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 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);
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 = __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_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_nlongs; 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 BIT(sn_partition_id));
802 * We must always use the nofault function regardless of whether we
803 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
804 * didn't, we'd never know that the other partition is down and would
805 * keep sending IRQs and amos to it until the heartbeat times out.
807 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
809 xp_nofault_PIOR_target));
811 local_irq_restore(irq_flags);
814 * Send activate IRQ to get other side to see that we've set our
815 * bit in their deactivate request amo.
817 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
819 part_sn2->activate_IRQ_nasid,
820 part_sn2->activate_IRQ_phys_cpuid);
824 xpc_cancel_partition_deactivation_request_sn2(struct xpc_partition *part)
826 unsigned long irq_flags;
827 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
828 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
829 sizeof(struct amo)));
831 local_irq_save(irq_flags);
833 /* clear bit corresponding to our partid in remote partition's amo */
834 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
835 ~BIT(sn_partition_id));
838 * We must always use the nofault function regardless of whether we
839 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
840 * didn't, we'd never know that the other partition is down and would
841 * keep sending IRQs and amos to it until the heartbeat times out.
843 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
845 xp_nofault_PIOR_target));
847 local_irq_restore(irq_flags);
851 xpc_partition_deactivation_requested_sn2(short partid)
853 struct amo *amo = xpc_vars_sn2->amos_page +
854 XPC_DEACTIVATE_REQUEST_AMO_SN2;
856 /* our partition's amo variable ANDed with partid mask */
857 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
862 * Update the remote partition's info.
865 xpc_update_partition_info_sn2(struct xpc_partition *part, u8 remote_rp_version,
866 unsigned long *remote_rp_ts_jiffies,
867 u64 remote_rp_pa, u64 remote_vars_pa,
868 struct xpc_vars_sn2 *remote_vars)
870 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
872 part->remote_rp_version = remote_rp_version;
873 dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n",
874 part->remote_rp_version);
876 part->remote_rp_ts_jiffies = *remote_rp_ts_jiffies;
877 dev_dbg(xpc_part, " remote_rp_ts_jiffies = 0x%016lx\n",
878 part->remote_rp_ts_jiffies);
880 part->remote_rp_pa = remote_rp_pa;
881 dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
883 part_sn2->remote_vars_pa = remote_vars_pa;
884 dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n",
885 part_sn2->remote_vars_pa);
887 part->last_heartbeat = remote_vars->heartbeat;
888 dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n",
889 part->last_heartbeat);
891 part_sn2->remote_vars_part_pa = remote_vars->vars_part_pa;
892 dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n",
893 part_sn2->remote_vars_part_pa);
895 part_sn2->activate_IRQ_nasid = remote_vars->activate_IRQ_nasid;
896 dev_dbg(xpc_part, " activate_IRQ_nasid = 0x%x\n",
897 part_sn2->activate_IRQ_nasid);
899 part_sn2->activate_IRQ_phys_cpuid =
900 remote_vars->activate_IRQ_phys_cpuid;
901 dev_dbg(xpc_part, " activate_IRQ_phys_cpuid = 0x%x\n",
902 part_sn2->activate_IRQ_phys_cpuid);
904 part_sn2->remote_amos_page_pa = remote_vars->amos_page_pa;
905 dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n",
906 part_sn2->remote_amos_page_pa);
908 part_sn2->remote_vars_version = remote_vars->version;
909 dev_dbg(xpc_part, " remote_vars_version = 0x%x\n",
910 part_sn2->remote_vars_version);
914 * Prior code has determined the nasid which generated a activate IRQ.
915 * Inspect that nasid to determine if its partition needs to be activated
918 * A partition is considered "awaiting activation" if our partition
919 * flags indicate it is not active and it has a heartbeat. A
920 * partition is considered "awaiting deactivation" if our partition
921 * flags indicate it is active but it has no heartbeat or it is not
922 * sending its heartbeat to us.
924 * To determine the heartbeat, the remote nasid must have a properly
925 * initialized reserved page.
928 xpc_identify_activate_IRQ_req_sn2(int nasid)
930 struct xpc_rsvd_page *remote_rp;
931 struct xpc_vars_sn2 *remote_vars;
934 int remote_rp_version;
936 unsigned long remote_rp_ts_jiffies = 0;
938 struct xpc_partition *part;
939 struct xpc_partition_sn2 *part_sn2;
942 /* pull over the reserved page structure */
944 remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer_sn2;
946 ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
947 if (ret != xpSuccess) {
948 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
949 "which sent interrupt, reason=%d\n", nasid, ret);
953 remote_vars_pa = remote_rp->sn.vars_pa;
954 remote_rp_version = remote_rp->version;
955 remote_rp_ts_jiffies = remote_rp->ts_jiffies;
957 partid = remote_rp->SAL_partid;
958 part = &xpc_partitions[partid];
959 part_sn2 = &part->sn.sn2;
961 /* pull over the cross partition variables */
963 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
965 ret = xpc_get_remote_vars_sn2(remote_vars_pa, remote_vars);
966 if (ret != xpSuccess) {
967 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
968 "which sent interrupt, reason=%d\n", nasid, ret);
970 XPC_DEACTIVATE_PARTITION(part, ret);
974 part->activate_IRQ_rcvd++;
976 dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
977 "%ld:0x%lx\n", (int)nasid, (int)partid, part->activate_IRQ_rcvd,
978 remote_vars->heartbeat, remote_vars->heartbeating_to_mask[0]);
980 if (xpc_partition_disengaged(part) &&
981 part->act_state == XPC_P_INACTIVE) {
983 xpc_update_partition_info_sn2(part, remote_rp_version,
984 &remote_rp_ts_jiffies,
985 remote_rp_pa, remote_vars_pa,
988 if (xpc_partition_deactivation_requested_sn2(partid)) {
990 * Other side is waiting on us to deactivate even though
996 xpc_activate_partition(part);
1000 DBUG_ON(part->remote_rp_version == 0);
1001 DBUG_ON(part_sn2->remote_vars_version == 0);
1003 if (remote_rp_ts_jiffies != part->remote_rp_ts_jiffies) {
1005 /* the other side rebooted */
1007 DBUG_ON(xpc_partition_engaged_sn2(partid));
1008 DBUG_ON(xpc_partition_deactivation_requested_sn2(partid));
1010 xpc_update_partition_info_sn2(part, remote_rp_version,
1011 &remote_rp_ts_jiffies,
1012 remote_rp_pa, remote_vars_pa,
1017 if (part->disengage_timeout > 0 && !xpc_partition_disengaged(part)) {
1018 /* still waiting on other side to disengage from us */
1023 XPC_DEACTIVATE_PARTITION(part, xpReactivating);
1024 else if (xpc_partition_deactivation_requested_sn2(partid))
1025 XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown);
1029 * Loop through the activation amo variables and process any bits
1030 * which are set. Each bit indicates a nasid sending a partition
1031 * activation or deactivation request.
1033 * Return #of IRQs detected.
1036 xpc_identify_activate_IRQ_sender_sn2(void)
1040 unsigned long nasid_mask_long;
1041 u64 nasid; /* remote nasid */
1042 int n_IRQs_detected = 0;
1043 struct amo *act_amos;
1045 act_amos = xpc_vars_sn2->amos_page + XPC_ACTIVATE_IRQ_AMOS_SN2;
1047 /* scan through activate amo variables looking for non-zero entries */
1048 for (l = 0; l < xpc_nasid_mask_nlongs; l++) {
1053 nasid_mask_long = xpc_receive_IRQ_amo_sn2(&act_amos[l]);
1055 b = find_first_bit(&nasid_mask_long, BITS_PER_LONG);
1056 if (b >= BITS_PER_LONG) {
1057 /* no IRQs from nasids in this amo variable */
1061 dev_dbg(xpc_part, "amo[%d] gave back 0x%lx\n", l,
1065 * If this nasid has been added to the machine since
1066 * our partition was reset, this will retain the
1067 * remote nasid in our reserved pages machine mask.
1068 * This is used in the event of module reload.
1070 xpc_mach_nasids[l] |= nasid_mask_long;
1072 /* locate the nasid(s) which sent interrupts */
1076 nasid = (l * BITS_PER_LONG + b) * 2;
1077 dev_dbg(xpc_part, "interrupt from nasid %ld\n", nasid);
1078 xpc_identify_activate_IRQ_req_sn2(nasid);
1080 b = find_next_bit(&nasid_mask_long, BITS_PER_LONG,
1082 } while (b < BITS_PER_LONG);
1084 return n_IRQs_detected;
1088 xpc_process_activate_IRQ_rcvd_sn2(int n_IRQs_expected)
1090 int n_IRQs_detected;
1092 n_IRQs_detected = xpc_identify_activate_IRQ_sender_sn2();
1093 if (n_IRQs_detected < n_IRQs_expected) {
1094 /* retry once to help avoid missing amo */
1095 (void)xpc_identify_activate_IRQ_sender_sn2();
1100 * Guarantee that the kzalloc'd memory is cacheline aligned.
1103 xpc_kzalloc_cacheline_aligned_sn2(size_t size, gfp_t flags, void **base)
1105 /* see if kzalloc will give us cachline aligned memory by default */
1106 *base = kzalloc(size, flags);
1110 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
1115 /* nope, we'll have to do it ourselves */
1116 *base = kzalloc(size + L1_CACHE_BYTES, flags);
1120 return (void *)L1_CACHE_ALIGN((u64)*base);
1124 * Setup the infrastructure necessary to support XPartition Communication
1125 * between the specified remote partition and the local one.
1127 static enum xp_retval
1128 xpc_setup_infrastructure_sn2(struct xpc_partition *part)
1130 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1131 enum xp_retval retval;
1135 struct xpc_channel *ch;
1136 struct timer_list *timer;
1137 short partid = XPC_PARTID(part);
1140 * Allocate all of the channel structures as a contiguous chunk of
1143 DBUG_ON(part->channels != NULL);
1144 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
1146 if (part->channels == NULL) {
1147 dev_err(xpc_chan, "can't get memory for channels\n");
1151 /* allocate all the required GET/PUT values */
1153 part_sn2->local_GPs =
1154 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1155 &part_sn2->local_GPs_base);
1156 if (part_sn2->local_GPs == NULL) {
1157 dev_err(xpc_chan, "can't get memory for local get/put "
1159 retval = xpNoMemory;
1163 part_sn2->remote_GPs =
1164 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1165 &part_sn2->remote_GPs_base);
1166 if (part_sn2->remote_GPs == NULL) {
1167 dev_err(xpc_chan, "can't get memory for remote get/put "
1169 retval = xpNoMemory;
1173 part_sn2->remote_GPs_pa = 0;
1175 /* allocate all the required open and close args */
1177 part->local_openclose_args =
1178 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1180 &part->local_openclose_args_base);
1181 if (part->local_openclose_args == NULL) {
1182 dev_err(xpc_chan, "can't get memory for local connect args\n");
1183 retval = xpNoMemory;
1187 part->remote_openclose_args =
1188 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1190 &part->remote_openclose_args_base);
1191 if (part->remote_openclose_args == NULL) {
1192 dev_err(xpc_chan, "can't get memory for remote connect args\n");
1193 retval = xpNoMemory;
1197 part_sn2->remote_openclose_args_pa = 0;
1199 part_sn2->local_chctl_amo_va = xpc_init_IRQ_amo_sn2(partid);
1200 part->chctl.all_flags = 0;
1201 spin_lock_init(&part->chctl_lock);
1203 part_sn2->notify_IRQ_nasid = 0;
1204 part_sn2->notify_IRQ_phys_cpuid = 0;
1205 part_sn2->remote_chctl_amo_va = NULL;
1207 atomic_set(&part->channel_mgr_requests, 1);
1208 init_waitqueue_head(&part->channel_mgr_wq);
1210 sprintf(part_sn2->notify_IRQ_owner, "xpc%02d", partid);
1211 ret = request_irq(SGI_XPC_NOTIFY, xpc_handle_notify_IRQ_sn2,
1212 IRQF_SHARED, part_sn2->notify_IRQ_owner,
1213 (void *)(u64)partid);
1215 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
1216 "errno=%d\n", -ret);
1217 retval = xpLackOfResources;
1221 /* Setup a timer to check for dropped notify IRQs */
1222 timer = &part_sn2->dropped_notify_IRQ_timer;
1225 (void (*)(unsigned long))xpc_check_for_dropped_notify_IRQ_sn2;
1226 timer->data = (unsigned long)part;
1227 timer->expires = jiffies + XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
1230 part->nchannels = XPC_MAX_NCHANNELS;
1232 atomic_set(&part->nchannels_active, 0);
1233 atomic_set(&part->nchannels_engaged, 0);
1235 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1236 ch = &part->channels[ch_number];
1238 ch->partid = partid;
1239 ch->number = ch_number;
1240 ch->flags = XPC_C_DISCONNECTED;
1242 ch->sn.sn2.local_GP = &part_sn2->local_GPs[ch_number];
1243 ch->local_openclose_args =
1244 &part->local_openclose_args[ch_number];
1246 atomic_set(&ch->kthreads_assigned, 0);
1247 atomic_set(&ch->kthreads_idle, 0);
1248 atomic_set(&ch->kthreads_active, 0);
1250 atomic_set(&ch->references, 0);
1251 atomic_set(&ch->n_to_notify, 0);
1253 spin_lock_init(&ch->lock);
1254 mutex_init(&ch->sn.sn2.msg_to_pull_mutex);
1255 init_completion(&ch->wdisconnect_wait);
1257 atomic_set(&ch->n_on_msg_allocate_wq, 0);
1258 init_waitqueue_head(&ch->msg_allocate_wq);
1259 init_waitqueue_head(&ch->idle_wq);
1263 * With the setting of the partition setup_state to XPC_P_SETUP, we're
1264 * declaring that this partition is ready to go.
1266 part->setup_state = XPC_P_SETUP;
1269 * Setup the per partition specific variables required by the
1270 * remote partition to establish channel connections with us.
1272 * The setting of the magic # indicates that these per partition
1273 * specific variables are ready to be used.
1275 xpc_vars_part_sn2[partid].GPs_pa = __pa(part_sn2->local_GPs);
1276 xpc_vars_part_sn2[partid].openclose_args_pa =
1277 __pa(part->local_openclose_args);
1278 xpc_vars_part_sn2[partid].chctl_amo_pa =
1279 __pa(part_sn2->local_chctl_amo_va);
1280 cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
1281 xpc_vars_part_sn2[partid].notify_IRQ_nasid = cpuid_to_nasid(cpuid);
1282 xpc_vars_part_sn2[partid].notify_IRQ_phys_cpuid =
1283 cpu_physical_id(cpuid);
1284 xpc_vars_part_sn2[partid].nchannels = part->nchannels;
1285 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC1;
1289 /* setup of infrastructure failed */
1291 kfree(part->remote_openclose_args_base);
1292 part->remote_openclose_args = NULL;
1294 kfree(part->local_openclose_args_base);
1295 part->local_openclose_args = NULL;
1297 kfree(part_sn2->remote_GPs_base);
1298 part_sn2->remote_GPs = NULL;
1300 kfree(part_sn2->local_GPs_base);
1301 part_sn2->local_GPs = NULL;
1303 kfree(part->channels);
1304 part->channels = NULL;
1309 * Teardown the infrastructure necessary to support XPartition Communication
1310 * between the specified remote partition and the local one.
1313 xpc_teardown_infrastructure_sn2(struct xpc_partition *part)
1315 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1316 short partid = XPC_PARTID(part);
1319 * We start off by making this partition inaccessible to local
1320 * processes by marking it as no longer setup. Then we make it
1321 * inaccessible to remote processes by clearing the XPC per partition
1322 * specific variable's magic # (which indicates that these variables
1323 * are no longer valid) and by ignoring all XPC notify IRQs sent to
1327 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1328 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1329 DBUG_ON(part->setup_state != XPC_P_SETUP);
1330 part->setup_state = XPC_P_WTEARDOWN;
1332 xpc_vars_part_sn2[partid].magic = 0;
1334 free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1337 * Before proceeding with the teardown we have to wait until all
1338 * existing references cease.
1340 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1342 /* now we can begin tearing down the infrastructure */
1344 part->setup_state = XPC_P_TORNDOWN;
1346 /* in case we've still got outstanding timers registered... */
1347 del_timer_sync(&part_sn2->dropped_notify_IRQ_timer);
1349 kfree(part->remote_openclose_args_base);
1350 part->remote_openclose_args = NULL;
1351 kfree(part->local_openclose_args_base);
1352 part->local_openclose_args = NULL;
1353 kfree(part_sn2->remote_GPs_base);
1354 part_sn2->remote_GPs = NULL;
1355 kfree(part_sn2->local_GPs_base);
1356 part_sn2->local_GPs = NULL;
1357 kfree(part->channels);
1358 part->channels = NULL;
1359 part_sn2->local_chctl_amo_va = NULL;
1363 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
1364 * (or multiple cachelines) from a remote partition.
1366 * src must be a cacheline aligned physical address on the remote partition.
1367 * dst must be a cacheline aligned virtual address on this partition.
1368 * cnt must be cacheline sized
1370 /* ??? Replace this function by call to xp_remote_memcpy() or bte_copy()? */
1371 static enum xp_retval
1372 xpc_pull_remote_cachelines_sn2(struct xpc_partition *part, void *dst,
1373 const void *src, size_t cnt)
1377 DBUG_ON((u64)src != L1_CACHE_ALIGN((u64)src));
1378 DBUG_ON((u64)dst != L1_CACHE_ALIGN((u64)dst));
1379 DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
1381 if (part->act_state == XPC_P_DEACTIVATING)
1382 return part->reason;
1384 ret = xp_remote_memcpy(dst, src, cnt);
1385 if (ret != xpSuccess) {
1386 dev_dbg(xpc_chan, "xp_remote_memcpy() from partition %d failed,"
1387 " ret=%d\n", XPC_PARTID(part), ret);
1393 * Pull the remote per partition specific variables from the specified
1396 static enum xp_retval
1397 xpc_pull_remote_vars_part_sn2(struct xpc_partition *part)
1399 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1400 u8 buffer[L1_CACHE_BYTES * 2];
1401 struct xpc_vars_part_sn2 *pulled_entry_cacheline =
1402 (struct xpc_vars_part_sn2 *)L1_CACHE_ALIGN((u64)buffer);
1403 struct xpc_vars_part_sn2 *pulled_entry;
1404 u64 remote_entry_cacheline_pa, remote_entry_pa;
1405 short partid = XPC_PARTID(part);
1408 /* pull the cacheline that contains the variables we're interested in */
1410 DBUG_ON(part_sn2->remote_vars_part_pa !=
1411 L1_CACHE_ALIGN(part_sn2->remote_vars_part_pa));
1412 DBUG_ON(sizeof(struct xpc_vars_part_sn2) != L1_CACHE_BYTES / 2);
1414 remote_entry_pa = part_sn2->remote_vars_part_pa +
1415 sn_partition_id * sizeof(struct xpc_vars_part_sn2);
1417 remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
1419 pulled_entry = (struct xpc_vars_part_sn2 *)((u64)pulled_entry_cacheline
1420 + (remote_entry_pa &
1421 (L1_CACHE_BYTES - 1)));
1423 ret = xpc_pull_remote_cachelines_sn2(part, pulled_entry_cacheline,
1424 (void *)remote_entry_cacheline_pa,
1426 if (ret != xpSuccess) {
1427 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
1428 "partition %d, ret=%d\n", partid, ret);
1432 /* see if they've been set up yet */
1434 if (pulled_entry->magic != XPC_VP_MAGIC1 &&
1435 pulled_entry->magic != XPC_VP_MAGIC2) {
1437 if (pulled_entry->magic != 0) {
1438 dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
1439 "partition %d has bad magic value (=0x%lx)\n",
1440 partid, sn_partition_id, pulled_entry->magic);
1444 /* they've not been initialized yet */
1448 if (xpc_vars_part_sn2[partid].magic == XPC_VP_MAGIC1) {
1450 /* validate the variables */
1452 if (pulled_entry->GPs_pa == 0 ||
1453 pulled_entry->openclose_args_pa == 0 ||
1454 pulled_entry->chctl_amo_pa == 0) {
1456 dev_err(xpc_chan, "partition %d's XPC vars_part for "
1457 "partition %d are not valid\n", partid,
1459 return xpInvalidAddress;
1462 /* the variables we imported look to be valid */
1464 part_sn2->remote_GPs_pa = pulled_entry->GPs_pa;
1465 part_sn2->remote_openclose_args_pa =
1466 pulled_entry->openclose_args_pa;
1467 part_sn2->remote_chctl_amo_va =
1468 (struct amo *)__va(pulled_entry->chctl_amo_pa);
1469 part_sn2->notify_IRQ_nasid = pulled_entry->notify_IRQ_nasid;
1470 part_sn2->notify_IRQ_phys_cpuid =
1471 pulled_entry->notify_IRQ_phys_cpuid;
1473 if (part->nchannels > pulled_entry->nchannels)
1474 part->nchannels = pulled_entry->nchannels;
1476 /* let the other side know that we've pulled their variables */
1478 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC2;
1481 if (pulled_entry->magic == XPC_VP_MAGIC1)
1488 * Establish first contact with the remote partititon. This involves pulling
1489 * the XPC per partition variables from the remote partition and waiting for
1490 * the remote partition to pull ours.
1492 static enum xp_retval
1493 xpc_make_first_contact_sn2(struct xpc_partition *part)
1495 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1499 * Register the remote partition's amos with SAL so it can handle
1500 * and cleanup errors within that address range should the remote
1501 * partition go down. We don't unregister this range because it is
1502 * difficult to tell when outstanding writes to the remote partition
1503 * are finished and thus when it is safe to unregister. This should
1504 * not result in wasted space in the SAL xp_addr_region table because
1505 * we should get the same page for remote_amos_page_pa after module
1506 * reloads and system reboots.
1508 if (sn_register_xp_addr_region(part_sn2->remote_amos_page_pa,
1509 PAGE_SIZE, 1) < 0) {
1510 dev_warn(xpc_part, "xpc_activating(%d) failed to register "
1511 "xp_addr region\n", XPC_PARTID(part));
1513 ret = xpPhysAddrRegFailed;
1514 XPC_DEACTIVATE_PARTITION(part, ret);
1519 * Send activate IRQ to get other side to activate if they've not
1520 * already begun to do so.
1522 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
1523 cnodeid_to_nasid(0),
1524 part_sn2->activate_IRQ_nasid,
1525 part_sn2->activate_IRQ_phys_cpuid);
1527 while ((ret = xpc_pull_remote_vars_part_sn2(part)) != xpSuccess) {
1528 if (ret != xpRetry) {
1529 XPC_DEACTIVATE_PARTITION(part, ret);
1533 dev_dbg(xpc_part, "waiting to make first contact with "
1534 "partition %d\n", XPC_PARTID(part));
1536 /* wait a 1/4 of a second or so */
1537 (void)msleep_interruptible(250);
1539 if (part->act_state == XPC_P_DEACTIVATING)
1540 return part->reason;
1547 * Get the chctl flags and pull the openclose args and/or remote GPs as needed.
1550 xpc_get_chctl_all_flags_sn2(struct xpc_partition *part)
1552 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1553 unsigned long irq_flags;
1554 union xpc_channel_ctl_flags chctl;
1558 * See if there are any chctl flags to be handled.
1561 spin_lock_irqsave(&part->chctl_lock, irq_flags);
1562 chctl = part->chctl;
1563 if (chctl.all_flags != 0)
1564 part->chctl.all_flags = 0;
1566 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
1568 if (xpc_any_openclose_chctl_flags_set(&chctl)) {
1569 ret = xpc_pull_remote_cachelines_sn2(part, part->
1570 remote_openclose_args,
1572 remote_openclose_args_pa,
1573 XPC_OPENCLOSE_ARGS_SIZE);
1574 if (ret != xpSuccess) {
1575 XPC_DEACTIVATE_PARTITION(part, ret);
1577 dev_dbg(xpc_chan, "failed to pull openclose args from "
1578 "partition %d, ret=%d\n", XPC_PARTID(part),
1581 /* don't bother processing chctl flags anymore */
1582 chctl.all_flags = 0;
1586 if (xpc_any_msg_chctl_flags_set(&chctl)) {
1587 ret = xpc_pull_remote_cachelines_sn2(part, part_sn2->remote_GPs,
1588 (void *)part_sn2->remote_GPs_pa,
1590 if (ret != xpSuccess) {
1591 XPC_DEACTIVATE_PARTITION(part, ret);
1593 dev_dbg(xpc_chan, "failed to pull GPs from partition "
1594 "%d, ret=%d\n", XPC_PARTID(part), ret);
1596 /* don't bother processing chctl flags anymore */
1597 chctl.all_flags = 0;
1601 return chctl.all_flags;
1605 * Allocate the local message queue and the notify queue.
1607 static enum xp_retval
1608 xpc_allocate_local_msgqueue_sn2(struct xpc_channel *ch)
1610 unsigned long irq_flags;
1614 for (nentries = ch->local_nentries; nentries > 0; nentries--) {
1616 nbytes = nentries * ch->msg_size;
1617 ch->local_msgqueue =
1618 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1619 &ch->local_msgqueue_base);
1620 if (ch->local_msgqueue == NULL)
1623 nbytes = nentries * sizeof(struct xpc_notify);
1624 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
1625 if (ch->notify_queue == NULL) {
1626 kfree(ch->local_msgqueue_base);
1627 ch->local_msgqueue = NULL;
1631 spin_lock_irqsave(&ch->lock, irq_flags);
1632 if (nentries < ch->local_nentries) {
1633 dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
1634 "partid=%d, channel=%d\n", nentries,
1635 ch->local_nentries, ch->partid, ch->number);
1637 ch->local_nentries = nentries;
1639 spin_unlock_irqrestore(&ch->lock, irq_flags);
1643 dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
1644 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
1649 * Allocate the cached remote message queue.
1651 static enum xp_retval
1652 xpc_allocate_remote_msgqueue_sn2(struct xpc_channel *ch)
1654 unsigned long irq_flags;
1658 DBUG_ON(ch->remote_nentries <= 0);
1660 for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
1662 nbytes = nentries * ch->msg_size;
1663 ch->remote_msgqueue =
1664 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1665 &ch->remote_msgqueue_base);
1666 if (ch->remote_msgqueue == NULL)
1669 spin_lock_irqsave(&ch->lock, irq_flags);
1670 if (nentries < ch->remote_nentries) {
1671 dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
1672 "partid=%d, channel=%d\n", nentries,
1673 ch->remote_nentries, ch->partid, ch->number);
1675 ch->remote_nentries = nentries;
1677 spin_unlock_irqrestore(&ch->lock, irq_flags);
1681 dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
1682 "partid=%d, channel=%d\n", ch->partid, ch->number);
1687 * Allocate message queues and other stuff associated with a channel.
1689 * Note: Assumes all of the channel sizes are filled in.
1691 static enum xp_retval
1692 xpc_allocate_msgqueues_sn2(struct xpc_channel *ch)
1696 DBUG_ON(ch->flags & XPC_C_SETUP);
1698 ret = xpc_allocate_local_msgqueue_sn2(ch);
1699 if (ret == xpSuccess) {
1701 ret = xpc_allocate_remote_msgqueue_sn2(ch);
1702 if (ret != xpSuccess) {
1703 kfree(ch->local_msgqueue_base);
1704 ch->local_msgqueue = NULL;
1705 kfree(ch->notify_queue);
1706 ch->notify_queue = NULL;
1713 * Free up message queues and other stuff that were allocated for the specified
1716 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
1717 * they're cleared when XPC_C_DISCONNECTED is cleared.
1720 xpc_free_msgqueues_sn2(struct xpc_channel *ch)
1722 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1724 DBUG_ON(!spin_is_locked(&ch->lock));
1725 DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
1727 ch->remote_msgqueue_pa = 0;
1731 ch->local_nentries = 0;
1732 ch->remote_nentries = 0;
1733 ch->kthreads_assigned_limit = 0;
1734 ch->kthreads_idle_limit = 0;
1736 ch_sn2->local_GP->get = 0;
1737 ch_sn2->local_GP->put = 0;
1738 ch_sn2->remote_GP.get = 0;
1739 ch_sn2->remote_GP.put = 0;
1740 ch_sn2->w_local_GP.get = 0;
1741 ch_sn2->w_local_GP.put = 0;
1742 ch_sn2->w_remote_GP.get = 0;
1743 ch_sn2->w_remote_GP.put = 0;
1744 ch_sn2->next_msg_to_pull = 0;
1746 if (ch->flags & XPC_C_SETUP) {
1747 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
1748 ch->flags, ch->partid, ch->number);
1750 kfree(ch->local_msgqueue_base);
1751 ch->local_msgqueue = NULL;
1752 kfree(ch->remote_msgqueue_base);
1753 ch->remote_msgqueue = NULL;
1754 kfree(ch->notify_queue);
1755 ch->notify_queue = NULL;
1760 * Notify those who wanted to be notified upon delivery of their message.
1763 xpc_notify_senders_sn2(struct xpc_channel *ch, enum xp_retval reason, s64 put)
1765 struct xpc_notify *notify;
1767 s64 get = ch->sn.sn2.w_remote_GP.get - 1;
1769 while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
1771 notify = &ch->notify_queue[get % ch->local_nentries];
1774 * See if the notify entry indicates it was associated with
1775 * a message who's sender wants to be notified. It is possible
1776 * that it is, but someone else is doing or has done the
1779 notify_type = notify->type;
1780 if (notify_type == 0 ||
1781 cmpxchg(¬ify->type, notify_type, 0) != notify_type) {
1785 DBUG_ON(notify_type != XPC_N_CALL);
1787 atomic_dec(&ch->n_to_notify);
1789 if (notify->func != NULL) {
1790 dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
1791 "msg_number=%ld, partid=%d, channel=%d\n",
1792 (void *)notify, get, ch->partid, ch->number);
1794 notify->func(reason, ch->partid, ch->number,
1797 dev_dbg(xpc_chan, "notify->func() returned, "
1798 "notify=0x%p, msg_number=%ld, partid=%d, "
1799 "channel=%d\n", (void *)notify, get,
1800 ch->partid, ch->number);
1806 xpc_notify_senders_of_disconnect_sn2(struct xpc_channel *ch)
1808 xpc_notify_senders_sn2(ch, ch->reason, ch->sn.sn2.w_local_GP.put);
1812 * Clear some of the msg flags in the local message queue.
1815 xpc_clear_local_msgqueue_flags_sn2(struct xpc_channel *ch)
1817 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1818 struct xpc_msg *msg;
1821 get = ch_sn2->w_remote_GP.get;
1823 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1824 (get % ch->local_nentries) *
1827 } while (++get < ch_sn2->remote_GP.get);
1831 * Clear some of the msg flags in the remote message queue.
1834 xpc_clear_remote_msgqueue_flags_sn2(struct xpc_channel *ch)
1836 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1837 struct xpc_msg *msg;
1840 put = ch_sn2->w_remote_GP.put;
1842 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1843 (put % ch->remote_nentries) *
1846 } while (++put < ch_sn2->remote_GP.put);
1850 xpc_process_msg_chctl_flags_sn2(struct xpc_partition *part, int ch_number)
1852 struct xpc_channel *ch = &part->channels[ch_number];
1853 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1856 ch_sn2->remote_GP = part->sn.sn2.remote_GPs[ch_number];
1858 /* See what, if anything, has changed for each connected channel */
1860 xpc_msgqueue_ref(ch);
1862 if (ch_sn2->w_remote_GP.get == ch_sn2->remote_GP.get &&
1863 ch_sn2->w_remote_GP.put == ch_sn2->remote_GP.put) {
1864 /* nothing changed since GPs were last pulled */
1865 xpc_msgqueue_deref(ch);
1869 if (!(ch->flags & XPC_C_CONNECTED)) {
1870 xpc_msgqueue_deref(ch);
1875 * First check to see if messages recently sent by us have been
1876 * received by the other side. (The remote GET value will have
1877 * changed since we last looked at it.)
1880 if (ch_sn2->w_remote_GP.get != ch_sn2->remote_GP.get) {
1883 * We need to notify any senders that want to be notified
1884 * that their sent messages have been received by their
1885 * intended recipients. We need to do this before updating
1886 * w_remote_GP.get so that we don't allocate the same message
1887 * queue entries prematurely (see xpc_allocate_msg()).
1889 if (atomic_read(&ch->n_to_notify) > 0) {
1891 * Notify senders that messages sent have been
1892 * received and delivered by the other side.
1894 xpc_notify_senders_sn2(ch, xpMsgDelivered,
1895 ch_sn2->remote_GP.get);
1899 * Clear msg->flags in previously sent messages, so that
1900 * they're ready for xpc_allocate_msg().
1902 xpc_clear_local_msgqueue_flags_sn2(ch);
1904 ch_sn2->w_remote_GP.get = ch_sn2->remote_GP.get;
1906 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1907 "channel=%d\n", ch_sn2->w_remote_GP.get, ch->partid,
1911 * If anyone was waiting for message queue entries to become
1912 * available, wake them up.
1914 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1915 wake_up(&ch->msg_allocate_wq);
1919 * Now check for newly sent messages by the other side. (The remote
1920 * PUT value will have changed since we last looked at it.)
1923 if (ch_sn2->w_remote_GP.put != ch_sn2->remote_GP.put) {
1925 * Clear msg->flags in previously received messages, so that
1926 * they're ready for xpc_get_deliverable_msg().
1928 xpc_clear_remote_msgqueue_flags_sn2(ch);
1930 ch_sn2->w_remote_GP.put = ch_sn2->remote_GP.put;
1932 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1933 "channel=%d\n", ch_sn2->w_remote_GP.put, ch->partid,
1936 nmsgs_sent = ch_sn2->w_remote_GP.put - ch_sn2->w_local_GP.get;
1937 if (nmsgs_sent > 0) {
1938 dev_dbg(xpc_chan, "msgs waiting to be copied and "
1939 "delivered=%d, partid=%d, channel=%d\n",
1940 nmsgs_sent, ch->partid, ch->number);
1942 if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1943 xpc_activate_kthreads(ch, nmsgs_sent);
1947 xpc_msgqueue_deref(ch);
1950 static struct xpc_msg *
1951 xpc_pull_remote_msg_sn2(struct xpc_channel *ch, s64 get)
1953 struct xpc_partition *part = &xpc_partitions[ch->partid];
1954 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1955 struct xpc_msg *remote_msg, *msg;
1956 u32 msg_index, nmsgs;
1960 if (mutex_lock_interruptible(&ch_sn2->msg_to_pull_mutex) != 0) {
1961 /* we were interrupted by a signal */
1965 while (get >= ch_sn2->next_msg_to_pull) {
1967 /* pull as many messages as are ready and able to be pulled */
1969 msg_index = ch_sn2->next_msg_to_pull % ch->remote_nentries;
1971 DBUG_ON(ch_sn2->next_msg_to_pull >= ch_sn2->w_remote_GP.put);
1972 nmsgs = ch_sn2->w_remote_GP.put - ch_sn2->next_msg_to_pull;
1973 if (msg_index + nmsgs > ch->remote_nentries) {
1974 /* ignore the ones that wrap the msg queue for now */
1975 nmsgs = ch->remote_nentries - msg_index;
1978 msg_offset = msg_index * ch->msg_size;
1979 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
1980 remote_msg = (struct xpc_msg *)(ch->remote_msgqueue_pa +
1983 ret = xpc_pull_remote_cachelines_sn2(part, msg, remote_msg,
1984 nmsgs * ch->msg_size);
1985 if (ret != xpSuccess) {
1987 dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
1988 " msg %ld from partition %d, channel=%d, "
1989 "ret=%d\n", nmsgs, ch_sn2->next_msg_to_pull,
1990 ch->partid, ch->number, ret);
1992 XPC_DEACTIVATE_PARTITION(part, ret);
1994 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
1998 ch_sn2->next_msg_to_pull += nmsgs;
2001 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
2003 /* return the message we were looking for */
2004 msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2005 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2011 xpc_n_of_deliverable_msgs_sn2(struct xpc_channel *ch)
2013 return ch->sn.sn2.w_remote_GP.put - ch->sn.sn2.w_local_GP.get;
2017 * Get a message to be delivered.
2019 static struct xpc_msg *
2020 xpc_get_deliverable_msg_sn2(struct xpc_channel *ch)
2022 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2023 struct xpc_msg *msg = NULL;
2027 if (ch->flags & XPC_C_DISCONNECTING)
2030 get = ch_sn2->w_local_GP.get;
2031 rmb(); /* guarantee that .get loads before .put */
2032 if (get == ch_sn2->w_remote_GP.put)
2035 /* There are messages waiting to be pulled and delivered.
2036 * We need to try to secure one for ourselves. We'll do this
2037 * by trying to increment w_local_GP.get and hope that no one
2038 * else beats us to it. If they do, we'll we'll simply have
2039 * to try again for the next one.
2042 if (cmpxchg(&ch_sn2->w_local_GP.get, get, get + 1) == get) {
2043 /* we got the entry referenced by get */
2045 dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2046 "partid=%d, channel=%d\n", get + 1,
2047 ch->partid, ch->number);
2049 /* pull the message from the remote partition */
2051 msg = xpc_pull_remote_msg_sn2(ch, get);
2053 DBUG_ON(msg != NULL && msg->number != get);
2054 DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2055 DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2066 * Now we actually send the messages that are ready to be sent by advancing
2067 * the local message queue's Put value and then send a chctl msgrequest to the
2068 * recipient partition.
2071 xpc_send_msgs_sn2(struct xpc_channel *ch, s64 initial_put)
2073 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2074 struct xpc_msg *msg;
2075 s64 put = initial_put + 1;
2076 int send_msgrequest = 0;
2081 if (put == ch_sn2->w_local_GP.put)
2084 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2085 (put % ch->local_nentries) *
2088 if (!(msg->flags & XPC_M_READY))
2094 if (put == initial_put) {
2095 /* nothing's changed */
2099 if (cmpxchg_rel(&ch_sn2->local_GP->put, initial_put, put) !=
2101 /* someone else beat us to it */
2102 DBUG_ON(ch_sn2->local_GP->put < initial_put);
2106 /* we just set the new value of local_GP->put */
2108 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
2109 "channel=%d\n", put, ch->partid, ch->number);
2111 send_msgrequest = 1;
2114 * We need to ensure that the message referenced by
2115 * local_GP->put is not XPC_M_READY or that local_GP->put
2116 * equals w_local_GP.put, so we'll go have a look.
2121 if (send_msgrequest)
2122 xpc_send_chctl_msgrequest_sn2(ch);
2126 * Allocate an entry for a message from the message queue associated with the
2127 * specified channel.
2129 static enum xp_retval
2130 xpc_allocate_msg_sn2(struct xpc_channel *ch, u32 flags,
2131 struct xpc_msg **address_of_msg)
2133 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2134 struct xpc_msg *msg;
2139 * Get the next available message entry from the local message queue.
2140 * If none are available, we'll make sure that we grab the latest
2147 put = ch_sn2->w_local_GP.put;
2148 rmb(); /* guarantee that .put loads before .get */
2149 if (put - ch_sn2->w_remote_GP.get < ch->local_nentries) {
2151 /* There are available message entries. We need to try
2152 * to secure one for ourselves. We'll do this by trying
2153 * to increment w_local_GP.put as long as someone else
2154 * doesn't beat us to it. If they do, we'll have to
2157 if (cmpxchg(&ch_sn2->w_local_GP.put, put, put + 1) ==
2159 /* we got the entry referenced by put */
2162 continue; /* try again */
2166 * There aren't any available msg entries at this time.
2168 * In waiting for a message entry to become available,
2169 * we set a timeout in case the other side is not sending
2170 * completion interrupts. This lets us fake a notify IRQ
2171 * that will cause the notify IRQ handler to fetch the latest
2172 * GP values as if an interrupt was sent by the other side.
2174 if (ret == xpTimeout)
2175 xpc_send_chctl_local_msgrequest_sn2(ch);
2177 if (flags & XPC_NOWAIT)
2180 ret = xpc_allocate_msg_wait(ch);
2181 if (ret != xpInterrupted && ret != xpTimeout)
2185 /* get the message's address and initialize it */
2186 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2187 (put % ch->local_nentries) * ch->msg_size);
2189 DBUG_ON(msg->flags != 0);
2192 dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
2193 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
2194 (void *)msg, msg->number, ch->partid, ch->number);
2196 *address_of_msg = msg;
2201 * Common code that does the actual sending of the message by advancing the
2202 * local message queue's Put value and sends a chctl msgrequest to the
2203 * partition the message is being sent to.
2205 static enum xp_retval
2206 xpc_send_msg_sn2(struct xpc_channel *ch, u32 flags, void *payload,
2207 u16 payload_size, u8 notify_type, xpc_notify_func func,
2210 enum xp_retval ret = xpSuccess;
2211 struct xpc_msg *msg = msg;
2212 struct xpc_notify *notify = notify;
2216 DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
2218 if (XPC_MSG_SIZE(payload_size) > ch->msg_size)
2219 return xpPayloadTooBig;
2221 xpc_msgqueue_ref(ch);
2223 if (ch->flags & XPC_C_DISCONNECTING) {
2227 if (!(ch->flags & XPC_C_CONNECTED)) {
2228 ret = xpNotConnected;
2232 ret = xpc_allocate_msg_sn2(ch, flags, &msg);
2233 if (ret != xpSuccess)
2236 msg_number = msg->number;
2238 if (notify_type != 0) {
2240 * Tell the remote side to send an ACK interrupt when the
2241 * message has been delivered.
2243 msg->flags |= XPC_M_INTERRUPT;
2245 atomic_inc(&ch->n_to_notify);
2247 notify = &ch->notify_queue[msg_number % ch->local_nentries];
2248 notify->func = func;
2250 notify->type = notify_type;
2252 /* ??? Is a mb() needed here? */
2254 if (ch->flags & XPC_C_DISCONNECTING) {
2256 * An error occurred between our last error check and
2257 * this one. We will try to clear the type field from
2258 * the notify entry. If we succeed then
2259 * xpc_disconnect_channel() didn't already process
2262 if (cmpxchg(¬ify->type, notify_type, 0) ==
2264 atomic_dec(&ch->n_to_notify);
2271 memcpy(&msg->payload, payload, payload_size);
2273 msg->flags |= XPC_M_READY;
2276 * The preceding store of msg->flags must occur before the following
2277 * load of local_GP->put.
2281 /* see if the message is next in line to be sent, if so send it */
2283 put = ch->sn.sn2.local_GP->put;
2284 if (put == msg_number)
2285 xpc_send_msgs_sn2(ch, put);
2288 xpc_msgqueue_deref(ch);
2293 * Now we actually acknowledge the messages that have been delivered and ack'd
2294 * by advancing the cached remote message queue's Get value and if requested
2295 * send a chctl msgrequest to the message sender's partition.
2298 xpc_acknowledge_msgs_sn2(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2300 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2301 struct xpc_msg *msg;
2302 s64 get = initial_get + 1;
2303 int send_msgrequest = 0;
2308 if (get == ch_sn2->w_local_GP.get)
2311 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2312 (get % ch->remote_nentries) *
2315 if (!(msg->flags & XPC_M_DONE))
2318 msg_flags |= msg->flags;
2322 if (get == initial_get) {
2323 /* nothing's changed */
2327 if (cmpxchg_rel(&ch_sn2->local_GP->get, initial_get, get) !=
2329 /* someone else beat us to it */
2330 DBUG_ON(ch_sn2->local_GP->get <= initial_get);
2334 /* we just set the new value of local_GP->get */
2336 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2337 "channel=%d\n", get, ch->partid, ch->number);
2339 send_msgrequest = (msg_flags & XPC_M_INTERRUPT);
2342 * We need to ensure that the message referenced by
2343 * local_GP->get is not XPC_M_DONE or that local_GP->get
2344 * equals w_local_GP.get, so we'll go have a look.
2349 if (send_msgrequest)
2350 xpc_send_chctl_msgrequest_sn2(ch);
2354 xpc_received_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg)
2357 s64 msg_number = msg->number;
2359 dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2360 (void *)msg, msg_number, ch->partid, ch->number);
2362 DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2363 msg_number % ch->remote_nentries);
2364 DBUG_ON(msg->flags & XPC_M_DONE);
2366 msg->flags |= XPC_M_DONE;
2369 * The preceding store of msg->flags must occur before the following
2370 * load of local_GP->get.
2375 * See if this message is next in line to be acknowledged as having
2378 get = ch->sn.sn2.local_GP->get;
2379 if (get == msg_number)
2380 xpc_acknowledge_msgs_sn2(ch, get, msg->flags);
2389 xpc_rsvd_page_init = xpc_rsvd_page_init_sn2;
2390 xpc_increment_heartbeat = xpc_increment_heartbeat_sn2;
2391 xpc_offline_heartbeat = xpc_offline_heartbeat_sn2;
2392 xpc_online_heartbeat = xpc_online_heartbeat_sn2;
2393 xpc_heartbeat_init = xpc_heartbeat_init_sn2;
2394 xpc_heartbeat_exit = xpc_heartbeat_exit_sn2;
2395 xpc_check_remote_hb = xpc_check_remote_hb_sn2;
2397 xpc_request_partition_activation = xpc_request_partition_activation_sn2;
2398 xpc_request_partition_reactivation =
2399 xpc_request_partition_reactivation_sn2;
2400 xpc_request_partition_deactivation =
2401 xpc_request_partition_deactivation_sn2;
2402 xpc_cancel_partition_deactivation_request =
2403 xpc_cancel_partition_deactivation_request_sn2;
2405 xpc_process_activate_IRQ_rcvd = xpc_process_activate_IRQ_rcvd_sn2;
2406 xpc_setup_infrastructure = xpc_setup_infrastructure_sn2;
2407 xpc_teardown_infrastructure = xpc_teardown_infrastructure_sn2;
2408 xpc_make_first_contact = xpc_make_first_contact_sn2;
2409 xpc_get_chctl_all_flags = xpc_get_chctl_all_flags_sn2;
2410 xpc_allocate_msgqueues = xpc_allocate_msgqueues_sn2;
2411 xpc_free_msgqueues = xpc_free_msgqueues_sn2;
2412 xpc_notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_sn2;
2413 xpc_process_msg_chctl_flags = xpc_process_msg_chctl_flags_sn2;
2414 xpc_n_of_deliverable_msgs = xpc_n_of_deliverable_msgs_sn2;
2415 xpc_get_deliverable_msg = xpc_get_deliverable_msg_sn2;
2417 xpc_indicate_partition_engaged = xpc_indicate_partition_engaged_sn2;
2418 xpc_partition_engaged = xpc_partition_engaged_sn2;
2419 xpc_any_partition_engaged = xpc_any_partition_engaged_sn2;
2420 xpc_indicate_partition_disengaged =
2421 xpc_indicate_partition_disengaged_sn2;
2422 xpc_assume_partition_disengaged = xpc_assume_partition_disengaged_sn2;
2424 xpc_send_chctl_closerequest = xpc_send_chctl_closerequest_sn2;
2425 xpc_send_chctl_closereply = xpc_send_chctl_closereply_sn2;
2426 xpc_send_chctl_openrequest = xpc_send_chctl_openrequest_sn2;
2427 xpc_send_chctl_openreply = xpc_send_chctl_openreply_sn2;
2429 xpc_send_msg = xpc_send_msg_sn2;
2430 xpc_received_msg = xpc_received_msg_sn2;
2432 buf_size = max(XPC_RP_VARS_SIZE,
2433 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES_SN2);
2434 xpc_remote_copy_buffer_sn2 = xpc_kmalloc_cacheline_aligned(buf_size,
2436 &xpc_remote_copy_buffer_base_sn2);
2437 if (xpc_remote_copy_buffer_sn2 == NULL) {
2438 dev_err(xpc_part, "can't get memory for remote copy buffer\n");
2442 /* open up protections for IPI and [potentially] amo operations */
2443 xpc_allow_IPI_ops_sn2();
2444 xpc_allow_amo_ops_shub_wars_1_1_sn2();
2447 * This is safe to do before the xpc_hb_checker thread has started
2448 * because the handler releases a wait queue. If an interrupt is
2449 * received before the thread is waiting, it will not go to sleep,
2450 * but rather immediately process the interrupt.
2452 ret = request_irq(SGI_XPC_ACTIVATE, xpc_handle_activate_IRQ_sn2, 0,
2455 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
2456 "errno=%d\n", -ret);
2457 xpc_disallow_IPI_ops_sn2();
2458 kfree(xpc_remote_copy_buffer_base_sn2);
2466 free_irq(SGI_XPC_ACTIVATE, NULL);
2467 xpc_disallow_IPI_ops_sn2();
2468 kfree(xpc_remote_copy_buffer_base_sn2);