2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
10 * Cross Partition Communication (XPC) sn2-based functions.
12 * Architecture specific implementation of common functions.
16 #include <linux/delay.h>
17 #include <asm/uncached.h>
18 #include <asm/sn/mspec.h>
19 #include <asm/sn/sn_sal.h>
23 * Define the number of u64s required to represent all the C-brick nasids
24 * as a bitmap. The cross-partition kernel modules deal only with
25 * C-brick nasids, thus the need for bitmaps which don't account for
26 * odd-numbered (non C-brick) nasids.
28 #define XPC_MAX_PHYSNODES_SN2 (MAX_NUMALINK_NODES / 2)
29 #define XP_NASID_MASK_BYTES_SN2 ((XPC_MAX_PHYSNODES_SN2 + 7) / 8)
30 #define XP_NASID_MASK_WORDS_SN2 ((XPC_MAX_PHYSNODES_SN2 + 63) / 64)
33 * Memory for XPC's amo variables is allocated by the MSPEC driver. These
34 * pages are located in the lowest granule. The lowest granule uses 4k pages
35 * for cached references and an alternate TLB handler to never provide a
36 * cacheable mapping for the entire region. This will prevent speculative
37 * reading of cached copies of our lines from being issued which will cause
38 * a PI FSB Protocol error to be generated by the SHUB. For XPC, we need 64
39 * amo variables (based on XP_MAX_NPARTITIONS_SN2) to identify the senders of
40 * NOTIFY IRQs, 128 amo variables (based on XP_NASID_MASK_WORDS_SN2) to identify
41 * the senders of ACTIVATE IRQs, 1 amo variable to identify which remote
42 * partitions (i.e., XPCs) consider themselves currently engaged with the
43 * local XPC and 1 amo variable to request partition deactivation.
45 #define XPC_NOTIFY_IRQ_AMOS_SN2 0
46 #define XPC_ACTIVATE_IRQ_AMOS_SN2 (XPC_NOTIFY_IRQ_AMOS_SN2 + \
47 XP_MAX_NPARTITIONS_SN2)
48 #define XPC_ENGAGED_PARTITIONS_AMO_SN2 (XPC_ACTIVATE_IRQ_AMOS_SN2 + \
49 XP_NASID_MASK_WORDS_SN2)
50 #define XPC_DEACTIVATE_REQUEST_AMO_SN2 (XPC_ENGAGED_PARTITIONS_AMO_SN2 + 1)
53 * Buffer used to store a local copy of portions of a remote partition's
54 * reserved page (either its header and part_nasids mask, or its vars).
56 static char *xpc_remote_copy_buffer_sn2;
57 static void *xpc_remote_copy_buffer_base_sn2;
59 static struct xpc_vars_sn2 *xpc_vars_sn2;
60 static struct xpc_vars_part_sn2 *xpc_vars_part_sn2;
62 /* SH_IPI_ACCESS shub register value on startup */
63 static u64 xpc_sh1_IPI_access_sn2;
64 static u64 xpc_sh2_IPI_access0_sn2;
65 static u64 xpc_sh2_IPI_access1_sn2;
66 static u64 xpc_sh2_IPI_access2_sn2;
67 static u64 xpc_sh2_IPI_access3_sn2;
70 * Change protections to allow IPI operations.
73 xpc_allow_IPI_ops_sn2(void)
78 /* !!! The following should get moved into SAL. */
80 xpc_sh2_IPI_access0_sn2 =
81 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS0));
82 xpc_sh2_IPI_access1_sn2 =
83 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS1));
84 xpc_sh2_IPI_access2_sn2 =
85 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS2));
86 xpc_sh2_IPI_access3_sn2 =
87 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH2_IPI_ACCESS3));
89 for_each_online_node(node) {
90 nasid = cnodeid_to_nasid(node);
91 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
93 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
95 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
97 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
101 xpc_sh1_IPI_access_sn2 =
102 (u64)HUB_L((u64 *)LOCAL_MMR_ADDR(SH1_IPI_ACCESS));
104 for_each_online_node(node) {
105 nasid = cnodeid_to_nasid(node);
106 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
113 * Restrict protections to disallow IPI operations.
116 xpc_disallow_IPI_ops_sn2(void)
121 /* !!! The following should get moved into SAL. */
123 for_each_online_node(node) {
124 nasid = cnodeid_to_nasid(node);
125 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
126 xpc_sh2_IPI_access0_sn2);
127 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
128 xpc_sh2_IPI_access1_sn2);
129 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
130 xpc_sh2_IPI_access2_sn2);
131 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
132 xpc_sh2_IPI_access3_sn2);
135 for_each_online_node(node) {
136 nasid = cnodeid_to_nasid(node);
137 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
138 xpc_sh1_IPI_access_sn2);
144 * The following set of functions are used for the sending and receiving of
145 * IRQs (also known as IPIs). There are two flavors of IRQs, one that is
146 * associated with partition activity (SGI_XPC_ACTIVATE) and the other that
147 * is associated with channel activity (SGI_XPC_NOTIFY).
151 xpc_receive_IRQ_amo_sn2(struct amo *amo)
153 return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_CLEAR);
156 static enum xp_retval
157 xpc_send_IRQ_sn2(struct amo *amo, u64 flag, int nasid, int phys_cpuid,
161 unsigned long irq_flags;
163 local_irq_save(irq_flags);
165 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR, flag);
166 sn_send_IPI_phys(nasid, phys_cpuid, vector, 0);
169 * We must always use the nofault function regardless of whether we
170 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
171 * didn't, we'd never know that the other partition is down and would
172 * keep sending IRQs and amos to it until the heartbeat times out.
174 ret = xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->variable),
175 xp_nofault_PIOR_target));
177 local_irq_restore(irq_flags);
179 return (ret == 0) ? xpSuccess : xpPioReadError;
183 xpc_init_IRQ_amo_sn2(int index)
185 struct amo *amo = xpc_vars_sn2->amos_page + index;
187 (void)xpc_receive_IRQ_amo_sn2(amo); /* clear amo variable */
192 * Functions associated with SGI_XPC_ACTIVATE IRQ.
196 * Notify the heartbeat check thread that an activate IRQ has been received.
199 xpc_handle_activate_IRQ_sn2(int irq, void *dev_id)
201 atomic_inc(&xpc_activate_IRQ_rcvd);
202 wake_up_interruptible(&xpc_activate_IRQ_wq);
207 * Flag the appropriate amo variable and send an IRQ to the specified node.
210 xpc_send_activate_IRQ_sn2(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_SN2);
289 if (xpc_part_ref(part)) {
290 xpc_check_for_sent_chctl_flags_sn2(part);
292 xpc_part_deref(part);
298 * Check to see if xpc_handle_notify_IRQ_sn2() dropped any IRQs on the floor
299 * because the write to their associated amo variable completed after the IRQ
303 xpc_check_for_dropped_notify_IRQ_sn2(struct xpc_partition *part)
305 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
307 if (xpc_part_ref(part)) {
308 xpc_check_for_sent_chctl_flags_sn2(part);
310 part_sn2->dropped_notify_IRQ_timer.expires = jiffies +
311 XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
312 add_timer(&part_sn2->dropped_notify_IRQ_timer);
313 xpc_part_deref(part);
318 * Send a notify IRQ to the remote partition that is associated with the
322 xpc_send_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
323 char *chctl_flag_string, unsigned long *irq_flags)
325 struct xpc_partition *part = &xpc_partitions[ch->partid];
326 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
327 union xpc_channel_ctl_flags chctl = { 0 };
330 if (likely(part->act_state != XPC_P_DEACTIVATING)) {
331 chctl.flags[ch->number] = chctl_flag;
332 ret = xpc_send_IRQ_sn2(part_sn2->remote_chctl_amo_va,
334 part_sn2->notify_IRQ_nasid,
335 part_sn2->notify_IRQ_phys_cpuid,
337 dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n",
338 chctl_flag_string, ch->partid, ch->number, ret);
339 if (unlikely(ret != xpSuccess)) {
340 if (irq_flags != NULL)
341 spin_unlock_irqrestore(&ch->lock, *irq_flags);
342 XPC_DEACTIVATE_PARTITION(part, ret);
343 if (irq_flags != NULL)
344 spin_lock_irqsave(&ch->lock, *irq_flags);
349 #define XPC_SEND_NOTIFY_IRQ_SN2(_ch, _ipi_f, _irq_f) \
350 xpc_send_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f, _irq_f)
353 * Make it look like the remote partition, which is associated with the
354 * specified channel, sent us a notify IRQ. This faked IRQ will be handled
355 * by xpc_check_for_dropped_notify_IRQ_sn2().
358 xpc_send_local_notify_IRQ_sn2(struct xpc_channel *ch, u8 chctl_flag,
359 char *chctl_flag_string)
361 struct xpc_partition *part = &xpc_partitions[ch->partid];
362 union xpc_channel_ctl_flags chctl = { 0 };
364 chctl.flags[ch->number] = chctl_flag;
365 FETCHOP_STORE_OP(TO_AMO((u64)&part->sn.sn2.local_chctl_amo_va->
366 variable), FETCHOP_OR, chctl.all_flags);
367 dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n",
368 chctl_flag_string, ch->partid, ch->number);
371 #define XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(_ch, _ipi_f) \
372 xpc_send_local_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f)
375 xpc_send_chctl_closerequest_sn2(struct xpc_channel *ch,
376 unsigned long *irq_flags)
378 struct xpc_openclose_args *args = ch->local_openclose_args;
380 args->reason = ch->reason;
381 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREQUEST, irq_flags);
385 xpc_send_chctl_closereply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
387 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_CLOSEREPLY, irq_flags);
391 xpc_send_chctl_openrequest_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
393 struct xpc_openclose_args *args = ch->local_openclose_args;
395 args->msg_size = ch->msg_size;
396 args->local_nentries = ch->local_nentries;
397 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREQUEST, irq_flags);
401 xpc_send_chctl_openreply_sn2(struct xpc_channel *ch, unsigned long *irq_flags)
403 struct xpc_openclose_args *args = ch->local_openclose_args;
405 args->remote_nentries = ch->remote_nentries;
406 args->local_nentries = ch->local_nentries;
407 args->local_msgqueue_pa = __pa(ch->local_msgqueue);
408 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_OPENREPLY, irq_flags);
412 xpc_send_chctl_msgrequest_sn2(struct xpc_channel *ch)
414 XPC_SEND_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST, NULL);
418 xpc_send_chctl_local_msgrequest_sn2(struct xpc_channel *ch)
420 XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(ch, XPC_CHCTL_MSGREQUEST);
424 * This next set of functions are used to keep track of when a partition is
425 * potentially engaged in accessing memory belonging to another partition.
429 xpc_indicate_partition_engaged_sn2(struct xpc_partition *part)
431 unsigned long irq_flags;
432 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
433 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
434 sizeof(struct amo)));
436 local_irq_save(irq_flags);
438 /* set bit corresponding to our partid in remote partition's amo */
439 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
440 BIT(sn_partition_id));
443 * We must always use the nofault function regardless of whether we
444 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
445 * didn't, we'd never know that the other partition is down and would
446 * keep sending IRQs and amos to it until the heartbeat times out.
448 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
450 xp_nofault_PIOR_target));
452 local_irq_restore(irq_flags);
456 xpc_indicate_partition_disengaged_sn2(struct xpc_partition *part)
458 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
459 unsigned long irq_flags;
460 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
461 (XPC_ENGAGED_PARTITIONS_AMO_SN2 *
462 sizeof(struct amo)));
464 local_irq_save(irq_flags);
466 /* clear bit corresponding to our partid in remote partition's amo */
467 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
468 ~BIT(sn_partition_id));
471 * We must always use the nofault function regardless of whether we
472 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
473 * didn't, we'd never know that the other partition is down and would
474 * keep sending IRQs and amos to it until the heartbeat times out.
476 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
478 xp_nofault_PIOR_target));
480 local_irq_restore(irq_flags);
483 * Send activate IRQ to get other side to see that we've cleared our
484 * bit in their engaged partitions amo.
486 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
488 part_sn2->activate_IRQ_nasid,
489 part_sn2->activate_IRQ_phys_cpuid);
493 xpc_partition_engaged_sn2(short partid)
495 struct amo *amo = xpc_vars_sn2->amos_page +
496 XPC_ENGAGED_PARTITIONS_AMO_SN2;
498 /* our partition's amo variable ANDed with partid mask */
499 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
504 xpc_any_partition_engaged_sn2(void)
506 struct amo *amo = xpc_vars_sn2->amos_page +
507 XPC_ENGAGED_PARTITIONS_AMO_SN2;
509 /* our partition's amo variable */
510 return FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) != 0;
514 xpc_assume_partition_disengaged_sn2(short partid)
516 struct amo *amo = xpc_vars_sn2->amos_page +
517 XPC_ENGAGED_PARTITIONS_AMO_SN2;
519 /* clear bit(s) based on partid mask in our partition's amo */
520 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
524 /* original protection values for each node */
525 static u64 xpc_prot_vec_sn2[MAX_NUMNODES];
528 * Change protections to allow amo operations on non-Shub 1.1 systems.
530 static enum xp_retval
531 xpc_allow_amo_ops_sn2(struct amo *amos_page)
537 * On SHUB 1.1, we cannot call sn_change_memprotect() since the BIST
538 * collides with memory operations. On those systems we call
539 * xpc_allow_amo_ops_shub_wars_1_1_sn2() instead.
541 if (!enable_shub_wars_1_1()) {
542 ret = sn_change_memprotect(ia64_tpa((u64)amos_page), PAGE_SIZE,
543 SN_MEMPROT_ACCESS_CLASS_1,
552 * Change protections to allow amo operations on Shub 1.1 systems.
555 xpc_allow_amo_ops_shub_wars_1_1_sn2(void)
560 if (!enable_shub_wars_1_1())
563 for_each_online_node(node) {
564 nasid = cnodeid_to_nasid(node);
565 /* save current protection values */
566 xpc_prot_vec_sn2[node] =
567 (u64)HUB_L((u64 *)GLOBAL_MMR_ADDR(nasid,
568 SH1_MD_DQLP_MMR_DIR_PRIVEC0));
569 /* open up everything */
570 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid,
571 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
573 HUB_S((u64 *)GLOBAL_MMR_ADDR(nasid,
574 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
579 static enum xp_retval
580 xpc_get_partition_rsvd_page_pa_sn2(u64 buf, u64 *cookie, u64 *paddr,
586 status = sn_partition_reserved_page_pa(buf, cookie, paddr, len);
587 if (status == SALRET_OK)
589 else if (status == SALRET_MORE_PASSES)
590 ret = xpNeedMoreInfo;
598 static enum xp_retval
599 xpc_rsvd_page_init_sn2(struct xpc_rsvd_page *rp)
601 struct amo *amos_page;
605 xpc_vars_sn2 = XPC_RP_VARS(rp);
607 rp->sn.vars_pa = __pa(xpc_vars_sn2);
609 /* vars_part array follows immediately after vars */
610 xpc_vars_part_sn2 = (struct xpc_vars_part_sn2 *)((u8 *)XPC_RP_VARS(rp) +
614 * Before clearing xpc_vars_sn2, see if a page of amos had been
615 * previously allocated. If not we'll need to allocate one and set
616 * permissions so that cross-partition amos are allowed.
618 * The allocated amo page needs MCA reporting to remain disabled after
619 * XPC has unloaded. To make this work, we keep a copy of the pointer
620 * to this page (i.e., amos_page) in the struct xpc_vars_sn2 structure,
621 * which is pointed to by the reserved page, and re-use that saved copy
622 * on subsequent loads of XPC. This amo page is never freed, and its
623 * memory protections are never restricted.
625 amos_page = xpc_vars_sn2->amos_page;
626 if (amos_page == NULL) {
627 amos_page = (struct amo *)TO_AMO(uncached_alloc_page(0, 1));
628 if (amos_page == NULL) {
629 dev_err(xpc_part, "can't allocate page of amos\n");
634 * Open up amo-R/W to cpu. This is done on Shub 1.1 systems
635 * when xpc_allow_amo_ops_shub_wars_1_1_sn2() is called.
637 ret = xpc_allow_amo_ops_sn2(amos_page);
638 if (ret != xpSuccess) {
639 dev_err(xpc_part, "can't allow amo operations\n");
640 uncached_free_page(__IA64_UNCACHED_OFFSET |
641 TO_PHYS((u64)amos_page), 1);
646 /* clear xpc_vars_sn2 */
647 memset(xpc_vars_sn2, 0, sizeof(struct xpc_vars_sn2));
649 xpc_vars_sn2->version = XPC_V_VERSION;
650 xpc_vars_sn2->activate_IRQ_nasid = cpuid_to_nasid(0);
651 xpc_vars_sn2->activate_IRQ_phys_cpuid = cpu_physical_id(0);
652 xpc_vars_sn2->vars_part_pa = __pa(xpc_vars_part_sn2);
653 xpc_vars_sn2->amos_page_pa = ia64_tpa((u64)amos_page);
654 xpc_vars_sn2->amos_page = amos_page; /* save for next load of XPC */
656 /* clear xpc_vars_part_sn2 */
657 memset((u64 *)xpc_vars_part_sn2, 0, sizeof(struct xpc_vars_part_sn2) *
658 XP_MAX_NPARTITIONS_SN2);
660 /* initialize the activate IRQ related amo variables */
661 for (i = 0; i < xpc_nasid_mask_nlongs; i++)
662 (void)xpc_init_IRQ_amo_sn2(XPC_ACTIVATE_IRQ_AMOS_SN2 + i);
664 /* initialize the engaged remote partitions related amo variables */
665 (void)xpc_init_IRQ_amo_sn2(XPC_ENGAGED_PARTITIONS_AMO_SN2);
666 (void)xpc_init_IRQ_amo_sn2(XPC_DEACTIVATE_REQUEST_AMO_SN2);
672 xpc_increment_heartbeat_sn2(void)
674 xpc_vars_sn2->heartbeat++;
678 xpc_offline_heartbeat_sn2(void)
680 xpc_increment_heartbeat_sn2();
681 xpc_vars_sn2->heartbeat_offline = 1;
685 xpc_online_heartbeat_sn2(void)
687 xpc_increment_heartbeat_sn2();
688 xpc_vars_sn2->heartbeat_offline = 0;
692 xpc_heartbeat_init_sn2(void)
694 DBUG_ON(xpc_vars_sn2 == NULL);
696 bitmap_zero(xpc_vars_sn2->heartbeating_to_mask, XP_MAX_NPARTITIONS_SN2);
697 xpc_heartbeating_to_mask = &xpc_vars_sn2->heartbeating_to_mask[0];
698 xpc_online_heartbeat_sn2();
702 xpc_heartbeat_exit_sn2(void)
704 xpc_offline_heartbeat_sn2();
708 * At periodic intervals, scan through all active partitions and ensure
709 * their heartbeat is still active. If not, the partition is deactivated.
712 xpc_check_remote_hb_sn2(void)
714 struct xpc_vars_sn2 *remote_vars;
715 struct xpc_partition *part;
719 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
721 for (partid = 0; partid < XP_MAX_NPARTITIONS_SN2; partid++) {
726 if (partid == sn_partition_id)
729 part = &xpc_partitions[partid];
731 if (part->act_state == XPC_P_INACTIVE ||
732 part->act_state == XPC_P_DEACTIVATING) {
736 /* pull the remote_hb cache line */
737 ret = xp_remote_memcpy(remote_vars,
738 (void *)part->sn.sn2.remote_vars_pa,
740 if (ret != xpSuccess) {
741 XPC_DEACTIVATE_PARTITION(part, ret);
745 dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
746 " = %ld, heartbeat_offline = %ld, HB_mask[0] = 0x%lx\n",
747 partid, remote_vars->heartbeat, part->last_heartbeat,
748 remote_vars->heartbeat_offline,
749 remote_vars->heartbeating_to_mask[0]);
751 if (((remote_vars->heartbeat == part->last_heartbeat) &&
752 (remote_vars->heartbeat_offline == 0)) ||
753 !xpc_hb_allowed(sn_partition_id,
754 &remote_vars->heartbeating_to_mask)) {
756 XPC_DEACTIVATE_PARTITION(part, xpNoHeartbeat);
760 part->last_heartbeat = remote_vars->heartbeat;
765 * Get a copy of the remote partition's XPC variables from the reserved page.
767 * remote_vars points to a buffer that is cacheline aligned for BTE copies and
768 * assumed to be of size XPC_RP_VARS_SIZE.
770 static enum xp_retval
771 xpc_get_remote_vars_sn2(u64 remote_vars_pa, struct xpc_vars_sn2 *remote_vars)
775 if (remote_vars_pa == 0)
778 /* pull over the cross partition variables */
779 ret = xp_remote_memcpy(remote_vars, (void *)remote_vars_pa,
781 if (ret != xpSuccess)
784 if (XPC_VERSION_MAJOR(remote_vars->version) !=
785 XPC_VERSION_MAJOR(XPC_V_VERSION)) {
793 xpc_request_partition_activation_sn2(struct xpc_rsvd_page *remote_rp,
794 u64 remote_rp_pa, int nasid)
796 xpc_send_local_activate_IRQ_sn2(nasid);
800 xpc_request_partition_reactivation_sn2(struct xpc_partition *part)
802 xpc_send_local_activate_IRQ_sn2(part->sn.sn2.activate_IRQ_nasid);
806 xpc_request_partition_deactivation_sn2(struct xpc_partition *part)
808 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
809 unsigned long irq_flags;
810 struct amo *amo = (struct amo *)__va(part_sn2->remote_amos_page_pa +
811 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
812 sizeof(struct amo)));
814 local_irq_save(irq_flags);
816 /* set bit corresponding to our partid in remote partition's amo */
817 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_OR,
818 BIT(sn_partition_id));
821 * We must always use the nofault function regardless of whether we
822 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
823 * didn't, we'd never know that the other partition is down and would
824 * keep sending IRQs and amos to it until the heartbeat times out.
826 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
828 xp_nofault_PIOR_target));
830 local_irq_restore(irq_flags);
833 * Send activate IRQ to get other side to see that we've set our
834 * bit in their deactivate request amo.
836 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
838 part_sn2->activate_IRQ_nasid,
839 part_sn2->activate_IRQ_phys_cpuid);
843 xpc_cancel_partition_deactivation_request_sn2(struct xpc_partition *part)
845 unsigned long irq_flags;
846 struct amo *amo = (struct amo *)__va(part->sn.sn2.remote_amos_page_pa +
847 (XPC_DEACTIVATE_REQUEST_AMO_SN2 *
848 sizeof(struct amo)));
850 local_irq_save(irq_flags);
852 /* clear bit corresponding to our partid in remote partition's amo */
853 FETCHOP_STORE_OP(TO_AMO((u64)&amo->variable), FETCHOP_AND,
854 ~BIT(sn_partition_id));
857 * We must always use the nofault function regardless of whether we
858 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
859 * didn't, we'd never know that the other partition is down and would
860 * keep sending IRQs and amos to it until the heartbeat times out.
862 (void)xp_nofault_PIOR((u64 *)GLOBAL_MMR_ADDR(NASID_GET(&amo->
864 xp_nofault_PIOR_target));
866 local_irq_restore(irq_flags);
870 xpc_partition_deactivation_requested_sn2(short partid)
872 struct amo *amo = xpc_vars_sn2->amos_page +
873 XPC_DEACTIVATE_REQUEST_AMO_SN2;
875 /* our partition's amo variable ANDed with partid mask */
876 return (FETCHOP_LOAD_OP(TO_AMO((u64)&amo->variable), FETCHOP_LOAD) &
881 * Update the remote partition's info.
884 xpc_update_partition_info_sn2(struct xpc_partition *part, u8 remote_rp_version,
885 unsigned long *remote_rp_ts_jiffies,
886 u64 remote_rp_pa, u64 remote_vars_pa,
887 struct xpc_vars_sn2 *remote_vars)
889 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
891 part->remote_rp_version = remote_rp_version;
892 dev_dbg(xpc_part, " remote_rp_version = 0x%016x\n",
893 part->remote_rp_version);
895 part->remote_rp_ts_jiffies = *remote_rp_ts_jiffies;
896 dev_dbg(xpc_part, " remote_rp_ts_jiffies = 0x%016lx\n",
897 part->remote_rp_ts_jiffies);
899 part->remote_rp_pa = remote_rp_pa;
900 dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
902 part_sn2->remote_vars_pa = remote_vars_pa;
903 dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n",
904 part_sn2->remote_vars_pa);
906 part->last_heartbeat = remote_vars->heartbeat;
907 dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n",
908 part->last_heartbeat);
910 part_sn2->remote_vars_part_pa = remote_vars->vars_part_pa;
911 dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n",
912 part_sn2->remote_vars_part_pa);
914 part_sn2->activate_IRQ_nasid = remote_vars->activate_IRQ_nasid;
915 dev_dbg(xpc_part, " activate_IRQ_nasid = 0x%x\n",
916 part_sn2->activate_IRQ_nasid);
918 part_sn2->activate_IRQ_phys_cpuid =
919 remote_vars->activate_IRQ_phys_cpuid;
920 dev_dbg(xpc_part, " activate_IRQ_phys_cpuid = 0x%x\n",
921 part_sn2->activate_IRQ_phys_cpuid);
923 part_sn2->remote_amos_page_pa = remote_vars->amos_page_pa;
924 dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n",
925 part_sn2->remote_amos_page_pa);
927 part_sn2->remote_vars_version = remote_vars->version;
928 dev_dbg(xpc_part, " remote_vars_version = 0x%x\n",
929 part_sn2->remote_vars_version);
933 * Prior code has determined the nasid which generated a activate IRQ.
934 * Inspect that nasid to determine if its partition needs to be activated
937 * A partition is considered "awaiting activation" if our partition
938 * flags indicate it is not active and it has a heartbeat. A
939 * partition is considered "awaiting deactivation" if our partition
940 * flags indicate it is active but it has no heartbeat or it is not
941 * sending its heartbeat to us.
943 * To determine the heartbeat, the remote nasid must have a properly
944 * initialized reserved page.
947 xpc_identify_activate_IRQ_req_sn2(int nasid)
949 struct xpc_rsvd_page *remote_rp;
950 struct xpc_vars_sn2 *remote_vars;
953 int remote_rp_version;
955 unsigned long remote_rp_ts_jiffies = 0;
957 struct xpc_partition *part;
958 struct xpc_partition_sn2 *part_sn2;
961 /* pull over the reserved page structure */
963 remote_rp = (struct xpc_rsvd_page *)xpc_remote_copy_buffer_sn2;
965 ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
966 if (ret != xpSuccess) {
967 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
968 "which sent interrupt, reason=%d\n", nasid, ret);
972 remote_vars_pa = remote_rp->sn.vars_pa;
973 remote_rp_version = remote_rp->version;
974 remote_rp_ts_jiffies = remote_rp->ts_jiffies;
976 partid = remote_rp->SAL_partid;
977 part = &xpc_partitions[partid];
978 part_sn2 = &part->sn.sn2;
980 /* pull over the cross partition variables */
982 remote_vars = (struct xpc_vars_sn2 *)xpc_remote_copy_buffer_sn2;
984 ret = xpc_get_remote_vars_sn2(remote_vars_pa, remote_vars);
985 if (ret != xpSuccess) {
986 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
987 "which sent interrupt, reason=%d\n", nasid, ret);
989 XPC_DEACTIVATE_PARTITION(part, ret);
993 part->activate_IRQ_rcvd++;
995 dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
996 "%ld:0x%lx\n", (int)nasid, (int)partid, part->activate_IRQ_rcvd,
997 remote_vars->heartbeat, remote_vars->heartbeating_to_mask[0]);
999 if (xpc_partition_disengaged(part) &&
1000 part->act_state == XPC_P_INACTIVE) {
1002 xpc_update_partition_info_sn2(part, remote_rp_version,
1003 &remote_rp_ts_jiffies,
1004 remote_rp_pa, remote_vars_pa,
1007 if (xpc_partition_deactivation_requested_sn2(partid)) {
1009 * Other side is waiting on us to deactivate even though
1015 xpc_activate_partition(part);
1019 DBUG_ON(part->remote_rp_version == 0);
1020 DBUG_ON(part_sn2->remote_vars_version == 0);
1022 if (remote_rp_ts_jiffies != part->remote_rp_ts_jiffies) {
1024 /* the other side rebooted */
1026 DBUG_ON(xpc_partition_engaged_sn2(partid));
1027 DBUG_ON(xpc_partition_deactivation_requested_sn2(partid));
1029 xpc_update_partition_info_sn2(part, remote_rp_version,
1030 &remote_rp_ts_jiffies,
1031 remote_rp_pa, remote_vars_pa,
1036 if (part->disengage_timeout > 0 && !xpc_partition_disengaged(part)) {
1037 /* still waiting on other side to disengage from us */
1042 XPC_DEACTIVATE_PARTITION(part, xpReactivating);
1043 else if (xpc_partition_deactivation_requested_sn2(partid))
1044 XPC_DEACTIVATE_PARTITION(part, xpOtherGoingDown);
1048 * Loop through the activation amo variables and process any bits
1049 * which are set. Each bit indicates a nasid sending a partition
1050 * activation or deactivation request.
1052 * Return #of IRQs detected.
1055 xpc_identify_activate_IRQ_sender_sn2(void)
1059 unsigned long nasid_mask_long;
1060 u64 nasid; /* remote nasid */
1061 int n_IRQs_detected = 0;
1062 struct amo *act_amos;
1064 act_amos = xpc_vars_sn2->amos_page + XPC_ACTIVATE_IRQ_AMOS_SN2;
1066 /* scan through activate amo variables looking for non-zero entries */
1067 for (l = 0; l < xpc_nasid_mask_nlongs; l++) {
1072 nasid_mask_long = xpc_receive_IRQ_amo_sn2(&act_amos[l]);
1074 b = find_first_bit(&nasid_mask_long, BITS_PER_LONG);
1075 if (b >= BITS_PER_LONG) {
1076 /* no IRQs from nasids in this amo variable */
1080 dev_dbg(xpc_part, "amo[%d] gave back 0x%lx\n", l,
1084 * If this nasid has been added to the machine since
1085 * our partition was reset, this will retain the
1086 * remote nasid in our reserved pages machine mask.
1087 * This is used in the event of module reload.
1089 xpc_mach_nasids[l] |= nasid_mask_long;
1091 /* locate the nasid(s) which sent interrupts */
1095 nasid = (l * BITS_PER_LONG + b) * 2;
1096 dev_dbg(xpc_part, "interrupt from nasid %ld\n", nasid);
1097 xpc_identify_activate_IRQ_req_sn2(nasid);
1099 b = find_next_bit(&nasid_mask_long, BITS_PER_LONG,
1101 } while (b < BITS_PER_LONG);
1103 return n_IRQs_detected;
1107 xpc_process_activate_IRQ_rcvd_sn2(int n_IRQs_expected)
1109 int n_IRQs_detected;
1111 n_IRQs_detected = xpc_identify_activate_IRQ_sender_sn2();
1112 if (n_IRQs_detected < n_IRQs_expected) {
1113 /* retry once to help avoid missing amo */
1114 (void)xpc_identify_activate_IRQ_sender_sn2();
1119 * Guarantee that the kzalloc'd memory is cacheline aligned.
1122 xpc_kzalloc_cacheline_aligned_sn2(size_t size, gfp_t flags, void **base)
1124 /* see if kzalloc will give us cachline aligned memory by default */
1125 *base = kzalloc(size, flags);
1129 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
1134 /* nope, we'll have to do it ourselves */
1135 *base = kzalloc(size + L1_CACHE_BYTES, flags);
1139 return (void *)L1_CACHE_ALIGN((u64)*base);
1143 * Setup the infrastructure necessary to support XPartition Communication
1144 * between the specified remote partition and the local one.
1146 static enum xp_retval
1147 xpc_setup_infrastructure_sn2(struct xpc_partition *part)
1149 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1150 enum xp_retval retval;
1154 struct xpc_channel *ch;
1155 struct timer_list *timer;
1156 short partid = XPC_PARTID(part);
1159 * Allocate all of the channel structures as a contiguous chunk of
1162 DBUG_ON(part->channels != NULL);
1163 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
1165 if (part->channels == NULL) {
1166 dev_err(xpc_chan, "can't get memory for channels\n");
1170 /* allocate all the required GET/PUT values */
1172 part_sn2->local_GPs =
1173 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1174 &part_sn2->local_GPs_base);
1175 if (part_sn2->local_GPs == NULL) {
1176 dev_err(xpc_chan, "can't get memory for local get/put "
1178 retval = xpNoMemory;
1182 part_sn2->remote_GPs =
1183 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE, GFP_KERNEL,
1184 &part_sn2->remote_GPs_base);
1185 if (part_sn2->remote_GPs == NULL) {
1186 dev_err(xpc_chan, "can't get memory for remote get/put "
1188 retval = xpNoMemory;
1192 part_sn2->remote_GPs_pa = 0;
1194 /* allocate all the required open and close args */
1196 part->local_openclose_args =
1197 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1199 &part->local_openclose_args_base);
1200 if (part->local_openclose_args == NULL) {
1201 dev_err(xpc_chan, "can't get memory for local connect args\n");
1202 retval = xpNoMemory;
1206 part->remote_openclose_args =
1207 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE,
1209 &part->remote_openclose_args_base);
1210 if (part->remote_openclose_args == NULL) {
1211 dev_err(xpc_chan, "can't get memory for remote connect args\n");
1212 retval = xpNoMemory;
1216 part_sn2->remote_openclose_args_pa = 0;
1218 part_sn2->local_chctl_amo_va = xpc_init_IRQ_amo_sn2(partid);
1219 part->chctl.all_flags = 0;
1220 spin_lock_init(&part->chctl_lock);
1222 part_sn2->notify_IRQ_nasid = 0;
1223 part_sn2->notify_IRQ_phys_cpuid = 0;
1224 part_sn2->remote_chctl_amo_va = NULL;
1226 atomic_set(&part->channel_mgr_requests, 1);
1227 init_waitqueue_head(&part->channel_mgr_wq);
1229 sprintf(part_sn2->notify_IRQ_owner, "xpc%02d", partid);
1230 ret = request_irq(SGI_XPC_NOTIFY, xpc_handle_notify_IRQ_sn2,
1231 IRQF_SHARED, part_sn2->notify_IRQ_owner,
1232 (void *)(u64)partid);
1234 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
1235 "errno=%d\n", -ret);
1236 retval = xpLackOfResources;
1240 /* Setup a timer to check for dropped notify IRQs */
1241 timer = &part_sn2->dropped_notify_IRQ_timer;
1244 (void (*)(unsigned long))xpc_check_for_dropped_notify_IRQ_sn2;
1245 timer->data = (unsigned long)part;
1246 timer->expires = jiffies + XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL;
1249 part->nchannels = XPC_MAX_NCHANNELS;
1251 atomic_set(&part->nchannels_active, 0);
1252 atomic_set(&part->nchannels_engaged, 0);
1254 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1255 ch = &part->channels[ch_number];
1257 ch->partid = partid;
1258 ch->number = ch_number;
1259 ch->flags = XPC_C_DISCONNECTED;
1261 ch->sn.sn2.local_GP = &part_sn2->local_GPs[ch_number];
1262 ch->local_openclose_args =
1263 &part->local_openclose_args[ch_number];
1265 atomic_set(&ch->kthreads_assigned, 0);
1266 atomic_set(&ch->kthreads_idle, 0);
1267 atomic_set(&ch->kthreads_active, 0);
1269 atomic_set(&ch->references, 0);
1270 atomic_set(&ch->n_to_notify, 0);
1272 spin_lock_init(&ch->lock);
1273 mutex_init(&ch->sn.sn2.msg_to_pull_mutex);
1274 init_completion(&ch->wdisconnect_wait);
1276 atomic_set(&ch->n_on_msg_allocate_wq, 0);
1277 init_waitqueue_head(&ch->msg_allocate_wq);
1278 init_waitqueue_head(&ch->idle_wq);
1282 * With the setting of the partition setup_state to XPC_P_SETUP, we're
1283 * declaring that this partition is ready to go.
1285 part->setup_state = XPC_P_SETUP;
1288 * Setup the per partition specific variables required by the
1289 * remote partition to establish channel connections with us.
1291 * The setting of the magic # indicates that these per partition
1292 * specific variables are ready to be used.
1294 xpc_vars_part_sn2[partid].GPs_pa = __pa(part_sn2->local_GPs);
1295 xpc_vars_part_sn2[partid].openclose_args_pa =
1296 __pa(part->local_openclose_args);
1297 xpc_vars_part_sn2[partid].chctl_amo_pa =
1298 __pa(part_sn2->local_chctl_amo_va);
1299 cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
1300 xpc_vars_part_sn2[partid].notify_IRQ_nasid = cpuid_to_nasid(cpuid);
1301 xpc_vars_part_sn2[partid].notify_IRQ_phys_cpuid =
1302 cpu_physical_id(cpuid);
1303 xpc_vars_part_sn2[partid].nchannels = part->nchannels;
1304 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC1;
1308 /* setup of infrastructure failed */
1310 kfree(part->remote_openclose_args_base);
1311 part->remote_openclose_args = NULL;
1313 kfree(part->local_openclose_args_base);
1314 part->local_openclose_args = NULL;
1316 kfree(part_sn2->remote_GPs_base);
1317 part_sn2->remote_GPs = NULL;
1319 kfree(part_sn2->local_GPs_base);
1320 part_sn2->local_GPs = NULL;
1322 kfree(part->channels);
1323 part->channels = NULL;
1328 * Teardown the infrastructure necessary to support XPartition Communication
1329 * between the specified remote partition and the local one.
1332 xpc_teardown_infrastructure_sn2(struct xpc_partition *part)
1334 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1335 short partid = XPC_PARTID(part);
1338 * We start off by making this partition inaccessible to local
1339 * processes by marking it as no longer setup. Then we make it
1340 * inaccessible to remote processes by clearing the XPC per partition
1341 * specific variable's magic # (which indicates that these variables
1342 * are no longer valid) and by ignoring all XPC notify IRQs sent to
1346 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1347 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1348 DBUG_ON(part->setup_state != XPC_P_SETUP);
1349 part->setup_state = XPC_P_WTEARDOWN;
1351 xpc_vars_part_sn2[partid].magic = 0;
1353 free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1356 * Before proceeding with the teardown we have to wait until all
1357 * existing references cease.
1359 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1361 /* now we can begin tearing down the infrastructure */
1363 part->setup_state = XPC_P_TORNDOWN;
1365 /* in case we've still got outstanding timers registered... */
1366 del_timer_sync(&part_sn2->dropped_notify_IRQ_timer);
1368 kfree(part->remote_openclose_args_base);
1369 part->remote_openclose_args = NULL;
1370 kfree(part->local_openclose_args_base);
1371 part->local_openclose_args = NULL;
1372 kfree(part_sn2->remote_GPs_base);
1373 part_sn2->remote_GPs = NULL;
1374 kfree(part_sn2->local_GPs_base);
1375 part_sn2->local_GPs = NULL;
1376 kfree(part->channels);
1377 part->channels = NULL;
1378 part_sn2->local_chctl_amo_va = NULL;
1382 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
1383 * (or multiple cachelines) from a remote partition.
1385 * src must be a cacheline aligned physical address on the remote partition.
1386 * dst must be a cacheline aligned virtual address on this partition.
1387 * cnt must be cacheline sized
1389 /* ??? Replace this function by call to xp_remote_memcpy() or bte_copy()? */
1390 static enum xp_retval
1391 xpc_pull_remote_cachelines_sn2(struct xpc_partition *part, void *dst,
1392 const void *src, size_t cnt)
1396 DBUG_ON((u64)src != L1_CACHE_ALIGN((u64)src));
1397 DBUG_ON((u64)dst != L1_CACHE_ALIGN((u64)dst));
1398 DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
1400 if (part->act_state == XPC_P_DEACTIVATING)
1401 return part->reason;
1403 ret = xp_remote_memcpy(dst, src, cnt);
1404 if (ret != xpSuccess) {
1405 dev_dbg(xpc_chan, "xp_remote_memcpy() from partition %d failed,"
1406 " ret=%d\n", XPC_PARTID(part), ret);
1412 * Pull the remote per partition specific variables from the specified
1415 static enum xp_retval
1416 xpc_pull_remote_vars_part_sn2(struct xpc_partition *part)
1418 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1419 u8 buffer[L1_CACHE_BYTES * 2];
1420 struct xpc_vars_part_sn2 *pulled_entry_cacheline =
1421 (struct xpc_vars_part_sn2 *)L1_CACHE_ALIGN((u64)buffer);
1422 struct xpc_vars_part_sn2 *pulled_entry;
1423 u64 remote_entry_cacheline_pa, remote_entry_pa;
1424 short partid = XPC_PARTID(part);
1427 /* pull the cacheline that contains the variables we're interested in */
1429 DBUG_ON(part_sn2->remote_vars_part_pa !=
1430 L1_CACHE_ALIGN(part_sn2->remote_vars_part_pa));
1431 DBUG_ON(sizeof(struct xpc_vars_part_sn2) != L1_CACHE_BYTES / 2);
1433 remote_entry_pa = part_sn2->remote_vars_part_pa +
1434 sn_partition_id * sizeof(struct xpc_vars_part_sn2);
1436 remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
1438 pulled_entry = (struct xpc_vars_part_sn2 *)((u64)pulled_entry_cacheline
1439 + (remote_entry_pa &
1440 (L1_CACHE_BYTES - 1)));
1442 ret = xpc_pull_remote_cachelines_sn2(part, pulled_entry_cacheline,
1443 (void *)remote_entry_cacheline_pa,
1445 if (ret != xpSuccess) {
1446 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
1447 "partition %d, ret=%d\n", partid, ret);
1451 /* see if they've been set up yet */
1453 if (pulled_entry->magic != XPC_VP_MAGIC1 &&
1454 pulled_entry->magic != XPC_VP_MAGIC2) {
1456 if (pulled_entry->magic != 0) {
1457 dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
1458 "partition %d has bad magic value (=0x%lx)\n",
1459 partid, sn_partition_id, pulled_entry->magic);
1463 /* they've not been initialized yet */
1467 if (xpc_vars_part_sn2[partid].magic == XPC_VP_MAGIC1) {
1469 /* validate the variables */
1471 if (pulled_entry->GPs_pa == 0 ||
1472 pulled_entry->openclose_args_pa == 0 ||
1473 pulled_entry->chctl_amo_pa == 0) {
1475 dev_err(xpc_chan, "partition %d's XPC vars_part for "
1476 "partition %d are not valid\n", partid,
1478 return xpInvalidAddress;
1481 /* the variables we imported look to be valid */
1483 part_sn2->remote_GPs_pa = pulled_entry->GPs_pa;
1484 part_sn2->remote_openclose_args_pa =
1485 pulled_entry->openclose_args_pa;
1486 part_sn2->remote_chctl_amo_va =
1487 (struct amo *)__va(pulled_entry->chctl_amo_pa);
1488 part_sn2->notify_IRQ_nasid = pulled_entry->notify_IRQ_nasid;
1489 part_sn2->notify_IRQ_phys_cpuid =
1490 pulled_entry->notify_IRQ_phys_cpuid;
1492 if (part->nchannels > pulled_entry->nchannels)
1493 part->nchannels = pulled_entry->nchannels;
1495 /* let the other side know that we've pulled their variables */
1497 xpc_vars_part_sn2[partid].magic = XPC_VP_MAGIC2;
1500 if (pulled_entry->magic == XPC_VP_MAGIC1)
1507 * Establish first contact with the remote partititon. This involves pulling
1508 * the XPC per partition variables from the remote partition and waiting for
1509 * the remote partition to pull ours.
1511 static enum xp_retval
1512 xpc_make_first_contact_sn2(struct xpc_partition *part)
1514 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1518 * Register the remote partition's amos with SAL so it can handle
1519 * and cleanup errors within that address range should the remote
1520 * partition go down. We don't unregister this range because it is
1521 * difficult to tell when outstanding writes to the remote partition
1522 * are finished and thus when it is safe to unregister. This should
1523 * not result in wasted space in the SAL xp_addr_region table because
1524 * we should get the same page for remote_amos_page_pa after module
1525 * reloads and system reboots.
1527 if (sn_register_xp_addr_region(part_sn2->remote_amos_page_pa,
1528 PAGE_SIZE, 1) < 0) {
1529 dev_warn(xpc_part, "xpc_activating(%d) failed to register "
1530 "xp_addr region\n", XPC_PARTID(part));
1532 ret = xpPhysAddrRegFailed;
1533 XPC_DEACTIVATE_PARTITION(part, ret);
1538 * Send activate IRQ to get other side to activate if they've not
1539 * already begun to do so.
1541 xpc_send_activate_IRQ_sn2(part_sn2->remote_amos_page_pa,
1542 cnodeid_to_nasid(0),
1543 part_sn2->activate_IRQ_nasid,
1544 part_sn2->activate_IRQ_phys_cpuid);
1546 while ((ret = xpc_pull_remote_vars_part_sn2(part)) != xpSuccess) {
1547 if (ret != xpRetry) {
1548 XPC_DEACTIVATE_PARTITION(part, ret);
1552 dev_dbg(xpc_part, "waiting to make first contact with "
1553 "partition %d\n", XPC_PARTID(part));
1555 /* wait a 1/4 of a second or so */
1556 (void)msleep_interruptible(250);
1558 if (part->act_state == XPC_P_DEACTIVATING)
1559 return part->reason;
1566 * Get the chctl flags and pull the openclose args and/or remote GPs as needed.
1569 xpc_get_chctl_all_flags_sn2(struct xpc_partition *part)
1571 struct xpc_partition_sn2 *part_sn2 = &part->sn.sn2;
1572 unsigned long irq_flags;
1573 union xpc_channel_ctl_flags chctl;
1577 * See if there are any chctl flags to be handled.
1580 spin_lock_irqsave(&part->chctl_lock, irq_flags);
1581 chctl = part->chctl;
1582 if (chctl.all_flags != 0)
1583 part->chctl.all_flags = 0;
1585 spin_unlock_irqrestore(&part->chctl_lock, irq_flags);
1587 if (xpc_any_openclose_chctl_flags_set(&chctl)) {
1588 ret = xpc_pull_remote_cachelines_sn2(part, part->
1589 remote_openclose_args,
1591 remote_openclose_args_pa,
1592 XPC_OPENCLOSE_ARGS_SIZE);
1593 if (ret != xpSuccess) {
1594 XPC_DEACTIVATE_PARTITION(part, ret);
1596 dev_dbg(xpc_chan, "failed to pull openclose args from "
1597 "partition %d, ret=%d\n", XPC_PARTID(part),
1600 /* don't bother processing chctl flags anymore */
1601 chctl.all_flags = 0;
1605 if (xpc_any_msg_chctl_flags_set(&chctl)) {
1606 ret = xpc_pull_remote_cachelines_sn2(part, part_sn2->remote_GPs,
1607 (void *)part_sn2->remote_GPs_pa,
1609 if (ret != xpSuccess) {
1610 XPC_DEACTIVATE_PARTITION(part, ret);
1612 dev_dbg(xpc_chan, "failed to pull GPs from partition "
1613 "%d, ret=%d\n", XPC_PARTID(part), ret);
1615 /* don't bother processing chctl flags anymore */
1616 chctl.all_flags = 0;
1620 return chctl.all_flags;
1624 * Allocate the local message queue and the notify queue.
1626 static enum xp_retval
1627 xpc_allocate_local_msgqueue_sn2(struct xpc_channel *ch)
1629 unsigned long irq_flags;
1633 for (nentries = ch->local_nentries; nentries > 0; nentries--) {
1635 nbytes = nentries * ch->msg_size;
1636 ch->local_msgqueue =
1637 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1638 &ch->local_msgqueue_base);
1639 if (ch->local_msgqueue == NULL)
1642 nbytes = nentries * sizeof(struct xpc_notify);
1643 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
1644 if (ch->notify_queue == NULL) {
1645 kfree(ch->local_msgqueue_base);
1646 ch->local_msgqueue = NULL;
1650 spin_lock_irqsave(&ch->lock, irq_flags);
1651 if (nentries < ch->local_nentries) {
1652 dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
1653 "partid=%d, channel=%d\n", nentries,
1654 ch->local_nentries, ch->partid, ch->number);
1656 ch->local_nentries = nentries;
1658 spin_unlock_irqrestore(&ch->lock, irq_flags);
1662 dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
1663 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
1668 * Allocate the cached remote message queue.
1670 static enum xp_retval
1671 xpc_allocate_remote_msgqueue_sn2(struct xpc_channel *ch)
1673 unsigned long irq_flags;
1677 DBUG_ON(ch->remote_nentries <= 0);
1679 for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
1681 nbytes = nentries * ch->msg_size;
1682 ch->remote_msgqueue =
1683 xpc_kzalloc_cacheline_aligned_sn2(nbytes, GFP_KERNEL,
1684 &ch->remote_msgqueue_base);
1685 if (ch->remote_msgqueue == NULL)
1688 spin_lock_irqsave(&ch->lock, irq_flags);
1689 if (nentries < ch->remote_nentries) {
1690 dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
1691 "partid=%d, channel=%d\n", nentries,
1692 ch->remote_nentries, ch->partid, ch->number);
1694 ch->remote_nentries = nentries;
1696 spin_unlock_irqrestore(&ch->lock, irq_flags);
1700 dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
1701 "partid=%d, channel=%d\n", ch->partid, ch->number);
1706 * Allocate message queues and other stuff associated with a channel.
1708 * Note: Assumes all of the channel sizes are filled in.
1710 static enum xp_retval
1711 xpc_allocate_msgqueues_sn2(struct xpc_channel *ch)
1715 DBUG_ON(ch->flags & XPC_C_SETUP);
1717 ret = xpc_allocate_local_msgqueue_sn2(ch);
1718 if (ret == xpSuccess) {
1720 ret = xpc_allocate_remote_msgqueue_sn2(ch);
1721 if (ret != xpSuccess) {
1722 kfree(ch->local_msgqueue_base);
1723 ch->local_msgqueue = NULL;
1724 kfree(ch->notify_queue);
1725 ch->notify_queue = NULL;
1732 * Free up message queues and other stuff that were allocated for the specified
1735 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
1736 * they're cleared when XPC_C_DISCONNECTED is cleared.
1739 xpc_free_msgqueues_sn2(struct xpc_channel *ch)
1741 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1743 DBUG_ON(!spin_is_locked(&ch->lock));
1744 DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
1746 ch->remote_msgqueue_pa = 0;
1750 ch->local_nentries = 0;
1751 ch->remote_nentries = 0;
1752 ch->kthreads_assigned_limit = 0;
1753 ch->kthreads_idle_limit = 0;
1755 ch_sn2->local_GP->get = 0;
1756 ch_sn2->local_GP->put = 0;
1757 ch_sn2->remote_GP.get = 0;
1758 ch_sn2->remote_GP.put = 0;
1759 ch_sn2->w_local_GP.get = 0;
1760 ch_sn2->w_local_GP.put = 0;
1761 ch_sn2->w_remote_GP.get = 0;
1762 ch_sn2->w_remote_GP.put = 0;
1763 ch_sn2->next_msg_to_pull = 0;
1765 if (ch->flags & XPC_C_SETUP) {
1766 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
1767 ch->flags, ch->partid, ch->number);
1769 kfree(ch->local_msgqueue_base);
1770 ch->local_msgqueue = NULL;
1771 kfree(ch->remote_msgqueue_base);
1772 ch->remote_msgqueue = NULL;
1773 kfree(ch->notify_queue);
1774 ch->notify_queue = NULL;
1779 * Notify those who wanted to be notified upon delivery of their message.
1782 xpc_notify_senders_sn2(struct xpc_channel *ch, enum xp_retval reason, s64 put)
1784 struct xpc_notify *notify;
1786 s64 get = ch->sn.sn2.w_remote_GP.get - 1;
1788 while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
1790 notify = &ch->notify_queue[get % ch->local_nentries];
1793 * See if the notify entry indicates it was associated with
1794 * a message who's sender wants to be notified. It is possible
1795 * that it is, but someone else is doing or has done the
1798 notify_type = notify->type;
1799 if (notify_type == 0 ||
1800 cmpxchg(¬ify->type, notify_type, 0) != notify_type) {
1804 DBUG_ON(notify_type != XPC_N_CALL);
1806 atomic_dec(&ch->n_to_notify);
1808 if (notify->func != NULL) {
1809 dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
1810 "msg_number=%ld, partid=%d, channel=%d\n",
1811 (void *)notify, get, ch->partid, ch->number);
1813 notify->func(reason, ch->partid, ch->number,
1816 dev_dbg(xpc_chan, "notify->func() returned, "
1817 "notify=0x%p, msg_number=%ld, partid=%d, "
1818 "channel=%d\n", (void *)notify, get,
1819 ch->partid, ch->number);
1825 xpc_notify_senders_of_disconnect_sn2(struct xpc_channel *ch)
1827 xpc_notify_senders_sn2(ch, ch->reason, ch->sn.sn2.w_local_GP.put);
1831 * Clear some of the msg flags in the local message queue.
1834 xpc_clear_local_msgqueue_flags_sn2(struct xpc_channel *ch)
1836 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1837 struct xpc_msg *msg;
1840 get = ch_sn2->w_remote_GP.get;
1842 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1843 (get % ch->local_nentries) *
1846 } while (++get < ch_sn2->remote_GP.get);
1850 * Clear some of the msg flags in the remote message queue.
1853 xpc_clear_remote_msgqueue_flags_sn2(struct xpc_channel *ch)
1855 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1856 struct xpc_msg *msg;
1859 put = ch_sn2->w_remote_GP.put;
1861 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1862 (put % ch->remote_nentries) *
1865 } while (++put < ch_sn2->remote_GP.put);
1869 xpc_process_msg_chctl_flags_sn2(struct xpc_partition *part, int ch_number)
1871 struct xpc_channel *ch = &part->channels[ch_number];
1872 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1875 ch_sn2->remote_GP = part->sn.sn2.remote_GPs[ch_number];
1877 /* See what, if anything, has changed for each connected channel */
1879 xpc_msgqueue_ref(ch);
1881 if (ch_sn2->w_remote_GP.get == ch_sn2->remote_GP.get &&
1882 ch_sn2->w_remote_GP.put == ch_sn2->remote_GP.put) {
1883 /* nothing changed since GPs were last pulled */
1884 xpc_msgqueue_deref(ch);
1888 if (!(ch->flags & XPC_C_CONNECTED)) {
1889 xpc_msgqueue_deref(ch);
1894 * First check to see if messages recently sent by us have been
1895 * received by the other side. (The remote GET value will have
1896 * changed since we last looked at it.)
1899 if (ch_sn2->w_remote_GP.get != ch_sn2->remote_GP.get) {
1902 * We need to notify any senders that want to be notified
1903 * that their sent messages have been received by their
1904 * intended recipients. We need to do this before updating
1905 * w_remote_GP.get so that we don't allocate the same message
1906 * queue entries prematurely (see xpc_allocate_msg()).
1908 if (atomic_read(&ch->n_to_notify) > 0) {
1910 * Notify senders that messages sent have been
1911 * received and delivered by the other side.
1913 xpc_notify_senders_sn2(ch, xpMsgDelivered,
1914 ch_sn2->remote_GP.get);
1918 * Clear msg->flags in previously sent messages, so that
1919 * they're ready for xpc_allocate_msg().
1921 xpc_clear_local_msgqueue_flags_sn2(ch);
1923 ch_sn2->w_remote_GP.get = ch_sn2->remote_GP.get;
1925 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1926 "channel=%d\n", ch_sn2->w_remote_GP.get, ch->partid,
1930 * If anyone was waiting for message queue entries to become
1931 * available, wake them up.
1933 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1934 wake_up(&ch->msg_allocate_wq);
1938 * Now check for newly sent messages by the other side. (The remote
1939 * PUT value will have changed since we last looked at it.)
1942 if (ch_sn2->w_remote_GP.put != ch_sn2->remote_GP.put) {
1944 * Clear msg->flags in previously received messages, so that
1945 * they're ready for xpc_get_deliverable_msg().
1947 xpc_clear_remote_msgqueue_flags_sn2(ch);
1949 ch_sn2->w_remote_GP.put = ch_sn2->remote_GP.put;
1951 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1952 "channel=%d\n", ch_sn2->w_remote_GP.put, ch->partid,
1955 nmsgs_sent = ch_sn2->w_remote_GP.put - ch_sn2->w_local_GP.get;
1956 if (nmsgs_sent > 0) {
1957 dev_dbg(xpc_chan, "msgs waiting to be copied and "
1958 "delivered=%d, partid=%d, channel=%d\n",
1959 nmsgs_sent, ch->partid, ch->number);
1961 if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1962 xpc_activate_kthreads(ch, nmsgs_sent);
1966 xpc_msgqueue_deref(ch);
1969 static struct xpc_msg *
1970 xpc_pull_remote_msg_sn2(struct xpc_channel *ch, s64 get)
1972 struct xpc_partition *part = &xpc_partitions[ch->partid];
1973 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
1974 struct xpc_msg *remote_msg, *msg;
1975 u32 msg_index, nmsgs;
1979 if (mutex_lock_interruptible(&ch_sn2->msg_to_pull_mutex) != 0) {
1980 /* we were interrupted by a signal */
1984 while (get >= ch_sn2->next_msg_to_pull) {
1986 /* pull as many messages as are ready and able to be pulled */
1988 msg_index = ch_sn2->next_msg_to_pull % ch->remote_nentries;
1990 DBUG_ON(ch_sn2->next_msg_to_pull >= ch_sn2->w_remote_GP.put);
1991 nmsgs = ch_sn2->w_remote_GP.put - ch_sn2->next_msg_to_pull;
1992 if (msg_index + nmsgs > ch->remote_nentries) {
1993 /* ignore the ones that wrap the msg queue for now */
1994 nmsgs = ch->remote_nentries - msg_index;
1997 msg_offset = msg_index * ch->msg_size;
1998 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
1999 remote_msg = (struct xpc_msg *)(ch->remote_msgqueue_pa +
2002 ret = xpc_pull_remote_cachelines_sn2(part, msg, remote_msg,
2003 nmsgs * ch->msg_size);
2004 if (ret != xpSuccess) {
2006 dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2007 " msg %ld from partition %d, channel=%d, "
2008 "ret=%d\n", nmsgs, ch_sn2->next_msg_to_pull,
2009 ch->partid, ch->number, ret);
2011 XPC_DEACTIVATE_PARTITION(part, ret);
2013 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
2017 ch_sn2->next_msg_to_pull += nmsgs;
2020 mutex_unlock(&ch_sn2->msg_to_pull_mutex);
2022 /* return the message we were looking for */
2023 msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2024 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2030 xpc_n_of_deliverable_msgs_sn2(struct xpc_channel *ch)
2032 return ch->sn.sn2.w_remote_GP.put - ch->sn.sn2.w_local_GP.get;
2036 * Get a message to be delivered.
2038 static struct xpc_msg *
2039 xpc_get_deliverable_msg_sn2(struct xpc_channel *ch)
2041 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2042 struct xpc_msg *msg = NULL;
2046 if (ch->flags & XPC_C_DISCONNECTING)
2049 get = ch_sn2->w_local_GP.get;
2050 rmb(); /* guarantee that .get loads before .put */
2051 if (get == ch_sn2->w_remote_GP.put)
2054 /* There are messages waiting to be pulled and delivered.
2055 * We need to try to secure one for ourselves. We'll do this
2056 * by trying to increment w_local_GP.get and hope that no one
2057 * else beats us to it. If they do, we'll we'll simply have
2058 * to try again for the next one.
2061 if (cmpxchg(&ch_sn2->w_local_GP.get, get, get + 1) == get) {
2062 /* we got the entry referenced by get */
2064 dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2065 "partid=%d, channel=%d\n", get + 1,
2066 ch->partid, ch->number);
2068 /* pull the message from the remote partition */
2070 msg = xpc_pull_remote_msg_sn2(ch, get);
2072 DBUG_ON(msg != NULL && msg->number != get);
2073 DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2074 DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2085 * Now we actually send the messages that are ready to be sent by advancing
2086 * the local message queue's Put value and then send a chctl msgrequest to the
2087 * recipient partition.
2090 xpc_send_msgs_sn2(struct xpc_channel *ch, s64 initial_put)
2092 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2093 struct xpc_msg *msg;
2094 s64 put = initial_put + 1;
2095 int send_msgrequest = 0;
2100 if (put == ch_sn2->w_local_GP.put)
2103 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2104 (put % ch->local_nentries) *
2107 if (!(msg->flags & XPC_M_READY))
2113 if (put == initial_put) {
2114 /* nothing's changed */
2118 if (cmpxchg_rel(&ch_sn2->local_GP->put, initial_put, put) !=
2120 /* someone else beat us to it */
2121 DBUG_ON(ch_sn2->local_GP->put < initial_put);
2125 /* we just set the new value of local_GP->put */
2127 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
2128 "channel=%d\n", put, ch->partid, ch->number);
2130 send_msgrequest = 1;
2133 * We need to ensure that the message referenced by
2134 * local_GP->put is not XPC_M_READY or that local_GP->put
2135 * equals w_local_GP.put, so we'll go have a look.
2140 if (send_msgrequest)
2141 xpc_send_chctl_msgrequest_sn2(ch);
2145 * Allocate an entry for a message from the message queue associated with the
2146 * specified channel.
2148 static enum xp_retval
2149 xpc_allocate_msg_sn2(struct xpc_channel *ch, u32 flags,
2150 struct xpc_msg **address_of_msg)
2152 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2153 struct xpc_msg *msg;
2158 * Get the next available message entry from the local message queue.
2159 * If none are available, we'll make sure that we grab the latest
2166 put = ch_sn2->w_local_GP.put;
2167 rmb(); /* guarantee that .put loads before .get */
2168 if (put - ch_sn2->w_remote_GP.get < ch->local_nentries) {
2170 /* There are available message entries. We need to try
2171 * to secure one for ourselves. We'll do this by trying
2172 * to increment w_local_GP.put as long as someone else
2173 * doesn't beat us to it. If they do, we'll have to
2176 if (cmpxchg(&ch_sn2->w_local_GP.put, put, put + 1) ==
2178 /* we got the entry referenced by put */
2181 continue; /* try again */
2185 * There aren't any available msg entries at this time.
2187 * In waiting for a message entry to become available,
2188 * we set a timeout in case the other side is not sending
2189 * completion interrupts. This lets us fake a notify IRQ
2190 * that will cause the notify IRQ handler to fetch the latest
2191 * GP values as if an interrupt was sent by the other side.
2193 if (ret == xpTimeout)
2194 xpc_send_chctl_local_msgrequest_sn2(ch);
2196 if (flags & XPC_NOWAIT)
2199 ret = xpc_allocate_msg_wait(ch);
2200 if (ret != xpInterrupted && ret != xpTimeout)
2204 /* get the message's address and initialize it */
2205 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
2206 (put % ch->local_nentries) * ch->msg_size);
2208 DBUG_ON(msg->flags != 0);
2211 dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
2212 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
2213 (void *)msg, msg->number, ch->partid, ch->number);
2215 *address_of_msg = msg;
2220 * Common code that does the actual sending of the message by advancing the
2221 * local message queue's Put value and sends a chctl msgrequest to the
2222 * partition the message is being sent to.
2224 static enum xp_retval
2225 xpc_send_msg_sn2(struct xpc_channel *ch, u32 flags, void *payload,
2226 u16 payload_size, u8 notify_type, xpc_notify_func func,
2229 enum xp_retval ret = xpSuccess;
2230 struct xpc_msg *msg = msg;
2231 struct xpc_notify *notify = notify;
2235 DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
2237 if (XPC_MSG_SIZE(payload_size) > ch->msg_size)
2238 return xpPayloadTooBig;
2240 xpc_msgqueue_ref(ch);
2242 if (ch->flags & XPC_C_DISCONNECTING) {
2246 if (!(ch->flags & XPC_C_CONNECTED)) {
2247 ret = xpNotConnected;
2251 ret = xpc_allocate_msg_sn2(ch, flags, &msg);
2252 if (ret != xpSuccess)
2255 msg_number = msg->number;
2257 if (notify_type != 0) {
2259 * Tell the remote side to send an ACK interrupt when the
2260 * message has been delivered.
2262 msg->flags |= XPC_M_INTERRUPT;
2264 atomic_inc(&ch->n_to_notify);
2266 notify = &ch->notify_queue[msg_number % ch->local_nentries];
2267 notify->func = func;
2269 notify->type = notify_type;
2271 /* ??? Is a mb() needed here? */
2273 if (ch->flags & XPC_C_DISCONNECTING) {
2275 * An error occurred between our last error check and
2276 * this one. We will try to clear the type field from
2277 * the notify entry. If we succeed then
2278 * xpc_disconnect_channel() didn't already process
2281 if (cmpxchg(¬ify->type, notify_type, 0) ==
2283 atomic_dec(&ch->n_to_notify);
2290 memcpy(&msg->payload, payload, payload_size);
2292 msg->flags |= XPC_M_READY;
2295 * The preceding store of msg->flags must occur before the following
2296 * load of local_GP->put.
2300 /* see if the message is next in line to be sent, if so send it */
2302 put = ch->sn.sn2.local_GP->put;
2303 if (put == msg_number)
2304 xpc_send_msgs_sn2(ch, put);
2307 xpc_msgqueue_deref(ch);
2312 * Now we actually acknowledge the messages that have been delivered and ack'd
2313 * by advancing the cached remote message queue's Get value and if requested
2314 * send a chctl msgrequest to the message sender's partition.
2317 xpc_acknowledge_msgs_sn2(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2319 struct xpc_channel_sn2 *ch_sn2 = &ch->sn.sn2;
2320 struct xpc_msg *msg;
2321 s64 get = initial_get + 1;
2322 int send_msgrequest = 0;
2327 if (get == ch_sn2->w_local_GP.get)
2330 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2331 (get % ch->remote_nentries) *
2334 if (!(msg->flags & XPC_M_DONE))
2337 msg_flags |= msg->flags;
2341 if (get == initial_get) {
2342 /* nothing's changed */
2346 if (cmpxchg_rel(&ch_sn2->local_GP->get, initial_get, get) !=
2348 /* someone else beat us to it */
2349 DBUG_ON(ch_sn2->local_GP->get <= initial_get);
2353 /* we just set the new value of local_GP->get */
2355 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2356 "channel=%d\n", get, ch->partid, ch->number);
2358 send_msgrequest = (msg_flags & XPC_M_INTERRUPT);
2361 * We need to ensure that the message referenced by
2362 * local_GP->get is not XPC_M_DONE or that local_GP->get
2363 * equals w_local_GP.get, so we'll go have a look.
2368 if (send_msgrequest)
2369 xpc_send_chctl_msgrequest_sn2(ch);
2373 xpc_received_msg_sn2(struct xpc_channel *ch, struct xpc_msg *msg)
2376 s64 msg_number = msg->number;
2378 dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2379 (void *)msg, msg_number, ch->partid, ch->number);
2381 DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2382 msg_number % ch->remote_nentries);
2383 DBUG_ON(msg->flags & XPC_M_DONE);
2385 msg->flags |= XPC_M_DONE;
2388 * The preceding store of msg->flags must occur before the following
2389 * load of local_GP->get.
2394 * See if this message is next in line to be acknowledged as having
2397 get = ch->sn.sn2.local_GP->get;
2398 if (get == msg_number)
2399 xpc_acknowledge_msgs_sn2(ch, get, msg->flags);
2408 xpc_get_partition_rsvd_page_pa = xpc_get_partition_rsvd_page_pa_sn2;
2409 xpc_rsvd_page_init = xpc_rsvd_page_init_sn2;
2410 xpc_increment_heartbeat = xpc_increment_heartbeat_sn2;
2411 xpc_offline_heartbeat = xpc_offline_heartbeat_sn2;
2412 xpc_online_heartbeat = xpc_online_heartbeat_sn2;
2413 xpc_heartbeat_init = xpc_heartbeat_init_sn2;
2414 xpc_heartbeat_exit = xpc_heartbeat_exit_sn2;
2415 xpc_check_remote_hb = xpc_check_remote_hb_sn2;
2417 xpc_request_partition_activation = xpc_request_partition_activation_sn2;
2418 xpc_request_partition_reactivation =
2419 xpc_request_partition_reactivation_sn2;
2420 xpc_request_partition_deactivation =
2421 xpc_request_partition_deactivation_sn2;
2422 xpc_cancel_partition_deactivation_request =
2423 xpc_cancel_partition_deactivation_request_sn2;
2425 xpc_process_activate_IRQ_rcvd = xpc_process_activate_IRQ_rcvd_sn2;
2426 xpc_setup_infrastructure = xpc_setup_infrastructure_sn2;
2427 xpc_teardown_infrastructure = xpc_teardown_infrastructure_sn2;
2428 xpc_make_first_contact = xpc_make_first_contact_sn2;
2429 xpc_get_chctl_all_flags = xpc_get_chctl_all_flags_sn2;
2430 xpc_allocate_msgqueues = xpc_allocate_msgqueues_sn2;
2431 xpc_free_msgqueues = xpc_free_msgqueues_sn2;
2432 xpc_notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_sn2;
2433 xpc_process_msg_chctl_flags = xpc_process_msg_chctl_flags_sn2;
2434 xpc_n_of_deliverable_msgs = xpc_n_of_deliverable_msgs_sn2;
2435 xpc_get_deliverable_msg = xpc_get_deliverable_msg_sn2;
2437 xpc_indicate_partition_engaged = xpc_indicate_partition_engaged_sn2;
2438 xpc_partition_engaged = xpc_partition_engaged_sn2;
2439 xpc_any_partition_engaged = xpc_any_partition_engaged_sn2;
2440 xpc_indicate_partition_disengaged =
2441 xpc_indicate_partition_disengaged_sn2;
2442 xpc_assume_partition_disengaged = xpc_assume_partition_disengaged_sn2;
2444 xpc_send_chctl_closerequest = xpc_send_chctl_closerequest_sn2;
2445 xpc_send_chctl_closereply = xpc_send_chctl_closereply_sn2;
2446 xpc_send_chctl_openrequest = xpc_send_chctl_openrequest_sn2;
2447 xpc_send_chctl_openreply = xpc_send_chctl_openreply_sn2;
2449 xpc_send_msg = xpc_send_msg_sn2;
2450 xpc_received_msg = xpc_received_msg_sn2;
2452 buf_size = max(XPC_RP_VARS_SIZE,
2453 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES_SN2);
2454 xpc_remote_copy_buffer_sn2 = xpc_kmalloc_cacheline_aligned(buf_size,
2456 &xpc_remote_copy_buffer_base_sn2);
2457 if (xpc_remote_copy_buffer_sn2 == NULL) {
2458 dev_err(xpc_part, "can't get memory for remote copy buffer\n");
2462 /* open up protections for IPI and [potentially] amo operations */
2463 xpc_allow_IPI_ops_sn2();
2464 xpc_allow_amo_ops_shub_wars_1_1_sn2();
2467 * This is safe to do before the xpc_hb_checker thread has started
2468 * because the handler releases a wait queue. If an interrupt is
2469 * received before the thread is waiting, it will not go to sleep,
2470 * but rather immediately process the interrupt.
2472 ret = request_irq(SGI_XPC_ACTIVATE, xpc_handle_activate_IRQ_sn2, 0,
2475 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
2476 "errno=%d\n", -ret);
2477 xpc_disallow_IPI_ops_sn2();
2478 kfree(xpc_remote_copy_buffer_base_sn2);
2486 free_irq(SGI_XPC_ACTIVATE, NULL);
2487 xpc_disallow_IPI_ops_sn2();
2488 kfree(xpc_remote_copy_buffer_base_sn2);