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/kernel.h>
17 #include <linux/delay.h>
18 #include <asm/uncached.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(u64 amos_page_pa, int from_nasid, int to_nasid,
213 int w_index = XPC_NASID_W_INDEX(from_nasid);
214 int b_index = XPC_NASID_B_INDEX(from_nasid);
215 struct amo *amos = (struct amo *)__va(amos_page_pa +
216 (XPC_ACTIVATE_IRQ_AMOS_SN2 *
217 sizeof(struct amo)));
219 (void)xpc_send_IRQ_sn2(&amos[w_index], (1UL << b_index), to_nasid,
220 to_phys_cpuid, SGI_XPC_ACTIVATE);
224 xpc_send_local_activate_IRQ_sn2(int from_nasid)
226 int w_index = XPC_NASID_W_INDEX(from_nasid);
227 int b_index = XPC_NASID_B_INDEX(from_nasid);
228 struct amo *amos = (struct amo *)__va(xpc_vars_sn2->amos_page_pa +
229 (XPC_ACTIVATE_IRQ_AMOS_SN2 *
230 sizeof(struct amo)));
232 /* fake the sending and receipt of an activate IRQ from remote nasid */
233 FETCHOP_STORE_OP(TO_AMO((u64)&amos[w_index].variable), FETCHOP_OR,
235 atomic_inc(&xpc_activate_IRQ_rcvd);
236 wake_up_interruptible(&xpc_activate_IRQ_wq);
240 * Functions associated with SGI_XPC_NOTIFY IRQ.
244 * Check to see if any chctl flags were sent from the specified partition.
247 xpc_check_for_sent_chctl_flags_sn2(struct xpc_partition *part)
249 union xpc_channel_ctl_flags chctl;
250 unsigned long irq_flags;
252 chctl.all_flags = xpc_receive_IRQ_amo_sn2(part->sn.sn2.
254 if (chctl.all_flags == 0)
257 spin_lock_irqsave(&part->chctl_lock, irq_flags);
258 part->chctl.all_flags |= chctl.all_flags;
259 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
261 dev_dbg(xpc_chan, "received notify IRQ from partid=%d, chctl.all_flags="
262 "0x%lx\n", XPC_PARTID(part), chctl.all_flags);
264 xpc_wakeup_channel_mgr(part);
268 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
269 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
270 * than one partition, we use an amo structure per partition to indicate
271 * whether a partition has sent an IRQ or not. If it has, then wake up the
272 * associated kthread to handle it.
274 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IRQs sent by XPC
275 * running on other partitions.
277 * Noteworthy Arguments:
279 * irq - Interrupt ReQuest number. NOT USED.
281 * dev_id - partid of IRQ's potential sender.
284 xpc_handle_notify_IRQ_sn2(int irq, void *dev_id)
286 short partid = (short)(u64)dev_id;
287 struct xpc_partition *part = &xpc_partitions[partid];
289 DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
291 if (xpc_part_ref(part)) {
292 xpc_check_for_sent_chctl_flags_sn2(part);
294 xpc_part_deref(part);
300 * Check to see if xpc_handle_notify_IRQ_sn2() dropped any IRQs on the floor
301 * because the write to their associated amo variable completed after the IRQ
305 xpc_check_for_dropped_notify_IRQ_sn2(struct xpc_partition *part)
307 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
309 if (xpc_part_ref(part)) {
310 xpc_check_for_sent_chctl_flags_sn2(part);
312 part_sn2->dropped_notify_IRQ_timer.expires = jiffies +
313 XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
314 add_timer(&part_sn2->dropped_notify_IRQ_timer);
315 xpc_part_deref(part);
320 * Send a notify IRQ to the remote partition that is associated with the
324 xpc_send_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
325 char *chctl_flag_string, unsigned long *irq_flags)
327 struct xpc_partition *part = &xpc_partitions[ch->partid];
328 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
329 union xpc_channel_ctl_flags chctl = { 0 };
332 if (likely(part->act_state != XPC_P_DEACTIVATING)) {
333 chctl.flags[ch->number] = chctl_flag;
334 ret = xpc_send_IRQ_sn2(part_sn2->remote_chctl_amo_va,
336 part_sn2->notify_IRQ_nasid,
337 part_sn2->notify_IRQ_phys_cpuid,
339 dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n",
340 chctl_flag_string, ch->partid, ch->number, ret);
341 if (unlikely(ret != xpSuccess)) {
342 if (irq_flags != NULL)
343 spin_unlock_irqrestore(&ch->lock, *irq_flags);
344 XPC_DEACTIVATE_PARTITION(part, ret);
345 if (irq_flags != NULL)
346 spin_lock_irqsave(&ch->lock, *irq_flags);
351 #define XPC_SEND_NOTIFY_IRQ_SN2(_ch, _ipi_f, _irq_f) \
352 xpc_send_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f, _irq_f)
355 * Make it look like the remote partition, which is associated with the
356 * specified channel, sent us a notify IRQ. This faked IRQ will be handled
357 * by xpc_check_for_dropped_notify_IRQ_sn2().
360 xpc_send_local_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
361 char *chctl_flag_string)
363 struct xpc_partition *part = &xpc_partitions[ch->partid];
364 union xpc_channel_ctl_flags chctl = { 0 };
366 chctl.flags[ch->number] = chctl_flag;
367 FETCHOP_STORE_OP(TO_AMO((u64)&part->sn.sn2.local_chctl_amo_va->
368 variable), FETCHOP_OR, chctl.all_flags);
369 dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n",
370 chctl_flag_string, ch->partid, ch->number);
373 #define XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(_ch, _ipi_f) \
374 xpc_send_local_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f)
377 xpc_send_chctl_closerequest_sn2(struct xpc_channel *ch,
378 unsigned long *irq_flags)
380 struct xpc_openclose_args *args = ch->local_openclose_args;
382 args->reason = ch->reason;
383 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREQUEST, irq_flags);
387 xpc_send_chctl_closereply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
389 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREPLY, irq_flags);
393 xpc_send_chctl_openrequest_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
395 struct xpc_openclose_args *args = ch->local_openclose_args;
397 args->msg_size = ch->msg_size;
398 args->local_nentries = ch->local_nentries;
399 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREQUEST, irq_flags);
403 xpc_send_chctl_openreply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
405 struct xpc_openclose_args *args = ch->local_openclose_args;
407 args->remote_nentries = ch->remote_nentries;
408 args->local_nentries = ch->local_nentries;
409 args->local_msgqueue_pa = __pa(ch->local_msgqueue);
410 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREPLY, irq_flags);
414 xpc_send_chctl_msgrequest_sn2(struct xpc_channel *ch)
416 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST, NULL);
420 xpc_send_chctl_local_msgrequest_sn2(struct xpc_channel *ch)
422 XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST);
426 * This next set of functions are used to keep track of when a partition is
427 * potentially engaged in accessing memory belonging to another partition.
431 xpc_indicate_partition_engaged_sn2(struct xpc_partition *part)
433 unsigned long irq_flags;
434 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
435 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
436 sizeof(struct amo)));
438 local_irq_save(irq_flags);
440 /* set bit corresponding to our partid in remote partition's amo */
441 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
442 (1UL << sn_partition_id));
444 * We must always use the nofault function regardless of whether we
445 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
446 * didn't, we'd never know that the other partition is down and would
447 * keep sending IRQs and amos to it until the heartbeat times out.
449 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
451 xp_nofault_PIOR_target));
453 local_irq_restore(irq_flags);
457 xpc_indicate_partition_disengaged_sn2(struct xpc_partition *part)
459 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
460 unsigned long irq_flags;
461 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
462 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
463 sizeof(struct amo)));
465 local_irq_save(irq_flags);
467 /* clear bit corresponding to our partid in remote partition's amo */
468 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
469 ~(1UL << 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) &
500 (1UL << partid)) != 0;
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_rsvd_page_init_sn2(struct xpc_rsvd_page *rp)
582 struct amo *amos_page;
586 xpc_vars_sn2 = XPC_RP_VARS(rp);
588 rp->sn.vars_pa = __pa(xpc_vars_sn2);
590 /* vars_part array follows immediately after vars */
591 xpc_vars_part_sn2 = (struct xpc_vars_part_sn2 *)((u8 *)XPC_RP_VARS(rp) +
595 * Before clearing xpc_vars_sn2, see if a page of amos had been
596 * previously allocated. If not we'll need to allocate one and set
597 * permissions so that cross-partition amos are allowed.
599 * The allocated amo page needs MCA reporting to remain disabled after
600 * XPC has unloaded. To make this work, we keep a copy of the pointer
601 * to this page (i.e., amos_page) in the struct xpc_vars_sn2 structure,
602 * which is pointed to by the reserved page, and re-use that saved copy
603 * on subsequent loads of XPC. This amo page is never freed, and its
604 * memory protections are never restricted.
606 amos_page = xpc_vars_sn2->amos_page;
607 if (amos_page == NULL) {
608 amos_page = (struct amo *)TO_AMO(uncached_alloc_page(0, 1));
609 if (amos_page == NULL) {
610 dev_err(xpc_part, "can't allocate page of amos\n");
615 * Open up amo-R/W to cpu. This is done on Shub 1.1 systems
616 * when xpc_allow_amo_ops_shub_wars_1_1_sn2() is called.
618 ret = xpc_allow_amo_ops_sn2(amos_page);
619 if (ret != xpSuccess) {
620 dev_err(xpc_part, "can't allow amo operations\n");
621 uncached_free_page(__IA64_UNCACHED_OFFSET |
622 TO_PHYS((u64)amos_page), 1);
627 /* clear xpc_vars_sn2 */
628 memset(xpc_vars_sn2, 0, sizeof(struct xpc_vars_sn2));
630 xpc_vars_sn2->version = XPC_V_VERSION;
631 xpc_vars_sn2->activate_IRQ_nasid = cpuid_to_nasid(0);
632 xpc_vars_sn2->activate_IRQ_phys_cpuid = cpu_physical_id(0);
633 xpc_vars_sn2->vars_part_pa = __pa(xpc_vars_part_sn2);
634 xpc_vars_sn2->amos_page_pa = ia64_tpa((u64)amos_page);
635 xpc_vars_sn2->amos_page = amos_page; /* save for next load of XPC */
637 /* clear xpc_vars_part_sn2 */
638 memset((u64 *)xpc_vars_part_sn2, 0, sizeof(struct xpc_vars_part_sn2) *
641 /* initialize the activate IRQ related amo variables */
642 for (i = 0; i < xpc_nasid_mask_words; i++)
643 (void)xpc_init_IRQ_amo_sn2(XPC_ACTIVATE_IRQ_AMOS_SN2 + i);
645 /* initialize the engaged remote partitions related amo variables */
646 (void)xpc_init_IRQ_amo_sn2(XPC_ENGAGED_PARTITIONS_AMO_SN2);
647 (void)xpc_init_IRQ_amo_sn2(XPC_DEACTIVATE_REQUEST_AMO_SN2);
653 xpc_increment_heartbeat_sn2(void)
655 xpc_vars_sn2->heartbeat++;
659 xpc_offline_heartbeat_sn2(void)
661 xpc_increment_heartbeat_sn2();
662 xpc_vars_sn2->heartbeat_offline = 1;
666 xpc_online_heartbeat_sn2(void)
668 xpc_increment_heartbeat_sn2();
669 xpc_vars_sn2->heartbeat_offline = 0;
673 xpc_heartbeat_init_sn2(void)
675 DBUG_ON(xpc_vars_sn2 == NULL);
677 bitmap_zero(xpc_vars_sn2->heartbeating_to_mask, XP_MAX_NPARTITIONS_SN2);
678 xpc_heartbeating_to_mask = &xpc_vars_sn2->heartbeating_to_mask[0];
679 xpc_online_heartbeat_sn2();
683 xpc_heartbeat_exit_sn2(void)
685 xpc_offline_heartbeat_sn2();
689 * At periodic intervals, scan through all active partitions and ensure
690 * their heartbeat is still active. If not, the partition is deactivated.
693 xpc_check_remote_hb_sn2(void)
695 struct xpc_vars_sn2 *remote_vars;
696 struct xpc_partition *part;
700 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
702 for (partid = 0; partid < xp_max_npartitions; partid++) {
707 if (partid == sn_partition_id)
710 part = &xpc_partitions[partid];
712 if (part->act_state == XPC_P_INACTIVE ||
713 part->act_state == XPC_P_DEACTIVATING) {
717 /* pull the remote_hb cache line */
718 ret = xp_remote_memcpy(remote_vars,
719 (void *)part->sn.sn2.remote_vars_pa,
721 if (ret != xpSuccess) {
722 XPC_DEACTIVATE_PARTITION(part, ret);
726 dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
727 " = %ld, heartbeat_offline = %ld, HB_mask[0] = 0x%lx\n",
728 partid, remote_vars->heartbeat, part->last_heartbeat,
729 remote_vars->heartbeat_offline,
730 remote_vars->heartbeating_to_mask[0]);
732 if (((remote_vars->heartbeat == part->last_heartbeat) &&
733 (remote_vars->heartbeat_offline == 0)) ||
734 !xpc_hb_allowed(sn_partition_id,
735 &remote_vars->heartbeating_to_mask)) {
737 XPC_DEACTIVATE_PARTITION(part, xpNoHeartbeat);
741 part->last_heartbeat = remote_vars->heartbeat;
746 * Get a copy of the remote partition's XPC variables from the reserved page.
748 * remote_vars points to a buffer that is cacheline aligned for BTE copies and
749 * assumed to be of size XPC_RP_VARS_SIZE.
751 static enum xp_retval
752 xpc_get_remote_vars_sn2(u64 remote_vars_pa, struct xpc_vars_sn2 *remote_vars)
756 if (remote_vars_pa == 0)
759 /* pull over the cross partition variables */
760 ret = xp_remote_memcpy(remote_vars, (void *)remote_vars_pa,
762 if (ret != xpSuccess)
765 if (XPC_VERSION_MAJOR(remote_vars->version) !=
766 XPC_VERSION_MAJOR(XPC_V_VERSION)) {
774 xpc_request_partition_activation_sn2(struct xpc_rsvd_page *remote_rp,
775 u64 remote_rp_pa, int nasid)
777 xpc_send_local_activate_IRQ_sn2(nasid);
781 xpc_request_partition_reactivation_sn2(struct xpc_partition *part)
783 xpc_send_local_activate_IRQ_sn2(part->sn.sn2.activate_IRQ_nasid);
787 xpc_request_partition_deactivation_sn2(struct xpc_partition *part)
789 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
790 unsigned long irq_flags;
791 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
792 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
793 sizeof(struct amo)));
795 local_irq_save(irq_flags);
797 /* set bit corresponding to our partid in remote partition's amo */
798 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
799 (1UL << sn_partition_id));
801 * We must always use the nofault function regardless of whether we
802 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
803 * didn't, we'd never know that the other partition is down and would
804 * keep sending IRQs and amos to it until the heartbeat times out.
806 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
808 xp_nofault_PIOR_target));
810 local_irq_restore(irq_flags);
813 * Send activate IRQ to get other side to see that we've set our
814 * bit in their deactivate request amo.
816 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
818 part_sn2->activate_IRQ_nasid,
819 part_sn2->activate_IRQ_phys_cpuid);
823 xpc_cancel_partition_deactivation_request_sn2(struct xpc_partition *part)
825 unsigned long irq_flags;
826 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
827 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
828 sizeof(struct amo)));
830 local_irq_save(irq_flags);
832 /* clear bit corresponding to our partid in remote partition's amo */
833 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
834 ~(1UL << sn_partition_id));
836 * We must always use the nofault function regardless of whether we
837 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
838 * didn't, we'd never know that the other partition is down and would
839 * keep sending IRQs and amos to it until the heartbeat times out.
841 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
843 xp_nofault_PIOR_target));
845 local_irq_restore(irq_flags);
849 xpc_partition_deactivation_requested_sn2(short partid)
851 struct amo *amo = xpc_vars_sn2->amos_page +
852 XPC_DEACTIVATE_REQUEST_AMO_SN2;
854 /* our partition's amo variable ANDed with partid mask */
855 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
856 (1UL << partid)) != 0;
860 * Update the remote partition's info.
863 xpc_update_partition_info_sn2(struct xpc_partition *part, u8 remote_rp_version,
864 unsigned long *remote_rp_stamp, u64 remote_rp_pa,
866 struct xpc_vars_sn2 *remote_vars)
868 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
870 part->remote_rp_version = remote_rp_version;
871 dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n",
872 part->remote_rp_version);
874 part->remote_rp_stamp = *remote_rp_stamp;
875 dev_dbg(xpc_part, " remote_rp_stamp = 0x%016lx\n",
876 part->remote_rp_stamp);
878 part->remote_rp_pa = remote_rp_pa;
879 dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
881 part_sn2->remote_vars_pa = remote_vars_pa;
882 dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n",
883 part_sn2->remote_vars_pa);
885 part->last_heartbeat = remote_vars->heartbeat;
886 dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n",
887 part->last_heartbeat);
889 part_sn2->remote_vars_part_pa = remote_vars->vars_part_pa;
890 dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n",
891 part_sn2->remote_vars_part_pa);
893 part_sn2->activate_IRQ_nasid = remote_vars->activate_IRQ_nasid;
894 dev_dbg(xpc_part, " activate_IRQ_nasid = 0x%x\n",
895 part_sn2->activate_IRQ_nasid);
897 part_sn2->activate_IRQ_phys_cpuid =
898 remote_vars->activate_IRQ_phys_cpuid;
899 dev_dbg(xpc_part, " activate_IRQ_phys_cpuid = 0x%x\n",
900 part_sn2->activate_IRQ_phys_cpuid);
902 part_sn2->remote_amos_page_pa = remote_vars->amos_page_pa;
903 dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n",
904 part_sn2->remote_amos_page_pa);
906 part_sn2->remote_vars_version = remote_vars->version;
907 dev_dbg(xpc_part, " remote_vars_version = 0x%x\n",
908 part_sn2->remote_vars_version);
912 * Prior code has determined the nasid which generated a activate IRQ.
913 * Inspect that nasid to determine if its partition needs to be activated
916 * A partition is considered "awaiting activation" if our partition
917 * flags indicate it is not active and it has a heartbeat. A
918 * partition is considered "awaiting deactivation" if our partition
919 * flags indicate it is active but it has no heartbeat or it is not
920 * sending its heartbeat to us.
922 * To determine the heartbeat, the remote nasid must have a properly
923 * initialized reserved page.
926 xpc_identify_activate_IRQ_req_sn2(int nasid)
928 struct xpc_rsvd_page *remote_rp;
929 struct xpc_vars_sn2 *remote_vars;
932 int remote_rp_version;
934 unsigned long remote_rp_stamp = 0;
936 struct xpc_partition *part;
937 struct xpc_partition_sn2 *part_sn2;
940 /* pull over the reserved page structure */
942 remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer_sn2;
944 ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
945 if (ret != xpSuccess) {
946 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
947 "which sent interrupt, reason=%d\n", nasid, ret);
951 remote_vars_pa = remote_rp->sn.vars_pa;
952 remote_rp_version = remote_rp->version;
953 remote_rp_stamp = remote_rp->stamp;
955 partid = remote_rp->SAL_partid;
956 part = &xpc_partitions[partid];
957 part_sn2 = &part->sn.sn2;
959 /* pull over the cross partition variables */
961 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
963 ret = xpc_get_remote_vars_sn2(remote_vars_pa, remote_vars);
964 if (ret != xpSuccess) {
965 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
966 "which sent interrupt, reason=%d\n", nasid, ret);
968 XPC_DEACTIVATE_PARTITION(part, ret);
972 part->activate_IRQ_rcvd++;
974 dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
975 "%ld:0x%lx\n", (int)nasid, (int)partid, part->activate_IRQ_rcvd,
976 remote_vars->heartbeat, remote_vars->heartbeating_to_mask[0]);
978 if (xpc_partition_disengaged(part) &&
979 part->act_state == XPC_P_INACTIVE) {
981 xpc_update_partition_info_sn2(part, remote_rp_version,
982 &remote_rp_stamp, remote_rp_pa,
983 remote_vars_pa, remote_vars);
985 if (xpc_partition_deactivation_requested_sn2(partid)) {
987 * Other side is waiting on us to deactivate even though
993 xpc_activate_partition(part);
997 DBUG_ON(part->remote_rp_version == 0);
998 DBUG_ON(part_sn2->remote_vars_version == 0);
1000 if (remote_rp_stamp != part->remote_rp_stamp) {
1002 /* the other side rebooted */
1004 DBUG_ON(xpc_partition_engaged_sn2(partid));
1005 DBUG_ON(xpc_partition_deactivation_requested_sn2(partid));
1007 xpc_update_partition_info_sn2(part, remote_rp_version,
1008 &remote_rp_stamp, remote_rp_pa,
1009 remote_vars_pa, remote_vars);
1013 if (part->disengage_timeout > 0 && !xpc_partition_disengaged(part)) {
1014 /* still waiting on other side to disengage from us */
1019 XPC_DEACTIVATE_PARTITION(part, xpReactivating);
1020 else if (xpc_partition_deactivation_requested_sn2(partid))
1021 XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown);
1025 * Loop through the activation amo variables and process any bits
1026 * which are set. Each bit indicates a nasid sending a partition
1027 * activation or deactivation request.
1029 * Return #of IRQs detected.
1032 xpc_identify_activate_IRQ_sender_sn2(void)
1036 u64 nasid; /* remote nasid */
1037 int n_IRQs_detected = 0;
1038 struct amo *act_amos;
1040 act_amos = xpc_vars_sn2->amos_page + XPC_ACTIVATE_IRQ_AMOS_SN2;
1042 /* scan through act amo variable looking for non-zero entries */
1043 for (word = 0; word < xpc_nasid_mask_words; word++) {
1048 nasid_mask = xpc_receive_IRQ_amo_sn2(&act_amos[word]);
1049 if (nasid_mask == 0) {
1050 /* no IRQs from nasids in this variable */
1054 dev_dbg(xpc_part, "amo[%d] gave back 0x%lx\n", word,
1058 * If this nasid has been added to the machine since
1059 * our partition was reset, this will retain the
1060 * remote nasid in our reserved pages machine mask.
1061 * This is used in the event of module reload.
1063 xpc_mach_nasids[word] |= nasid_mask;
1065 /* locate the nasid(s) which sent interrupts */
1067 for (bit = 0; bit < (8 * sizeof(u64)); bit++) {
1068 if (nasid_mask & (1UL << bit)) {
1070 nasid = XPC_NASID_FROM_W_B(word, bit);
1071 dev_dbg(xpc_part, "interrupt from nasid %ld\n",
1073 xpc_identify_activate_IRQ_req_sn2(nasid);
1077 return n_IRQs_detected;
1081 xpc_process_activate_IRQ_rcvd_sn2(int n_IRQs_expected)
1083 int n_IRQs_detected;
1085 n_IRQs_detected = xpc_identify_activate_IRQ_sender_sn2();
1086 if (n_IRQs_detected < n_IRQs_expected) {
1087 /* retry once to help avoid missing amo */
1088 (void)xpc_identify_activate_IRQ_sender_sn2();
1093 * Guarantee that the kzalloc'd memory is cacheline aligned.
1096 xpc_kzalloc_cacheline_aligned_sn2(size_t size, gfp_t flags, void **base)
1098 /* see if kzalloc will give us cachline aligned memory by default */
1099 *base = kzalloc(size, flags);
1103 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
1108 /* nope, we'll have to do it ourselves */
1109 *base = kzalloc(size + L1_CACHE_BYTES, flags);
1113 return (void *)L1_CACHE_ALIGN((u64)*base);
1117 * Setup the infrastructure necessary to support XPartition Communication
1118 * between the specified remote partition and the local one.
1120 static enum xp_retval
1121 xpc_setup_infrastructure_sn2(struct xpc_partition *part)
1123 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1124 enum xp_retval retval;
1128 struct xpc_channel *ch;
1129 struct timer_list *timer;
1130 short partid = XPC_PARTID(part);
1133 * Allocate all of the channel structures as a contiguous chunk of
1136 DBUG_ON(part->channels != NULL);
1137 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
1139 if (part->channels == NULL) {
1140 dev_err(xpc_chan, "can't get memory for channels\n");
1144 /* allocate all the required GET/PUT values */
1146 part_sn2->local_GPs =
1147 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1148 &part_sn2->local_GPs_base);
1149 if (part_sn2->local_GPs == NULL) {
1150 dev_err(xpc_chan, "can't get memory for local get/put "
1152 retval = xpNoMemory;
1156 part_sn2->remote_GPs =
1157 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1158 &part_sn2->remote_GPs_base);
1159 if (part_sn2->remote_GPs == NULL) {
1160 dev_err(xpc_chan, "can't get memory for remote get/put "
1162 retval = xpNoMemory;
1166 part_sn2->remote_GPs_pa = 0;
1168 /* allocate all the required open and close args */
1170 part->local_openclose_args =
1171 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1173 &part->local_openclose_args_base);
1174 if (part->local_openclose_args == NULL) {
1175 dev_err(xpc_chan, "can't get memory for local connect args\n");
1176 retval = xpNoMemory;
1180 part->remote_openclose_args =
1181 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1183 &part->remote_openclose_args_base);
1184 if (part->remote_openclose_args == NULL) {
1185 dev_err(xpc_chan, "can't get memory for remote connect args\n");
1186 retval = xpNoMemory;
1190 part_sn2->remote_openclose_args_pa = 0;
1192 part_sn2->local_chctl_amo_va = xpc_init_IRQ_amo_sn2(partid);
1193 part->chctl.all_flags = 0;
1194 spin_lock_init(&part->chctl_lock);
1196 part_sn2->notify_IRQ_nasid = 0;
1197 part_sn2->notify_IRQ_phys_cpuid = 0;
1198 part_sn2->remote_chctl_amo_va = NULL;
1200 atomic_set(&part->channel_mgr_requests, 1);
1201 init_waitqueue_head(&part->channel_mgr_wq);
1203 sprintf(part_sn2->notify_IRQ_owner, "xpc%02d", partid);
1204 ret = request_irq(SGI_XPC_NOTIFY, xpc_handle_notify_IRQ_sn2,
1205 IRQF_SHARED, part_sn2->notify_IRQ_owner,
1206 (void *)(u64)partid);
1208 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
1209 "errno=%d\n", -ret);
1210 retval = xpLackOfResources;
1214 /* Setup a timer to check for dropped notify IRQs */
1215 timer = &part_sn2->dropped_notify_IRQ_timer;
1218 (void (*)(unsigned long))xpc_check_for_dropped_notify_IRQ_sn2;
1219 timer->data = (unsigned long)part;
1220 timer->expires = jiffies + XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
1223 part->nchannels = XPC_MAX_NCHANNELS;
1225 atomic_set(&part->nchannels_active, 0);
1226 atomic_set(&part->nchannels_engaged, 0);
1228 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1229 ch = &part->channels[ch_number];
1231 ch->partid = partid;
1232 ch->number = ch_number;
1233 ch->flags = XPC_C_DISCONNECTED;
1235 ch->sn.sn2.local_GP = &part_sn2->local_GPs[ch_number];
1236 ch->local_openclose_args =
1237 &part->local_openclose_args[ch_number];
1239 atomic_set(&ch->kthreads_assigned, 0);
1240 atomic_set(&ch->kthreads_idle, 0);
1241 atomic_set(&ch->kthreads_active, 0);
1243 atomic_set(&ch->references, 0);
1244 atomic_set(&ch->n_to_notify, 0);
1246 spin_lock_init(&ch->lock);
1247 mutex_init(&ch->sn.sn2.msg_to_pull_mutex);
1248 init_completion(&ch->wdisconnect_wait);
1250 atomic_set(&ch->n_on_msg_allocate_wq, 0);
1251 init_waitqueue_head(&ch->msg_allocate_wq);
1252 init_waitqueue_head(&ch->idle_wq);
1256 * With the setting of the partition setup_state to XPC_P_SETUP, we're
1257 * declaring that this partition is ready to go.
1259 part->setup_state = XPC_P_SETUP;
1262 * Setup the per partition specific variables required by the
1263 * remote partition to establish channel connections with us.
1265 * The setting of the magic # indicates that these per partition
1266 * specific variables are ready to be used.
1268 xpc_vars_part_sn2[partid].GPs_pa = __pa(part_sn2->local_GPs);
1269 xpc_vars_part_sn2[partid].openclose_args_pa =
1270 __pa(part->local_openclose_args);
1271 xpc_vars_part_sn2[partid].chctl_amo_pa =
1272 __pa(part_sn2->local_chctl_amo_va);
1273 cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
1274 xpc_vars_part_sn2[partid].notify_IRQ_nasid = cpuid_to_nasid(cpuid);
1275 xpc_vars_part_sn2[partid].notify_IRQ_phys_cpuid =
1276 cpu_physical_id(cpuid);
1277 xpc_vars_part_sn2[partid].nchannels = part->nchannels;
1278 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC1;
1282 /* setup of infrastructure failed */
1284 kfree(part->remote_openclose_args_base);
1285 part->remote_openclose_args = NULL;
1287 kfree(part->local_openclose_args_base);
1288 part->local_openclose_args = NULL;
1290 kfree(part_sn2->remote_GPs_base);
1291 part_sn2->remote_GPs = NULL;
1293 kfree(part_sn2->local_GPs_base);
1294 part_sn2->local_GPs = NULL;
1296 kfree(part->channels);
1297 part->channels = NULL;
1302 * Teardown the infrastructure necessary to support XPartition Communication
1303 * between the specified remote partition and the local one.
1306 xpc_teardown_infrastructure_sn2(struct xpc_partition *part)
1308 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1309 short partid = XPC_PARTID(part);
1312 * We start off by making this partition inaccessible to local
1313 * processes by marking it as no longer setup. Then we make it
1314 * inaccessible to remote processes by clearing the XPC per partition
1315 * specific variable's magic # (which indicates that these variables
1316 * are no longer valid) and by ignoring all XPC notify IRQs sent to
1320 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1321 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1322 DBUG_ON(part->setup_state != XPC_P_SETUP);
1323 part->setup_state = XPC_P_WTEARDOWN;
1325 xpc_vars_part_sn2[partid].magic = 0;
1327 free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1330 * Before proceeding with the teardown we have to wait until all
1331 * existing references cease.
1333 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1335 /* now we can begin tearing down the infrastructure */
1337 part->setup_state = XPC_P_TORNDOWN;
1339 /* in case we've still got outstanding timers registered... */
1340 del_timer_sync(&part_sn2->dropped_notify_IRQ_timer);
1342 kfree(part->remote_openclose_args_base);
1343 part->remote_openclose_args = NULL;
1344 kfree(part->local_openclose_args_base);
1345 part->local_openclose_args = NULL;
1346 kfree(part_sn2->remote_GPs_base);
1347 part_sn2->remote_GPs = NULL;
1348 kfree(part_sn2->local_GPs_base);
1349 part_sn2->local_GPs = NULL;
1350 kfree(part->channels);
1351 part->channels = NULL;
1352 part_sn2->local_chctl_amo_va = NULL;
1356 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
1357 * (or multiple cachelines) from a remote partition.
1359 * src must be a cacheline aligned physical address on the remote partition.
1360 * dst must be a cacheline aligned virtual address on this partition.
1361 * cnt must be cacheline sized
1363 /* ??? Replace this function by call to xp_remote_memcpy() or bte_copy()? */
1364 static enum xp_retval
1365 xpc_pull_remote_cachelines_sn2(struct xpc_partition *part, void *dst,
1366 const void *src, size_t cnt)
1370 DBUG_ON((u64)src != L1_CACHE_ALIGN((u64)src));
1371 DBUG_ON((u64)dst != L1_CACHE_ALIGN((u64)dst));
1372 DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
1374 if (part->act_state == XPC_P_DEACTIVATING)
1375 return part->reason;
1377 ret = xp_remote_memcpy(dst, src, cnt);
1378 if (ret != xpSuccess) {
1379 dev_dbg(xpc_chan, "xp_remote_memcpy() from partition %d failed,"
1380 " ret=%d\n", XPC_PARTID(part), ret);
1386 * Pull the remote per partition specific variables from the specified
1389 static enum xp_retval
1390 xpc_pull_remote_vars_part_sn2(struct xpc_partition *part)
1392 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1393 u8 buffer[L1_CACHE_BYTES * 2];
1394 struct xpc_vars_part_sn2 *pulled_entry_cacheline =
1395 (struct xpc_vars_part_sn2 *)L1_CACHE_ALIGN((u64)buffer);
1396 struct xpc_vars_part_sn2 *pulled_entry;
1397 u64 remote_entry_cacheline_pa, remote_entry_pa;
1398 short partid = XPC_PARTID(part);
1401 /* pull the cacheline that contains the variables we're interested in */
1403 DBUG_ON(part_sn2->remote_vars_part_pa !=
1404 L1_CACHE_ALIGN(part_sn2->remote_vars_part_pa));
1405 DBUG_ON(sizeof(struct xpc_vars_part_sn2) != L1_CACHE_BYTES / 2);
1407 remote_entry_pa = part_sn2->remote_vars_part_pa +
1408 sn_partition_id * sizeof(struct xpc_vars_part_sn2);
1410 remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
1412 pulled_entry = (struct xpc_vars_part_sn2 *)((u64)pulled_entry_cacheline
1413 + (remote_entry_pa &
1414 (L1_CACHE_BYTES - 1)));
1416 ret = xpc_pull_remote_cachelines_sn2(part, pulled_entry_cacheline,
1417 (void *)remote_entry_cacheline_pa,
1419 if (ret != xpSuccess) {
1420 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
1421 "partition %d, ret=%d\n", partid, ret);
1425 /* see if they've been set up yet */
1427 if (pulled_entry->magic != XPC_VP_MAGIC1 &&
1428 pulled_entry->magic != XPC_VP_MAGIC2) {
1430 if (pulled_entry->magic != 0) {
1431 dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
1432 "partition %d has bad magic value (=0x%lx)\n",
1433 partid, sn_partition_id, pulled_entry->magic);
1437 /* they've not been initialized yet */
1441 if (xpc_vars_part_sn2[partid].magic == XPC_VP_MAGIC1) {
1443 /* validate the variables */
1445 if (pulled_entry->GPs_pa == 0 ||
1446 pulled_entry->openclose_args_pa == 0 ||
1447 pulled_entry->chctl_amo_pa == 0) {
1449 dev_err(xpc_chan, "partition %d's XPC vars_part for "
1450 "partition %d are not valid\n", partid,
1452 return xpInvalidAddress;
1455 /* the variables we imported look to be valid */
1457 part_sn2->remote_GPs_pa = pulled_entry->GPs_pa;
1458 part_sn2->remote_openclose_args_pa =
1459 pulled_entry->openclose_args_pa;
1460 part_sn2->remote_chctl_amo_va =
1461 (struct amo *)__va(pulled_entry->chctl_amo_pa);
1462 part_sn2->notify_IRQ_nasid = pulled_entry->notify_IRQ_nasid;
1463 part_sn2->notify_IRQ_phys_cpuid =
1464 pulled_entry->notify_IRQ_phys_cpuid;
1466 if (part->nchannels > pulled_entry->nchannels)
1467 part->nchannels = pulled_entry->nchannels;
1469 /* let the other side know that we've pulled their variables */
1471 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC2;
1474 if (pulled_entry->magic == XPC_VP_MAGIC1)
1481 * Establish first contact with the remote partititon. This involves pulling
1482 * the XPC per partition variables from the remote partition and waiting for
1483 * the remote partition to pull ours.
1485 static enum xp_retval
1486 xpc_make_first_contact_sn2(struct xpc_partition *part)
1488 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1492 * Register the remote partition's amos with SAL so it can handle
1493 * and cleanup errors within that address range should the remote
1494 * partition go down. We don't unregister this range because it is
1495 * difficult to tell when outstanding writes to the remote partition
1496 * are finished and thus when it is safe to unregister. This should
1497 * not result in wasted space in the SAL xp_addr_region table because
1498 * we should get the same page for remote_amos_page_pa after module
1499 * reloads and system reboots.
1501 if (sn_register_xp_addr_region(part_sn2->remote_amos_page_pa,
1502 PAGE_SIZE, 1) < 0) {
1503 dev_warn(xpc_part, "xpc_activating(%d) failed to register "
1504 "xp_addr region\n", XPC_PARTID(part));
1506 ret = xpPhysAddrRegFailed;
1507 XPC_DEACTIVATE_PARTITION(part, ret);
1512 * Send activate IRQ to get other side to activate if they've not
1513 * already begun to do so.
1515 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
1516 cnodeid_to_nasid(0),
1517 part_sn2->activate_IRQ_nasid,
1518 part_sn2->activate_IRQ_phys_cpuid);
1520 while ((ret = xpc_pull_remote_vars_part_sn2(part)) != xpSuccess) {
1521 if (ret != xpRetry) {
1522 XPC_DEACTIVATE_PARTITION(part, ret);
1526 dev_dbg(xpc_part, "waiting to make first contact with "
1527 "partition %d\n", XPC_PARTID(part));
1529 /* wait a 1/4 of a second or so */
1530 (void)msleep_interruptible(250);
1532 if (part->act_state == XPC_P_DEACTIVATING)
1533 return part->reason;
1540 * Get the chctl flags and pull the openclose args and/or remote GPs as needed.
1543 xpc_get_chctl_all_flags_sn2(struct xpc_partition *part)
1545 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1546 unsigned long irq_flags;
1547 union xpc_channel_ctl_flags chctl;
1551 * See if there are any chctl flags to be handled.
1554 spin_lock_irqsave(&part->chctl_lock, irq_flags);
1555 chctl = part->chctl;
1556 if (chctl.all_flags != 0)
1557 part->chctl.all_flags = 0;
1559 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
1561 if (xpc_any_openclose_chctl_flags_set(&chctl)) {
1562 ret = xpc_pull_remote_cachelines_sn2(part, part->
1563 remote_openclose_args,
1565 remote_openclose_args_pa,
1566 XPC_OPENCLOSE_ARGS_SIZE);
1567 if (ret != xpSuccess) {
1568 XPC_DEACTIVATE_PARTITION(part, ret);
1570 dev_dbg(xpc_chan, "failed to pull openclose args from "
1571 "partition %d, ret=%d\n", XPC_PARTID(part),
1574 /* don't bother processing chctl flags anymore */
1575 chctl.all_flags = 0;
1579 if (xpc_any_msg_chctl_flags_set(&chctl)) {
1580 ret = xpc_pull_remote_cachelines_sn2(part, part_sn2->remote_GPs,
1581 (void *)part_sn2->remote_GPs_pa,
1583 if (ret != xpSuccess) {
1584 XPC_DEACTIVATE_PARTITION(part, ret);
1586 dev_dbg(xpc_chan, "failed to pull GPs from partition "
1587 "%d, ret=%d\n", XPC_PARTID(part), ret);
1589 /* don't bother processing chctl flags anymore */
1590 chctl.all_flags = 0;
1594 return chctl.all_flags;
1598 * Allocate the local message queue and the notify queue.
1600 static enum xp_retval
1601 xpc_allocate_local_msgqueue_sn2(struct xpc_channel *ch)
1603 unsigned long irq_flags;
1607 for (nentries = ch->local_nentries; nentries > 0; nentries--) {
1609 nbytes = nentries * ch->msg_size;
1610 ch->local_msgqueue =
1611 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1612 &ch->local_msgqueue_base);
1613 if (ch->local_msgqueue == NULL)
1616 nbytes = nentries * sizeof(struct xpc_notify);
1617 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
1618 if (ch->notify_queue == NULL) {
1619 kfree(ch->local_msgqueue_base);
1620 ch->local_msgqueue = NULL;
1624 spin_lock_irqsave(&ch->lock, irq_flags);
1625 if (nentries < ch->local_nentries) {
1626 dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
1627 "partid=%d, channel=%d\n", nentries,
1628 ch->local_nentries, ch->partid, ch->number);
1630 ch->local_nentries = nentries;
1632 spin_unlock_irqrestore(&ch->lock, irq_flags);
1636 dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
1637 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
1642 * Allocate the cached remote message queue.
1644 static enum xp_retval
1645 xpc_allocate_remote_msgqueue_sn2(struct xpc_channel *ch)
1647 unsigned long irq_flags;
1651 DBUG_ON(ch->remote_nentries <= 0);
1653 for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
1655 nbytes = nentries * ch->msg_size;
1656 ch->remote_msgqueue =
1657 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1658 &ch->remote_msgqueue_base);
1659 if (ch->remote_msgqueue == NULL)
1662 spin_lock_irqsave(&ch->lock, irq_flags);
1663 if (nentries < ch->remote_nentries) {
1664 dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
1665 "partid=%d, channel=%d\n", nentries,
1666 ch->remote_nentries, ch->partid, ch->number);
1668 ch->remote_nentries = nentries;
1670 spin_unlock_irqrestore(&ch->lock, irq_flags);
1674 dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
1675 "partid=%d, channel=%d\n", ch->partid, ch->number);
1680 * Allocate message queues and other stuff associated with a channel.
1682 * Note: Assumes all of the channel sizes are filled in.
1684 static enum xp_retval
1685 xpc_allocate_msgqueues_sn2(struct xpc_channel *ch)
1689 DBUG_ON(ch->flags & XPC_C_SETUP);
1691 ret = xpc_allocate_local_msgqueue_sn2(ch);
1692 if (ret == xpSuccess) {
1694 ret = xpc_allocate_remote_msgqueue_sn2(ch);
1695 if (ret != xpSuccess) {
1696 kfree(ch->local_msgqueue_base);
1697 ch->local_msgqueue = NULL;
1698 kfree(ch->notify_queue);
1699 ch->notify_queue = NULL;
1706 * Free up message queues and other stuff that were allocated for the specified
1709 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
1710 * they're cleared when XPC_C_DISCONNECTED is cleared.
1713 xpc_free_msgqueues_sn2(struct xpc_channel *ch)
1715 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1717 DBUG_ON(!spin_is_locked(&ch->lock));
1718 DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
1720 ch->remote_msgqueue_pa = 0;
1724 ch->local_nentries = 0;
1725 ch->remote_nentries = 0;
1726 ch->kthreads_assigned_limit = 0;
1727 ch->kthreads_idle_limit = 0;
1729 ch_sn2->local_GP->get = 0;
1730 ch_sn2->local_GP->put = 0;
1731 ch_sn2->remote_GP.get = 0;
1732 ch_sn2->remote_GP.put = 0;
1733 ch_sn2->w_local_GP.get = 0;
1734 ch_sn2->w_local_GP.put = 0;
1735 ch_sn2->w_remote_GP.get = 0;
1736 ch_sn2->w_remote_GP.put = 0;
1737 ch_sn2->next_msg_to_pull = 0;
1739 if (ch->flags & XPC_C_SETUP) {
1740 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
1741 ch->flags, ch->partid, ch->number);
1743 kfree(ch->local_msgqueue_base);
1744 ch->local_msgqueue = NULL;
1745 kfree(ch->remote_msgqueue_base);
1746 ch->remote_msgqueue = NULL;
1747 kfree(ch->notify_queue);
1748 ch->notify_queue = NULL;
1753 * Notify those who wanted to be notified upon delivery of their message.
1756 xpc_notify_senders_sn2(struct xpc_channel *ch, enum xp_retval reason, s64 put)
1758 struct xpc_notify *notify;
1760 s64 get = ch->sn.sn2.w_remote_GP.get - 1;
1762 while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
1764 notify = &ch->notify_queue[get % ch->local_nentries];
1767 * See if the notify entry indicates it was associated with
1768 * a message who's sender wants to be notified. It is possible
1769 * that it is, but someone else is doing or has done the
1772 notify_type = notify->type;
1773 if (notify_type == 0 ||
1774 cmpxchg(¬ify->type, notify_type, 0) != notify_type) {
1778 DBUG_ON(notify_type != XPC_N_CALL);
1780 atomic_dec(&ch->n_to_notify);
1782 if (notify->func != NULL) {
1783 dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
1784 "msg_number=%ld, partid=%d, channel=%d\n",
1785 (void *)notify, get, ch->partid, ch->number);
1787 notify->func(reason, ch->partid, ch->number,
1790 dev_dbg(xpc_chan, "notify->func() returned, "
1791 "notify=0x%p, msg_number=%ld, partid=%d, "
1792 "channel=%d\n", (void *)notify, get,
1793 ch->partid, ch->number);
1799 xpc_notify_senders_of_disconnect_sn2(struct xpc_channel *ch)
1801 xpc_notify_senders_sn2(ch, ch->reason, ch->sn.sn2.w_local_GP.put);
1805 * Clear some of the msg flags in the local message queue.
1808 xpc_clear_local_msgqueue_flags_sn2(struct xpc_channel *ch)
1810 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1811 struct xpc_msg *msg;
1814 get = ch_sn2->w_remote_GP.get;
1816 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1817 (get % ch->local_nentries) *
1820 } while (++get < ch_sn2->remote_GP.get);
1824 * Clear some of the msg flags in the remote message queue.
1827 xpc_clear_remote_msgqueue_flags_sn2(struct xpc_channel *ch)
1829 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1830 struct xpc_msg *msg;
1833 put = ch_sn2->w_remote_GP.put;
1835 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1836 (put % ch->remote_nentries) *
1839 } while (++put < ch_sn2->remote_GP.put);
1843 xpc_process_msg_chctl_flags_sn2(struct xpc_partition *part, int ch_number)
1845 struct xpc_channel *ch = &part->channels[ch_number];
1846 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1849 ch_sn2->remote_GP = part->sn.sn2.remote_GPs[ch_number];
1851 /* See what, if anything, has changed for each connected channel */
1853 xpc_msgqueue_ref(ch);
1855 if (ch_sn2->w_remote_GP.get == ch_sn2->remote_GP.get &&
1856 ch_sn2->w_remote_GP.put == ch_sn2->remote_GP.put) {
1857 /* nothing changed since GPs were last pulled */
1858 xpc_msgqueue_deref(ch);
1862 if (!(ch->flags & XPC_C_CONNECTED)) {
1863 xpc_msgqueue_deref(ch);
1868 * First check to see if messages recently sent by us have been
1869 * received by the other side. (The remote GET value will have
1870 * changed since we last looked at it.)
1873 if (ch_sn2->w_remote_GP.get != ch_sn2->remote_GP.get) {
1876 * We need to notify any senders that want to be notified
1877 * that their sent messages have been received by their
1878 * intended recipients. We need to do this before updating
1879 * w_remote_GP.get so that we don't allocate the same message
1880 * queue entries prematurely (see xpc_allocate_msg()).
1882 if (atomic_read(&ch->n_to_notify) > 0) {
1884 * Notify senders that messages sent have been
1885 * received and delivered by the other side.
1887 xpc_notify_senders_sn2(ch, xpMsgDelivered,
1888 ch_sn2->remote_GP.get);
1892 * Clear msg->flags in previously sent messages, so that
1893 * they're ready for xpc_allocate_msg().
1895 xpc_clear_local_msgqueue_flags_sn2(ch);
1897 ch_sn2->w_remote_GP.get = ch_sn2->remote_GP.get;
1899 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1900 "channel=%d\n", ch_sn2->w_remote_GP.get, ch->partid,
1904 * If anyone was waiting for message queue entries to become
1905 * available, wake them up.
1907 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1908 wake_up(&ch->msg_allocate_wq);
1912 * Now check for newly sent messages by the other side. (The remote
1913 * PUT value will have changed since we last looked at it.)
1916 if (ch_sn2->w_remote_GP.put != ch_sn2->remote_GP.put) {
1918 * Clear msg->flags in previously received messages, so that
1919 * they're ready for xpc_get_deliverable_msg().
1921 xpc_clear_remote_msgqueue_flags_sn2(ch);
1923 ch_sn2->w_remote_GP.put = ch_sn2->remote_GP.put;
1925 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1926 "channel=%d\n", ch_sn2->w_remote_GP.put, ch->partid,
1929 nmsgs_sent = ch_sn2->w_remote_GP.put - ch_sn2->w_local_GP.get;
1930 if (nmsgs_sent > 0) {
1931 dev_dbg(xpc_chan, "msgs waiting to be copied and "
1932 "delivered=%d, partid=%d, channel=%d\n",
1933 nmsgs_sent, ch->partid, ch->number);
1935 if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1936 xpc_activate_kthreads(ch, nmsgs_sent);
1940 xpc_msgqueue_deref(ch);
1943 static struct xpc_msg *
1944 xpc_pull_remote_msg_sn2(struct xpc_channel *ch, s64 get)
1946 struct xpc_partition *part = &xpc_partitions[ch->partid];
1947 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1948 struct xpc_msg *remote_msg, *msg;
1949 u32 msg_index, nmsgs;
1953 if (mutex_lock_interruptible(&ch_sn2->msg_to_pull_mutex) != 0) {
1954 /* we were interrupted by a signal */
1958 while (get >= ch_sn2->next_msg_to_pull) {
1960 /* pull as many messages as are ready and able to be pulled */
1962 msg_index = ch_sn2->next_msg_to_pull % ch->remote_nentries;
1964 DBUG_ON(ch_sn2->next_msg_to_pull >= ch_sn2->w_remote_GP.put);
1965 nmsgs = ch_sn2->w_remote_GP.put - ch_sn2->next_msg_to_pull;
1966 if (msg_index + nmsgs > ch->remote_nentries) {
1967 /* ignore the ones that wrap the msg queue for now */
1968 nmsgs = ch->remote_nentries - msg_index;
1971 msg_offset = msg_index * ch->msg_size;
1972 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
1973 remote_msg = (struct xpc_msg *)(ch->remote_msgqueue_pa +
1976 ret = xpc_pull_remote_cachelines_sn2(part, msg, remote_msg,
1977 nmsgs * ch->msg_size);
1978 if (ret != xpSuccess) {
1980 dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
1981 " msg %ld from partition %d, channel=%d, "
1982 "ret=%d\n", nmsgs, ch_sn2->next_msg_to_pull,
1983 ch->partid, ch->number, ret);
1985 XPC_DEACTIVATE_PARTITION(part, ret);
1987 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
1991 ch_sn2->next_msg_to_pull += nmsgs;
1994 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
1996 /* return the message we were looking for */
1997 msg_offset = (get % ch->remote_nentries) * ch->msg_size;
1998 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2004 xpc_n_of_deliverable_msgs_sn2(struct xpc_channel *ch)
2006 return ch->sn.sn2.w_remote_GP.put - ch->sn.sn2.w_local_GP.get;
2010 * Get a message to be delivered.
2012 static struct xpc_msg *
2013 xpc_get_deliverable_msg_sn2(struct xpc_channel *ch)
2015 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2016 struct xpc_msg *msg = NULL;
2020 if (ch->flags & XPC_C_DISCONNECTING)
2023 get = ch_sn2->w_local_GP.get;
2024 rmb(); /* guarantee that .get loads before .put */
2025 if (get == ch_sn2->w_remote_GP.put)
2028 /* There are messages waiting to be pulled and delivered.
2029 * We need to try to secure one for ourselves. We'll do this
2030 * by trying to increment w_local_GP.get and hope that no one
2031 * else beats us to it. If they do, we'll we'll simply have
2032 * to try again for the next one.
2035 if (cmpxchg(&ch_sn2->w_local_GP.get, get, get + 1) == get) {
2036 /* we got the entry referenced by get */
2038 dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2039 "partid=%d, channel=%d\n", get + 1,
2040 ch->partid, ch->number);
2042 /* pull the message from the remote partition */
2044 msg = xpc_pull_remote_msg_sn2(ch, get);
2046 DBUG_ON(msg != NULL && msg->number != get);
2047 DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2048 DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2059 * Now we actually send the messages that are ready to be sent by advancing
2060 * the local message queue's Put value and then send a chctl msgrequest to the
2061 * recipient partition.
2064 xpc_send_msgs_sn2(struct xpc_channel *ch, s64 initial_put)
2066 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2067 struct xpc_msg *msg;
2068 s64 put = initial_put + 1;
2069 int send_msgrequest = 0;
2074 if (put == ch_sn2->w_local_GP.put)
2077 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2078 (put % ch->local_nentries) *
2081 if (!(msg->flags & XPC_M_READY))
2087 if (put == initial_put) {
2088 /* nothing's changed */
2092 if (cmpxchg_rel(&ch_sn2->local_GP->put, initial_put, put) !=
2094 /* someone else beat us to it */
2095 DBUG_ON(ch_sn2->local_GP->put < initial_put);
2099 /* we just set the new value of local_GP->put */
2101 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
2102 "channel=%d\n", put, ch->partid, ch->number);
2104 send_msgrequest = 1;
2107 * We need to ensure that the message referenced by
2108 * local_GP->put is not XPC_M_READY or that local_GP->put
2109 * equals w_local_GP.put, so we'll go have a look.
2114 if (send_msgrequest)
2115 xpc_send_chctl_msgrequest_sn2(ch);
2119 * Allocate an entry for a message from the message queue associated with the
2120 * specified channel.
2122 static enum xp_retval
2123 xpc_allocate_msg_sn2(struct xpc_channel *ch, u32 flags,
2124 struct xpc_msg **address_of_msg)
2126 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2127 struct xpc_msg *msg;
2132 * Get the next available message entry from the local message queue.
2133 * If none are available, we'll make sure that we grab the latest
2140 put = ch_sn2->w_local_GP.put;
2141 rmb(); /* guarantee that .put loads before .get */
2142 if (put - ch_sn2->w_remote_GP.get < ch->local_nentries) {
2144 /* There are available message entries. We need to try
2145 * to secure one for ourselves. We'll do this by trying
2146 * to increment w_local_GP.put as long as someone else
2147 * doesn't beat us to it. If they do, we'll have to
2150 if (cmpxchg(&ch_sn2->w_local_GP.put, put, put + 1) ==
2152 /* we got the entry referenced by put */
2155 continue; /* try again */
2159 * There aren't any available msg entries at this time.
2161 * In waiting for a message entry to become available,
2162 * we set a timeout in case the other side is not sending
2163 * completion interrupts. This lets us fake a notify IRQ
2164 * that will cause the notify IRQ handler to fetch the latest
2165 * GP values as if an interrupt was sent by the other side.
2167 if (ret == xpTimeout)
2168 xpc_send_chctl_local_msgrequest_sn2(ch);
2170 if (flags & XPC_NOWAIT)
2173 ret = xpc_allocate_msg_wait(ch);
2174 if (ret != xpInterrupted && ret != xpTimeout)
2178 /* get the message's address and initialize it */
2179 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2180 (put % ch->local_nentries) * ch->msg_size);
2182 DBUG_ON(msg->flags != 0);
2185 dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
2186 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
2187 (void *)msg, msg->number, ch->partid, ch->number);
2189 *address_of_msg = msg;
2194 * Common code that does the actual sending of the message by advancing the
2195 * local message queue's Put value and sends a chctl msgrequest to the
2196 * partition the message is being sent to.
2198 static enum xp_retval
2199 xpc_send_msg_sn2(struct xpc_channel *ch, u32 flags, void *payload,
2200 u16 payload_size, u8 notify_type, xpc_notify_func func,
2203 enum xp_retval ret = xpSuccess;
2204 struct xpc_msg *msg = msg;
2205 struct xpc_notify *notify = notify;
2209 DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
2211 if (XPC_MSG_SIZE(payload_size) > ch->msg_size)
2212 return xpPayloadTooBig;
2214 xpc_msgqueue_ref(ch);
2216 if (ch->flags & XPC_C_DISCONNECTING) {
2220 if (!(ch->flags & XPC_C_CONNECTED)) {
2221 ret = xpNotConnected;
2225 ret = xpc_allocate_msg_sn2(ch, flags, &msg);
2226 if (ret != xpSuccess)
2229 msg_number = msg->number;
2231 if (notify_type != 0) {
2233 * Tell the remote side to send an ACK interrupt when the
2234 * message has been delivered.
2236 msg->flags |= XPC_M_INTERRUPT;
2238 atomic_inc(&ch->n_to_notify);
2240 notify = &ch->notify_queue[msg_number % ch->local_nentries];
2241 notify->func = func;
2243 notify->type = notify_type;
2245 /* ??? Is a mb() needed here? */
2247 if (ch->flags & XPC_C_DISCONNECTING) {
2249 * An error occurred between our last error check and
2250 * this one. We will try to clear the type field from
2251 * the notify entry. If we succeed then
2252 * xpc_disconnect_channel() didn't already process
2255 if (cmpxchg(¬ify->type, notify_type, 0) ==
2257 atomic_dec(&ch->n_to_notify);
2264 memcpy(&msg->payload, payload, payload_size);
2266 msg->flags |= XPC_M_READY;
2269 * The preceding store of msg->flags must occur before the following
2270 * load of local_GP->put.
2274 /* see if the message is next in line to be sent, if so send it */
2276 put = ch->sn.sn2.local_GP->put;
2277 if (put == msg_number)
2278 xpc_send_msgs_sn2(ch, put);
2281 xpc_msgqueue_deref(ch);
2286 * Now we actually acknowledge the messages that have been delivered and ack'd
2287 * by advancing the cached remote message queue's Get value and if requested
2288 * send a chctl msgrequest to the message sender's partition.
2291 xpc_acknowledge_msgs_sn2(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2293 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2294 struct xpc_msg *msg;
2295 s64 get = initial_get + 1;
2296 int send_msgrequest = 0;
2301 if (get == ch_sn2->w_local_GP.get)
2304 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2305 (get % ch->remote_nentries) *
2308 if (!(msg->flags & XPC_M_DONE))
2311 msg_flags |= msg->flags;
2315 if (get == initial_get) {
2316 /* nothing's changed */
2320 if (cmpxchg_rel(&ch_sn2->local_GP->get, initial_get, get) !=
2322 /* someone else beat us to it */
2323 DBUG_ON(ch_sn2->local_GP->get <= initial_get);
2327 /* we just set the new value of local_GP->get */
2329 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2330 "channel=%d\n", get, ch->partid, ch->number);
2332 send_msgrequest = (msg_flags & XPC_M_INTERRUPT);
2335 * We need to ensure that the message referenced by
2336 * local_GP->get is not XPC_M_DONE or that local_GP->get
2337 * equals w_local_GP.get, so we'll go have a look.
2342 if (send_msgrequest)
2343 xpc_send_chctl_msgrequest_sn2(ch);
2347 xpc_received_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg)
2350 s64 msg_number = msg->number;
2352 dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2353 (void *)msg, msg_number, ch->partid, ch->number);
2355 DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2356 msg_number % ch->remote_nentries);
2357 DBUG_ON(msg->flags & XPC_M_DONE);
2359 msg->flags |= XPC_M_DONE;
2362 * The preceding store of msg->flags must occur before the following
2363 * load of local_GP->get.
2368 * See if this message is next in line to be acknowledged as having
2371 get = ch->sn.sn2.local_GP->get;
2372 if (get == msg_number)
2373 xpc_acknowledge_msgs_sn2(ch, get, msg->flags);
2382 xpc_rsvd_page_init = xpc_rsvd_page_init_sn2;
2383 xpc_increment_heartbeat = xpc_increment_heartbeat_sn2;
2384 xpc_offline_heartbeat = xpc_offline_heartbeat_sn2;
2385 xpc_online_heartbeat = xpc_online_heartbeat_sn2;
2386 xpc_heartbeat_init = xpc_heartbeat_init_sn2;
2387 xpc_heartbeat_exit = xpc_heartbeat_exit_sn2;
2388 xpc_check_remote_hb = xpc_check_remote_hb_sn2;
2390 xpc_request_partition_activation = xpc_request_partition_activation_sn2;
2391 xpc_request_partition_reactivation =
2392 xpc_request_partition_reactivation_sn2;
2393 xpc_request_partition_deactivation =
2394 xpc_request_partition_deactivation_sn2;
2395 xpc_cancel_partition_deactivation_request =
2396 xpc_cancel_partition_deactivation_request_sn2;
2398 xpc_process_activate_IRQ_rcvd = xpc_process_activate_IRQ_rcvd_sn2;
2399 xpc_setup_infrastructure = xpc_setup_infrastructure_sn2;
2400 xpc_teardown_infrastructure = xpc_teardown_infrastructure_sn2;
2401 xpc_make_first_contact = xpc_make_first_contact_sn2;
2402 xpc_get_chctl_all_flags = xpc_get_chctl_all_flags_sn2;
2403 xpc_allocate_msgqueues = xpc_allocate_msgqueues_sn2;
2404 xpc_free_msgqueues = xpc_free_msgqueues_sn2;
2405 xpc_notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_sn2;
2406 xpc_process_msg_chctl_flags = xpc_process_msg_chctl_flags_sn2;
2407 xpc_n_of_deliverable_msgs = xpc_n_of_deliverable_msgs_sn2;
2408 xpc_get_deliverable_msg = xpc_get_deliverable_msg_sn2;
2410 xpc_indicate_partition_engaged = xpc_indicate_partition_engaged_sn2;
2411 xpc_partition_engaged = xpc_partition_engaged_sn2;
2412 xpc_any_partition_engaged = xpc_any_partition_engaged_sn2;
2413 xpc_indicate_partition_disengaged =
2414 xpc_indicate_partition_disengaged_sn2;
2415 xpc_assume_partition_disengaged = xpc_assume_partition_disengaged_sn2;
2417 xpc_send_chctl_closerequest = xpc_send_chctl_closerequest_sn2;
2418 xpc_send_chctl_closereply = xpc_send_chctl_closereply_sn2;
2419 xpc_send_chctl_openrequest = xpc_send_chctl_openrequest_sn2;
2420 xpc_send_chctl_openreply = xpc_send_chctl_openreply_sn2;
2422 xpc_send_msg = xpc_send_msg_sn2;
2423 xpc_received_msg = xpc_received_msg_sn2;
2425 buf_size = max(XPC_RP_VARS_SIZE,
2426 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES_SN2);
2427 xpc_remote_copy_buffer_sn2 = xpc_kmalloc_cacheline_aligned(buf_size,
2429 &xpc_remote_copy_buffer_base_sn2);
2430 if (xpc_remote_copy_buffer_sn2 == NULL) {
2431 dev_err(xpc_part, "can't get memory for remote copy buffer\n");
2435 /* open up protections for IPI and [potentially] amo operations */
2436 xpc_allow_IPI_ops_sn2();
2437 xpc_allow_amo_ops_shub_wars_1_1_sn2();
2440 * This is safe to do before the xpc_hb_checker thread has started
2441 * because the handler releases a wait queue. If an interrupt is
2442 * received before the thread is waiting, it will not go to sleep,
2443 * but rather immediately process the interrupt.
2445 ret = request_irq(SGI_XPC_ACTIVATE, xpc_handle_activate_IRQ_sn2, 0,
2448 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
2449 "errno=%d\n", -ret);
2450 xpc_disallow_IPI_ops_sn2();
2451 kfree(xpc_remote_copy_buffer_base_sn2);
2459 free_irq(SGI_XPC_ACTIVATE, NULL);
2460 xpc_disallow_IPI_ops_sn2();
2461 kfree(xpc_remote_copy_buffer_base_sn2);