2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
10 * Cross Partition Communication (XPC) sn2-based functions.
12 * Architecture specific implementation of common functions.
16 #include <linux/delay.h>
17 #include <asm/uncached.h>
18 #include <asm/sn/mspec.h>
19 #include <asm/sn/sn_sal.h>
23 * Define the number of u64s required to represent all the C-brick nasids
24 * as a bitmap. The cross-partition kernel modules deal only with
25 * C-brick nasids, thus the need for bitmaps which don't account for
26 * odd-numbered (non C-brick) nasids.
28 #define XPC_MAX_PHYSNODES_SN2 (MAX_NUMALINK_NODES / 2)
29 #define XP_NASID_MASK_BYTES_SN2 ((XPC_MAX_PHYSNODES_SN2 + 7) / 8)
30 #define XP_NASID_MASK_WORDS_SN2 ((XPC_MAX_PHYSNODES_SN2 + 63) / 64)
33 * Memory for XPC's amo variables is allocated by the MSPEC driver. These
34 * pages are located in the lowest granule. The lowest granule uses 4k pages
35 * for cached references and an alternate TLB handler to never provide a
36 * cacheable mapping for the entire region. This will prevent speculative
37 * reading of cached copies of our lines from being issued which will cause
38 * a PI FSB Protocol error to be generated by the SHUB. For XPC, we need 64
39 * amo variables (based on XP_MAX_NPARTITIONS_SN2) to identify the senders of
40 * NOTIFY IRQs, 128 amo variables (based on XP_NASID_MASK_WORDS_SN2) to identify
41 * the senders of ACTIVATE IRQs, 1 amo variable to identify which remote
42 * partitions (i.e., XPCs) consider themselves currently engaged with the
43 * local XPC and 1 amo variable to request partition deactivation.
45 #define XPC_NOTIFY_IRQ_AMOS_SN2 0
46 #define XPC_ACTIVATE_IRQ_AMOS_SN2 (XPC_NOTIFY_IRQ_AMOS_SN2 + \
47 XP_MAX_NPARTITIONS_SN2)
48 #define XPC_ENGAGED_PARTITIONS_AMO_SN2 (XPC_ACTIVATE_IRQ_AMOS_SN2 + \
49 XP_NASID_MASK_WORDS_SN2)
50 #define XPC_DEACTIVATE_REQUEST_AMO_SN2 (XPC_ENGAGED_PARTITIONS_AMO_SN2 + 1)
53 * Buffer used to store a local copy of portions of a remote partition's
54 * reserved page (either its header and part_nasids mask, or its vars).
56 static char *xpc_remote_copy_buffer_sn2;
57 static void *xpc_remote_copy_buffer_base_sn2;
59 static struct xpc_vars_sn2 *xpc_vars_sn2;
60 static struct xpc_vars_part_sn2 *xpc_vars_part_sn2;
62 /* SH_IPI_ACCESS shub register value on startup */
63 static u64 xpc_sh1_IPI_access_sn2;
64 static u64 xpc_sh2_IPI_access0_sn2;
65 static u64 xpc_sh2_IPI_access1_sn2;
66 static u64 xpc_sh2_IPI_access2_sn2;
67 static u64 xpc_sh2_IPI_access3_sn2;
70 * Change protections to allow IPI operations.
73 xpc_allow_IPI_ops_sn2(void)
78 /* !!! The following should get moved into SAL. */
80 xpc_sh2_IPI_access0_sn2 =
81 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS0));
82 xpc_sh2_IPI_access1_sn2 =
83 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS1));
84 xpc_sh2_IPI_access2_sn2 =
85 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS2));
86 xpc_sh2_IPI_access3_sn2 =
87 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS3));
89 for_each_online_node(node) {
90 nasid = cnodeid_to_nasid(node);
91 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
93 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
95 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
97 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
101 xpc_sh1_IPI_access_sn2 =
102 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH1_IPI_ACCESS));
104 for_each_online_node(node) {
105 nasid = cnodeid_to_nasid(node);
106 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
113 * Restrict protections to disallow IPI operations.
116 xpc_disallow_IPI_ops_sn2(void)
121 /* !!! The following should get moved into SAL. */
123 for_each_online_node(node) {
124 nasid = cnodeid_to_nasid(node);
125 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
126 xpc_sh2_IPI_access0_sn2);
127 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
128 xpc_sh2_IPI_access1_sn2);
129 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
130 xpc_sh2_IPI_access2_sn2);
131 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
132 xpc_sh2_IPI_access3_sn2);
135 for_each_online_node(node) {
136 nasid = cnodeid_to_nasid(node);
137 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
138 xpc_sh1_IPI_access_sn2);
144 * The following set of functions are used for the sending and receiving of
145 * IRQs (also known as IPIs). There are two flavors of IRQs, one that is
146 * associated with partition activity (SGI_XPC_ACTIVATE) and the other that
147 * is associated with channel activity (SGI_XPC_NOTIFY).
151 xpc_receive_IRQ_amo_sn2(struct amo *amo)
153 return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_CLEAR);
156 static enum xp_retval
157 xpc_send_IRQ_sn2(struct amo *amo, u64 flag, int nasid, int phys_cpuid,
161 unsigned long irq_flags;
163 local_irq_save(irq_flags);
165 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR, flag);
166 sn_send_IPI_phys(nasid, phys_cpuid, vector, 0);
169 * We must always use the nofault function regardless of whether we
170 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
171 * didn't, we'd never know that the other partition is down and would
172 * keep sending IRQs and amos to it until the heartbeat times out.
174 ret = xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->variable),
175 xp_nofault_PIOR_target));
177 local_irq_restore(irq_flags);
179 return (ret == 0) ? xpSuccess : xpPioReadError;
183 xpc_init_IRQ_amo_sn2(int index)
185 struct amo *amo = xpc_vars_sn2->amos_page + index;
187 (void)xpc_receive_IRQ_amo_sn2(amo); /* clear amo variable */
192 * Functions associated with SGI_XPC_ACTIVATE IRQ.
196 * Notify the heartbeat check thread that an activate IRQ has been received.
199 xpc_handle_activate_IRQ_sn2(int irq, void *dev_id)
201 atomic_inc(&xpc_activate_IRQ_rcvd);
202 wake_up_interruptible(&xpc_activate_IRQ_wq);
207 * Flag the appropriate amo variable and send an IRQ to the specified node.
210 xpc_send_activate_IRQ_sn2(unsigned long amos_page_pa, int from_nasid,
211 int to_nasid, int to_phys_cpuid)
213 struct amo *amos = (struct amo *)__va(amos_page_pa +
214 (XPC_ACTIVATE_IRQ_AMOS_SN2 *
215 sizeof(struct amo)));
217 (void)xpc_send_IRQ_sn2(&amos[BIT_WORD(from_nasid / 2)],
218 BIT_MASK(from_nasid / 2), to_nasid,
219 to_phys_cpuid, SGI_XPC_ACTIVATE);
223 xpc_send_local_activate_IRQ_sn2(int from_nasid)
225 struct amo *amos = (struct amo *)__va(xpc_vars_sn2->amos_page_pa +
226 (XPC_ACTIVATE_IRQ_AMOS_SN2 *
227 sizeof(struct amo)));
229 /* fake the sending and receipt of an activate IRQ from remote nasid */
230 FETCHOP_STORE_OP(TO_AMO((u64)&amos[BIT_WORD(from_nasid / 2)].variable),
231 FETCHOP_OR, BIT_MASK(from_nasid / 2));
233 atomic_inc(&xpc_activate_IRQ_rcvd);
234 wake_up_interruptible(&xpc_activate_IRQ_wq);
238 * Functions associated with SGI_XPC_NOTIFY IRQ.
242 * Check to see if any chctl flags were sent from the specified partition.
245 xpc_check_for_sent_chctl_flags_sn2(struct xpc_partition *part)
247 union xpc_channel_ctl_flags chctl;
248 unsigned long irq_flags;
250 chctl.all_flags = xpc_receive_IRQ_amo_sn2(part->sn.sn2.
252 if (chctl.all_flags == 0)
255 spin_lock_irqsave(&part->chctl_lock, irq_flags);
256 part->chctl.all_flags |= chctl.all_flags;
257 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
259 dev_dbg(xpc_chan, "received notify IRQ from partid=%d, chctl.all_flags="
260 "0x%lx\n", XPC_PARTID(part), chctl.all_flags);
262 xpc_wakeup_channel_mgr(part);
266 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
267 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
268 * than one partition, we use an amo structure per partition to indicate
269 * whether a partition has sent an IRQ or not. If it has, then wake up the
270 * associated kthread to handle it.
272 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IRQs sent by XPC
273 * running on other partitions.
275 * Noteworthy Arguments:
277 * irq - Interrupt ReQuest number. NOT USED.
279 * dev_id - partid of IRQ's potential sender.
282 xpc_handle_notify_IRQ_sn2(int irq, void *dev_id)
284 short partid = (short)(u64)dev_id;
285 struct xpc_partition *part = &xpc_partitions[partid];
287 DBUG_ON(partid < 0 || partid >= XP_MAX_NPARTITIONS_SN2);
289 if (xpc_part_ref(part)) {
290 xpc_check_for_sent_chctl_flags_sn2(part);
292 xpc_part_deref(part);
298 * Check to see if xpc_handle_notify_IRQ_sn2() dropped any IRQs on the floor
299 * because the write to their associated amo variable completed after the IRQ
303 xpc_check_for_dropped_notify_IRQ_sn2(struct xpc_partition *part)
305 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
307 if (xpc_part_ref(part)) {
308 xpc_check_for_sent_chctl_flags_sn2(part);
310 part_sn2->dropped_notify_IRQ_timer.expires = jiffies +
311 XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
312 add_timer(&part_sn2->dropped_notify_IRQ_timer);
313 xpc_part_deref(part);
318 * Send a notify IRQ to the remote partition that is associated with the
322 xpc_send_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
323 char *chctl_flag_string, unsigned long *irq_flags)
325 struct xpc_partition *part = &xpc_partitions[ch->partid];
326 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
327 union xpc_channel_ctl_flags chctl = { 0 };
330 if (likely(part->act_state != XPC_P_DEACTIVATING)) {
331 chctl.flags[ch->number] = chctl_flag;
332 ret = xpc_send_IRQ_sn2(part_sn2->remote_chctl_amo_va,
334 part_sn2->notify_IRQ_nasid,
335 part_sn2->notify_IRQ_phys_cpuid,
337 dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n",
338 chctl_flag_string, ch->partid, ch->number, ret);
339 if (unlikely(ret != xpSuccess)) {
340 if (irq_flags != NULL)
341 spin_unlock_irqrestore(&ch->lock, *irq_flags);
342 XPC_DEACTIVATE_PARTITION(part, ret);
343 if (irq_flags != NULL)
344 spin_lock_irqsave(&ch->lock, *irq_flags);
349 #define XPC_SEND_NOTIFY_IRQ_SN2(_ch, _ipi_f, _irq_f) \
350 xpc_send_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f, _irq_f)
353 * Make it look like the remote partition, which is associated with the
354 * specified channel, sent us a notify IRQ. This faked IRQ will be handled
355 * by xpc_check_for_dropped_notify_IRQ_sn2().
358 xpc_send_local_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
359 char *chctl_flag_string)
361 struct xpc_partition *part = &xpc_partitions[ch->partid];
362 union xpc_channel_ctl_flags chctl = { 0 };
364 chctl.flags[ch->number] = chctl_flag;
365 FETCHOP_STORE_OP(TO_AMO((u64)&part->sn.sn2.local_chctl_amo_va->
366 variable), FETCHOP_OR, chctl.all_flags);
367 dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n",
368 chctl_flag_string, ch->partid, ch->number);
371 #define XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(_ch, _ipi_f) \
372 xpc_send_local_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f)
375 xpc_send_chctl_closerequest_sn2(struct xpc_channel *ch,
376 unsigned long *irq_flags)
378 struct xpc_openclose_args *args = ch->local_openclose_args;
380 args->reason = ch->reason;
381 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREQUEST, irq_flags);
385 xpc_send_chctl_closereply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
387 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREPLY, irq_flags);
391 xpc_send_chctl_openrequest_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
393 struct xpc_openclose_args *args = ch->local_openclose_args;
395 args->msg_size = ch->msg_size;
396 args->local_nentries = ch->local_nentries;
397 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREQUEST, irq_flags);
401 xpc_send_chctl_openreply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
403 struct xpc_openclose_args *args = ch->local_openclose_args;
405 args->remote_nentries = ch->remote_nentries;
406 args->local_nentries = ch->local_nentries;
407 args->local_msgqueue_pa = xp_pa(ch->local_msgqueue);
408 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREPLY, irq_flags);
412 xpc_send_chctl_msgrequest_sn2(struct xpc_channel *ch)
414 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST, NULL);
418 xpc_send_chctl_local_msgrequest_sn2(struct xpc_channel *ch)
420 XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST);
424 * This next set of functions are used to keep track of when a partition is
425 * potentially engaged in accessing memory belonging to another partition.
429 xpc_indicate_partition_engaged_sn2(struct xpc_partition *part)
431 unsigned long irq_flags;
432 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
433 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
434 sizeof(struct amo)));
436 local_irq_save(irq_flags);
438 /* set bit corresponding to our partid in remote partition's amo */
439 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
440 BIT(sn_partition_id));
443 * We must always use the nofault function regardless of whether we
444 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
445 * didn't, we'd never know that the other partition is down and would
446 * keep sending IRQs and amos to it until the heartbeat times out.
448 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
450 xp_nofault_PIOR_target));
452 local_irq_restore(irq_flags);
456 xpc_indicate_partition_disengaged_sn2(struct xpc_partition *part)
458 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
459 unsigned long irq_flags;
460 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
461 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
462 sizeof(struct amo)));
464 local_irq_save(irq_flags);
466 /* clear bit corresponding to our partid in remote partition's amo */
467 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
468 ~BIT(sn_partition_id));
471 * We must always use the nofault function regardless of whether we
472 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
473 * didn't, we'd never know that the other partition is down and would
474 * keep sending IRQs and amos to it until the heartbeat times out.
476 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
478 xp_nofault_PIOR_target));
480 local_irq_restore(irq_flags);
483 * Send activate IRQ to get other side to see that we've cleared our
484 * bit in their engaged partitions amo.
486 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
488 part_sn2->activate_IRQ_nasid,
489 part_sn2->activate_IRQ_phys_cpuid);
493 xpc_partition_engaged_sn2(short partid)
495 struct amo *amo = xpc_vars_sn2->amos_page +
496 XPC_ENGAGED_PARTITIONS_AMO_SN2;
498 /* our partition's amo variable ANDed with partid mask */
499 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
504 xpc_any_partition_engaged_sn2(void)
506 struct amo *amo = xpc_vars_sn2->amos_page +
507 XPC_ENGAGED_PARTITIONS_AMO_SN2;
509 /* our partition's amo variable */
510 return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) != 0;
514 xpc_assume_partition_disengaged_sn2(short partid)
516 struct amo *amo = xpc_vars_sn2->amos_page +
517 XPC_ENGAGED_PARTITIONS_AMO_SN2;
519 /* clear bit(s) based on partid mask in our partition's amo */
520 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
524 /* original protection values for each node */
525 static u64 xpc_prot_vec_sn2[MAX_NUMNODES];
528 * Change protections to allow amo operations on non-Shub 1.1 systems.
530 static enum xp_retval
531 xpc_allow_amo_ops_sn2(struct amo *amos_page)
537 * On SHUB 1.1, we cannot call sn_change_memprotect() since the BIST
538 * collides with memory operations. On those systems we call
539 * xpc_allow_amo_ops_shub_wars_1_1_sn2() instead.
541 if (!enable_shub_wars_1_1()) {
542 ret = sn_change_memprotect(ia64_tpa((u64)amos_page), PAGE_SIZE,
543 SN_MEMPROT_ACCESS_CLASS_1,
552 * Change protections to allow amo operations on Shub 1.1 systems.
555 xpc_allow_amo_ops_shub_wars_1_1_sn2(void)
560 if (!enable_shub_wars_1_1())
563 for_each_online_node(node) {
564 nasid = cnodeid_to_nasid(node);
565 /* save current protection values */
566 xpc_prot_vec_sn2[node] =
567 (u64)HUB_L((u64 *)GLOBAL_MMR_ADDR(nasid,
568 SH1_MD_DQLP_MMR_DIR_PRIVEC0));
569 /* open up everything */
570 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid,
571 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
573 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid,
574 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
579 static enum xp_retval
580 xpc_get_partition_rsvd_page_pa_sn2(void *buf, u64 *cookie, unsigned long *rp_pa,
586 status = sn_partition_reserved_page_pa((u64)buf, cookie, rp_pa, len);
587 if (status == SALRET_OK)
589 else if (status == SALRET_MORE_PASSES)
590 ret = xpNeedMoreInfo;
598 static enum xp_retval
599 xpc_rsvd_page_init_sn2(struct xpc_rsvd_page *rp)
601 struct amo *amos_page;
605 xpc_vars_sn2 = XPC_RP_VARS(rp);
607 rp->sn.vars_pa = xp_pa(xpc_vars_sn2);
609 /* vars_part array follows immediately after vars */
610 xpc_vars_part_sn2 = (struct xpc_vars_part_sn2 *)((u8 *)XPC_RP_VARS(rp) +
614 * Before clearing xpc_vars_sn2, see if a page of amos had been
615 * previously allocated. If not we'll need to allocate one and set
616 * permissions so that cross-partition amos are allowed.
618 * The allocated amo page needs MCA reporting to remain disabled after
619 * XPC has unloaded. To make this work, we keep a copy of the pointer
620 * to this page (i.e., amos_page) in the struct xpc_vars_sn2 structure,
621 * which is pointed to by the reserved page, and re-use that saved copy
622 * on subsequent loads of XPC. This amo page is never freed, and its
623 * memory protections are never restricted.
625 amos_page = xpc_vars_sn2->amos_page;
626 if (amos_page == NULL) {
627 amos_page = (struct amo *)TO_AMO(uncached_alloc_page(0, 1));
628 if (amos_page == NULL) {
629 dev_err(xpc_part, "can't allocate page of amos\n");
634 * Open up amo-R/W to cpu. This is done on Shub 1.1 systems
635 * when xpc_allow_amo_ops_shub_wars_1_1_sn2() is called.
637 ret = xpc_allow_amo_ops_sn2(amos_page);
638 if (ret != xpSuccess) {
639 dev_err(xpc_part, "can't allow amo operations\n");
640 uncached_free_page(__IA64_UNCACHED_OFFSET |
641 TO_PHYS((u64)amos_page), 1);
646 /* clear xpc_vars_sn2 */
647 memset(xpc_vars_sn2, 0, sizeof(struct xpc_vars_sn2));
649 xpc_vars_sn2->version = XPC_V_VERSION;
650 xpc_vars_sn2->activate_IRQ_nasid = cpuid_to_nasid(0);
651 xpc_vars_sn2->activate_IRQ_phys_cpuid = cpu_physical_id(0);
652 xpc_vars_sn2->vars_part_pa = xp_pa(xpc_vars_part_sn2);
653 xpc_vars_sn2->amos_page_pa = ia64_tpa((u64)amos_page);
654 xpc_vars_sn2->amos_page = amos_page; /* save for next load of XPC */
656 /* clear xpc_vars_part_sn2 */
657 memset((u64 *)xpc_vars_part_sn2, 0, sizeof(struct xpc_vars_part_sn2) *
658 XP_MAX_NPARTITIONS_SN2);
660 /* initialize the activate IRQ related amo variables */
661 for (i = 0; i < xpc_nasid_mask_nlongs; i++)
662 (void)xpc_init_IRQ_amo_sn2(XPC_ACTIVATE_IRQ_AMOS_SN2 + i);
664 /* initialize the engaged remote partitions related amo variables */
665 (void)xpc_init_IRQ_amo_sn2(XPC_ENGAGED_PARTITIONS_AMO_SN2);
666 (void)xpc_init_IRQ_amo_sn2(XPC_DEACTIVATE_REQUEST_AMO_SN2);
672 xpc_increment_heartbeat_sn2(void)
674 xpc_vars_sn2->heartbeat++;
678 xpc_offline_heartbeat_sn2(void)
680 xpc_increment_heartbeat_sn2();
681 xpc_vars_sn2->heartbeat_offline = 1;
685 xpc_online_heartbeat_sn2(void)
687 xpc_increment_heartbeat_sn2();
688 xpc_vars_sn2->heartbeat_offline = 0;
692 xpc_heartbeat_init_sn2(void)
694 DBUG_ON(xpc_vars_sn2 == NULL);
696 bitmap_zero(xpc_vars_sn2->heartbeating_to_mask, XP_MAX_NPARTITIONS_SN2);
697 xpc_heartbeating_to_mask = &xpc_vars_sn2->heartbeating_to_mask[0];
698 xpc_online_heartbeat_sn2();
702 xpc_heartbeat_exit_sn2(void)
704 xpc_offline_heartbeat_sn2();
708 * At periodic intervals, scan through all active partitions and ensure
709 * their heartbeat is still active. If not, the partition is deactivated.
712 xpc_check_remote_hb_sn2(void)
714 struct xpc_vars_sn2 *remote_vars;
715 struct xpc_partition *part;
719 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
721 for (partid = 0; partid < XP_MAX_NPARTITIONS_SN2; partid++) {
726 if (partid == sn_partition_id)
729 part = &xpc_partitions[partid];
731 if (part->act_state == XPC_P_INACTIVE ||
732 part->act_state == XPC_P_DEACTIVATING) {
736 /* pull the remote_hb cache line */
737 ret = xp_remote_memcpy(xp_pa(remote_vars),
738 part->sn.sn2.remote_vars_pa,
740 if (ret != xpSuccess) {
741 XPC_DEACTIVATE_PARTITION(part, ret);
745 dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
746 " = %ld, heartbeat_offline = %ld, HB_mask[0] = 0x%lx\n",
747 partid, remote_vars->heartbeat, part->last_heartbeat,
748 remote_vars->heartbeat_offline,
749 remote_vars->heartbeating_to_mask[0]);
751 if (((remote_vars->heartbeat == part->last_heartbeat) &&
752 (remote_vars->heartbeat_offline == 0)) ||
753 !xpc_hb_allowed(sn_partition_id,
754 &remote_vars->heartbeating_to_mask)) {
756 XPC_DEACTIVATE_PARTITION(part, xpNoHeartbeat);
760 part->last_heartbeat = remote_vars->heartbeat;
765 * Get a copy of the remote partition's XPC variables from the reserved page.
767 * remote_vars points to a buffer that is cacheline aligned for BTE copies and
768 * assumed to be of size XPC_RP_VARS_SIZE.
770 static enum xp_retval
771 xpc_get_remote_vars_sn2(unsigned long remote_vars_pa,
772 struct xpc_vars_sn2 *remote_vars)
776 if (remote_vars_pa == 0)
779 /* pull over the cross partition variables */
780 ret = xp_remote_memcpy(xp_pa(remote_vars), remote_vars_pa,
782 if (ret != xpSuccess)
785 if (XPC_VERSION_MAJOR(remote_vars->version) !=
786 XPC_VERSION_MAJOR(XPC_V_VERSION)) {
794 xpc_request_partition_activation_sn2(struct xpc_rsvd_page *remote_rp,
795 unsigned long remote_rp_pa, int nasid)
797 xpc_send_local_activate_IRQ_sn2(nasid);
801 xpc_request_partition_reactivation_sn2(struct xpc_partition *part)
803 xpc_send_local_activate_IRQ_sn2(part->sn.sn2.activate_IRQ_nasid);
807 xpc_request_partition_deactivation_sn2(struct xpc_partition *part)
809 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
810 unsigned long irq_flags;
811 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
812 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
813 sizeof(struct amo)));
815 local_irq_save(irq_flags);
817 /* set bit corresponding to our partid in remote partition's amo */
818 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
819 BIT(sn_partition_id));
822 * We must always use the nofault function regardless of whether we
823 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
824 * didn't, we'd never know that the other partition is down and would
825 * keep sending IRQs and amos to it until the heartbeat times out.
827 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
829 xp_nofault_PIOR_target));
831 local_irq_restore(irq_flags);
834 * Send activate IRQ to get other side to see that we've set our
835 * bit in their deactivate request amo.
837 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
839 part_sn2->activate_IRQ_nasid,
840 part_sn2->activate_IRQ_phys_cpuid);
844 xpc_cancel_partition_deactivation_request_sn2(struct xpc_partition *part)
846 unsigned long irq_flags;
847 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
848 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
849 sizeof(struct amo)));
851 local_irq_save(irq_flags);
853 /* clear bit corresponding to our partid in remote partition's amo */
854 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
855 ~BIT(sn_partition_id));
858 * We must always use the nofault function regardless of whether we
859 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
860 * didn't, we'd never know that the other partition is down and would
861 * keep sending IRQs and amos to it until the heartbeat times out.
863 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
865 xp_nofault_PIOR_target));
867 local_irq_restore(irq_flags);
871 xpc_partition_deactivation_requested_sn2(short partid)
873 struct amo *amo = xpc_vars_sn2->amos_page +
874 XPC_DEACTIVATE_REQUEST_AMO_SN2;
876 /* our partition's amo variable ANDed with partid mask */
877 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
882 * Update the remote partition's info.
885 xpc_update_partition_info_sn2(struct xpc_partition *part, u8 remote_rp_version,
886 unsigned long *remote_rp_ts_jiffies,
887 unsigned long remote_rp_pa,
888 unsigned long remote_vars_pa,
889 struct xpc_vars_sn2 *remote_vars)
891 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
893 part->remote_rp_version = remote_rp_version;
894 dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n",
895 part->remote_rp_version);
897 part->remote_rp_ts_jiffies = *remote_rp_ts_jiffies;
898 dev_dbg(xpc_part, " remote_rp_ts_jiffies = 0x%016lx\n",
899 part->remote_rp_ts_jiffies);
901 part->remote_rp_pa = remote_rp_pa;
902 dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
904 part_sn2->remote_vars_pa = remote_vars_pa;
905 dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n",
906 part_sn2->remote_vars_pa);
908 part->last_heartbeat = remote_vars->heartbeat;
909 dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n",
910 part->last_heartbeat);
912 part_sn2->remote_vars_part_pa = remote_vars->vars_part_pa;
913 dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n",
914 part_sn2->remote_vars_part_pa);
916 part_sn2->activate_IRQ_nasid = remote_vars->activate_IRQ_nasid;
917 dev_dbg(xpc_part, " activate_IRQ_nasid = 0x%x\n",
918 part_sn2->activate_IRQ_nasid);
920 part_sn2->activate_IRQ_phys_cpuid =
921 remote_vars->activate_IRQ_phys_cpuid;
922 dev_dbg(xpc_part, " activate_IRQ_phys_cpuid = 0x%x\n",
923 part_sn2->activate_IRQ_phys_cpuid);
925 part_sn2->remote_amos_page_pa = remote_vars->amos_page_pa;
926 dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n",
927 part_sn2->remote_amos_page_pa);
929 part_sn2->remote_vars_version = remote_vars->version;
930 dev_dbg(xpc_part, " remote_vars_version = 0x%x\n",
931 part_sn2->remote_vars_version);
935 * Prior code has determined the nasid which generated a activate IRQ.
936 * Inspect that nasid to determine if its partition needs to be activated
939 * A partition is considered "awaiting activation" if our partition
940 * flags indicate it is not active and it has a heartbeat. A
941 * partition is considered "awaiting deactivation" if our partition
942 * flags indicate it is active but it has no heartbeat or it is not
943 * sending its heartbeat to us.
945 * To determine the heartbeat, the remote nasid must have a properly
946 * initialized reserved page.
949 xpc_identify_activate_IRQ_req_sn2(int nasid)
951 struct xpc_rsvd_page *remote_rp;
952 struct xpc_vars_sn2 *remote_vars;
953 unsigned long remote_rp_pa;
954 unsigned long remote_vars_pa;
955 int remote_rp_version;
957 unsigned long remote_rp_ts_jiffies = 0;
959 struct xpc_partition *part;
960 struct xpc_partition_sn2 *part_sn2;
963 /* pull over the reserved page structure */
965 remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer_sn2;
967 ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
968 if (ret != xpSuccess) {
969 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
970 "which sent interrupt, reason=%d\n", nasid, ret);
974 remote_vars_pa = remote_rp->sn.vars_pa;
975 remote_rp_version = remote_rp->version;
976 remote_rp_ts_jiffies = remote_rp->ts_jiffies;
978 partid = remote_rp->SAL_partid;
979 part = &xpc_partitions[partid];
980 part_sn2 = &part->sn.sn2;
982 /* pull over the cross partition variables */
984 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
986 ret = xpc_get_remote_vars_sn2(remote_vars_pa, remote_vars);
987 if (ret != xpSuccess) {
988 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
989 "which sent interrupt, reason=%d\n", nasid, ret);
991 XPC_DEACTIVATE_PARTITION(part, ret);
995 part->activate_IRQ_rcvd++;
997 dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
998 "%ld:0x%lx\n", (int)nasid, (int)partid, part->activate_IRQ_rcvd,
999 remote_vars->heartbeat, remote_vars->heartbeating_to_mask[0]);
1001 if (xpc_partition_disengaged(part) &&
1002 part->act_state == XPC_P_INACTIVE) {
1004 xpc_update_partition_info_sn2(part, remote_rp_version,
1005 &remote_rp_ts_jiffies,
1006 remote_rp_pa, remote_vars_pa,
1009 if (xpc_partition_deactivation_requested_sn2(partid)) {
1011 * Other side is waiting on us to deactivate even though
1017 xpc_activate_partition(part);
1021 DBUG_ON(part->remote_rp_version == 0);
1022 DBUG_ON(part_sn2->remote_vars_version == 0);
1024 if (remote_rp_ts_jiffies != part->remote_rp_ts_jiffies) {
1026 /* the other side rebooted */
1028 DBUG_ON(xpc_partition_engaged_sn2(partid));
1029 DBUG_ON(xpc_partition_deactivation_requested_sn2(partid));
1031 xpc_update_partition_info_sn2(part, remote_rp_version,
1032 &remote_rp_ts_jiffies,
1033 remote_rp_pa, remote_vars_pa,
1038 if (part->disengage_timeout > 0 && !xpc_partition_disengaged(part)) {
1039 /* still waiting on other side to disengage from us */
1044 XPC_DEACTIVATE_PARTITION(part, xpReactivating);
1045 else if (xpc_partition_deactivation_requested_sn2(partid))
1046 XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown);
1050 * Loop through the activation amo variables and process any bits
1051 * which are set. Each bit indicates a nasid sending a partition
1052 * activation or deactivation request.
1054 * Return #of IRQs detected.
1057 xpc_identify_activate_IRQ_sender_sn2(void)
1061 unsigned long nasid_mask_long;
1062 u64 nasid; /* remote nasid */
1063 int n_IRQs_detected = 0;
1064 struct amo *act_amos;
1066 act_amos = xpc_vars_sn2->amos_page + XPC_ACTIVATE_IRQ_AMOS_SN2;
1068 /* scan through activate amo variables looking for non-zero entries */
1069 for (l = 0; l < xpc_nasid_mask_nlongs; l++) {
1074 nasid_mask_long = xpc_receive_IRQ_amo_sn2(&act_amos[l]);
1076 b = find_first_bit(&nasid_mask_long, BITS_PER_LONG);
1077 if (b >= BITS_PER_LONG) {
1078 /* no IRQs from nasids in this amo variable */
1082 dev_dbg(xpc_part, "amo[%d] gave back 0x%lx\n", l,
1086 * If this nasid has been added to the machine since
1087 * our partition was reset, this will retain the
1088 * remote nasid in our reserved pages machine mask.
1089 * This is used in the event of module reload.
1091 xpc_mach_nasids[l] |= nasid_mask_long;
1093 /* locate the nasid(s) which sent interrupts */
1097 nasid = (l * BITS_PER_LONG + b) * 2;
1098 dev_dbg(xpc_part, "interrupt from nasid %ld\n", nasid);
1099 xpc_identify_activate_IRQ_req_sn2(nasid);
1101 b = find_next_bit(&nasid_mask_long, BITS_PER_LONG,
1103 } while (b < BITS_PER_LONG);
1105 return n_IRQs_detected;
1109 xpc_process_activate_IRQ_rcvd_sn2(int n_IRQs_expected)
1111 int n_IRQs_detected;
1113 n_IRQs_detected = xpc_identify_activate_IRQ_sender_sn2();
1114 if (n_IRQs_detected < n_IRQs_expected) {
1115 /* retry once to help avoid missing amo */
1116 (void)xpc_identify_activate_IRQ_sender_sn2();
1121 * Guarantee that the kzalloc'd memory is cacheline aligned.
1124 xpc_kzalloc_cacheline_aligned_sn2(size_t size, gfp_t flags, void **base)
1126 /* see if kzalloc will give us cachline aligned memory by default */
1127 *base = kzalloc(size, flags);
1131 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
1136 /* nope, we'll have to do it ourselves */
1137 *base = kzalloc(size + L1_CACHE_BYTES, flags);
1141 return (void *)L1_CACHE_ALIGN((u64)*base);
1145 * Setup the infrastructure necessary to support XPartition Communication
1146 * between the specified remote partition and the local one.
1148 static enum xp_retval
1149 xpc_setup_infrastructure_sn2(struct xpc_partition *part)
1151 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1152 enum xp_retval retval;
1156 struct xpc_channel *ch;
1157 struct timer_list *timer;
1158 short partid = XPC_PARTID(part);
1161 * Allocate all of the channel structures as a contiguous chunk of
1164 DBUG_ON(part->channels != NULL);
1165 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
1167 if (part->channels == NULL) {
1168 dev_err(xpc_chan, "can't get memory for channels\n");
1172 /* allocate all the required GET/PUT values */
1174 part_sn2->local_GPs =
1175 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1176 &part_sn2->local_GPs_base);
1177 if (part_sn2->local_GPs == NULL) {
1178 dev_err(xpc_chan, "can't get memory for local get/put "
1180 retval = xpNoMemory;
1184 part_sn2->remote_GPs =
1185 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1186 &part_sn2->remote_GPs_base);
1187 if (part_sn2->remote_GPs == NULL) {
1188 dev_err(xpc_chan, "can't get memory for remote get/put "
1190 retval = xpNoMemory;
1194 part_sn2->remote_GPs_pa = 0;
1196 /* allocate all the required open and close args */
1198 part->local_openclose_args =
1199 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1201 &part->local_openclose_args_base);
1202 if (part->local_openclose_args == NULL) {
1203 dev_err(xpc_chan, "can't get memory for local connect args\n");
1204 retval = xpNoMemory;
1208 part->remote_openclose_args =
1209 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1211 &part->remote_openclose_args_base);
1212 if (part->remote_openclose_args == NULL) {
1213 dev_err(xpc_chan, "can't get memory for remote connect args\n");
1214 retval = xpNoMemory;
1218 part_sn2->remote_openclose_args_pa = 0;
1220 part_sn2->local_chctl_amo_va = xpc_init_IRQ_amo_sn2(partid);
1221 part->chctl.all_flags = 0;
1222 spin_lock_init(&part->chctl_lock);
1224 part_sn2->notify_IRQ_nasid = 0;
1225 part_sn2->notify_IRQ_phys_cpuid = 0;
1226 part_sn2->remote_chctl_amo_va = NULL;
1228 atomic_set(&part->channel_mgr_requests, 1);
1229 init_waitqueue_head(&part->channel_mgr_wq);
1231 sprintf(part_sn2->notify_IRQ_owner, "xpc%02d", partid);
1232 ret = request_irq(SGI_XPC_NOTIFY, xpc_handle_notify_IRQ_sn2,
1233 IRQF_SHARED, part_sn2->notify_IRQ_owner,
1234 (void *)(u64)partid);
1236 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
1237 "errno=%d\n", -ret);
1238 retval = xpLackOfResources;
1242 /* Setup a timer to check for dropped notify IRQs */
1243 timer = &part_sn2->dropped_notify_IRQ_timer;
1246 (void (*)(unsigned long))xpc_check_for_dropped_notify_IRQ_sn2;
1247 timer->data = (unsigned long)part;
1248 timer->expires = jiffies + XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
1251 part->nchannels = XPC_MAX_NCHANNELS;
1253 atomic_set(&part->nchannels_active, 0);
1254 atomic_set(&part->nchannels_engaged, 0);
1256 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1257 ch = &part->channels[ch_number];
1259 ch->partid = partid;
1260 ch->number = ch_number;
1261 ch->flags = XPC_C_DISCONNECTED;
1263 ch->sn.sn2.local_GP = &part_sn2->local_GPs[ch_number];
1264 ch->local_openclose_args =
1265 &part->local_openclose_args[ch_number];
1267 atomic_set(&ch->kthreads_assigned, 0);
1268 atomic_set(&ch->kthreads_idle, 0);
1269 atomic_set(&ch->kthreads_active, 0);
1271 atomic_set(&ch->references, 0);
1272 atomic_set(&ch->n_to_notify, 0);
1274 spin_lock_init(&ch->lock);
1275 mutex_init(&ch->sn.sn2.msg_to_pull_mutex);
1276 init_completion(&ch->wdisconnect_wait);
1278 atomic_set(&ch->n_on_msg_allocate_wq, 0);
1279 init_waitqueue_head(&ch->msg_allocate_wq);
1280 init_waitqueue_head(&ch->idle_wq);
1284 * With the setting of the partition setup_state to XPC_P_SETUP, we're
1285 * declaring that this partition is ready to go.
1287 part->setup_state = XPC_P_SETUP;
1290 * Setup the per partition specific variables required by the
1291 * remote partition to establish channel connections with us.
1293 * The setting of the magic # indicates that these per partition
1294 * specific variables are ready to be used.
1296 xpc_vars_part_sn2[partid].GPs_pa = xp_pa(part_sn2->local_GPs);
1297 xpc_vars_part_sn2[partid].openclose_args_pa =
1298 xp_pa(part->local_openclose_args);
1299 xpc_vars_part_sn2[partid].chctl_amo_pa =
1300 xp_pa(part_sn2->local_chctl_amo_va);
1301 cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
1302 xpc_vars_part_sn2[partid].notify_IRQ_nasid = cpuid_to_nasid(cpuid);
1303 xpc_vars_part_sn2[partid].notify_IRQ_phys_cpuid =
1304 cpu_physical_id(cpuid);
1305 xpc_vars_part_sn2[partid].nchannels = part->nchannels;
1306 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC1;
1310 /* setup of infrastructure failed */
1312 kfree(part->remote_openclose_args_base);
1313 part->remote_openclose_args = NULL;
1315 kfree(part->local_openclose_args_base);
1316 part->local_openclose_args = NULL;
1318 kfree(part_sn2->remote_GPs_base);
1319 part_sn2->remote_GPs = NULL;
1321 kfree(part_sn2->local_GPs_base);
1322 part_sn2->local_GPs = NULL;
1324 kfree(part->channels);
1325 part->channels = NULL;
1330 * Teardown the infrastructure necessary to support XPartition Communication
1331 * between the specified remote partition and the local one.
1334 xpc_teardown_infrastructure_sn2(struct xpc_partition *part)
1336 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1337 short partid = XPC_PARTID(part);
1340 * We start off by making this partition inaccessible to local
1341 * processes by marking it as no longer setup. Then we make it
1342 * inaccessible to remote processes by clearing the XPC per partition
1343 * specific variable's magic # (which indicates that these variables
1344 * are no longer valid) and by ignoring all XPC notify IRQs sent to
1348 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1349 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1350 DBUG_ON(part->setup_state != XPC_P_SETUP);
1351 part->setup_state = XPC_P_WTEARDOWN;
1353 xpc_vars_part_sn2[partid].magic = 0;
1355 free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1358 * Before proceeding with the teardown we have to wait until all
1359 * existing references cease.
1361 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1363 /* now we can begin tearing down the infrastructure */
1365 part->setup_state = XPC_P_TORNDOWN;
1367 /* in case we've still got outstanding timers registered... */
1368 del_timer_sync(&part_sn2->dropped_notify_IRQ_timer);
1370 kfree(part->remote_openclose_args_base);
1371 part->remote_openclose_args = NULL;
1372 kfree(part->local_openclose_args_base);
1373 part->local_openclose_args = NULL;
1374 kfree(part_sn2->remote_GPs_base);
1375 part_sn2->remote_GPs = NULL;
1376 kfree(part_sn2->local_GPs_base);
1377 part_sn2->local_GPs = NULL;
1378 kfree(part->channels);
1379 part->channels = NULL;
1380 part_sn2->local_chctl_amo_va = NULL;
1384 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
1385 * (or multiple cachelines) from a remote partition.
1387 * src_pa must be a cacheline aligned physical address on the remote partition.
1388 * dst must be a cacheline aligned virtual address on this partition.
1389 * cnt must be cacheline sized
1391 /* ??? Replace this function by call to xp_remote_memcpy() or bte_copy()? */
1392 static enum xp_retval
1393 xpc_pull_remote_cachelines_sn2(struct xpc_partition *part, void *dst,
1394 const unsigned long src_pa, size_t cnt)
1398 DBUG_ON(src_pa != L1_CACHE_ALIGN(src_pa));
1399 DBUG_ON((unsigned long)dst != L1_CACHE_ALIGN((unsigned long)dst));
1400 DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
1402 if (part->act_state == XPC_P_DEACTIVATING)
1403 return part->reason;
1405 ret = xp_remote_memcpy(xp_pa(dst), src_pa, cnt);
1406 if (ret != xpSuccess) {
1407 dev_dbg(xpc_chan, "xp_remote_memcpy() from partition %d failed,"
1408 " ret=%d\n", XPC_PARTID(part), ret);
1414 * Pull the remote per partition specific variables from the specified
1417 static enum xp_retval
1418 xpc_pull_remote_vars_part_sn2(struct xpc_partition *part)
1420 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1421 u8 buffer[L1_CACHE_BYTES * 2];
1422 struct xpc_vars_part_sn2 *pulled_entry_cacheline =
1423 (struct xpc_vars_part_sn2 *)L1_CACHE_ALIGN((u64)buffer);
1424 struct xpc_vars_part_sn2 *pulled_entry;
1425 unsigned long remote_entry_cacheline_pa;
1426 unsigned long remote_entry_pa;
1427 short partid = XPC_PARTID(part);
1430 /* pull the cacheline that contains the variables we're interested in */
1432 DBUG_ON(part_sn2->remote_vars_part_pa !=
1433 L1_CACHE_ALIGN(part_sn2->remote_vars_part_pa));
1434 DBUG_ON(sizeof(struct xpc_vars_part_sn2) != L1_CACHE_BYTES / 2);
1436 remote_entry_pa = part_sn2->remote_vars_part_pa +
1437 sn_partition_id * sizeof(struct xpc_vars_part_sn2);
1439 remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
1441 pulled_entry = (struct xpc_vars_part_sn2 *)((u64)pulled_entry_cacheline
1442 + (remote_entry_pa &
1443 (L1_CACHE_BYTES - 1)));
1445 ret = xpc_pull_remote_cachelines_sn2(part, pulled_entry_cacheline,
1446 remote_entry_cacheline_pa,
1448 if (ret != xpSuccess) {
1449 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
1450 "partition %d, ret=%d\n", partid, ret);
1454 /* see if they've been set up yet */
1456 if (pulled_entry->magic != XPC_VP_MAGIC1 &&
1457 pulled_entry->magic != XPC_VP_MAGIC2) {
1459 if (pulled_entry->magic != 0) {
1460 dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
1461 "partition %d has bad magic value (=0x%lx)\n",
1462 partid, sn_partition_id, pulled_entry->magic);
1466 /* they've not been initialized yet */
1470 if (xpc_vars_part_sn2[partid].magic == XPC_VP_MAGIC1) {
1472 /* validate the variables */
1474 if (pulled_entry->GPs_pa == 0 ||
1475 pulled_entry->openclose_args_pa == 0 ||
1476 pulled_entry->chctl_amo_pa == 0) {
1478 dev_err(xpc_chan, "partition %d's XPC vars_part for "
1479 "partition %d are not valid\n", partid,
1481 return xpInvalidAddress;
1484 /* the variables we imported look to be valid */
1486 part_sn2->remote_GPs_pa = pulled_entry->GPs_pa;
1487 part_sn2->remote_openclose_args_pa =
1488 pulled_entry->openclose_args_pa;
1489 part_sn2->remote_chctl_amo_va =
1490 (struct amo *)__va(pulled_entry->chctl_amo_pa);
1491 part_sn2->notify_IRQ_nasid = pulled_entry->notify_IRQ_nasid;
1492 part_sn2->notify_IRQ_phys_cpuid =
1493 pulled_entry->notify_IRQ_phys_cpuid;
1495 if (part->nchannels > pulled_entry->nchannels)
1496 part->nchannels = pulled_entry->nchannels;
1498 /* let the other side know that we've pulled their variables */
1500 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC2;
1503 if (pulled_entry->magic == XPC_VP_MAGIC1)
1510 * Establish first contact with the remote partititon. This involves pulling
1511 * the XPC per partition variables from the remote partition and waiting for
1512 * the remote partition to pull ours.
1514 static enum xp_retval
1515 xpc_make_first_contact_sn2(struct xpc_partition *part)
1517 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1521 * Register the remote partition's amos with SAL so it can handle
1522 * and cleanup errors within that address range should the remote
1523 * partition go down. We don't unregister this range because it is
1524 * difficult to tell when outstanding writes to the remote partition
1525 * are finished and thus when it is safe to unregister. This should
1526 * not result in wasted space in the SAL xp_addr_region table because
1527 * we should get the same page for remote_amos_page_pa after module
1528 * reloads and system reboots.
1530 if (sn_register_xp_addr_region(part_sn2->remote_amos_page_pa,
1531 PAGE_SIZE, 1) < 0) {
1532 dev_warn(xpc_part, "xpc_activating(%d) failed to register "
1533 "xp_addr region\n", XPC_PARTID(part));
1535 ret = xpPhysAddrRegFailed;
1536 XPC_DEACTIVATE_PARTITION(part, ret);
1541 * Send activate IRQ to get other side to activate if they've not
1542 * already begun to do so.
1544 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
1545 cnodeid_to_nasid(0),
1546 part_sn2->activate_IRQ_nasid,
1547 part_sn2->activate_IRQ_phys_cpuid);
1549 while ((ret = xpc_pull_remote_vars_part_sn2(part)) != xpSuccess) {
1550 if (ret != xpRetry) {
1551 XPC_DEACTIVATE_PARTITION(part, ret);
1555 dev_dbg(xpc_part, "waiting to make first contact with "
1556 "partition %d\n", XPC_PARTID(part));
1558 /* wait a 1/4 of a second or so */
1559 (void)msleep_interruptible(250);
1561 if (part->act_state == XPC_P_DEACTIVATING)
1562 return part->reason;
1569 * Get the chctl flags and pull the openclose args and/or remote GPs as needed.
1572 xpc_get_chctl_all_flags_sn2(struct xpc_partition *part)
1574 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1575 unsigned long irq_flags;
1576 union xpc_channel_ctl_flags chctl;
1580 * See if there are any chctl flags to be handled.
1583 spin_lock_irqsave(&part->chctl_lock, irq_flags);
1584 chctl = part->chctl;
1585 if (chctl.all_flags != 0)
1586 part->chctl.all_flags = 0;
1588 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
1590 if (xpc_any_openclose_chctl_flags_set(&chctl)) {
1591 ret = xpc_pull_remote_cachelines_sn2(part, part->
1592 remote_openclose_args,
1594 remote_openclose_args_pa,
1595 XPC_OPENCLOSE_ARGS_SIZE);
1596 if (ret != xpSuccess) {
1597 XPC_DEACTIVATE_PARTITION(part, ret);
1599 dev_dbg(xpc_chan, "failed to pull openclose args from "
1600 "partition %d, ret=%d\n", XPC_PARTID(part),
1603 /* don't bother processing chctl flags anymore */
1604 chctl.all_flags = 0;
1608 if (xpc_any_msg_chctl_flags_set(&chctl)) {
1609 ret = xpc_pull_remote_cachelines_sn2(part, part_sn2->remote_GPs,
1610 part_sn2->remote_GPs_pa,
1612 if (ret != xpSuccess) {
1613 XPC_DEACTIVATE_PARTITION(part, ret);
1615 dev_dbg(xpc_chan, "failed to pull GPs from partition "
1616 "%d, ret=%d\n", XPC_PARTID(part), ret);
1618 /* don't bother processing chctl flags anymore */
1619 chctl.all_flags = 0;
1623 return chctl.all_flags;
1627 * Allocate the local message queue and the notify queue.
1629 static enum xp_retval
1630 xpc_allocate_local_msgqueue_sn2(struct xpc_channel *ch)
1632 unsigned long irq_flags;
1636 for (nentries = ch->local_nentries; nentries > 0; nentries--) {
1638 nbytes = nentries * ch->msg_size;
1639 ch->local_msgqueue =
1640 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1641 &ch->local_msgqueue_base);
1642 if (ch->local_msgqueue == NULL)
1645 nbytes = nentries * sizeof(struct xpc_notify);
1646 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
1647 if (ch->notify_queue == NULL) {
1648 kfree(ch->local_msgqueue_base);
1649 ch->local_msgqueue = NULL;
1653 spin_lock_irqsave(&ch->lock, irq_flags);
1654 if (nentries < ch->local_nentries) {
1655 dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
1656 "partid=%d, channel=%d\n", nentries,
1657 ch->local_nentries, ch->partid, ch->number);
1659 ch->local_nentries = nentries;
1661 spin_unlock_irqrestore(&ch->lock, irq_flags);
1665 dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
1666 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
1671 * Allocate the cached remote message queue.
1673 static enum xp_retval
1674 xpc_allocate_remote_msgqueue_sn2(struct xpc_channel *ch)
1676 unsigned long irq_flags;
1680 DBUG_ON(ch->remote_nentries <= 0);
1682 for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
1684 nbytes = nentries * ch->msg_size;
1685 ch->remote_msgqueue =
1686 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1687 &ch->remote_msgqueue_base);
1688 if (ch->remote_msgqueue == NULL)
1691 spin_lock_irqsave(&ch->lock, irq_flags);
1692 if (nentries < ch->remote_nentries) {
1693 dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
1694 "partid=%d, channel=%d\n", nentries,
1695 ch->remote_nentries, ch->partid, ch->number);
1697 ch->remote_nentries = nentries;
1699 spin_unlock_irqrestore(&ch->lock, irq_flags);
1703 dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
1704 "partid=%d, channel=%d\n", ch->partid, ch->number);
1709 * Allocate message queues and other stuff associated with a channel.
1711 * Note: Assumes all of the channel sizes are filled in.
1713 static enum xp_retval
1714 xpc_allocate_msgqueues_sn2(struct xpc_channel *ch)
1718 DBUG_ON(ch->flags & XPC_C_SETUP);
1720 ret = xpc_allocate_local_msgqueue_sn2(ch);
1721 if (ret == xpSuccess) {
1723 ret = xpc_allocate_remote_msgqueue_sn2(ch);
1724 if (ret != xpSuccess) {
1725 kfree(ch->local_msgqueue_base);
1726 ch->local_msgqueue = NULL;
1727 kfree(ch->notify_queue);
1728 ch->notify_queue = NULL;
1735 * Free up message queues and other stuff that were allocated for the specified
1738 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
1739 * they're cleared when XPC_C_DISCONNECTED is cleared.
1742 xpc_free_msgqueues_sn2(struct xpc_channel *ch)
1744 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1746 DBUG_ON(!spin_is_locked(&ch->lock));
1747 DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
1749 ch->remote_msgqueue_pa = 0;
1753 ch->local_nentries = 0;
1754 ch->remote_nentries = 0;
1755 ch->kthreads_assigned_limit = 0;
1756 ch->kthreads_idle_limit = 0;
1758 ch_sn2->local_GP->get = 0;
1759 ch_sn2->local_GP->put = 0;
1760 ch_sn2->remote_GP.get = 0;
1761 ch_sn2->remote_GP.put = 0;
1762 ch_sn2->w_local_GP.get = 0;
1763 ch_sn2->w_local_GP.put = 0;
1764 ch_sn2->w_remote_GP.get = 0;
1765 ch_sn2->w_remote_GP.put = 0;
1766 ch_sn2->next_msg_to_pull = 0;
1768 if (ch->flags & XPC_C_SETUP) {
1769 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
1770 ch->flags, ch->partid, ch->number);
1772 kfree(ch->local_msgqueue_base);
1773 ch->local_msgqueue = NULL;
1774 kfree(ch->remote_msgqueue_base);
1775 ch->remote_msgqueue = NULL;
1776 kfree(ch->notify_queue);
1777 ch->notify_queue = NULL;
1782 * Notify those who wanted to be notified upon delivery of their message.
1785 xpc_notify_senders_sn2(struct xpc_channel *ch, enum xp_retval reason, s64 put)
1787 struct xpc_notify *notify;
1789 s64 get = ch->sn.sn2.w_remote_GP.get - 1;
1791 while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
1793 notify = &ch->notify_queue[get % ch->local_nentries];
1796 * See if the notify entry indicates it was associated with
1797 * a message who's sender wants to be notified. It is possible
1798 * that it is, but someone else is doing or has done the
1801 notify_type = notify->type;
1802 if (notify_type == 0 ||
1803 cmpxchg(¬ify->type, notify_type, 0) != notify_type) {
1807 DBUG_ON(notify_type != XPC_N_CALL);
1809 atomic_dec(&ch->n_to_notify);
1811 if (notify->func != NULL) {
1812 dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
1813 "msg_number=%ld, partid=%d, channel=%d\n",
1814 (void *)notify, get, ch->partid, ch->number);
1816 notify->func(reason, ch->partid, ch->number,
1819 dev_dbg(xpc_chan, "notify->func() returned, "
1820 "notify=0x%p, msg_number=%ld, partid=%d, "
1821 "channel=%d\n", (void *)notify, get,
1822 ch->partid, ch->number);
1828 xpc_notify_senders_of_disconnect_sn2(struct xpc_channel *ch)
1830 xpc_notify_senders_sn2(ch, ch->reason, ch->sn.sn2.w_local_GP.put);
1834 * Clear some of the msg flags in the local message queue.
1837 xpc_clear_local_msgqueue_flags_sn2(struct xpc_channel *ch)
1839 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1840 struct xpc_msg *msg;
1843 get = ch_sn2->w_remote_GP.get;
1845 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1846 (get % ch->local_nentries) *
1849 } while (++get < ch_sn2->remote_GP.get);
1853 * Clear some of the msg flags in the remote message queue.
1856 xpc_clear_remote_msgqueue_flags_sn2(struct xpc_channel *ch)
1858 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1859 struct xpc_msg *msg;
1862 put = ch_sn2->w_remote_GP.put;
1864 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1865 (put % ch->remote_nentries) *
1868 } while (++put < ch_sn2->remote_GP.put);
1872 xpc_process_msg_chctl_flags_sn2(struct xpc_partition *part, int ch_number)
1874 struct xpc_channel *ch = &part->channels[ch_number];
1875 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1878 ch_sn2->remote_GP = part->sn.sn2.remote_GPs[ch_number];
1880 /* See what, if anything, has changed for each connected channel */
1882 xpc_msgqueue_ref(ch);
1884 if (ch_sn2->w_remote_GP.get == ch_sn2->remote_GP.get &&
1885 ch_sn2->w_remote_GP.put == ch_sn2->remote_GP.put) {
1886 /* nothing changed since GPs were last pulled */
1887 xpc_msgqueue_deref(ch);
1891 if (!(ch->flags & XPC_C_CONNECTED)) {
1892 xpc_msgqueue_deref(ch);
1897 * First check to see if messages recently sent by us have been
1898 * received by the other side. (The remote GET value will have
1899 * changed since we last looked at it.)
1902 if (ch_sn2->w_remote_GP.get != ch_sn2->remote_GP.get) {
1905 * We need to notify any senders that want to be notified
1906 * that their sent messages have been received by their
1907 * intended recipients. We need to do this before updating
1908 * w_remote_GP.get so that we don't allocate the same message
1909 * queue entries prematurely (see xpc_allocate_msg()).
1911 if (atomic_read(&ch->n_to_notify) > 0) {
1913 * Notify senders that messages sent have been
1914 * received and delivered by the other side.
1916 xpc_notify_senders_sn2(ch, xpMsgDelivered,
1917 ch_sn2->remote_GP.get);
1921 * Clear msg->flags in previously sent messages, so that
1922 * they're ready for xpc_allocate_msg().
1924 xpc_clear_local_msgqueue_flags_sn2(ch);
1926 ch_sn2->w_remote_GP.get = ch_sn2->remote_GP.get;
1928 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1929 "channel=%d\n", ch_sn2->w_remote_GP.get, ch->partid,
1933 * If anyone was waiting for message queue entries to become
1934 * available, wake them up.
1936 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1937 wake_up(&ch->msg_allocate_wq);
1941 * Now check for newly sent messages by the other side. (The remote
1942 * PUT value will have changed since we last looked at it.)
1945 if (ch_sn2->w_remote_GP.put != ch_sn2->remote_GP.put) {
1947 * Clear msg->flags in previously received messages, so that
1948 * they're ready for xpc_get_deliverable_msg().
1950 xpc_clear_remote_msgqueue_flags_sn2(ch);
1952 ch_sn2->w_remote_GP.put = ch_sn2->remote_GP.put;
1954 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1955 "channel=%d\n", ch_sn2->w_remote_GP.put, ch->partid,
1958 nmsgs_sent = ch_sn2->w_remote_GP.put - ch_sn2->w_local_GP.get;
1959 if (nmsgs_sent > 0) {
1960 dev_dbg(xpc_chan, "msgs waiting to be copied and "
1961 "delivered=%d, partid=%d, channel=%d\n",
1962 nmsgs_sent, ch->partid, ch->number);
1964 if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1965 xpc_activate_kthreads(ch, nmsgs_sent);
1969 xpc_msgqueue_deref(ch);
1972 static struct xpc_msg *
1973 xpc_pull_remote_msg_sn2(struct xpc_channel *ch, s64 get)
1975 struct xpc_partition *part = &xpc_partitions[ch->partid];
1976 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1977 unsigned long remote_msg_pa;
1978 struct xpc_msg *msg;
1984 if (mutex_lock_interruptible(&ch_sn2->msg_to_pull_mutex) != 0) {
1985 /* we were interrupted by a signal */
1989 while (get >= ch_sn2->next_msg_to_pull) {
1991 /* pull as many messages as are ready and able to be pulled */
1993 msg_index = ch_sn2->next_msg_to_pull % ch->remote_nentries;
1995 DBUG_ON(ch_sn2->next_msg_to_pull >= ch_sn2->w_remote_GP.put);
1996 nmsgs = ch_sn2->w_remote_GP.put - ch_sn2->next_msg_to_pull;
1997 if (msg_index + nmsgs > ch->remote_nentries) {
1998 /* ignore the ones that wrap the msg queue for now */
1999 nmsgs = ch->remote_nentries - msg_index;
2002 msg_offset = msg_index * ch->msg_size;
2003 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2004 remote_msg_pa = ch->remote_msgqueue_pa + msg_offset;
2006 ret = xpc_pull_remote_cachelines_sn2(part, msg, remote_msg_pa,
2007 nmsgs * ch->msg_size);
2008 if (ret != xpSuccess) {
2010 dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2011 " msg %ld from partition %d, channel=%d, "
2012 "ret=%d\n", nmsgs, ch_sn2->next_msg_to_pull,
2013 ch->partid, ch->number, ret);
2015 XPC_DEACTIVATE_PARTITION(part, ret);
2017 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
2021 ch_sn2->next_msg_to_pull += nmsgs;
2024 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
2026 /* return the message we were looking for */
2027 msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2028 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2034 xpc_n_of_deliverable_msgs_sn2(struct xpc_channel *ch)
2036 return ch->sn.sn2.w_remote_GP.put - ch->sn.sn2.w_local_GP.get;
2040 * Get a message to be delivered.
2042 static struct xpc_msg *
2043 xpc_get_deliverable_msg_sn2(struct xpc_channel *ch)
2045 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2046 struct xpc_msg *msg = NULL;
2050 if (ch->flags & XPC_C_DISCONNECTING)
2053 get = ch_sn2->w_local_GP.get;
2054 rmb(); /* guarantee that .get loads before .put */
2055 if (get == ch_sn2->w_remote_GP.put)
2058 /* There are messages waiting to be pulled and delivered.
2059 * We need to try to secure one for ourselves. We'll do this
2060 * by trying to increment w_local_GP.get and hope that no one
2061 * else beats us to it. If they do, we'll we'll simply have
2062 * to try again for the next one.
2065 if (cmpxchg(&ch_sn2->w_local_GP.get, get, get + 1) == get) {
2066 /* we got the entry referenced by get */
2068 dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2069 "partid=%d, channel=%d\n", get + 1,
2070 ch->partid, ch->number);
2072 /* pull the message from the remote partition */
2074 msg = xpc_pull_remote_msg_sn2(ch, get);
2076 DBUG_ON(msg != NULL && msg->number != get);
2077 DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2078 DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2089 * Now we actually send the messages that are ready to be sent by advancing
2090 * the local message queue's Put value and then send a chctl msgrequest to the
2091 * recipient partition.
2094 xpc_send_msgs_sn2(struct xpc_channel *ch, s64 initial_put)
2096 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2097 struct xpc_msg *msg;
2098 s64 put = initial_put + 1;
2099 int send_msgrequest = 0;
2104 if (put == ch_sn2->w_local_GP.put)
2107 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2108 (put % ch->local_nentries) *
2111 if (!(msg->flags & XPC_M_READY))
2117 if (put == initial_put) {
2118 /* nothing's changed */
2122 if (cmpxchg_rel(&ch_sn2->local_GP->put, initial_put, put) !=
2124 /* someone else beat us to it */
2125 DBUG_ON(ch_sn2->local_GP->put < initial_put);
2129 /* we just set the new value of local_GP->put */
2131 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
2132 "channel=%d\n", put, ch->partid, ch->number);
2134 send_msgrequest = 1;
2137 * We need to ensure that the message referenced by
2138 * local_GP->put is not XPC_M_READY or that local_GP->put
2139 * equals w_local_GP.put, so we'll go have a look.
2144 if (send_msgrequest)
2145 xpc_send_chctl_msgrequest_sn2(ch);
2149 * Allocate an entry for a message from the message queue associated with the
2150 * specified channel.
2152 static enum xp_retval
2153 xpc_allocate_msg_sn2(struct xpc_channel *ch, u32 flags,
2154 struct xpc_msg **address_of_msg)
2156 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2157 struct xpc_msg *msg;
2162 * Get the next available message entry from the local message queue.
2163 * If none are available, we'll make sure that we grab the latest
2170 put = ch_sn2->w_local_GP.put;
2171 rmb(); /* guarantee that .put loads before .get */
2172 if (put - ch_sn2->w_remote_GP.get < ch->local_nentries) {
2174 /* There are available message entries. We need to try
2175 * to secure one for ourselves. We'll do this by trying
2176 * to increment w_local_GP.put as long as someone else
2177 * doesn't beat us to it. If they do, we'll have to
2180 if (cmpxchg(&ch_sn2->w_local_GP.put, put, put + 1) ==
2182 /* we got the entry referenced by put */
2185 continue; /* try again */
2189 * There aren't any available msg entries at this time.
2191 * In waiting for a message entry to become available,
2192 * we set a timeout in case the other side is not sending
2193 * completion interrupts. This lets us fake a notify IRQ
2194 * that will cause the notify IRQ handler to fetch the latest
2195 * GP values as if an interrupt was sent by the other side.
2197 if (ret == xpTimeout)
2198 xpc_send_chctl_local_msgrequest_sn2(ch);
2200 if (flags & XPC_NOWAIT)
2203 ret = xpc_allocate_msg_wait(ch);
2204 if (ret != xpInterrupted && ret != xpTimeout)
2208 /* get the message's address and initialize it */
2209 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2210 (put % ch->local_nentries) * ch->msg_size);
2212 DBUG_ON(msg->flags != 0);
2215 dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
2216 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
2217 (void *)msg, msg->number, ch->partid, ch->number);
2219 *address_of_msg = msg;
2224 * Common code that does the actual sending of the message by advancing the
2225 * local message queue's Put value and sends a chctl msgrequest to the
2226 * partition the message is being sent to.
2228 static enum xp_retval
2229 xpc_send_msg_sn2(struct xpc_channel *ch, u32 flags, void *payload,
2230 u16 payload_size, u8 notify_type, xpc_notify_func func,
2233 enum xp_retval ret = xpSuccess;
2234 struct xpc_msg *msg = msg;
2235 struct xpc_notify *notify = notify;
2239 DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
2241 if (XPC_MSG_SIZE(payload_size) > ch->msg_size)
2242 return xpPayloadTooBig;
2244 xpc_msgqueue_ref(ch);
2246 if (ch->flags & XPC_C_DISCONNECTING) {
2250 if (!(ch->flags & XPC_C_CONNECTED)) {
2251 ret = xpNotConnected;
2255 ret = xpc_allocate_msg_sn2(ch, flags, &msg);
2256 if (ret != xpSuccess)
2259 msg_number = msg->number;
2261 if (notify_type != 0) {
2263 * Tell the remote side to send an ACK interrupt when the
2264 * message has been delivered.
2266 msg->flags |= XPC_M_INTERRUPT;
2268 atomic_inc(&ch->n_to_notify);
2270 notify = &ch->notify_queue[msg_number % ch->local_nentries];
2271 notify->func = func;
2273 notify->type = notify_type;
2275 /* ??? Is a mb() needed here? */
2277 if (ch->flags & XPC_C_DISCONNECTING) {
2279 * An error occurred between our last error check and
2280 * this one. We will try to clear the type field from
2281 * the notify entry. If we succeed then
2282 * xpc_disconnect_channel() didn't already process
2285 if (cmpxchg(¬ify->type, notify_type, 0) ==
2287 atomic_dec(&ch->n_to_notify);
2294 memcpy(&msg->payload, payload, payload_size);
2296 msg->flags |= XPC_M_READY;
2299 * The preceding store of msg->flags must occur before the following
2300 * load of local_GP->put.
2304 /* see if the message is next in line to be sent, if so send it */
2306 put = ch->sn.sn2.local_GP->put;
2307 if (put == msg_number)
2308 xpc_send_msgs_sn2(ch, put);
2311 xpc_msgqueue_deref(ch);
2316 * Now we actually acknowledge the messages that have been delivered and ack'd
2317 * by advancing the cached remote message queue's Get value and if requested
2318 * send a chctl msgrequest to the message sender's partition.
2321 xpc_acknowledge_msgs_sn2(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2323 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2324 struct xpc_msg *msg;
2325 s64 get = initial_get + 1;
2326 int send_msgrequest = 0;
2331 if (get == ch_sn2->w_local_GP.get)
2334 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2335 (get % ch->remote_nentries) *
2338 if (!(msg->flags & XPC_M_DONE))
2341 msg_flags |= msg->flags;
2345 if (get == initial_get) {
2346 /* nothing's changed */
2350 if (cmpxchg_rel(&ch_sn2->local_GP->get, initial_get, get) !=
2352 /* someone else beat us to it */
2353 DBUG_ON(ch_sn2->local_GP->get <= initial_get);
2357 /* we just set the new value of local_GP->get */
2359 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2360 "channel=%d\n", get, ch->partid, ch->number);
2362 send_msgrequest = (msg_flags & XPC_M_INTERRUPT);
2365 * We need to ensure that the message referenced by
2366 * local_GP->get is not XPC_M_DONE or that local_GP->get
2367 * equals w_local_GP.get, so we'll go have a look.
2372 if (send_msgrequest)
2373 xpc_send_chctl_msgrequest_sn2(ch);
2377 xpc_received_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg)
2380 s64 msg_number = msg->number;
2382 dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2383 (void *)msg, msg_number, ch->partid, ch->number);
2385 DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2386 msg_number % ch->remote_nentries);
2387 DBUG_ON(msg->flags & XPC_M_DONE);
2389 msg->flags |= XPC_M_DONE;
2392 * The preceding store of msg->flags must occur before the following
2393 * load of local_GP->get.
2398 * See if this message is next in line to be acknowledged as having
2401 get = ch->sn.sn2.local_GP->get;
2402 if (get == msg_number)
2403 xpc_acknowledge_msgs_sn2(ch, get, msg->flags);
2412 xpc_get_partition_rsvd_page_pa = xpc_get_partition_rsvd_page_pa_sn2;
2413 xpc_rsvd_page_init = xpc_rsvd_page_init_sn2;
2414 xpc_increment_heartbeat = xpc_increment_heartbeat_sn2;
2415 xpc_offline_heartbeat = xpc_offline_heartbeat_sn2;
2416 xpc_online_heartbeat = xpc_online_heartbeat_sn2;
2417 xpc_heartbeat_init = xpc_heartbeat_init_sn2;
2418 xpc_heartbeat_exit = xpc_heartbeat_exit_sn2;
2419 xpc_check_remote_hb = xpc_check_remote_hb_sn2;
2421 xpc_request_partition_activation = xpc_request_partition_activation_sn2;
2422 xpc_request_partition_reactivation =
2423 xpc_request_partition_reactivation_sn2;
2424 xpc_request_partition_deactivation =
2425 xpc_request_partition_deactivation_sn2;
2426 xpc_cancel_partition_deactivation_request =
2427 xpc_cancel_partition_deactivation_request_sn2;
2429 xpc_process_activate_IRQ_rcvd = xpc_process_activate_IRQ_rcvd_sn2;
2430 xpc_setup_infrastructure = xpc_setup_infrastructure_sn2;
2431 xpc_teardown_infrastructure = xpc_teardown_infrastructure_sn2;
2432 xpc_make_first_contact = xpc_make_first_contact_sn2;
2433 xpc_get_chctl_all_flags = xpc_get_chctl_all_flags_sn2;
2434 xpc_allocate_msgqueues = xpc_allocate_msgqueues_sn2;
2435 xpc_free_msgqueues = xpc_free_msgqueues_sn2;
2436 xpc_notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_sn2;
2437 xpc_process_msg_chctl_flags = xpc_process_msg_chctl_flags_sn2;
2438 xpc_n_of_deliverable_msgs = xpc_n_of_deliverable_msgs_sn2;
2439 xpc_get_deliverable_msg = xpc_get_deliverable_msg_sn2;
2441 xpc_indicate_partition_engaged = xpc_indicate_partition_engaged_sn2;
2442 xpc_partition_engaged = xpc_partition_engaged_sn2;
2443 xpc_any_partition_engaged = xpc_any_partition_engaged_sn2;
2444 xpc_indicate_partition_disengaged =
2445 xpc_indicate_partition_disengaged_sn2;
2446 xpc_assume_partition_disengaged = xpc_assume_partition_disengaged_sn2;
2448 xpc_send_chctl_closerequest = xpc_send_chctl_closerequest_sn2;
2449 xpc_send_chctl_closereply = xpc_send_chctl_closereply_sn2;
2450 xpc_send_chctl_openrequest = xpc_send_chctl_openrequest_sn2;
2451 xpc_send_chctl_openreply = xpc_send_chctl_openreply_sn2;
2453 xpc_send_msg = xpc_send_msg_sn2;
2454 xpc_received_msg = xpc_received_msg_sn2;
2456 buf_size = max(XPC_RP_VARS_SIZE,
2457 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES_SN2);
2458 xpc_remote_copy_buffer_sn2 = xpc_kmalloc_cacheline_aligned(buf_size,
2460 &xpc_remote_copy_buffer_base_sn2);
2461 if (xpc_remote_copy_buffer_sn2 == NULL) {
2462 dev_err(xpc_part, "can't get memory for remote copy buffer\n");
2466 /* open up protections for IPI and [potentially] amo operations */
2467 xpc_allow_IPI_ops_sn2();
2468 xpc_allow_amo_ops_shub_wars_1_1_sn2();
2471 * This is safe to do before the xpc_hb_checker thread has started
2472 * because the handler releases a wait queue. If an interrupt is
2473 * received before the thread is waiting, it will not go to sleep,
2474 * but rather immediately process the interrupt.
2476 ret = request_irq(SGI_XPC_ACTIVATE, xpc_handle_activate_IRQ_sn2, 0,
2479 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
2480 "errno=%d\n", -ret);
2481 xpc_disallow_IPI_ops_sn2();
2482 kfree(xpc_remote_copy_buffer_base_sn2);
2490 free_irq(SGI_XPC_ACTIVATE, NULL);
2491 xpc_disallow_IPI_ops_sn2();
2492 kfree(xpc_remote_copy_buffer_base_sn2);