2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
10 * Cross Partition Communication (XPC) sn2-based functions.
12 * Architecture specific implementation of common functions.
16 #include <linux/delay.h>
17 #include <asm/uncached.h>
18 #include <asm/sn/mspec.h>
19 #include <asm/sn/sn_sal.h>
23 * Define the number of u64s required to represent all the C-brick nasids
24 * as a bitmap. The cross-partition kernel modules deal only with
25 * C-brick nasids, thus the need for bitmaps which don't account for
26 * odd-numbered (non C-brick) nasids.
28 #define XPC_MAX_PHYSNODES_SN2 (MAX_NUMALINK_NODES / 2)
29 #define XP_NASID_MASK_BYTES_SN2 ((XPC_MAX_PHYSNODES_SN2 + 7) / 8)
30 #define XP_NASID_MASK_WORDS_SN2 ((XPC_MAX_PHYSNODES_SN2 + 63) / 64)
33 * Memory for XPC's amo variables is allocated by the MSPEC driver. These
34 * pages are located in the lowest granule. The lowest granule uses 4k pages
35 * for cached references and an alternate TLB handler to never provide a
36 * cacheable mapping for the entire region. This will prevent speculative
37 * reading of cached copies of our lines from being issued which will cause
38 * a PI FSB Protocol error to be generated by the SHUB. For XPC, we need 64
39 * amo variables (based on XP_MAX_NPARTITIONS_SN2) to identify the senders of
40 * NOTIFY IRQs, 128 amo variables (based on XP_NASID_MASK_WORDS_SN2) to identify
41 * the senders of ACTIVATE IRQs, 1 amo variable to identify which remote
42 * partitions (i.e., XPCs) consider themselves currently engaged with the
43 * local XPC and 1 amo variable to request partition deactivation.
45 #define XPC_NOTIFY_IRQ_AMOS_SN2 0
46 #define XPC_ACTIVATE_IRQ_AMOS_SN2 (XPC_NOTIFY_IRQ_AMOS_SN2 + \
47 XP_MAX_NPARTITIONS_SN2)
48 #define XPC_ENGAGED_PARTITIONS_AMO_SN2 (XPC_ACTIVATE_IRQ_AMOS_SN2 + \
49 XP_NASID_MASK_WORDS_SN2)
50 #define XPC_DEACTIVATE_REQUEST_AMO_SN2 (XPC_ENGAGED_PARTITIONS_AMO_SN2 + 1)
53 * Buffer used to store a local copy of portions of a remote partition's
54 * reserved page (either its header and part_nasids mask, or its vars).
56 static char *xpc_remote_copy_buffer_sn2
;
57 static void *xpc_remote_copy_buffer_base_sn2
;
59 static struct xpc_vars_sn2
*xpc_vars_sn2
;
60 static struct xpc_vars_part_sn2
*xpc_vars_part_sn2
;
62 /* SH_IPI_ACCESS shub register value on startup */
63 static u64 xpc_sh1_IPI_access_sn2
;
64 static u64 xpc_sh2_IPI_access0_sn2
;
65 static u64 xpc_sh2_IPI_access1_sn2
;
66 static u64 xpc_sh2_IPI_access2_sn2
;
67 static u64 xpc_sh2_IPI_access3_sn2
;
70 * Change protections to allow IPI operations.
73 xpc_allow_IPI_ops_sn2(void)
78 /* !!! The following should get moved into SAL. */
80 xpc_sh2_IPI_access0_sn2
=
81 (u64
)HUB_L((u64
*)LOCAL_MMR_ADDR(SH2_IPI_ACCESS0
));
82 xpc_sh2_IPI_access1_sn2
=
83 (u64
)HUB_L((u64
*)LOCAL_MMR_ADDR(SH2_IPI_ACCESS1
));
84 xpc_sh2_IPI_access2_sn2
=
85 (u64
)HUB_L((u64
*)LOCAL_MMR_ADDR(SH2_IPI_ACCESS2
));
86 xpc_sh2_IPI_access3_sn2
=
87 (u64
)HUB_L((u64
*)LOCAL_MMR_ADDR(SH2_IPI_ACCESS3
));
89 for_each_online_node(node
) {
90 nasid
= cnodeid_to_nasid(node
);
91 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS0
),
93 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS1
),
95 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS2
),
97 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS3
),
101 xpc_sh1_IPI_access_sn2
=
102 (u64
)HUB_L((u64
*)LOCAL_MMR_ADDR(SH1_IPI_ACCESS
));
104 for_each_online_node(node
) {
105 nasid
= cnodeid_to_nasid(node
);
106 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH1_IPI_ACCESS
),
113 * Restrict protections to disallow IPI operations.
116 xpc_disallow_IPI_ops_sn2(void)
121 /* !!! The following should get moved into SAL. */
123 for_each_online_node(node
) {
124 nasid
= cnodeid_to_nasid(node
);
125 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS0
),
126 xpc_sh2_IPI_access0_sn2
);
127 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS1
),
128 xpc_sh2_IPI_access1_sn2
);
129 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS2
),
130 xpc_sh2_IPI_access2_sn2
);
131 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH2_IPI_ACCESS3
),
132 xpc_sh2_IPI_access3_sn2
);
135 for_each_online_node(node
) {
136 nasid
= cnodeid_to_nasid(node
);
137 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
, SH1_IPI_ACCESS
),
138 xpc_sh1_IPI_access_sn2
);
144 * The following set of functions are used for the sending and receiving of
145 * IRQs (also known as IPIs). There are two flavors of IRQs, one that is
146 * associated with partition activity (SGI_XPC_ACTIVATE) and the other that
147 * is associated with channel activity (SGI_XPC_NOTIFY).
151 xpc_receive_IRQ_amo_sn2(struct amo
*amo
)
153 return FETCHOP_LOAD_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_CLEAR
);
156 static enum xp_retval
157 xpc_send_IRQ_sn2(struct amo
*amo
, u64 flag
, int nasid
, int phys_cpuid
,
161 unsigned long irq_flags
;
163 local_irq_save(irq_flags
);
165 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_OR
, flag
);
166 sn_send_IPI_phys(nasid
, phys_cpuid
, vector
, 0);
169 * We must always use the nofault function regardless of whether we
170 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
171 * didn't, we'd never know that the other partition is down and would
172 * keep sending IRQs and amos to it until the heartbeat times out.
174 ret
= xp_nofault_PIOR((u64
*)GLOBAL_MMR_ADDR(NASID_GET(&amo
->variable
),
175 xp_nofault_PIOR_target
));
177 local_irq_restore(irq_flags
);
179 return (ret
== 0) ? xpSuccess
: xpPioReadError
;
183 xpc_init_IRQ_amo_sn2(int index
)
185 struct amo
*amo
= xpc_vars_sn2
->amos_page
+ index
;
187 (void)xpc_receive_IRQ_amo_sn2(amo
); /* clear amo variable */
192 * Functions associated with SGI_XPC_ACTIVATE IRQ.
196 * Notify the heartbeat check thread that an activate IRQ has been received.
199 xpc_handle_activate_IRQ_sn2(int irq
, void *dev_id
)
201 atomic_inc(&xpc_activate_IRQ_rcvd
);
202 wake_up_interruptible(&xpc_activate_IRQ_wq
);
207 * Flag the appropriate amo variable and send an IRQ to the specified node.
210 xpc_send_activate_IRQ_sn2(unsigned long amos_page_pa
, int from_nasid
,
211 int to_nasid
, int to_phys_cpuid
)
213 struct amo
*amos
= (struct amo
*)__va(amos_page_pa
+
214 (XPC_ACTIVATE_IRQ_AMOS_SN2
*
215 sizeof(struct amo
)));
217 (void)xpc_send_IRQ_sn2(&amos
[BIT_WORD(from_nasid
/ 2)],
218 BIT_MASK(from_nasid
/ 2), to_nasid
,
219 to_phys_cpuid
, SGI_XPC_ACTIVATE
);
223 xpc_send_local_activate_IRQ_sn2(int from_nasid
)
225 struct amo
*amos
= (struct amo
*)__va(xpc_vars_sn2
->amos_page_pa
+
226 (XPC_ACTIVATE_IRQ_AMOS_SN2
*
227 sizeof(struct amo
)));
229 /* fake the sending and receipt of an activate IRQ from remote nasid */
230 FETCHOP_STORE_OP(TO_AMO((u64
)&amos
[BIT_WORD(from_nasid
/ 2)].variable
),
231 FETCHOP_OR
, BIT_MASK(from_nasid
/ 2));
233 atomic_inc(&xpc_activate_IRQ_rcvd
);
234 wake_up_interruptible(&xpc_activate_IRQ_wq
);
238 * Functions associated with SGI_XPC_NOTIFY IRQ.
242 * Check to see if any chctl flags were sent from the specified partition.
245 xpc_check_for_sent_chctl_flags_sn2(struct xpc_partition
*part
)
247 union xpc_channel_ctl_flags chctl
;
248 unsigned long irq_flags
;
250 chctl
.all_flags
= xpc_receive_IRQ_amo_sn2(part
->sn
.sn2
.
252 if (chctl
.all_flags
== 0)
255 spin_lock_irqsave(&part
->chctl_lock
, irq_flags
);
256 part
->chctl
.all_flags
|= chctl
.all_flags
;
257 spin_unlock_irqrestore(&part
->chctl_lock
, irq_flags
);
259 dev_dbg(xpc_chan
, "received notify IRQ from partid=%d, chctl.all_flags="
260 "0x%lx\n", XPC_PARTID(part
), chctl
.all_flags
);
262 xpc_wakeup_channel_mgr(part
);
266 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
267 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
268 * than one partition, we use an amo structure per partition to indicate
269 * whether a partition has sent an IRQ or not. If it has, then wake up the
270 * associated kthread to handle it.
272 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IRQs sent by XPC
273 * running on other partitions.
275 * Noteworthy Arguments:
277 * irq - Interrupt ReQuest number. NOT USED.
279 * dev_id - partid of IRQ's potential sender.
282 xpc_handle_notify_IRQ_sn2(int irq
, void *dev_id
)
284 short partid
= (short)(u64
)dev_id
;
285 struct xpc_partition
*part
= &xpc_partitions
[partid
];
287 DBUG_ON(partid
< 0 || partid
>= XP_MAX_NPARTITIONS_SN2
);
289 if (xpc_part_ref(part
)) {
290 xpc_check_for_sent_chctl_flags_sn2(part
);
292 xpc_part_deref(part
);
298 * Check to see if xpc_handle_notify_IRQ_sn2() dropped any IRQs on the floor
299 * because the write to their associated amo variable completed after the IRQ
303 xpc_check_for_dropped_notify_IRQ_sn2(struct xpc_partition
*part
)
305 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
307 if (xpc_part_ref(part
)) {
308 xpc_check_for_sent_chctl_flags_sn2(part
);
310 part_sn2
->dropped_notify_IRQ_timer
.expires
= jiffies
+
311 XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL
;
312 add_timer(&part_sn2
->dropped_notify_IRQ_timer
);
313 xpc_part_deref(part
);
318 * Send a notify IRQ to the remote partition that is associated with the
322 xpc_send_notify_IRQ_sn2(struct xpc_channel
*ch
, u8 chctl_flag
,
323 char *chctl_flag_string
, unsigned long *irq_flags
)
325 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
326 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
327 union xpc_channel_ctl_flags chctl
= { 0 };
330 if (likely(part
->act_state
!= XPC_P_AS_DEACTIVATING
)) {
331 chctl
.flags
[ch
->number
] = chctl_flag
;
332 ret
= xpc_send_IRQ_sn2(part_sn2
->remote_chctl_amo_va
,
334 part_sn2
->notify_IRQ_nasid
,
335 part_sn2
->notify_IRQ_phys_cpuid
,
337 dev_dbg(xpc_chan
, "%s sent to partid=%d, channel=%d, ret=%d\n",
338 chctl_flag_string
, ch
->partid
, ch
->number
, ret
);
339 if (unlikely(ret
!= xpSuccess
)) {
340 if (irq_flags
!= NULL
)
341 spin_unlock_irqrestore(&ch
->lock
, *irq_flags
);
342 XPC_DEACTIVATE_PARTITION(part
, ret
);
343 if (irq_flags
!= NULL
)
344 spin_lock_irqsave(&ch
->lock
, *irq_flags
);
349 #define XPC_SEND_NOTIFY_IRQ_SN2(_ch, _ipi_f, _irq_f) \
350 xpc_send_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f, _irq_f)
353 * Make it look like the remote partition, which is associated with the
354 * specified channel, sent us a notify IRQ. This faked IRQ will be handled
355 * by xpc_check_for_dropped_notify_IRQ_sn2().
358 xpc_send_local_notify_IRQ_sn2(struct xpc_channel
*ch
, u8 chctl_flag
,
359 char *chctl_flag_string
)
361 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
362 union xpc_channel_ctl_flags chctl
= { 0 };
364 chctl
.flags
[ch
->number
] = chctl_flag
;
365 FETCHOP_STORE_OP(TO_AMO((u64
)&part
->sn
.sn2
.local_chctl_amo_va
->
366 variable
), FETCHOP_OR
, chctl
.all_flags
);
367 dev_dbg(xpc_chan
, "%s sent local from partid=%d, channel=%d\n",
368 chctl_flag_string
, ch
->partid
, ch
->number
);
371 #define XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(_ch, _ipi_f) \
372 xpc_send_local_notify_IRQ_sn2(_ch, _ipi_f, #_ipi_f)
375 xpc_send_chctl_closerequest_sn2(struct xpc_channel
*ch
,
376 unsigned long *irq_flags
)
378 struct xpc_openclose_args
*args
= ch
->local_openclose_args
;
380 args
->reason
= ch
->reason
;
381 XPC_SEND_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_CLOSEREQUEST
, irq_flags
);
385 xpc_send_chctl_closereply_sn2(struct xpc_channel
*ch
, unsigned long *irq_flags
)
387 XPC_SEND_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_CLOSEREPLY
, irq_flags
);
391 xpc_send_chctl_openrequest_sn2(struct xpc_channel
*ch
, unsigned long *irq_flags
)
393 struct xpc_openclose_args
*args
= ch
->local_openclose_args
;
395 args
->msg_size
= ch
->msg_size
;
396 args
->local_nentries
= ch
->local_nentries
;
397 XPC_SEND_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_OPENREQUEST
, irq_flags
);
401 xpc_send_chctl_openreply_sn2(struct xpc_channel
*ch
, unsigned long *irq_flags
)
403 struct xpc_openclose_args
*args
= ch
->local_openclose_args
;
405 args
->remote_nentries
= ch
->remote_nentries
;
406 args
->local_nentries
= ch
->local_nentries
;
407 args
->local_msgqueue_pa
= xp_pa(ch
->local_msgqueue
);
408 XPC_SEND_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_OPENREPLY
, irq_flags
);
412 xpc_send_chctl_msgrequest_sn2(struct xpc_channel
*ch
)
414 XPC_SEND_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_MSGREQUEST
, NULL
);
418 xpc_send_chctl_local_msgrequest_sn2(struct xpc_channel
*ch
)
420 XPC_SEND_LOCAL_NOTIFY_IRQ_SN2(ch
, XPC_CHCTL_MSGREQUEST
);
424 * This next set of functions are used to keep track of when a partition is
425 * potentially engaged in accessing memory belonging to another partition.
429 xpc_indicate_partition_engaged_sn2(struct xpc_partition
*part
)
431 unsigned long irq_flags
;
432 struct amo
*amo
= (struct amo
*)__va(part
->sn
.sn2
.remote_amos_page_pa
+
433 (XPC_ENGAGED_PARTITIONS_AMO_SN2
*
434 sizeof(struct amo
)));
436 local_irq_save(irq_flags
);
438 /* set bit corresponding to our partid in remote partition's amo */
439 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_OR
,
440 BIT(sn_partition_id
));
443 * We must always use the nofault function regardless of whether we
444 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
445 * didn't, we'd never know that the other partition is down and would
446 * keep sending IRQs and amos to it until the heartbeat times out.
448 (void)xp_nofault_PIOR((u64
*)GLOBAL_MMR_ADDR(NASID_GET(&amo
->
450 xp_nofault_PIOR_target
));
452 local_irq_restore(irq_flags
);
456 xpc_indicate_partition_disengaged_sn2(struct xpc_partition
*part
)
458 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
459 unsigned long irq_flags
;
460 struct amo
*amo
= (struct amo
*)__va(part_sn2
->remote_amos_page_pa
+
461 (XPC_ENGAGED_PARTITIONS_AMO_SN2
*
462 sizeof(struct amo
)));
464 local_irq_save(irq_flags
);
466 /* clear bit corresponding to our partid in remote partition's amo */
467 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_AND
,
468 ~BIT(sn_partition_id
));
471 * We must always use the nofault function regardless of whether we
472 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
473 * didn't, we'd never know that the other partition is down and would
474 * keep sending IRQs and amos to it until the heartbeat times out.
476 (void)xp_nofault_PIOR((u64
*)GLOBAL_MMR_ADDR(NASID_GET(&amo
->
478 xp_nofault_PIOR_target
));
480 local_irq_restore(irq_flags
);
483 * Send activate IRQ to get other side to see that we've cleared our
484 * bit in their engaged partitions amo.
486 xpc_send_activate_IRQ_sn2(part_sn2
->remote_amos_page_pa
,
488 part_sn2
->activate_IRQ_nasid
,
489 part_sn2
->activate_IRQ_phys_cpuid
);
493 xpc_partition_engaged_sn2(short partid
)
495 struct amo
*amo
= xpc_vars_sn2
->amos_page
+
496 XPC_ENGAGED_PARTITIONS_AMO_SN2
;
498 /* our partition's amo variable ANDed with partid mask */
499 return (FETCHOP_LOAD_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_LOAD
) &
504 xpc_any_partition_engaged_sn2(void)
506 struct amo
*amo
= xpc_vars_sn2
->amos_page
+
507 XPC_ENGAGED_PARTITIONS_AMO_SN2
;
509 /* our partition's amo variable */
510 return FETCHOP_LOAD_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_LOAD
) != 0;
514 xpc_assume_partition_disengaged_sn2(short partid
)
516 struct amo
*amo
= xpc_vars_sn2
->amos_page
+
517 XPC_ENGAGED_PARTITIONS_AMO_SN2
;
519 /* clear bit(s) based on partid mask in our partition's amo */
520 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_AND
,
524 /* original protection values for each node */
525 static u64 xpc_prot_vec_sn2
[MAX_NUMNODES
];
528 * Change protections to allow amo operations on non-Shub 1.1 systems.
530 static enum xp_retval
531 xpc_allow_amo_ops_sn2(struct amo
*amos_page
)
537 * On SHUB 1.1, we cannot call sn_change_memprotect() since the BIST
538 * collides with memory operations. On those systems we call
539 * xpc_allow_amo_ops_shub_wars_1_1_sn2() instead.
541 if (!enable_shub_wars_1_1()) {
542 ret
= sn_change_memprotect(ia64_tpa((u64
)amos_page
), PAGE_SIZE
,
543 SN_MEMPROT_ACCESS_CLASS_1
,
552 * Change protections to allow amo operations on Shub 1.1 systems.
555 xpc_allow_amo_ops_shub_wars_1_1_sn2(void)
560 if (!enable_shub_wars_1_1())
563 for_each_online_node(node
) {
564 nasid
= cnodeid_to_nasid(node
);
565 /* save current protection values */
566 xpc_prot_vec_sn2
[node
] =
567 (u64
)HUB_L((u64
*)GLOBAL_MMR_ADDR(nasid
,
568 SH1_MD_DQLP_MMR_DIR_PRIVEC0
));
569 /* open up everything */
570 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
,
571 SH1_MD_DQLP_MMR_DIR_PRIVEC0
),
573 HUB_S((u64
*)GLOBAL_MMR_ADDR(nasid
,
574 SH1_MD_DQRP_MMR_DIR_PRIVEC0
),
579 static enum xp_retval
580 xpc_get_partition_rsvd_page_pa_sn2(void *buf
, u64
*cookie
, unsigned long *rp_pa
,
586 status
= sn_partition_reserved_page_pa((u64
)buf
, cookie
, rp_pa
, len
);
587 if (status
== SALRET_OK
)
589 else if (status
== SALRET_MORE_PASSES
)
590 ret
= xpNeedMoreInfo
;
598 static enum xp_retval
599 xpc_rsvd_page_init_sn2(struct xpc_rsvd_page
*rp
)
601 struct amo
*amos_page
;
605 xpc_vars_sn2
= XPC_RP_VARS(rp
);
607 rp
->sn
.vars_pa
= xp_pa(xpc_vars_sn2
);
609 /* vars_part array follows immediately after vars */
610 xpc_vars_part_sn2
= (struct xpc_vars_part_sn2
*)((u8
*)XPC_RP_VARS(rp
) +
614 * Before clearing xpc_vars_sn2, see if a page of amos had been
615 * previously allocated. If not we'll need to allocate one and set
616 * permissions so that cross-partition amos are allowed.
618 * The allocated amo page needs MCA reporting to remain disabled after
619 * XPC has unloaded. To make this work, we keep a copy of the pointer
620 * to this page (i.e., amos_page) in the struct xpc_vars_sn2 structure,
621 * which is pointed to by the reserved page, and re-use that saved copy
622 * on subsequent loads of XPC. This amo page is never freed, and its
623 * memory protections are never restricted.
625 amos_page
= xpc_vars_sn2
->amos_page
;
626 if (amos_page
== NULL
) {
627 amos_page
= (struct amo
*)TO_AMO(uncached_alloc_page(0, 1));
628 if (amos_page
== NULL
) {
629 dev_err(xpc_part
, "can't allocate page of amos\n");
634 * Open up amo-R/W to cpu. This is done on Shub 1.1 systems
635 * when xpc_allow_amo_ops_shub_wars_1_1_sn2() is called.
637 ret
= xpc_allow_amo_ops_sn2(amos_page
);
638 if (ret
!= xpSuccess
) {
639 dev_err(xpc_part
, "can't allow amo operations\n");
640 uncached_free_page(__IA64_UNCACHED_OFFSET
|
641 TO_PHYS((u64
)amos_page
), 1);
646 /* clear xpc_vars_sn2 */
647 memset(xpc_vars_sn2
, 0, sizeof(struct xpc_vars_sn2
));
649 xpc_vars_sn2
->version
= XPC_V_VERSION
;
650 xpc_vars_sn2
->activate_IRQ_nasid
= cpuid_to_nasid(0);
651 xpc_vars_sn2
->activate_IRQ_phys_cpuid
= cpu_physical_id(0);
652 xpc_vars_sn2
->vars_part_pa
= xp_pa(xpc_vars_part_sn2
);
653 xpc_vars_sn2
->amos_page_pa
= ia64_tpa((u64
)amos_page
);
654 xpc_vars_sn2
->amos_page
= amos_page
; /* save for next load of XPC */
656 /* clear xpc_vars_part_sn2 */
657 memset((u64
*)xpc_vars_part_sn2
, 0, sizeof(struct xpc_vars_part_sn2
) *
658 XP_MAX_NPARTITIONS_SN2
);
660 /* initialize the activate IRQ related amo variables */
661 for (i
= 0; i
< xpc_nasid_mask_nlongs
; i
++)
662 (void)xpc_init_IRQ_amo_sn2(XPC_ACTIVATE_IRQ_AMOS_SN2
+ i
);
664 /* initialize the engaged remote partitions related amo variables */
665 (void)xpc_init_IRQ_amo_sn2(XPC_ENGAGED_PARTITIONS_AMO_SN2
);
666 (void)xpc_init_IRQ_amo_sn2(XPC_DEACTIVATE_REQUEST_AMO_SN2
);
672 xpc_increment_heartbeat_sn2(void)
674 xpc_vars_sn2
->heartbeat
++;
678 xpc_offline_heartbeat_sn2(void)
680 xpc_increment_heartbeat_sn2();
681 xpc_vars_sn2
->heartbeat_offline
= 1;
685 xpc_online_heartbeat_sn2(void)
687 xpc_increment_heartbeat_sn2();
688 xpc_vars_sn2
->heartbeat_offline
= 0;
692 xpc_heartbeat_init_sn2(void)
694 DBUG_ON(xpc_vars_sn2
== NULL
);
696 bitmap_zero(xpc_vars_sn2
->heartbeating_to_mask
, XP_MAX_NPARTITIONS_SN2
);
697 xpc_heartbeating_to_mask
= &xpc_vars_sn2
->heartbeating_to_mask
[0];
698 xpc_online_heartbeat_sn2();
702 xpc_heartbeat_exit_sn2(void)
704 xpc_offline_heartbeat_sn2();
707 static enum xp_retval
708 xpc_get_remote_heartbeat_sn2(struct xpc_partition
*part
)
710 struct xpc_vars_sn2
*remote_vars
;
713 remote_vars
= (struct xpc_vars_sn2
*)xpc_remote_copy_buffer_sn2
;
715 /* pull the remote vars structure that contains the heartbeat */
716 ret
= xp_remote_memcpy(xp_pa(remote_vars
),
717 part
->sn
.sn2
.remote_vars_pa
,
719 if (ret
!= xpSuccess
)
722 dev_dbg(xpc_part
, "partid=%d, heartbeat=%ld, last_heartbeat=%ld, "
723 "heartbeat_offline=%ld, HB_mask[0]=0x%lx\n", XPC_PARTID(part
),
724 remote_vars
->heartbeat
, part
->last_heartbeat
,
725 remote_vars
->heartbeat_offline
,
726 remote_vars
->heartbeating_to_mask
[0]);
728 if ((remote_vars
->heartbeat
== part
->last_heartbeat
&&
729 remote_vars
->heartbeat_offline
== 0) ||
730 !xpc_hb_allowed(sn_partition_id
,
731 &remote_vars
->heartbeating_to_mask
)) {
734 part
->last_heartbeat
= remote_vars
->heartbeat
;
741 * Get a copy of the remote partition's XPC variables from the reserved page.
743 * remote_vars points to a buffer that is cacheline aligned for BTE copies and
744 * assumed to be of size XPC_RP_VARS_SIZE.
746 static enum xp_retval
747 xpc_get_remote_vars_sn2(unsigned long remote_vars_pa
,
748 struct xpc_vars_sn2
*remote_vars
)
752 if (remote_vars_pa
== 0)
755 /* pull over the cross partition variables */
756 ret
= xp_remote_memcpy(xp_pa(remote_vars
), remote_vars_pa
,
758 if (ret
!= xpSuccess
)
761 if (XPC_VERSION_MAJOR(remote_vars
->version
) !=
762 XPC_VERSION_MAJOR(XPC_V_VERSION
)) {
770 xpc_request_partition_activation_sn2(struct xpc_rsvd_page
*remote_rp
,
771 unsigned long remote_rp_pa
, int nasid
)
773 xpc_send_local_activate_IRQ_sn2(nasid
);
777 xpc_request_partition_reactivation_sn2(struct xpc_partition
*part
)
779 xpc_send_local_activate_IRQ_sn2(part
->sn
.sn2
.activate_IRQ_nasid
);
783 xpc_request_partition_deactivation_sn2(struct xpc_partition
*part
)
785 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
786 unsigned long irq_flags
;
787 struct amo
*amo
= (struct amo
*)__va(part_sn2
->remote_amos_page_pa
+
788 (XPC_DEACTIVATE_REQUEST_AMO_SN2
*
789 sizeof(struct amo
)));
791 local_irq_save(irq_flags
);
793 /* set bit corresponding to our partid in remote partition's amo */
794 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_OR
,
795 BIT(sn_partition_id
));
798 * We must always use the nofault function regardless of whether we
799 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
800 * didn't, we'd never know that the other partition is down and would
801 * keep sending IRQs and amos to it until the heartbeat times out.
803 (void)xp_nofault_PIOR((u64
*)GLOBAL_MMR_ADDR(NASID_GET(&amo
->
805 xp_nofault_PIOR_target
));
807 local_irq_restore(irq_flags
);
810 * Send activate IRQ to get other side to see that we've set our
811 * bit in their deactivate request amo.
813 xpc_send_activate_IRQ_sn2(part_sn2
->remote_amos_page_pa
,
815 part_sn2
->activate_IRQ_nasid
,
816 part_sn2
->activate_IRQ_phys_cpuid
);
820 xpc_cancel_partition_deactivation_request_sn2(struct xpc_partition
*part
)
822 unsigned long irq_flags
;
823 struct amo
*amo
= (struct amo
*)__va(part
->sn
.sn2
.remote_amos_page_pa
+
824 (XPC_DEACTIVATE_REQUEST_AMO_SN2
*
825 sizeof(struct amo
)));
827 local_irq_save(irq_flags
);
829 /* clear bit corresponding to our partid in remote partition's amo */
830 FETCHOP_STORE_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_AND
,
831 ~BIT(sn_partition_id
));
834 * We must always use the nofault function regardless of whether we
835 * are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
836 * didn't, we'd never know that the other partition is down and would
837 * keep sending IRQs and amos to it until the heartbeat times out.
839 (void)xp_nofault_PIOR((u64
*)GLOBAL_MMR_ADDR(NASID_GET(&amo
->
841 xp_nofault_PIOR_target
));
843 local_irq_restore(irq_flags
);
847 xpc_partition_deactivation_requested_sn2(short partid
)
849 struct amo
*amo
= xpc_vars_sn2
->amos_page
+
850 XPC_DEACTIVATE_REQUEST_AMO_SN2
;
852 /* our partition's amo variable ANDed with partid mask */
853 return (FETCHOP_LOAD_OP(TO_AMO((u64
)&amo
->variable
), FETCHOP_LOAD
) &
858 * Update the remote partition's info.
861 xpc_update_partition_info_sn2(struct xpc_partition
*part
, u8 remote_rp_version
,
862 unsigned long *remote_rp_ts_jiffies
,
863 unsigned long remote_rp_pa
,
864 unsigned long remote_vars_pa
,
865 struct xpc_vars_sn2
*remote_vars
)
867 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
869 part
->remote_rp_version
= remote_rp_version
;
870 dev_dbg(xpc_part
, " remote_rp_version = 0x%016x\n",
871 part
->remote_rp_version
);
873 part
->remote_rp_ts_jiffies
= *remote_rp_ts_jiffies
;
874 dev_dbg(xpc_part
, " remote_rp_ts_jiffies = 0x%016lx\n",
875 part
->remote_rp_ts_jiffies
);
877 part
->remote_rp_pa
= remote_rp_pa
;
878 dev_dbg(xpc_part
, " remote_rp_pa = 0x%016lx\n", part
->remote_rp_pa
);
880 part_sn2
->remote_vars_pa
= remote_vars_pa
;
881 dev_dbg(xpc_part
, " remote_vars_pa = 0x%016lx\n",
882 part_sn2
->remote_vars_pa
);
884 part
->last_heartbeat
= remote_vars
->heartbeat
;
885 dev_dbg(xpc_part
, " last_heartbeat = 0x%016lx\n",
886 part
->last_heartbeat
);
888 part_sn2
->remote_vars_part_pa
= remote_vars
->vars_part_pa
;
889 dev_dbg(xpc_part
, " remote_vars_part_pa = 0x%016lx\n",
890 part_sn2
->remote_vars_part_pa
);
892 part_sn2
->activate_IRQ_nasid
= remote_vars
->activate_IRQ_nasid
;
893 dev_dbg(xpc_part
, " activate_IRQ_nasid = 0x%x\n",
894 part_sn2
->activate_IRQ_nasid
);
896 part_sn2
->activate_IRQ_phys_cpuid
=
897 remote_vars
->activate_IRQ_phys_cpuid
;
898 dev_dbg(xpc_part
, " activate_IRQ_phys_cpuid = 0x%x\n",
899 part_sn2
->activate_IRQ_phys_cpuid
);
901 part_sn2
->remote_amos_page_pa
= remote_vars
->amos_page_pa
;
902 dev_dbg(xpc_part
, " remote_amos_page_pa = 0x%lx\n",
903 part_sn2
->remote_amos_page_pa
);
905 part_sn2
->remote_vars_version
= remote_vars
->version
;
906 dev_dbg(xpc_part
, " remote_vars_version = 0x%x\n",
907 part_sn2
->remote_vars_version
);
911 * Prior code has determined the nasid which generated a activate IRQ.
912 * Inspect that nasid to determine if its partition needs to be activated
915 * A partition is considered "awaiting activation" if our partition
916 * flags indicate it is not active and it has a heartbeat. A
917 * partition is considered "awaiting deactivation" if our partition
918 * flags indicate it is active but it has no heartbeat or it is not
919 * sending its heartbeat to us.
921 * To determine the heartbeat, the remote nasid must have a properly
922 * initialized reserved page.
925 xpc_identify_activate_IRQ_req_sn2(int nasid
)
927 struct xpc_rsvd_page
*remote_rp
;
928 struct xpc_vars_sn2
*remote_vars
;
929 unsigned long remote_rp_pa
;
930 unsigned long remote_vars_pa
;
931 int remote_rp_version
;
933 unsigned long remote_rp_ts_jiffies
= 0;
935 struct xpc_partition
*part
;
936 struct xpc_partition_sn2
*part_sn2
;
939 /* pull over the reserved page structure */
941 remote_rp
= (struct xpc_rsvd_page
*)xpc_remote_copy_buffer_sn2
;
943 ret
= xpc_get_remote_rp(nasid
, NULL
, remote_rp
, &remote_rp_pa
);
944 if (ret
!= xpSuccess
) {
945 dev_warn(xpc_part
, "unable to get reserved page from nasid %d, "
946 "which sent interrupt, reason=%d\n", nasid
, ret
);
950 remote_vars_pa
= remote_rp
->sn
.vars_pa
;
951 remote_rp_version
= remote_rp
->version
;
952 remote_rp_ts_jiffies
= remote_rp
->ts_jiffies
;
954 partid
= remote_rp
->SAL_partid
;
955 part
= &xpc_partitions
[partid
];
956 part_sn2
= &part
->sn
.sn2
;
958 /* pull over the cross partition variables */
960 remote_vars
= (struct xpc_vars_sn2
*)xpc_remote_copy_buffer_sn2
;
962 ret
= xpc_get_remote_vars_sn2(remote_vars_pa
, remote_vars
);
963 if (ret
!= xpSuccess
) {
964 dev_warn(xpc_part
, "unable to get XPC variables from nasid %d, "
965 "which sent interrupt, reason=%d\n", nasid
, ret
);
967 XPC_DEACTIVATE_PARTITION(part
, ret
);
971 part
->activate_IRQ_rcvd
++;
973 dev_dbg(xpc_part
, "partid for nasid %d is %d; IRQs = %d; HB = "
974 "%ld:0x%lx\n", (int)nasid
, (int)partid
, part
->activate_IRQ_rcvd
,
975 remote_vars
->heartbeat
, remote_vars
->heartbeating_to_mask
[0]);
977 if (xpc_partition_disengaged(part
) &&
978 part
->act_state
== XPC_P_AS_INACTIVE
) {
980 xpc_update_partition_info_sn2(part
, remote_rp_version
,
981 &remote_rp_ts_jiffies
,
982 remote_rp_pa
, remote_vars_pa
,
985 if (xpc_partition_deactivation_requested_sn2(partid
)) {
987 * Other side is waiting on us to deactivate even though
993 xpc_activate_partition(part
);
997 DBUG_ON(part
->remote_rp_version
== 0);
998 DBUG_ON(part_sn2
->remote_vars_version
== 0);
1000 if (remote_rp_ts_jiffies
!= part
->remote_rp_ts_jiffies
) {
1002 /* the other side rebooted */
1004 DBUG_ON(xpc_partition_engaged_sn2(partid
));
1005 DBUG_ON(xpc_partition_deactivation_requested_sn2(partid
));
1007 xpc_update_partition_info_sn2(part
, remote_rp_version
,
1008 &remote_rp_ts_jiffies
,
1009 remote_rp_pa
, remote_vars_pa
,
1014 if (part
->disengage_timeout
> 0 && !xpc_partition_disengaged(part
)) {
1015 /* still waiting on other side to disengage from us */
1020 XPC_DEACTIVATE_PARTITION(part
, xpReactivating
);
1021 else if (xpc_partition_deactivation_requested_sn2(partid
))
1022 XPC_DEACTIVATE_PARTITION(part
, xpOtherGoingDown
);
1026 * Loop through the activation amo variables and process any bits
1027 * which are set. Each bit indicates a nasid sending a partition
1028 * activation or deactivation request.
1030 * Return #of IRQs detected.
1033 xpc_identify_activate_IRQ_sender_sn2(void)
1037 unsigned long nasid_mask_long
;
1038 u64 nasid
; /* remote nasid */
1039 int n_IRQs_detected
= 0;
1040 struct amo
*act_amos
;
1042 act_amos
= xpc_vars_sn2
->amos_page
+ XPC_ACTIVATE_IRQ_AMOS_SN2
;
1044 /* scan through activate amo variables looking for non-zero entries */
1045 for (l
= 0; l
< xpc_nasid_mask_nlongs
; l
++) {
1050 nasid_mask_long
= xpc_receive_IRQ_amo_sn2(&act_amos
[l
]);
1052 b
= find_first_bit(&nasid_mask_long
, BITS_PER_LONG
);
1053 if (b
>= BITS_PER_LONG
) {
1054 /* no IRQs from nasids in this amo variable */
1058 dev_dbg(xpc_part
, "amo[%d] gave back 0x%lx\n", l
,
1062 * If this nasid has been added to the machine since
1063 * our partition was reset, this will retain the
1064 * remote nasid in our reserved pages machine mask.
1065 * This is used in the event of module reload.
1067 xpc_mach_nasids
[l
] |= nasid_mask_long
;
1069 /* locate the nasid(s) which sent interrupts */
1073 nasid
= (l
* BITS_PER_LONG
+ b
) * 2;
1074 dev_dbg(xpc_part
, "interrupt from nasid %ld\n", nasid
);
1075 xpc_identify_activate_IRQ_req_sn2(nasid
);
1077 b
= find_next_bit(&nasid_mask_long
, BITS_PER_LONG
,
1079 } while (b
< BITS_PER_LONG
);
1081 return n_IRQs_detected
;
1085 xpc_process_activate_IRQ_rcvd_sn2(int n_IRQs_expected
)
1087 int n_IRQs_detected
;
1089 n_IRQs_detected
= xpc_identify_activate_IRQ_sender_sn2();
1090 if (n_IRQs_detected
< n_IRQs_expected
) {
1091 /* retry once to help avoid missing amo */
1092 (void)xpc_identify_activate_IRQ_sender_sn2();
1097 * Guarantee that the kzalloc'd memory is cacheline aligned.
1100 xpc_kzalloc_cacheline_aligned_sn2(size_t size
, gfp_t flags
, void **base
)
1102 /* see if kzalloc will give us cachline aligned memory by default */
1103 *base
= kzalloc(size
, flags
);
1107 if ((u64
)*base
== L1_CACHE_ALIGN((u64
)*base
))
1112 /* nope, we'll have to do it ourselves */
1113 *base
= kzalloc(size
+ L1_CACHE_BYTES
, flags
);
1117 return (void *)L1_CACHE_ALIGN((u64
)*base
);
1121 * Setup the infrastructure necessary to support XPartition Communication
1122 * between the specified remote partition and the local one.
1124 static enum xp_retval
1125 xpc_setup_infrastructure_sn2(struct xpc_partition
*part
)
1127 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
1128 enum xp_retval retval
;
1132 struct xpc_channel
*ch
;
1133 struct timer_list
*timer
;
1134 short partid
= XPC_PARTID(part
);
1137 * Allocate all of the channel structures as a contiguous chunk of
1140 DBUG_ON(part
->channels
!= NULL
);
1141 part
->channels
= kzalloc(sizeof(struct xpc_channel
) * XPC_MAX_NCHANNELS
,
1143 if (part
->channels
== NULL
) {
1144 dev_err(xpc_chan
, "can't get memory for channels\n");
1148 /* allocate all the required GET/PUT values */
1150 part_sn2
->local_GPs
=
1151 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE
, GFP_KERNEL
,
1152 &part_sn2
->local_GPs_base
);
1153 if (part_sn2
->local_GPs
== NULL
) {
1154 dev_err(xpc_chan
, "can't get memory for local get/put "
1156 retval
= xpNoMemory
;
1160 part_sn2
->remote_GPs
=
1161 xpc_kzalloc_cacheline_aligned_sn2(XPC_GP_SIZE
, GFP_KERNEL
,
1162 &part_sn2
->remote_GPs_base
);
1163 if (part_sn2
->remote_GPs
== NULL
) {
1164 dev_err(xpc_chan
, "can't get memory for remote get/put "
1166 retval
= xpNoMemory
;
1170 part_sn2
->remote_GPs_pa
= 0;
1172 /* allocate all the required open and close args */
1174 part
->local_openclose_args
=
1175 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE
,
1177 &part
->local_openclose_args_base
);
1178 if (part
->local_openclose_args
== NULL
) {
1179 dev_err(xpc_chan
, "can't get memory for local connect args\n");
1180 retval
= xpNoMemory
;
1184 part
->remote_openclose_args
=
1185 xpc_kzalloc_cacheline_aligned_sn2(XPC_OPENCLOSE_ARGS_SIZE
,
1187 &part
->remote_openclose_args_base
);
1188 if (part
->remote_openclose_args
== NULL
) {
1189 dev_err(xpc_chan
, "can't get memory for remote connect args\n");
1190 retval
= xpNoMemory
;
1194 part_sn2
->remote_openclose_args_pa
= 0;
1196 part_sn2
->local_chctl_amo_va
= xpc_init_IRQ_amo_sn2(partid
);
1197 part
->chctl
.all_flags
= 0;
1198 spin_lock_init(&part
->chctl_lock
);
1200 part_sn2
->notify_IRQ_nasid
= 0;
1201 part_sn2
->notify_IRQ_phys_cpuid
= 0;
1202 part_sn2
->remote_chctl_amo_va
= NULL
;
1204 atomic_set(&part
->channel_mgr_requests
, 1);
1205 init_waitqueue_head(&part
->channel_mgr_wq
);
1207 sprintf(part_sn2
->notify_IRQ_owner
, "xpc%02d", partid
);
1208 ret
= request_irq(SGI_XPC_NOTIFY
, xpc_handle_notify_IRQ_sn2
,
1209 IRQF_SHARED
, part_sn2
->notify_IRQ_owner
,
1210 (void *)(u64
)partid
);
1212 dev_err(xpc_chan
, "can't register NOTIFY IRQ handler, "
1213 "errno=%d\n", -ret
);
1214 retval
= xpLackOfResources
;
1218 /* Setup a timer to check for dropped notify IRQs */
1219 timer
= &part_sn2
->dropped_notify_IRQ_timer
;
1222 (void (*)(unsigned long))xpc_check_for_dropped_notify_IRQ_sn2
;
1223 timer
->data
= (unsigned long)part
;
1224 timer
->expires
= jiffies
+ XPC_DROPPED_NOTIFY_IRQ_WAIT_INTERVAL
;
1227 part
->nchannels
= XPC_MAX_NCHANNELS
;
1229 atomic_set(&part
->nchannels_active
, 0);
1230 atomic_set(&part
->nchannels_engaged
, 0);
1232 for (ch_number
= 0; ch_number
< part
->nchannels
; ch_number
++) {
1233 ch
= &part
->channels
[ch_number
];
1235 ch
->partid
= partid
;
1236 ch
->number
= ch_number
;
1237 ch
->flags
= XPC_C_DISCONNECTED
;
1239 ch
->sn
.sn2
.local_GP
= &part_sn2
->local_GPs
[ch_number
];
1240 ch
->local_openclose_args
=
1241 &part
->local_openclose_args
[ch_number
];
1243 atomic_set(&ch
->kthreads_assigned
, 0);
1244 atomic_set(&ch
->kthreads_idle
, 0);
1245 atomic_set(&ch
->kthreads_active
, 0);
1247 atomic_set(&ch
->references
, 0);
1248 atomic_set(&ch
->n_to_notify
, 0);
1250 spin_lock_init(&ch
->lock
);
1251 mutex_init(&ch
->sn
.sn2
.msg_to_pull_mutex
);
1252 init_completion(&ch
->wdisconnect_wait
);
1254 atomic_set(&ch
->n_on_msg_allocate_wq
, 0);
1255 init_waitqueue_head(&ch
->msg_allocate_wq
);
1256 init_waitqueue_head(&ch
->idle_wq
);
1260 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
1261 * we're declaring that this partition is ready to go.
1263 part
->setup_state
= XPC_P_SS_SETUP
;
1266 * Setup the per partition specific variables required by the
1267 * remote partition to establish channel connections with us.
1269 * The setting of the magic # indicates that these per partition
1270 * specific variables are ready to be used.
1272 xpc_vars_part_sn2
[partid
].GPs_pa
= xp_pa(part_sn2
->local_GPs
);
1273 xpc_vars_part_sn2
[partid
].openclose_args_pa
=
1274 xp_pa(part
->local_openclose_args
);
1275 xpc_vars_part_sn2
[partid
].chctl_amo_pa
=
1276 xp_pa(part_sn2
->local_chctl_amo_va
);
1277 cpuid
= raw_smp_processor_id(); /* any CPU in this partition will do */
1278 xpc_vars_part_sn2
[partid
].notify_IRQ_nasid
= cpuid_to_nasid(cpuid
);
1279 xpc_vars_part_sn2
[partid
].notify_IRQ_phys_cpuid
=
1280 cpu_physical_id(cpuid
);
1281 xpc_vars_part_sn2
[partid
].nchannels
= part
->nchannels
;
1282 xpc_vars_part_sn2
[partid
].magic
= XPC_VP_MAGIC1
;
1286 /* setup of infrastructure failed */
1288 kfree(part
->remote_openclose_args_base
);
1289 part
->remote_openclose_args
= NULL
;
1291 kfree(part
->local_openclose_args_base
);
1292 part
->local_openclose_args
= NULL
;
1294 kfree(part_sn2
->remote_GPs_base
);
1295 part_sn2
->remote_GPs
= NULL
;
1297 kfree(part_sn2
->local_GPs_base
);
1298 part_sn2
->local_GPs
= NULL
;
1300 kfree(part
->channels
);
1301 part
->channels
= NULL
;
1306 * Teardown the infrastructure necessary to support XPartition Communication
1307 * between the specified remote partition and the local one.
1310 xpc_teardown_infrastructure_sn2(struct xpc_partition
*part
)
1312 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
1313 short partid
= XPC_PARTID(part
);
1316 * We start off by making this partition inaccessible to local
1317 * processes by marking it as no longer setup. Then we make it
1318 * inaccessible to remote processes by clearing the XPC per partition
1319 * specific variable's magic # (which indicates that these variables
1320 * are no longer valid) and by ignoring all XPC notify IRQs sent to
1324 DBUG_ON(atomic_read(&part
->nchannels_engaged
) != 0);
1325 DBUG_ON(atomic_read(&part
->nchannels_active
) != 0);
1326 DBUG_ON(part
->setup_state
!= XPC_P_SS_SETUP
);
1327 part
->setup_state
= XPC_P_SS_WTEARDOWN
;
1329 xpc_vars_part_sn2
[partid
].magic
= 0;
1331 free_irq(SGI_XPC_NOTIFY
, (void *)(u64
)partid
);
1334 * Before proceeding with the teardown we have to wait until all
1335 * existing references cease.
1337 wait_event(part
->teardown_wq
, (atomic_read(&part
->references
) == 0));
1339 /* now we can begin tearing down the infrastructure */
1341 part
->setup_state
= XPC_P_SS_TORNDOWN
;
1343 /* in case we've still got outstanding timers registered... */
1344 del_timer_sync(&part_sn2
->dropped_notify_IRQ_timer
);
1346 kfree(part
->remote_openclose_args_base
);
1347 part
->remote_openclose_args
= NULL
;
1348 kfree(part
->local_openclose_args_base
);
1349 part
->local_openclose_args
= NULL
;
1350 kfree(part_sn2
->remote_GPs_base
);
1351 part_sn2
->remote_GPs
= NULL
;
1352 kfree(part_sn2
->local_GPs_base
);
1353 part_sn2
->local_GPs
= NULL
;
1354 kfree(part
->channels
);
1355 part
->channels
= NULL
;
1356 part_sn2
->local_chctl_amo_va
= NULL
;
1360 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
1361 * (or multiple cachelines) from a remote partition.
1363 * src_pa must be a cacheline aligned physical address on the remote partition.
1364 * dst must be a cacheline aligned virtual address on this partition.
1365 * cnt must be cacheline sized
1367 /* ??? Replace this function by call to xp_remote_memcpy() or bte_copy()? */
1368 static enum xp_retval
1369 xpc_pull_remote_cachelines_sn2(struct xpc_partition
*part
, void *dst
,
1370 const unsigned long src_pa
, size_t cnt
)
1374 DBUG_ON(src_pa
!= L1_CACHE_ALIGN(src_pa
));
1375 DBUG_ON((unsigned long)dst
!= L1_CACHE_ALIGN((unsigned long)dst
));
1376 DBUG_ON(cnt
!= L1_CACHE_ALIGN(cnt
));
1378 if (part
->act_state
== XPC_P_AS_DEACTIVATING
)
1379 return part
->reason
;
1381 ret
= xp_remote_memcpy(xp_pa(dst
), src_pa
, cnt
);
1382 if (ret
!= xpSuccess
) {
1383 dev_dbg(xpc_chan
, "xp_remote_memcpy() from partition %d failed,"
1384 " ret=%d\n", XPC_PARTID(part
), ret
);
1390 * Pull the remote per partition specific variables from the specified
1393 static enum xp_retval
1394 xpc_pull_remote_vars_part_sn2(struct xpc_partition
*part
)
1396 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
1397 u8 buffer
[L1_CACHE_BYTES
* 2];
1398 struct xpc_vars_part_sn2
*pulled_entry_cacheline
=
1399 (struct xpc_vars_part_sn2
*)L1_CACHE_ALIGN((u64
)buffer
);
1400 struct xpc_vars_part_sn2
*pulled_entry
;
1401 unsigned long remote_entry_cacheline_pa
;
1402 unsigned long remote_entry_pa
;
1403 short partid
= XPC_PARTID(part
);
1406 /* pull the cacheline that contains the variables we're interested in */
1408 DBUG_ON(part_sn2
->remote_vars_part_pa
!=
1409 L1_CACHE_ALIGN(part_sn2
->remote_vars_part_pa
));
1410 DBUG_ON(sizeof(struct xpc_vars_part_sn2
) != L1_CACHE_BYTES
/ 2);
1412 remote_entry_pa
= part_sn2
->remote_vars_part_pa
+
1413 sn_partition_id
* sizeof(struct xpc_vars_part_sn2
);
1415 remote_entry_cacheline_pa
= (remote_entry_pa
& ~(L1_CACHE_BYTES
- 1));
1417 pulled_entry
= (struct xpc_vars_part_sn2
*)((u64
)pulled_entry_cacheline
1418 + (remote_entry_pa
&
1419 (L1_CACHE_BYTES
- 1)));
1421 ret
= xpc_pull_remote_cachelines_sn2(part
, pulled_entry_cacheline
,
1422 remote_entry_cacheline_pa
,
1424 if (ret
!= xpSuccess
) {
1425 dev_dbg(xpc_chan
, "failed to pull XPC vars_part from "
1426 "partition %d, ret=%d\n", partid
, ret
);
1430 /* see if they've been set up yet */
1432 if (pulled_entry
->magic
!= XPC_VP_MAGIC1
&&
1433 pulled_entry
->magic
!= XPC_VP_MAGIC2
) {
1435 if (pulled_entry
->magic
!= 0) {
1436 dev_dbg(xpc_chan
, "partition %d's XPC vars_part for "
1437 "partition %d has bad magic value (=0x%lx)\n",
1438 partid
, sn_partition_id
, pulled_entry
->magic
);
1442 /* they've not been initialized yet */
1446 if (xpc_vars_part_sn2
[partid
].magic
== XPC_VP_MAGIC1
) {
1448 /* validate the variables */
1450 if (pulled_entry
->GPs_pa
== 0 ||
1451 pulled_entry
->openclose_args_pa
== 0 ||
1452 pulled_entry
->chctl_amo_pa
== 0) {
1454 dev_err(xpc_chan
, "partition %d's XPC vars_part for "
1455 "partition %d are not valid\n", partid
,
1457 return xpInvalidAddress
;
1460 /* the variables we imported look to be valid */
1462 part_sn2
->remote_GPs_pa
= pulled_entry
->GPs_pa
;
1463 part_sn2
->remote_openclose_args_pa
=
1464 pulled_entry
->openclose_args_pa
;
1465 part_sn2
->remote_chctl_amo_va
=
1466 (struct amo
*)__va(pulled_entry
->chctl_amo_pa
);
1467 part_sn2
->notify_IRQ_nasid
= pulled_entry
->notify_IRQ_nasid
;
1468 part_sn2
->notify_IRQ_phys_cpuid
=
1469 pulled_entry
->notify_IRQ_phys_cpuid
;
1471 if (part
->nchannels
> pulled_entry
->nchannels
)
1472 part
->nchannels
= pulled_entry
->nchannels
;
1474 /* let the other side know that we've pulled their variables */
1476 xpc_vars_part_sn2
[partid
].magic
= XPC_VP_MAGIC2
;
1479 if (pulled_entry
->magic
== XPC_VP_MAGIC1
)
1486 * Establish first contact with the remote partititon. This involves pulling
1487 * the XPC per partition variables from the remote partition and waiting for
1488 * the remote partition to pull ours.
1490 static enum xp_retval
1491 xpc_make_first_contact_sn2(struct xpc_partition
*part
)
1493 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
1497 * Register the remote partition's amos with SAL so it can handle
1498 * and cleanup errors within that address range should the remote
1499 * partition go down. We don't unregister this range because it is
1500 * difficult to tell when outstanding writes to the remote partition
1501 * are finished and thus when it is safe to unregister. This should
1502 * not result in wasted space in the SAL xp_addr_region table because
1503 * we should get the same page for remote_amos_page_pa after module
1504 * reloads and system reboots.
1506 if (sn_register_xp_addr_region(part_sn2
->remote_amos_page_pa
,
1507 PAGE_SIZE
, 1) < 0) {
1508 dev_warn(xpc_part
, "xpc_activating(%d) failed to register "
1509 "xp_addr region\n", XPC_PARTID(part
));
1511 ret
= xpPhysAddrRegFailed
;
1512 XPC_DEACTIVATE_PARTITION(part
, ret
);
1517 * Send activate IRQ to get other side to activate if they've not
1518 * already begun to do so.
1520 xpc_send_activate_IRQ_sn2(part_sn2
->remote_amos_page_pa
,
1521 cnodeid_to_nasid(0),
1522 part_sn2
->activate_IRQ_nasid
,
1523 part_sn2
->activate_IRQ_phys_cpuid
);
1525 while ((ret
= xpc_pull_remote_vars_part_sn2(part
)) != xpSuccess
) {
1526 if (ret
!= xpRetry
) {
1527 XPC_DEACTIVATE_PARTITION(part
, ret
);
1531 dev_dbg(xpc_part
, "waiting to make first contact with "
1532 "partition %d\n", XPC_PARTID(part
));
1534 /* wait a 1/4 of a second or so */
1535 (void)msleep_interruptible(250);
1537 if (part
->act_state
== XPC_P_AS_DEACTIVATING
)
1538 return part
->reason
;
1545 * Get the chctl flags and pull the openclose args and/or remote GPs as needed.
1548 xpc_get_chctl_all_flags_sn2(struct xpc_partition
*part
)
1550 struct xpc_partition_sn2
*part_sn2
= &part
->sn
.sn2
;
1551 unsigned long irq_flags
;
1552 union xpc_channel_ctl_flags chctl
;
1556 * See if there are any chctl flags to be handled.
1559 spin_lock_irqsave(&part
->chctl_lock
, irq_flags
);
1560 chctl
= part
->chctl
;
1561 if (chctl
.all_flags
!= 0)
1562 part
->chctl
.all_flags
= 0;
1564 spin_unlock_irqrestore(&part
->chctl_lock
, irq_flags
);
1566 if (xpc_any_openclose_chctl_flags_set(&chctl
)) {
1567 ret
= xpc_pull_remote_cachelines_sn2(part
, part
->
1568 remote_openclose_args
,
1570 remote_openclose_args_pa
,
1571 XPC_OPENCLOSE_ARGS_SIZE
);
1572 if (ret
!= xpSuccess
) {
1573 XPC_DEACTIVATE_PARTITION(part
, ret
);
1575 dev_dbg(xpc_chan
, "failed to pull openclose args from "
1576 "partition %d, ret=%d\n", XPC_PARTID(part
),
1579 /* don't bother processing chctl flags anymore */
1580 chctl
.all_flags
= 0;
1584 if (xpc_any_msg_chctl_flags_set(&chctl
)) {
1585 ret
= xpc_pull_remote_cachelines_sn2(part
, part_sn2
->remote_GPs
,
1586 part_sn2
->remote_GPs_pa
,
1588 if (ret
!= xpSuccess
) {
1589 XPC_DEACTIVATE_PARTITION(part
, ret
);
1591 dev_dbg(xpc_chan
, "failed to pull GPs from partition "
1592 "%d, ret=%d\n", XPC_PARTID(part
), ret
);
1594 /* don't bother processing chctl flags anymore */
1595 chctl
.all_flags
= 0;
1599 return chctl
.all_flags
;
1603 * Allocate the local message queue and the notify queue.
1605 static enum xp_retval
1606 xpc_allocate_local_msgqueue_sn2(struct xpc_channel
*ch
)
1608 unsigned long irq_flags
;
1612 for (nentries
= ch
->local_nentries
; nentries
> 0; nentries
--) {
1614 nbytes
= nentries
* ch
->msg_size
;
1615 ch
->local_msgqueue
=
1616 xpc_kzalloc_cacheline_aligned_sn2(nbytes
, GFP_KERNEL
,
1617 &ch
->local_msgqueue_base
);
1618 if (ch
->local_msgqueue
== NULL
)
1621 nbytes
= nentries
* sizeof(struct xpc_notify
);
1622 ch
->notify_queue
= kzalloc(nbytes
, GFP_KERNEL
);
1623 if (ch
->notify_queue
== NULL
) {
1624 kfree(ch
->local_msgqueue_base
);
1625 ch
->local_msgqueue
= NULL
;
1629 spin_lock_irqsave(&ch
->lock
, irq_flags
);
1630 if (nentries
< ch
->local_nentries
) {
1631 dev_dbg(xpc_chan
, "nentries=%d local_nentries=%d, "
1632 "partid=%d, channel=%d\n", nentries
,
1633 ch
->local_nentries
, ch
->partid
, ch
->number
);
1635 ch
->local_nentries
= nentries
;
1637 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
1641 dev_dbg(xpc_chan
, "can't get memory for local message queue and notify "
1642 "queue, partid=%d, channel=%d\n", ch
->partid
, ch
->number
);
1647 * Allocate the cached remote message queue.
1649 static enum xp_retval
1650 xpc_allocate_remote_msgqueue_sn2(struct xpc_channel
*ch
)
1652 unsigned long irq_flags
;
1656 DBUG_ON(ch
->remote_nentries
<= 0);
1658 for (nentries
= ch
->remote_nentries
; nentries
> 0; nentries
--) {
1660 nbytes
= nentries
* ch
->msg_size
;
1661 ch
->remote_msgqueue
=
1662 xpc_kzalloc_cacheline_aligned_sn2(nbytes
, GFP_KERNEL
,
1663 &ch
->remote_msgqueue_base
);
1664 if (ch
->remote_msgqueue
== NULL
)
1667 spin_lock_irqsave(&ch
->lock
, irq_flags
);
1668 if (nentries
< ch
->remote_nentries
) {
1669 dev_dbg(xpc_chan
, "nentries=%d remote_nentries=%d, "
1670 "partid=%d, channel=%d\n", nentries
,
1671 ch
->remote_nentries
, ch
->partid
, ch
->number
);
1673 ch
->remote_nentries
= nentries
;
1675 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
1679 dev_dbg(xpc_chan
, "can't get memory for cached remote message queue, "
1680 "partid=%d, channel=%d\n", ch
->partid
, ch
->number
);
1685 * Allocate message queues and other stuff associated with a channel.
1687 * Note: Assumes all of the channel sizes are filled in.
1689 static enum xp_retval
1690 xpc_allocate_msgqueues_sn2(struct xpc_channel
*ch
)
1694 DBUG_ON(ch
->flags
& XPC_C_SETUP
);
1696 ret
= xpc_allocate_local_msgqueue_sn2(ch
);
1697 if (ret
== xpSuccess
) {
1699 ret
= xpc_allocate_remote_msgqueue_sn2(ch
);
1700 if (ret
!= xpSuccess
) {
1701 kfree(ch
->local_msgqueue_base
);
1702 ch
->local_msgqueue
= NULL
;
1703 kfree(ch
->notify_queue
);
1704 ch
->notify_queue
= NULL
;
1711 * Free up message queues and other stuff that were allocated for the specified
1714 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
1715 * they're cleared when XPC_C_DISCONNECTED is cleared.
1718 xpc_free_msgqueues_sn2(struct xpc_channel
*ch
)
1720 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
1722 DBUG_ON(!spin_is_locked(&ch
->lock
));
1723 DBUG_ON(atomic_read(&ch
->n_to_notify
) != 0);
1725 ch
->remote_msgqueue_pa
= 0;
1729 ch
->local_nentries
= 0;
1730 ch
->remote_nentries
= 0;
1731 ch
->kthreads_assigned_limit
= 0;
1732 ch
->kthreads_idle_limit
= 0;
1734 ch_sn2
->local_GP
->get
= 0;
1735 ch_sn2
->local_GP
->put
= 0;
1736 ch_sn2
->remote_GP
.get
= 0;
1737 ch_sn2
->remote_GP
.put
= 0;
1738 ch_sn2
->w_local_GP
.get
= 0;
1739 ch_sn2
->w_local_GP
.put
= 0;
1740 ch_sn2
->w_remote_GP
.get
= 0;
1741 ch_sn2
->w_remote_GP
.put
= 0;
1742 ch_sn2
->next_msg_to_pull
= 0;
1744 if (ch
->flags
& XPC_C_SETUP
) {
1745 dev_dbg(xpc_chan
, "ch->flags=0x%x, partid=%d, channel=%d\n",
1746 ch
->flags
, ch
->partid
, ch
->number
);
1748 kfree(ch
->local_msgqueue_base
);
1749 ch
->local_msgqueue
= NULL
;
1750 kfree(ch
->remote_msgqueue_base
);
1751 ch
->remote_msgqueue
= NULL
;
1752 kfree(ch
->notify_queue
);
1753 ch
->notify_queue
= NULL
;
1758 * Notify those who wanted to be notified upon delivery of their message.
1761 xpc_notify_senders_sn2(struct xpc_channel
*ch
, enum xp_retval reason
, s64 put
)
1763 struct xpc_notify
*notify
;
1765 s64 get
= ch
->sn
.sn2
.w_remote_GP
.get
- 1;
1767 while (++get
< put
&& atomic_read(&ch
->n_to_notify
) > 0) {
1769 notify
= &ch
->notify_queue
[get
% ch
->local_nentries
];
1772 * See if the notify entry indicates it was associated with
1773 * a message who's sender wants to be notified. It is possible
1774 * that it is, but someone else is doing or has done the
1777 notify_type
= notify
->type
;
1778 if (notify_type
== 0 ||
1779 cmpxchg(¬ify
->type
, notify_type
, 0) != notify_type
) {
1783 DBUG_ON(notify_type
!= XPC_N_CALL
);
1785 atomic_dec(&ch
->n_to_notify
);
1787 if (notify
->func
!= NULL
) {
1788 dev_dbg(xpc_chan
, "notify->func() called, notify=0x%p, "
1789 "msg_number=%ld, partid=%d, channel=%d\n",
1790 (void *)notify
, get
, ch
->partid
, ch
->number
);
1792 notify
->func(reason
, ch
->partid
, ch
->number
,
1795 dev_dbg(xpc_chan
, "notify->func() returned, "
1796 "notify=0x%p, msg_number=%ld, partid=%d, "
1797 "channel=%d\n", (void *)notify
, get
,
1798 ch
->partid
, ch
->number
);
1804 xpc_notify_senders_of_disconnect_sn2(struct xpc_channel
*ch
)
1806 xpc_notify_senders_sn2(ch
, ch
->reason
, ch
->sn
.sn2
.w_local_GP
.put
);
1810 * Clear some of the msg flags in the local message queue.
1813 xpc_clear_local_msgqueue_flags_sn2(struct xpc_channel
*ch
)
1815 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
1816 struct xpc_msg
*msg
;
1819 get
= ch_sn2
->w_remote_GP
.get
;
1821 msg
= (struct xpc_msg
*)((u64
)ch
->local_msgqueue
+
1822 (get
% ch
->local_nentries
) *
1825 } while (++get
< ch_sn2
->remote_GP
.get
);
1829 * Clear some of the msg flags in the remote message queue.
1832 xpc_clear_remote_msgqueue_flags_sn2(struct xpc_channel
*ch
)
1834 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
1835 struct xpc_msg
*msg
;
1838 put
= ch_sn2
->w_remote_GP
.put
;
1840 msg
= (struct xpc_msg
*)((u64
)ch
->remote_msgqueue
+
1841 (put
% ch
->remote_nentries
) *
1844 } while (++put
< ch_sn2
->remote_GP
.put
);
1848 xpc_process_msg_chctl_flags_sn2(struct xpc_partition
*part
, int ch_number
)
1850 struct xpc_channel
*ch
= &part
->channels
[ch_number
];
1851 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
1854 ch_sn2
->remote_GP
= part
->sn
.sn2
.remote_GPs
[ch_number
];
1856 /* See what, if anything, has changed for each connected channel */
1858 xpc_msgqueue_ref(ch
);
1860 if (ch_sn2
->w_remote_GP
.get
== ch_sn2
->remote_GP
.get
&&
1861 ch_sn2
->w_remote_GP
.put
== ch_sn2
->remote_GP
.put
) {
1862 /* nothing changed since GPs were last pulled */
1863 xpc_msgqueue_deref(ch
);
1867 if (!(ch
->flags
& XPC_C_CONNECTED
)) {
1868 xpc_msgqueue_deref(ch
);
1873 * First check to see if messages recently sent by us have been
1874 * received by the other side. (The remote GET value will have
1875 * changed since we last looked at it.)
1878 if (ch_sn2
->w_remote_GP
.get
!= ch_sn2
->remote_GP
.get
) {
1881 * We need to notify any senders that want to be notified
1882 * that their sent messages have been received by their
1883 * intended recipients. We need to do this before updating
1884 * w_remote_GP.get so that we don't allocate the same message
1885 * queue entries prematurely (see xpc_allocate_msg()).
1887 if (atomic_read(&ch
->n_to_notify
) > 0) {
1889 * Notify senders that messages sent have been
1890 * received and delivered by the other side.
1892 xpc_notify_senders_sn2(ch
, xpMsgDelivered
,
1893 ch_sn2
->remote_GP
.get
);
1897 * Clear msg->flags in previously sent messages, so that
1898 * they're ready for xpc_allocate_msg().
1900 xpc_clear_local_msgqueue_flags_sn2(ch
);
1902 ch_sn2
->w_remote_GP
.get
= ch_sn2
->remote_GP
.get
;
1904 dev_dbg(xpc_chan
, "w_remote_GP.get changed to %ld, partid=%d, "
1905 "channel=%d\n", ch_sn2
->w_remote_GP
.get
, ch
->partid
,
1909 * If anyone was waiting for message queue entries to become
1910 * available, wake them up.
1912 if (atomic_read(&ch
->n_on_msg_allocate_wq
) > 0)
1913 wake_up(&ch
->msg_allocate_wq
);
1917 * Now check for newly sent messages by the other side. (The remote
1918 * PUT value will have changed since we last looked at it.)
1921 if (ch_sn2
->w_remote_GP
.put
!= ch_sn2
->remote_GP
.put
) {
1923 * Clear msg->flags in previously received messages, so that
1924 * they're ready for xpc_get_deliverable_msg().
1926 xpc_clear_remote_msgqueue_flags_sn2(ch
);
1928 ch_sn2
->w_remote_GP
.put
= ch_sn2
->remote_GP
.put
;
1930 dev_dbg(xpc_chan
, "w_remote_GP.put changed to %ld, partid=%d, "
1931 "channel=%d\n", ch_sn2
->w_remote_GP
.put
, ch
->partid
,
1934 nmsgs_sent
= ch_sn2
->w_remote_GP
.put
- ch_sn2
->w_local_GP
.get
;
1935 if (nmsgs_sent
> 0) {
1936 dev_dbg(xpc_chan
, "msgs waiting to be copied and "
1937 "delivered=%d, partid=%d, channel=%d\n",
1938 nmsgs_sent
, ch
->partid
, ch
->number
);
1940 if (ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
)
1941 xpc_activate_kthreads(ch
, nmsgs_sent
);
1945 xpc_msgqueue_deref(ch
);
1948 static struct xpc_msg
*
1949 xpc_pull_remote_msg_sn2(struct xpc_channel
*ch
, s64 get
)
1951 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
1952 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
1953 unsigned long remote_msg_pa
;
1954 struct xpc_msg
*msg
;
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_pa
= ch
->remote_msgqueue_pa
+ msg_offset
;
1982 ret
= xpc_pull_remote_cachelines_sn2(part
, msg
, remote_msg_pa
,
1983 nmsgs
* ch
->msg_size
);
1984 if (ret
!= xpSuccess
) {
1986 dev_dbg(xpc_chan
, "failed to pull %d msgs starting with"
1987 " msg %ld from partition %d, channel=%d, "
1988 "ret=%d\n", nmsgs
, ch_sn2
->next_msg_to_pull
,
1989 ch
->partid
, ch
->number
, ret
);
1991 XPC_DEACTIVATE_PARTITION(part
, ret
);
1993 mutex_unlock(&ch_sn2
->msg_to_pull_mutex
);
1997 ch_sn2
->next_msg_to_pull
+= nmsgs
;
2000 mutex_unlock(&ch_sn2
->msg_to_pull_mutex
);
2002 /* return the message we were looking for */
2003 msg_offset
= (get
% ch
->remote_nentries
) * ch
->msg_size
;
2004 msg
= (struct xpc_msg
*)((u64
)ch
->remote_msgqueue
+ msg_offset
);
2010 xpc_n_of_deliverable_msgs_sn2(struct xpc_channel
*ch
)
2012 return ch
->sn
.sn2
.w_remote_GP
.put
- ch
->sn
.sn2
.w_local_GP
.get
;
2016 * Get a message to be delivered.
2018 static struct xpc_msg
*
2019 xpc_get_deliverable_msg_sn2(struct xpc_channel
*ch
)
2021 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
2022 struct xpc_msg
*msg
= NULL
;
2026 if (ch
->flags
& XPC_C_DISCONNECTING
)
2029 get
= ch_sn2
->w_local_GP
.get
;
2030 rmb(); /* guarantee that .get loads before .put */
2031 if (get
== ch_sn2
->w_remote_GP
.put
)
2034 /* There are messages waiting to be pulled and delivered.
2035 * We need to try to secure one for ourselves. We'll do this
2036 * by trying to increment w_local_GP.get and hope that no one
2037 * else beats us to it. If they do, we'll we'll simply have
2038 * to try again for the next one.
2041 if (cmpxchg(&ch_sn2
->w_local_GP
.get
, get
, get
+ 1) == get
) {
2042 /* we got the entry referenced by get */
2044 dev_dbg(xpc_chan
, "w_local_GP.get changed to %ld, "
2045 "partid=%d, channel=%d\n", get
+ 1,
2046 ch
->partid
, ch
->number
);
2048 /* pull the message from the remote partition */
2050 msg
= xpc_pull_remote_msg_sn2(ch
, get
);
2052 DBUG_ON(msg
!= NULL
&& msg
->number
!= get
);
2053 DBUG_ON(msg
!= NULL
&& (msg
->flags
& XPC_M_DONE
));
2054 DBUG_ON(msg
!= NULL
&& !(msg
->flags
& XPC_M_READY
));
2065 * Now we actually send the messages that are ready to be sent by advancing
2066 * the local message queue's Put value and then send a chctl msgrequest to the
2067 * recipient partition.
2070 xpc_send_msgs_sn2(struct xpc_channel
*ch
, s64 initial_put
)
2072 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
2073 struct xpc_msg
*msg
;
2074 s64 put
= initial_put
+ 1;
2075 int send_msgrequest
= 0;
2080 if (put
== ch_sn2
->w_local_GP
.put
)
2083 msg
= (struct xpc_msg
*)((u64
)ch
->local_msgqueue
+
2084 (put
% ch
->local_nentries
) *
2087 if (!(msg
->flags
& XPC_M_READY
))
2093 if (put
== initial_put
) {
2094 /* nothing's changed */
2098 if (cmpxchg_rel(&ch_sn2
->local_GP
->put
, initial_put
, put
) !=
2100 /* someone else beat us to it */
2101 DBUG_ON(ch_sn2
->local_GP
->put
< initial_put
);
2105 /* we just set the new value of local_GP->put */
2107 dev_dbg(xpc_chan
, "local_GP->put changed to %ld, partid=%d, "
2108 "channel=%d\n", put
, ch
->partid
, ch
->number
);
2110 send_msgrequest
= 1;
2113 * We need to ensure that the message referenced by
2114 * local_GP->put is not XPC_M_READY or that local_GP->put
2115 * equals w_local_GP.put, so we'll go have a look.
2120 if (send_msgrequest
)
2121 xpc_send_chctl_msgrequest_sn2(ch
);
2125 * Allocate an entry for a message from the message queue associated with the
2126 * specified channel.
2128 static enum xp_retval
2129 xpc_allocate_msg_sn2(struct xpc_channel
*ch
, u32 flags
,
2130 struct xpc_msg
**address_of_msg
)
2132 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
2133 struct xpc_msg
*msg
;
2138 * Get the next available message entry from the local message queue.
2139 * If none are available, we'll make sure that we grab the latest
2146 put
= ch_sn2
->w_local_GP
.put
;
2147 rmb(); /* guarantee that .put loads before .get */
2148 if (put
- ch_sn2
->w_remote_GP
.get
< ch
->local_nentries
) {
2150 /* There are available message entries. We need to try
2151 * to secure one for ourselves. We'll do this by trying
2152 * to increment w_local_GP.put as long as someone else
2153 * doesn't beat us to it. If they do, we'll have to
2156 if (cmpxchg(&ch_sn2
->w_local_GP
.put
, put
, put
+ 1) ==
2158 /* we got the entry referenced by put */
2161 continue; /* try again */
2165 * There aren't any available msg entries at this time.
2167 * In waiting for a message entry to become available,
2168 * we set a timeout in case the other side is not sending
2169 * completion interrupts. This lets us fake a notify IRQ
2170 * that will cause the notify IRQ handler to fetch the latest
2171 * GP values as if an interrupt was sent by the other side.
2173 if (ret
== xpTimeout
)
2174 xpc_send_chctl_local_msgrequest_sn2(ch
);
2176 if (flags
& XPC_NOWAIT
)
2179 ret
= xpc_allocate_msg_wait(ch
);
2180 if (ret
!= xpInterrupted
&& ret
!= xpTimeout
)
2184 /* get the message's address and initialize it */
2185 msg
= (struct xpc_msg
*)((u64
)ch
->local_msgqueue
+
2186 (put
% ch
->local_nentries
) * ch
->msg_size
);
2188 DBUG_ON(msg
->flags
!= 0);
2191 dev_dbg(xpc_chan
, "w_local_GP.put changed to %ld; msg=0x%p, "
2192 "msg_number=%ld, partid=%d, channel=%d\n", put
+ 1,
2193 (void *)msg
, msg
->number
, ch
->partid
, ch
->number
);
2195 *address_of_msg
= msg
;
2200 * Common code that does the actual sending of the message by advancing the
2201 * local message queue's Put value and sends a chctl msgrequest to the
2202 * partition the message is being sent to.
2204 static enum xp_retval
2205 xpc_send_msg_sn2(struct xpc_channel
*ch
, u32 flags
, void *payload
,
2206 u16 payload_size
, u8 notify_type
, xpc_notify_func func
,
2209 enum xp_retval ret
= xpSuccess
;
2210 struct xpc_msg
*msg
= msg
;
2211 struct xpc_notify
*notify
= notify
;
2215 DBUG_ON(notify_type
== XPC_N_CALL
&& func
== NULL
);
2217 if (XPC_MSG_SIZE(payload_size
) > ch
->msg_size
)
2218 return xpPayloadTooBig
;
2220 xpc_msgqueue_ref(ch
);
2222 if (ch
->flags
& XPC_C_DISCONNECTING
) {
2226 if (!(ch
->flags
& XPC_C_CONNECTED
)) {
2227 ret
= xpNotConnected
;
2231 ret
= xpc_allocate_msg_sn2(ch
, flags
, &msg
);
2232 if (ret
!= xpSuccess
)
2235 msg_number
= msg
->number
;
2237 if (notify_type
!= 0) {
2239 * Tell the remote side to send an ACK interrupt when the
2240 * message has been delivered.
2242 msg
->flags
|= XPC_M_INTERRUPT
;
2244 atomic_inc(&ch
->n_to_notify
);
2246 notify
= &ch
->notify_queue
[msg_number
% ch
->local_nentries
];
2247 notify
->func
= func
;
2249 notify
->type
= notify_type
;
2251 /* ??? Is a mb() needed here? */
2253 if (ch
->flags
& XPC_C_DISCONNECTING
) {
2255 * An error occurred between our last error check and
2256 * this one. We will try to clear the type field from
2257 * the notify entry. If we succeed then
2258 * xpc_disconnect_channel() didn't already process
2261 if (cmpxchg(¬ify
->type
, notify_type
, 0) ==
2263 atomic_dec(&ch
->n_to_notify
);
2270 memcpy(&msg
->payload
, payload
, payload_size
);
2272 msg
->flags
|= XPC_M_READY
;
2275 * The preceding store of msg->flags must occur before the following
2276 * load of local_GP->put.
2280 /* see if the message is next in line to be sent, if so send it */
2282 put
= ch
->sn
.sn2
.local_GP
->put
;
2283 if (put
== msg_number
)
2284 xpc_send_msgs_sn2(ch
, put
);
2287 xpc_msgqueue_deref(ch
);
2292 * Now we actually acknowledge the messages that have been delivered and ack'd
2293 * by advancing the cached remote message queue's Get value and if requested
2294 * send a chctl msgrequest to the message sender's partition.
2297 xpc_acknowledge_msgs_sn2(struct xpc_channel
*ch
, s64 initial_get
, u8 msg_flags
)
2299 struct xpc_channel_sn2
*ch_sn2
= &ch
->sn
.sn2
;
2300 struct xpc_msg
*msg
;
2301 s64 get
= initial_get
+ 1;
2302 int send_msgrequest
= 0;
2307 if (get
== ch_sn2
->w_local_GP
.get
)
2310 msg
= (struct xpc_msg
*)((u64
)ch
->remote_msgqueue
+
2311 (get
% ch
->remote_nentries
) *
2314 if (!(msg
->flags
& XPC_M_DONE
))
2317 msg_flags
|= msg
->flags
;
2321 if (get
== initial_get
) {
2322 /* nothing's changed */
2326 if (cmpxchg_rel(&ch_sn2
->local_GP
->get
, initial_get
, get
) !=
2328 /* someone else beat us to it */
2329 DBUG_ON(ch_sn2
->local_GP
->get
<= initial_get
);
2333 /* we just set the new value of local_GP->get */
2335 dev_dbg(xpc_chan
, "local_GP->get changed to %ld, partid=%d, "
2336 "channel=%d\n", get
, ch
->partid
, ch
->number
);
2338 send_msgrequest
= (msg_flags
& XPC_M_INTERRUPT
);
2341 * We need to ensure that the message referenced by
2342 * local_GP->get is not XPC_M_DONE or that local_GP->get
2343 * equals w_local_GP.get, so we'll go have a look.
2348 if (send_msgrequest
)
2349 xpc_send_chctl_msgrequest_sn2(ch
);
2353 xpc_received_msg_sn2(struct xpc_channel
*ch
, struct xpc_msg
*msg
)
2356 s64 msg_number
= msg
->number
;
2358 dev_dbg(xpc_chan
, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2359 (void *)msg
, msg_number
, ch
->partid
, ch
->number
);
2361 DBUG_ON((((u64
)msg
- (u64
)ch
->remote_msgqueue
) / ch
->msg_size
) !=
2362 msg_number
% ch
->remote_nentries
);
2363 DBUG_ON(msg
->flags
& XPC_M_DONE
);
2365 msg
->flags
|= XPC_M_DONE
;
2368 * The preceding store of msg->flags must occur before the following
2369 * load of local_GP->get.
2374 * See if this message is next in line to be acknowledged as having
2377 get
= ch
->sn
.sn2
.local_GP
->get
;
2378 if (get
== msg_number
)
2379 xpc_acknowledge_msgs_sn2(ch
, get
, msg
->flags
);
2388 xpc_get_partition_rsvd_page_pa
= xpc_get_partition_rsvd_page_pa_sn2
;
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_get_remote_heartbeat
= xpc_get_remote_heartbeat_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
);