lock, irq_flags); - part->act_state = XPC_P_INACTIVE; - XPC_SET_REASON(part, xpPhysAddrRegFailed, __LINE__); - spin_unlock_irqrestore(&part->act_lock, irq_flags); - part->remote_rp_pa = 0; - return 0; - } - - xpc_allow_hb(partid, xpc_vars); - xpc_IPI_send_activated(part); + xpc_allow_hb(partid); if (xpc_setup_infrastructure(part) == xpSuccess) { (void)xpc_part_ref(part); /* this will always succeed */ @@ -440,12 +473,12 @@ xpc_activating(void *__partid) xpc_teardown_infrastructure(part); } - xpc_disallow_hb(partid, xpc_vars); + xpc_disallow_hb(partid); xpc_mark_partition_inactive(part); if (part->reason == xpReactivating) { /* interrupting ourselves results in activating partition */ - xpc_IPI_send_reactivate(part); + xpc_IPI_send_local_reactivate(part->reactivate_nasid); } return 0; @@ -478,6 +511,32 @@ xpc_activate_partition(struct xpc_partit } /* + * Check to see if there is any channel activity to/from the specified + * partition. + */ +static void +xpc_check_for_channel_activity(struct xpc_partition *part) +{ + u64 IPI_amo; + unsigned long irq_flags; + +/* this needs to be uncommented, but I'm thinking this function and the */ +/* ones that call it need to be moved into xpc_sn2.c... */ + IPI_amo = 0; /* = xpc_IPI_receive(part->local_IPI_amo_va); */ + if (IPI_amo == 0) + return; + + spin_lock_irqsave(&part->IPI_lock, irq_flags); + part->local_IPI_amo |= IPI_amo; + spin_unlock_irqrestore(&part->IPI_lock, irq_flags); + + dev_dbg(xpc_chan, "received IPI from partid=%d, IPI_amo=0x%lx\n", + XPC_PARTID(part), IPI_amo); + + xpc_wakeup_channel_mgr(part); +} + +/* * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more * than one partition, we use an AMO_t structure per partition to indicate @@ -902,14 +961,11 @@ xpc_do_exit(enum xp_retval reason) } while (1); DBUG_ON(xpc_partition_engaged(-1UL)); + DBUG_ON(xpc_any_hbs_allowed() != 0); /* indicate to others that our reserved page is uninitialized */ xpc_rsvd_page->stamp = ZERO_STAMP; - /* now it's time to eliminate our heartbeat */ - del_timer_sync(&xpc_hb_timer); - DBUG_ON(xpc_vars->heartbeating_to_mask != 0); - if (reason == xpUnloading) { (void)unregister_die_notifier(&xpc_die_notifier); (void)unregister_reboot_notifier(&xpc_reboot_notifier); @@ -968,7 +1024,7 @@ xpc_die_disengage(void) /* keep xpc_hb_checker thread from doing anything (just in case) */ xpc_exiting = 1; - xpc_vars->heartbeating_to_mask = 0; /* indicate we're deactivated */ + xpc_disallow_all_hbs(); /*indicate we're deactivated */ for (partid = 0; partid < xp_max_npartitions; partid++) { part = &xpc_partitions[partid]; @@ -1054,8 +1110,7 @@ xpc_system_die(struct notifier_block *nb /* fall through */ case DIE_MCA_MONARCH_ENTER: case DIE_INIT_MONARCH_ENTER: - xpc_vars->heartbeat++; - xpc_vars->heartbeat_offline = 1; + xpc_offline_heartbeat(); break; case DIE_KDEBUG_LEAVE: @@ -1066,8 +1121,7 @@ xpc_system_die(struct notifier_block *nb /* fall through */ case DIE_MCA_MONARCH_LEAVE: case DIE_INIT_MONARCH_LEAVE: - xpc_vars->heartbeat++; - xpc_vars->heartbeat_offline = 0; + xpc_online_heartbeat(); break; } @@ -1202,9 +1256,6 @@ xpc_init(void) if (ret != 0) dev_warn(xpc_part, "can't register die notifier\n"); - init_timer(&xpc_hb_timer); - xpc_hb_timer.function = xpc_hb_beater; - /* * The real work-horse behind xpc. This processes incoming * interrupts and monitors remote heartbeats. @@ -1246,7 +1297,6 @@ out_4: /* indicate to others that our reserved page is uninitialized */ xpc_rsvd_page->stamp = ZERO_STAMP; - del_timer_sync(&xpc_hb_timer); (void)unregister_die_notifier(&xpc_die_notifier); (void)unregister_reboot_notifier(&xpc_reboot_notifier); out_3: Index: linux-2.6/drivers/misc/sgi-xp/xpc_uv.c =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc_uv.c 2008-05-28 07:03:09.444499121 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc_uv.c 2008-05-28 07:03:15.353233496 -0500 @@ -19,15 +19,22 @@ /* >>> uv_gpa() is defined in */ #define uv_gpa(_a) ((unsigned long)_a) -/* >>> temporarily define next three items for xpc.h */ -#define SGI_XPC_ACTIVATE 23 -#define SGI_XPC_NOTIFY 24 -#define sn_send_IPI_phys(_a, _b, _c, _d) - #include "xpc.h" +static DECLARE_BITMAP(xpc_heartbeating_to_mask_uv, XP_MAX_NPARTITIONS_UV); + static void *xpc_activate_mq; +static void +xpc_IPI_send_local_activate_uv(struct xpc_partition *part) +{ + /* + * >>> make our side think that the remote parition sent an activate + * >>> message our way. Also do what the activate IRQ handler would + * >>> do had one really been sent. + */ +} + static enum xp_retval xpc_rsvd_page_init_uv(struct xpc_rsvd_page *rp) { @@ -36,6 +43,41 @@ xpc_rsvd_page_init_uv(struct xpc_rsvd_pa return xpSuccess; } +static void +xpc_increment_heartbeat_uv(void) +{ + /* >>> send heartbeat msg to xpc_heartbeating_to_mask partids */ +} + +static void +xpc_heartbeat_init_uv(void) +{ + bitmap_zero(xpc_heartbeating_to_mask_uv, XP_MAX_NPARTITIONS_UV); + xpc_heartbeating_to_mask = &xpc_heartbeating_to_mask_uv[0]; +} + +static void +xpc_heartbeat_exit_uv(void) +{ + /* >>> send heartbeat_offline msg to xpc_heartbeating_to_mask partids */ +} + +static void +xpc_initiate_partition_activation_uv(struct xpc_rsvd_page *remote_rp, + u64 remote_rp_pa, int nasid) +{ + short partid = remote_rp->SAL_partid; + struct xpc_partition *part = &xpc_partitions[partid]; + +/* + * >>> setup part structure with the bits of info we can glean from the rp + * >>> part->remote_rp_pa = remote_rp_pa; + * >>> part->sn.uv.activate_mq_gpa = remote_rp->sn.activate_mq_gpa; + */ + + xpc_IPI_send_local_activate_uv(part); +} + /* * Setup the infrastructure necessary to support XPartition Communication * between the specified remote partition and the local one. @@ -83,6 +125,11 @@ void xpc_init_uv(void) { xpc_rsvd_page_init = xpc_rsvd_page_init_uv; + xpc_increment_heartbeat = xpc_increment_heartbeat_uv; + xpc_heartbeat_init = xpc_heartbeat_init_uv; + xpc_heartbeat_exit = xpc_heartbeat_exit_uv; + xpc_initiate_partition_activation = + xpc_initiate_partition_activation_uv; xpc_setup_infrastructure = xpc_setup_infrastructure_uv; xpc_teardown_infrastructure = xpc_teardown_infrastructure_uv; xpc_make_first_contact = xpc_make_first_contact_uv; Index: linux-2.6/drivers/misc/sgi-xp/xpc_channel.c =================================================================== --- linux-2.6.orig/drivers/misc/sgi-xp/xpc_channel.c 2008-05-28 07:03:09.384491663 -0500 +++ linux-2.6/drivers/misc/sgi-xp/xpc_channel.c 2008-05-28 07:03:15.369235484 -0500 @@ -1165,7 +1165,7 @@ xpc_disconnect_callout(struct xpc_channe * Wait for a message entry to become available for the specified channel, * but don't wait any longer than 1 jiffy. */ -static enum xp_retval +enum xp_retval xpc_allocate_msg_wait(struct xpc_channel *ch) { enum xp_retval ret; @@ -1193,96 +1193,6 @@ xpc_allocate_msg_wait(struct xpc_channel /* * Allocate an entry for a message from the message queue associated with the - * specified channel. - */ -static enum xp_retval -xpc_allocate_msg(struct xpc_channel *ch, u32 flags, - struct xpc_msg **address_of_msg) -{ - struct xpc_msg *msg; - enum xp_retval ret; - s64 put; - - /* this reference will be dropped in xpc_send_msg() */ - xpc_msgqueue_ref(ch); - - if (ch->flags & XPC_C_DISCONNECTING) { - xpc_msgqueue_deref(ch); - return ch->reason; - } - if (!(ch->flags & XPC_C_CONNECTED)) { - xpc_msgqueue_deref(ch); - return xpNotConnected; - } - - /* - * Get the next available message entry from the local message queue. - * If none are available, we'll make sure that we grab the latest - * GP values. - */ - ret = xpTimeout; - - while (1) { - - put = ch->w_local_GP.put; - rmb(); /* guarantee that .put loads before .get */ - if (put - ch->w_remote_GP.get < ch->local_nentries) { - - /* There are available message entries. We need to try - * to secure one for ourselves. We'll do this by trying - * to increment w_local_GP.put as long as someone else - * doesn't beat us to it. If they do, we'll have to - * try again. - */ - if (cmpxchg(&ch->w_local_GP.put, put, put + 1) == put) { - /* we got the entry referenced by put */ - break; - } - continue; /* try again */ - } - - /* - * There aren't any available msg entries at this time. - * - * In waiting for a message entry to become available, - * we set a timeout in case the other side is not - * sending completion IPIs. This lets us fake an IPI - * that will cause the IPI handler to fetch the latest - * GP values as if an IPI was sent by the other side. - */ - if (ret == xpTimeout) - xpc_IPI_send_local_msgrequest(ch); - - if (flags & XPC_NOWAIT) { - xpc_msgqueue_deref(ch); - return xpNoWait; - } - - ret = xpc_allocate_msg_wait(ch); - if (ret != xpInterrupted && ret != xpTimeout) { - xpc_msgqueue_deref(ch); - return ret; - } - } - - /* get the message's address and initialize it */ - msg = (struct xpc_msg *)((u64)ch->local_msgqueue + - (put % ch->local_nentries) * ch->msg_size); - - DBUG_ON(msg->flags != 0); - msg->number = put; - - dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, " - "msg_number=%ld, partid=%d, channel=%d\n", put + 1, - (void *)msg, msg->number, ch->partid, ch->number); - - *address_of_msg = msg; - - return xpSuccess; -} - -/* - * Allocate an entry for a message from the message queue associated with the * specified channel. NOTE that this routine can sleep waiting for a message * entry to become available. To not sleep, pass in the XPC_NOWAIT flag. * @@ -1318,144 +1228,6 @@ xpc_initiate_allocate(short partid, int } /* - * Now we actually send the messages that are ready to be sent by advancing - * the local message queue's Put value and then send an IPI to the recipient - * partition. - */ -static void -xpc_send_msgs(struct xpc_channel *ch, s64 initial_put) -{ - struct xpc_msg *msg; - s64 put = initial_put + 1; - int send_IPI = 0; - - while (1) { - - while (1) { - if (put == ch->w_local_GP.put) - break; - - msg = (struct xpc_msg *)((u64)ch->local_msgqueue + - (put % ch->local_nentries) * - ch->msg_size); - - if (!(msg->flags & XPC_M_READY)) - break; - - put++; - } - - if (put == initial_put) { - /* nothing's changed */ - break; - } - - if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) != - initial_put) { - /* someone else beat us to it */ - DBUG_ON(ch->local_GP->put < initial_put); - break; - } - - /* we just set the new value of local_GP->put */ - - dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, " - "channel=%d\n", put, ch->partid, ch->number); - - send_IPI = 1; - - /* - * We need to ensure that the message referenced by - * local_GP->put is not XPC_M_READY or that local_GP->put - * equals w_local_GP.put, so we'll go have a look. - */ - initial_put = put; - } - - if (send_IPI) - xpc_IPI_send_msgrequest(ch); -} - -/* - * Common code that does the actual sending of the message by advancing the - * local message queue's Put value and sends an IPI to the partition the - * message is being sent to. - */ -static enum xp_retval -xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type, - xpc_notify_func func, void *key) -{ - enum xp_retval ret = xpSuccess; - struct xpc_notify *notify = notify; - s64 put, msg_number = msg->number; - - DBUG_ON(notify_type == XPC_N_CALL && func == NULL); - DBUG_ON((((u64)msg - (u64)ch->local_msgqueue) / ch->msg_size) != - msg_number % ch->local_nentries); - DBUG_ON(msg->flags & XPC_M_READY); - - if (ch->flags & XPC_C_DISCONNECTING) { - /* drop the reference grabbed in xpc_allocate_msg() */ - xpc_msgqueue_deref(ch); - return ch->reason; - } - - if (notify_type != 0) { - /* - * Tell the remote side to send an ACK interrupt when the - * message has been delivered. - */ - msg->flags |= XPC_M_INTERRUPT; - - atomic_inc(&ch->n_to_notify); - - notify = &ch->notify_queue[msg_number % ch->local_nentries]; - notify->func = func; - notify->key = key; - notify->type = notify_type; - - /* >>> is a mb() needed here? */ - - if (ch->flags & XPC_C_DISCONNECTING) { - /* - * An error occurred between our last error check and - * this one. We will try to clear the type field from - * the notify entry. If we succeed then - * xpc_disconnect_channel() didn't already process - * the notify entry. - */ - if (cmpxchg(¬ify->type, notify_type, 0) == - notify_type) { - atomic_dec(&ch->n_to_notify); - ret = ch->reason; - } - - /* drop the reference grabbed in xpc_allocate_msg() */ - xpc_msgqueue_deref(ch); - return ret; - } - } - - msg->flags |= XPC_M_READY; - - /* - * The preceding store of msg->flags must occur before the following - * load of ch->local_GP->put. - */ - mb(); - - /* see if the message is next in line to be sent, if so send it */ - - put = ch->local_GP->put; - if (put == msg_number) - xpc_send_msgs(ch, put); - - /* drop the reference grabbed in xpc_allocate_msg() */ - xpc_msgqueue_deref(ch); - return ret; -} - -/* * Send a message previously allocated using xpc_initiate_allocate() on the * specified channel connected to the specified partition. * @@ -1586,66 +1358,6 @@ xpc_deliver_msg(struct xpc_channel *ch) } /* - * Now we actually acknowledge the messages that have been delivered and ack'd - * by advancing the cached remote message queue's Get value and if requested - * send an IPI to the message sender's partition. - */ -static void -xpc_acknowledge_msgs(struct xpc_channel *ch, s64 initial_get, u8 msg_flags) -{ - struct xpc_msg *msg; - s64 get = initial_get + 1; - int send_IPI = 0; - - while (1) { - - while (1) { - if (get == ch->w_local_GP.get) - break; - - msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + - (get % ch->remote_nentries) * - ch->msg_size); - - if (!(msg->flags & XPC_M_DONE)) - break; - - msg_flags |= msg->flags; - get++; - } - - if (get == initial_get) { - /* nothing's changed */ - break; - } - - if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) != - initial_get) { - /* someone else beat us to it */ - DBUG_ON(ch->local_GP->get <= initial_get); - break; - } - - /* we just set the new value of local_GP->get */ - - dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, " - "channel=%d\n", get, ch->partid, ch->number); - - send_IPI = (msg_flags & XPC_M_INTERRUPT); - - /* - * We need to ensure that the message referenced by - * local_GP->get is not XPC_M_DONE or that local_GP->get - * equals w_local_GP.get, so we'll go have a look. - */ - initial_get = get; - } - - if (send_IPI) - xpc_IPI_send_msgrequest(ch); -} - -/* * Acknowledge receipt of a delivered message. * * If a message has XPC_M_INTERRUPT set, send an interrupt to the partition @@ -1668,35 +1380,12 @@ xpc_initiate_received(short partid, int struct xpc_partition *part = &xpc_partitions[partid]; struct xpc_channel *ch; struct xpc_msg *msg = XPC_MSG_ADDRESS(payload); - s64 get, msg_number = msg->number; DBUG_ON(partid < 0 || partid >= xp_max_npartitions); DBUG_ON(ch_number < 0 || ch_number >= part->nchannels); ch = &part->channels[ch_number]; - - dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n", - (void *)msg, msg_number, ch->partid, ch->number); - - DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) != - msg_number % ch->remote_nentries); - DBUG_ON(msg->flags & XPC_M_DONE); - - msg->flags |= XPC_M_DONE; - - /* - * The preceding store of msg->flags must occur before the following - * load of ch->local_GP->get. - */ - mb(); - - /* - * See if this message is next in line to be acknowledged as having - * been delivered. - */ - get = ch->local_GP->get; - if (get == msg_number) - xpc_acknowledge_msgs(ch, get, msg->flags); + xpc_received_msg(ch, msg); /* the call to xpc_msgqueue_ref() was done by xpc_deliver_msg() */ xpc_msgqueue_deref(ch); -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/