]>
Commit | Line | Data |
---|---|---|
89eb8eb9 DN |
1 | /* |
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 | |
4 | * for more details. | |
5 | * | |
6 | * Copyright (c) 2004-2005 Silicon Graphics, Inc. All Rights Reserved. | |
7 | */ | |
8 | ||
9 | ||
10 | /* | |
11 | * Cross Partition Communication (XPC) support - standard version. | |
12 | * | |
13 | * XPC provides a message passing capability that crosses partition | |
14 | * boundaries. This module is made up of two parts: | |
15 | * | |
16 | * partition This part detects the presence/absence of other | |
17 | * partitions. It provides a heartbeat and monitors | |
18 | * the heartbeats of other partitions. | |
19 | * | |
20 | * channel This part manages the channels and sends/receives | |
21 | * messages across them to/from other partitions. | |
22 | * | |
23 | * There are a couple of additional functions residing in XP, which | |
24 | * provide an interface to XPC for its users. | |
25 | * | |
26 | * | |
27 | * Caveats: | |
28 | * | |
29 | * . We currently have no way to determine which nasid an IPI came | |
30 | * from. Thus, xpc_IPI_send() does a remote AMO write followed by | |
31 | * an IPI. The AMO indicates where data is to be pulled from, so | |
32 | * after the IPI arrives, the remote partition checks the AMO word. | |
33 | * The IPI can actually arrive before the AMO however, so other code | |
34 | * must periodically check for this case. Also, remote AMO operations | |
35 | * do not reliably time out. Thus we do a remote PIO read solely to | |
36 | * know whether the remote partition is down and whether we should | |
37 | * stop sending IPIs to it. This remote PIO read operation is set up | |
38 | * in a special nofault region so SAL knows to ignore (and cleanup) | |
39 | * any errors due to the remote AMO write, PIO read, and/or PIO | |
40 | * write operations. | |
41 | * | |
42 | * If/when new hardware solves this IPI problem, we should abandon | |
43 | * the current approach. | |
44 | * | |
45 | */ | |
46 | ||
47 | ||
48 | #include <linux/kernel.h> | |
49 | #include <linux/module.h> | |
50 | #include <linux/init.h> | |
51 | #include <linux/sched.h> | |
52 | #include <linux/syscalls.h> | |
53 | #include <linux/cache.h> | |
54 | #include <linux/interrupt.h> | |
55 | #include <linux/slab.h> | |
69913927 | 56 | #include <linux/delay.h> |
89eb8eb9 DN |
57 | #include <asm/sn/intr.h> |
58 | #include <asm/sn/sn_sal.h> | |
59 | #include <asm/uaccess.h> | |
60 | #include "xpc.h" | |
61 | ||
62 | ||
63 | /* define two XPC debug device structures to be used with dev_dbg() et al */ | |
64 | ||
65 | struct device_driver xpc_dbg_name = { | |
66 | .name = "xpc" | |
67 | }; | |
68 | ||
69 | struct device xpc_part_dbg_subname = { | |
70 | .bus_id = {0}, /* set to "part" at xpc_init() time */ | |
71 | .driver = &xpc_dbg_name | |
72 | }; | |
73 | ||
74 | struct device xpc_chan_dbg_subname = { | |
75 | .bus_id = {0}, /* set to "chan" at xpc_init() time */ | |
76 | .driver = &xpc_dbg_name | |
77 | }; | |
78 | ||
79 | struct device *xpc_part = &xpc_part_dbg_subname; | |
80 | struct device *xpc_chan = &xpc_chan_dbg_subname; | |
81 | ||
82 | ||
83 | /* systune related variables for /proc/sys directories */ | |
84 | ||
85 | static int xpc_hb_min = 1; | |
86 | static int xpc_hb_max = 10; | |
87 | ||
88 | static int xpc_hb_check_min = 10; | |
89 | static int xpc_hb_check_max = 120; | |
90 | ||
91 | static ctl_table xpc_sys_xpc_hb_dir[] = { | |
92 | { | |
93 | 1, | |
94 | "hb_interval", | |
95 | &xpc_hb_interval, | |
96 | sizeof(int), | |
97 | 0644, | |
98 | NULL, | |
99 | &proc_dointvec_minmax, | |
100 | &sysctl_intvec, | |
101 | NULL, | |
102 | &xpc_hb_min, &xpc_hb_max | |
103 | }, | |
104 | { | |
105 | 2, | |
106 | "hb_check_interval", | |
107 | &xpc_hb_check_interval, | |
108 | sizeof(int), | |
109 | 0644, | |
110 | NULL, | |
111 | &proc_dointvec_minmax, | |
112 | &sysctl_intvec, | |
113 | NULL, | |
114 | &xpc_hb_check_min, &xpc_hb_check_max | |
115 | }, | |
116 | {0} | |
117 | }; | |
118 | static ctl_table xpc_sys_xpc_dir[] = { | |
119 | { | |
120 | 1, | |
121 | "hb", | |
122 | NULL, | |
123 | 0, | |
124 | 0555, | |
125 | xpc_sys_xpc_hb_dir | |
126 | }, | |
127 | {0} | |
128 | }; | |
129 | static ctl_table xpc_sys_dir[] = { | |
130 | { | |
131 | 1, | |
132 | "xpc", | |
133 | NULL, | |
134 | 0, | |
135 | 0555, | |
136 | xpc_sys_xpc_dir | |
137 | }, | |
138 | {0} | |
139 | }; | |
140 | static struct ctl_table_header *xpc_sysctl; | |
141 | ||
142 | ||
143 | /* #of IRQs received */ | |
144 | static atomic_t xpc_act_IRQ_rcvd; | |
145 | ||
146 | /* IRQ handler notifies this wait queue on receipt of an IRQ */ | |
147 | static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq); | |
148 | ||
149 | static unsigned long xpc_hb_check_timeout; | |
150 | ||
151 | /* xpc_hb_checker thread exited notification */ | |
152 | static DECLARE_MUTEX_LOCKED(xpc_hb_checker_exited); | |
153 | ||
154 | /* xpc_discovery thread exited notification */ | |
155 | static DECLARE_MUTEX_LOCKED(xpc_discovery_exited); | |
156 | ||
157 | ||
158 | static struct timer_list xpc_hb_timer; | |
159 | ||
160 | ||
161 | static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *); | |
162 | ||
163 | ||
164 | /* | |
165 | * Notify the heartbeat check thread that an IRQ has been received. | |
166 | */ | |
167 | static irqreturn_t | |
168 | xpc_act_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs) | |
169 | { | |
170 | atomic_inc(&xpc_act_IRQ_rcvd); | |
171 | wake_up_interruptible(&xpc_act_IRQ_wq); | |
172 | return IRQ_HANDLED; | |
173 | } | |
174 | ||
175 | ||
176 | /* | |
177 | * Timer to produce the heartbeat. The timer structures function is | |
178 | * already set when this is initially called. A tunable is used to | |
179 | * specify when the next timeout should occur. | |
180 | */ | |
181 | static void | |
182 | xpc_hb_beater(unsigned long dummy) | |
183 | { | |
184 | xpc_vars->heartbeat++; | |
185 | ||
186 | if (jiffies >= xpc_hb_check_timeout) { | |
187 | wake_up_interruptible(&xpc_act_IRQ_wq); | |
188 | } | |
189 | ||
190 | xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ); | |
191 | add_timer(&xpc_hb_timer); | |
192 | } | |
193 | ||
194 | ||
195 | /* | |
196 | * This thread is responsible for nearly all of the partition | |
197 | * activation/deactivation. | |
198 | */ | |
199 | static int | |
200 | xpc_hb_checker(void *ignore) | |
201 | { | |
202 | int last_IRQ_count = 0; | |
203 | int new_IRQ_count; | |
204 | int force_IRQ=0; | |
205 | ||
206 | ||
207 | /* this thread was marked active by xpc_hb_init() */ | |
208 | ||
209 | daemonize(XPC_HB_CHECK_THREAD_NAME); | |
210 | ||
211 | set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU)); | |
212 | ||
213 | xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ); | |
214 | ||
215 | while (!(volatile int) xpc_exiting) { | |
216 | ||
217 | /* wait for IRQ or timeout */ | |
218 | (void) wait_event_interruptible(xpc_act_IRQ_wq, | |
219 | (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) || | |
220 | jiffies >= xpc_hb_check_timeout || | |
221 | (volatile int) xpc_exiting)); | |
222 | ||
223 | dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have " | |
224 | "been received\n", | |
225 | (int) (xpc_hb_check_timeout - jiffies), | |
226 | atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count); | |
227 | ||
228 | ||
229 | /* checking of remote heartbeats is skewed by IRQ handling */ | |
230 | if (jiffies >= xpc_hb_check_timeout) { | |
231 | dev_dbg(xpc_part, "checking remote heartbeats\n"); | |
232 | xpc_check_remote_hb(); | |
233 | ||
234 | /* | |
235 | * We need to periodically recheck to ensure no | |
236 | * IPI/AMO pairs have been missed. That check | |
237 | * must always reset xpc_hb_check_timeout. | |
238 | */ | |
239 | force_IRQ = 1; | |
240 | } | |
241 | ||
242 | ||
243 | new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd); | |
244 | if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) { | |
245 | force_IRQ = 0; | |
246 | ||
247 | dev_dbg(xpc_part, "found an IRQ to process; will be " | |
248 | "resetting xpc_hb_check_timeout\n"); | |
249 | ||
250 | last_IRQ_count += xpc_identify_act_IRQ_sender(); | |
251 | if (last_IRQ_count < new_IRQ_count) { | |
252 | /* retry once to help avoid missing AMO */ | |
253 | (void) xpc_identify_act_IRQ_sender(); | |
254 | } | |
255 | last_IRQ_count = new_IRQ_count; | |
256 | ||
257 | xpc_hb_check_timeout = jiffies + | |
258 | (xpc_hb_check_interval * HZ); | |
259 | } | |
260 | } | |
261 | ||
262 | dev_dbg(xpc_part, "heartbeat checker is exiting\n"); | |
263 | ||
264 | ||
265 | /* mark this thread as inactive */ | |
266 | up(&xpc_hb_checker_exited); | |
267 | return 0; | |
268 | } | |
269 | ||
270 | ||
271 | /* | |
272 | * This thread will attempt to discover other partitions to activate | |
273 | * based on info provided by SAL. This new thread is short lived and | |
274 | * will exit once discovery is complete. | |
275 | */ | |
276 | static int | |
277 | xpc_initiate_discovery(void *ignore) | |
278 | { | |
279 | daemonize(XPC_DISCOVERY_THREAD_NAME); | |
280 | ||
281 | xpc_discovery(); | |
282 | ||
283 | dev_dbg(xpc_part, "discovery thread is exiting\n"); | |
284 | ||
285 | /* mark this thread as inactive */ | |
286 | up(&xpc_discovery_exited); | |
287 | return 0; | |
288 | } | |
289 | ||
290 | ||
291 | /* | |
292 | * Establish first contact with the remote partititon. This involves pulling | |
293 | * the XPC per partition variables from the remote partition and waiting for | |
294 | * the remote partition to pull ours. | |
295 | */ | |
296 | static enum xpc_retval | |
297 | xpc_make_first_contact(struct xpc_partition *part) | |
298 | { | |
299 | enum xpc_retval ret; | |
300 | ||
301 | ||
302 | while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) { | |
303 | if (ret != xpcRetry) { | |
304 | XPC_DEACTIVATE_PARTITION(part, ret); | |
305 | return ret; | |
306 | } | |
307 | ||
308 | dev_dbg(xpc_chan, "waiting to make first contact with " | |
309 | "partition %d\n", XPC_PARTID(part)); | |
310 | ||
311 | /* wait a 1/4 of a second or so */ | |
69913927 | 312 | msleep_interruptible(250); |
89eb8eb9 DN |
313 | |
314 | if (part->act_state == XPC_P_DEACTIVATING) { | |
315 | return part->reason; | |
316 | } | |
317 | } | |
318 | ||
319 | return xpc_mark_partition_active(part); | |
320 | } | |
321 | ||
322 | ||
323 | /* | |
324 | * The first kthread assigned to a newly activated partition is the one | |
325 | * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to | |
326 | * that kthread until the partition is brought down, at which time that kthread | |
327 | * returns back to XPC HB. (The return of that kthread will signify to XPC HB | |
328 | * that XPC has dismantled all communication infrastructure for the associated | |
329 | * partition.) This kthread becomes the channel manager for that partition. | |
330 | * | |
331 | * Each active partition has a channel manager, who, besides connecting and | |
332 | * disconnecting channels, will ensure that each of the partition's connected | |
333 | * channels has the required number of assigned kthreads to get the work done. | |
334 | */ | |
335 | static void | |
336 | xpc_channel_mgr(struct xpc_partition *part) | |
337 | { | |
338 | while (part->act_state != XPC_P_DEACTIVATING || | |
339 | atomic_read(&part->nchannels_active) > 0) { | |
340 | ||
341 | xpc_process_channel_activity(part); | |
342 | ||
343 | ||
344 | /* | |
345 | * Wait until we've been requested to activate kthreads or | |
346 | * all of the channel's message queues have been torn down or | |
347 | * a signal is pending. | |
348 | * | |
349 | * The channel_mgr_requests is set to 1 after being awakened, | |
350 | * This is done to prevent the channel mgr from making one pass | |
351 | * through the loop for each request, since he will | |
352 | * be servicing all the requests in one pass. The reason it's | |
353 | * set to 1 instead of 0 is so that other kthreads will know | |
354 | * that the channel mgr is running and won't bother trying to | |
355 | * wake him up. | |
356 | */ | |
357 | atomic_dec(&part->channel_mgr_requests); | |
358 | (void) wait_event_interruptible(part->channel_mgr_wq, | |
359 | (atomic_read(&part->channel_mgr_requests) > 0 || | |
360 | (volatile u64) part->local_IPI_amo != 0 || | |
361 | ((volatile u8) part->act_state == | |
362 | XPC_P_DEACTIVATING && | |
363 | atomic_read(&part->nchannels_active) == 0))); | |
364 | atomic_set(&part->channel_mgr_requests, 1); | |
365 | ||
366 | // >>> Does it need to wakeup periodically as well? In case we | |
367 | // >>> miscalculated the #of kthreads to wakeup or create? | |
368 | } | |
369 | } | |
370 | ||
371 | ||
372 | /* | |
373 | * When XPC HB determines that a partition has come up, it will create a new | |
374 | * kthread and that kthread will call this function to attempt to set up the | |
375 | * basic infrastructure used for Cross Partition Communication with the newly | |
376 | * upped partition. | |
377 | * | |
378 | * The kthread that was created by XPC HB and which setup the XPC | |
379 | * infrastructure will remain assigned to the partition until the partition | |
380 | * goes down. At which time the kthread will teardown the XPC infrastructure | |
381 | * and then exit. | |
382 | * | |
383 | * XPC HB will put the remote partition's XPC per partition specific variables | |
384 | * physical address into xpc_partitions[partid].remote_vars_part_pa prior to | |
385 | * calling xpc_partition_up(). | |
386 | */ | |
387 | static void | |
388 | xpc_partition_up(struct xpc_partition *part) | |
389 | { | |
390 | DBUG_ON(part->channels != NULL); | |
391 | ||
392 | dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part)); | |
393 | ||
394 | if (xpc_setup_infrastructure(part) != xpcSuccess) { | |
395 | return; | |
396 | } | |
397 | ||
398 | /* | |
399 | * The kthread that XPC HB called us with will become the | |
400 | * channel manager for this partition. It will not return | |
401 | * back to XPC HB until the partition's XPC infrastructure | |
402 | * has been dismantled. | |
403 | */ | |
404 | ||
405 | (void) xpc_part_ref(part); /* this will always succeed */ | |
406 | ||
407 | if (xpc_make_first_contact(part) == xpcSuccess) { | |
408 | xpc_channel_mgr(part); | |
409 | } | |
410 | ||
411 | xpc_part_deref(part); | |
412 | ||
413 | xpc_teardown_infrastructure(part); | |
414 | } | |
415 | ||
416 | ||
417 | static int | |
418 | xpc_activating(void *__partid) | |
419 | { | |
420 | partid_t partid = (u64) __partid; | |
421 | struct xpc_partition *part = &xpc_partitions[partid]; | |
422 | unsigned long irq_flags; | |
d46523ea | 423 | struct sched_param param = { sched_priority: MAX_RT_PRIO - 1 }; |
89eb8eb9 DN |
424 | int ret; |
425 | ||
426 | ||
427 | DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS); | |
428 | ||
429 | spin_lock_irqsave(&part->act_lock, irq_flags); | |
430 | ||
431 | if (part->act_state == XPC_P_DEACTIVATING) { | |
432 | part->act_state = XPC_P_INACTIVE; | |
433 | spin_unlock_irqrestore(&part->act_lock, irq_flags); | |
434 | part->remote_rp_pa = 0; | |
435 | return 0; | |
436 | } | |
437 | ||
438 | /* indicate the thread is activating */ | |
439 | DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ); | |
440 | part->act_state = XPC_P_ACTIVATING; | |
441 | ||
442 | XPC_SET_REASON(part, 0, 0); | |
443 | spin_unlock_irqrestore(&part->act_lock, irq_flags); | |
444 | ||
445 | dev_dbg(xpc_part, "bringing partition %d up\n", partid); | |
446 | ||
447 | daemonize("xpc%02d", partid); | |
448 | ||
449 | /* | |
450 | * This thread needs to run at a realtime priority to prevent a | |
451 | * significant performance degradation. | |
452 | */ | |
453 | ret = sched_setscheduler(current, SCHED_FIFO, ¶m); | |
454 | if (ret != 0) { | |
455 | dev_warn(xpc_part, "unable to set pid %d to a realtime " | |
456 | "priority, ret=%d\n", current->pid, ret); | |
457 | } | |
458 | ||
459 | /* allow this thread and its children to run on any CPU */ | |
460 | set_cpus_allowed(current, CPU_MASK_ALL); | |
461 | ||
462 | /* | |
463 | * Register the remote partition's AMOs with SAL so it can handle | |
464 | * and cleanup errors within that address range should the remote | |
465 | * partition go down. We don't unregister this range because it is | |
466 | * difficult to tell when outstanding writes to the remote partition | |
467 | * are finished and thus when it is safe to unregister. This should | |
468 | * not result in wasted space in the SAL xp_addr_region table because | |
469 | * we should get the same page for remote_amos_page_pa after module | |
470 | * reloads and system reboots. | |
471 | */ | |
472 | if (sn_register_xp_addr_region(part->remote_amos_page_pa, | |
473 | PAGE_SIZE, 1) < 0) { | |
474 | dev_warn(xpc_part, "xpc_partition_up(%d) failed to register " | |
475 | "xp_addr region\n", partid); | |
476 | ||
477 | spin_lock_irqsave(&part->act_lock, irq_flags); | |
478 | part->act_state = XPC_P_INACTIVE; | |
479 | XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__); | |
480 | spin_unlock_irqrestore(&part->act_lock, irq_flags); | |
481 | part->remote_rp_pa = 0; | |
482 | return 0; | |
483 | } | |
484 | ||
485 | XPC_ALLOW_HB(partid, xpc_vars); | |
486 | xpc_IPI_send_activated(part); | |
487 | ||
488 | ||
489 | /* | |
490 | * xpc_partition_up() holds this thread and marks this partition as | |
491 | * XPC_P_ACTIVE by calling xpc_hb_mark_active(). | |
492 | */ | |
493 | (void) xpc_partition_up(part); | |
494 | ||
495 | xpc_mark_partition_inactive(part); | |
496 | ||
497 | if (part->reason == xpcReactivating) { | |
498 | /* interrupting ourselves results in activating partition */ | |
499 | xpc_IPI_send_reactivate(part); | |
500 | } | |
501 | ||
502 | return 0; | |
503 | } | |
504 | ||
505 | ||
506 | void | |
507 | xpc_activate_partition(struct xpc_partition *part) | |
508 | { | |
509 | partid_t partid = XPC_PARTID(part); | |
510 | unsigned long irq_flags; | |
511 | pid_t pid; | |
512 | ||
513 | ||
514 | spin_lock_irqsave(&part->act_lock, irq_flags); | |
515 | ||
516 | pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0); | |
517 | ||
518 | DBUG_ON(part->act_state != XPC_P_INACTIVE); | |
519 | ||
520 | if (pid > 0) { | |
521 | part->act_state = XPC_P_ACTIVATION_REQ; | |
522 | XPC_SET_REASON(part, xpcCloneKThread, __LINE__); | |
523 | } else { | |
524 | XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__); | |
525 | } | |
526 | ||
527 | spin_unlock_irqrestore(&part->act_lock, irq_flags); | |
528 | } | |
529 | ||
530 | ||
531 | /* | |
532 | * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified | |
533 | * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more | |
534 | * than one partition, we use an AMO_t structure per partition to indicate | |
535 | * whether a partition has sent an IPI or not. >>> If it has, then wake up the | |
536 | * associated kthread to handle it. | |
537 | * | |
538 | * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC | |
539 | * running on other partitions. | |
540 | * | |
541 | * Noteworthy Arguments: | |
542 | * | |
543 | * irq - Interrupt ReQuest number. NOT USED. | |
544 | * | |
545 | * dev_id - partid of IPI's potential sender. | |
546 | * | |
547 | * regs - processor's context before the processor entered | |
548 | * interrupt code. NOT USED. | |
549 | */ | |
550 | irqreturn_t | |
551 | xpc_notify_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs) | |
552 | { | |
553 | partid_t partid = (partid_t) (u64) dev_id; | |
554 | struct xpc_partition *part = &xpc_partitions[partid]; | |
555 | ||
556 | ||
557 | DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS); | |
558 | ||
559 | if (xpc_part_ref(part)) { | |
560 | xpc_check_for_channel_activity(part); | |
561 | ||
562 | xpc_part_deref(part); | |
563 | } | |
564 | return IRQ_HANDLED; | |
565 | } | |
566 | ||
567 | ||
568 | /* | |
569 | * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor | |
570 | * because the write to their associated IPI amo completed after the IRQ/IPI | |
571 | * was received. | |
572 | */ | |
573 | void | |
574 | xpc_dropped_IPI_check(struct xpc_partition *part) | |
575 | { | |
576 | if (xpc_part_ref(part)) { | |
577 | xpc_check_for_channel_activity(part); | |
578 | ||
579 | part->dropped_IPI_timer.expires = jiffies + | |
580 | XPC_P_DROPPED_IPI_WAIT; | |
581 | add_timer(&part->dropped_IPI_timer); | |
582 | xpc_part_deref(part); | |
583 | } | |
584 | } | |
585 | ||
586 | ||
587 | void | |
588 | xpc_activate_kthreads(struct xpc_channel *ch, int needed) | |
589 | { | |
590 | int idle = atomic_read(&ch->kthreads_idle); | |
591 | int assigned = atomic_read(&ch->kthreads_assigned); | |
592 | int wakeup; | |
593 | ||
594 | ||
595 | DBUG_ON(needed <= 0); | |
596 | ||
597 | if (idle > 0) { | |
598 | wakeup = (needed > idle) ? idle : needed; | |
599 | needed -= wakeup; | |
600 | ||
601 | dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, " | |
602 | "channel=%d\n", wakeup, ch->partid, ch->number); | |
603 | ||
604 | /* only wakeup the requested number of kthreads */ | |
605 | wake_up_nr(&ch->idle_wq, wakeup); | |
606 | } | |
607 | ||
608 | if (needed <= 0) { | |
609 | return; | |
610 | } | |
611 | ||
612 | if (needed + assigned > ch->kthreads_assigned_limit) { | |
613 | needed = ch->kthreads_assigned_limit - assigned; | |
614 | // >>>should never be less than 0 | |
615 | if (needed <= 0) { | |
616 | return; | |
617 | } | |
618 | } | |
619 | ||
620 | dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n", | |
621 | needed, ch->partid, ch->number); | |
622 | ||
623 | xpc_create_kthreads(ch, needed); | |
624 | } | |
625 | ||
626 | ||
627 | /* | |
628 | * This function is where XPC's kthreads wait for messages to deliver. | |
629 | */ | |
630 | static void | |
631 | xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch) | |
632 | { | |
633 | do { | |
634 | /* deliver messages to their intended recipients */ | |
635 | ||
636 | while ((volatile s64) ch->w_local_GP.get < | |
637 | (volatile s64) ch->w_remote_GP.put && | |
638 | !((volatile u32) ch->flags & | |
639 | XPC_C_DISCONNECTING)) { | |
640 | xpc_deliver_msg(ch); | |
641 | } | |
642 | ||
643 | if (atomic_inc_return(&ch->kthreads_idle) > | |
644 | ch->kthreads_idle_limit) { | |
645 | /* too many idle kthreads on this channel */ | |
646 | atomic_dec(&ch->kthreads_idle); | |
647 | break; | |
648 | } | |
649 | ||
650 | dev_dbg(xpc_chan, "idle kthread calling " | |
651 | "wait_event_interruptible_exclusive()\n"); | |
652 | ||
653 | (void) wait_event_interruptible_exclusive(ch->idle_wq, | |
654 | ((volatile s64) ch->w_local_GP.get < | |
655 | (volatile s64) ch->w_remote_GP.put || | |
656 | ((volatile u32) ch->flags & | |
657 | XPC_C_DISCONNECTING))); | |
658 | ||
659 | atomic_dec(&ch->kthreads_idle); | |
660 | ||
661 | } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING)); | |
662 | } | |
663 | ||
664 | ||
665 | static int | |
666 | xpc_daemonize_kthread(void *args) | |
667 | { | |
668 | partid_t partid = XPC_UNPACK_ARG1(args); | |
669 | u16 ch_number = XPC_UNPACK_ARG2(args); | |
670 | struct xpc_partition *part = &xpc_partitions[partid]; | |
671 | struct xpc_channel *ch; | |
672 | int n_needed; | |
673 | ||
674 | ||
675 | daemonize("xpc%02dc%d", partid, ch_number); | |
676 | ||
677 | dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n", | |
678 | partid, ch_number); | |
679 | ||
680 | ch = &part->channels[ch_number]; | |
681 | ||
682 | if (!(ch->flags & XPC_C_DISCONNECTING)) { | |
683 | DBUG_ON(!(ch->flags & XPC_C_CONNECTED)); | |
684 | ||
685 | /* let registerer know that connection has been established */ | |
686 | ||
687 | if (atomic_read(&ch->kthreads_assigned) == 1) { | |
688 | xpc_connected_callout(ch); | |
689 | ||
690 | /* | |
691 | * It is possible that while the callout was being | |
692 | * made that the remote partition sent some messages. | |
693 | * If that is the case, we may need to activate | |
694 | * additional kthreads to help deliver them. We only | |
695 | * need one less than total #of messages to deliver. | |
696 | */ | |
697 | n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1; | |
698 | if (n_needed > 0 && | |
699 | !(ch->flags & XPC_C_DISCONNECTING)) { | |
700 | xpc_activate_kthreads(ch, n_needed); | |
701 | } | |
702 | } | |
703 | ||
704 | xpc_kthread_waitmsgs(part, ch); | |
705 | } | |
706 | ||
707 | if (atomic_dec_return(&ch->kthreads_assigned) == 0 && | |
708 | ((ch->flags & XPC_C_CONNECTCALLOUT) || | |
709 | (ch->reason != xpcUnregistering && | |
710 | ch->reason != xpcOtherUnregistering))) { | |
711 | xpc_disconnected_callout(ch); | |
712 | } | |
713 | ||
714 | ||
715 | xpc_msgqueue_deref(ch); | |
716 | ||
717 | dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n", | |
718 | partid, ch_number); | |
719 | ||
720 | xpc_part_deref(part); | |
721 | return 0; | |
722 | } | |
723 | ||
724 | ||
725 | /* | |
726 | * For each partition that XPC has established communications with, there is | |
727 | * a minimum of one kernel thread assigned to perform any operation that | |
728 | * may potentially sleep or block (basically the callouts to the asynchronous | |
729 | * functions registered via xpc_connect()). | |
730 | * | |
731 | * Additional kthreads are created and destroyed by XPC as the workload | |
732 | * demands. | |
733 | * | |
734 | * A kthread is assigned to one of the active channels that exists for a given | |
735 | * partition. | |
736 | */ | |
737 | void | |
738 | xpc_create_kthreads(struct xpc_channel *ch, int needed) | |
739 | { | |
740 | unsigned long irq_flags; | |
741 | pid_t pid; | |
742 | u64 args = XPC_PACK_ARGS(ch->partid, ch->number); | |
743 | ||
744 | ||
745 | while (needed-- > 0) { | |
746 | pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0); | |
747 | if (pid < 0) { | |
748 | /* the fork failed */ | |
749 | ||
750 | if (atomic_read(&ch->kthreads_assigned) < | |
751 | ch->kthreads_idle_limit) { | |
752 | /* | |
753 | * Flag this as an error only if we have an | |
754 | * insufficient #of kthreads for the channel | |
755 | * to function. | |
756 | * | |
757 | * No xpc_msgqueue_ref() is needed here since | |
758 | * the channel mgr is doing this. | |
759 | */ | |
760 | spin_lock_irqsave(&ch->lock, irq_flags); | |
761 | XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources, | |
762 | &irq_flags); | |
763 | spin_unlock_irqrestore(&ch->lock, irq_flags); | |
764 | } | |
765 | break; | |
766 | } | |
767 | ||
768 | /* | |
769 | * The following is done on behalf of the newly created | |
770 | * kthread. That kthread is responsible for doing the | |
771 | * counterpart to the following before it exits. | |
772 | */ | |
773 | (void) xpc_part_ref(&xpc_partitions[ch->partid]); | |
774 | xpc_msgqueue_ref(ch); | |
775 | atomic_inc(&ch->kthreads_assigned); | |
776 | ch->kthreads_created++; // >>> temporary debug only!!! | |
777 | } | |
778 | } | |
779 | ||
780 | ||
781 | void | |
782 | xpc_disconnect_wait(int ch_number) | |
783 | { | |
784 | partid_t partid; | |
785 | struct xpc_partition *part; | |
786 | struct xpc_channel *ch; | |
787 | ||
788 | ||
789 | /* now wait for all callouts to the caller's function to cease */ | |
790 | for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) { | |
791 | part = &xpc_partitions[partid]; | |
792 | ||
793 | if (xpc_part_ref(part)) { | |
794 | ch = &part->channels[ch_number]; | |
795 | ||
796 | // >>> how do we keep from falling into the window between our check and going | |
797 | // >>> down and coming back up where sema is re-inited? | |
798 | if (ch->flags & XPC_C_SETUP) { | |
799 | (void) down(&ch->teardown_sema); | |
800 | } | |
801 | ||
802 | xpc_part_deref(part); | |
803 | } | |
804 | } | |
805 | } | |
806 | ||
807 | ||
808 | static void | |
809 | xpc_do_exit(void) | |
810 | { | |
811 | partid_t partid; | |
812 | int active_part_count; | |
813 | struct xpc_partition *part; | |
814 | ||
815 | ||
816 | /* now it's time to eliminate our heartbeat */ | |
817 | del_timer_sync(&xpc_hb_timer); | |
818 | xpc_vars->heartbeating_to_mask = 0; | |
819 | ||
820 | /* indicate to others that our reserved page is uninitialized */ | |
821 | xpc_rsvd_page->vars_pa = 0; | |
822 | ||
823 | /* | |
824 | * Ignore all incoming interrupts. Without interupts the heartbeat | |
825 | * checker won't activate any new partitions that may come up. | |
826 | */ | |
827 | free_irq(SGI_XPC_ACTIVATE, NULL); | |
828 | ||
829 | /* | |
830 | * Cause the heartbeat checker and the discovery threads to exit. | |
831 | * We don't want them attempting to activate new partitions as we | |
832 | * try to deactivate the existing ones. | |
833 | */ | |
834 | xpc_exiting = 1; | |
835 | wake_up_interruptible(&xpc_act_IRQ_wq); | |
836 | ||
837 | /* wait for the heartbeat checker thread to mark itself inactive */ | |
838 | down(&xpc_hb_checker_exited); | |
839 | ||
840 | /* wait for the discovery thread to mark itself inactive */ | |
841 | down(&xpc_discovery_exited); | |
842 | ||
843 | ||
69913927 | 844 | msleep_interruptible(300); |
89eb8eb9 DN |
845 | |
846 | ||
847 | /* wait for all partitions to become inactive */ | |
848 | ||
849 | do { | |
850 | active_part_count = 0; | |
851 | ||
852 | for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) { | |
853 | part = &xpc_partitions[partid]; | |
854 | if (part->act_state != XPC_P_INACTIVE) { | |
855 | active_part_count++; | |
856 | ||
857 | XPC_DEACTIVATE_PARTITION(part, xpcUnloading); | |
858 | } | |
859 | } | |
860 | ||
69913927 NA |
861 | if (active_part_count) |
862 | msleep_interruptible(300); | |
89eb8eb9 DN |
863 | } while (active_part_count > 0); |
864 | ||
865 | ||
866 | /* close down protections for IPI operations */ | |
867 | xpc_restrict_IPI_ops(); | |
868 | ||
869 | ||
870 | /* clear the interface to XPC's functions */ | |
871 | xpc_clear_interface(); | |
872 | ||
873 | if (xpc_sysctl) { | |
874 | unregister_sysctl_table(xpc_sysctl); | |
875 | } | |
876 | } | |
877 | ||
878 | ||
879 | int __init | |
880 | xpc_init(void) | |
881 | { | |
882 | int ret; | |
883 | partid_t partid; | |
884 | struct xpc_partition *part; | |
885 | pid_t pid; | |
886 | ||
887 | ||
408865ce DN |
888 | if (!ia64_platform_is("sn2")) { |
889 | return -ENODEV; | |
890 | } | |
891 | ||
89eb8eb9 DN |
892 | /* |
893 | * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng | |
894 | * both a partition's reserved page and its XPC variables. Its size was | |
895 | * based on the size of a reserved page. So we need to ensure that the | |
896 | * XPC variables will fit as well. | |
897 | */ | |
898 | if (XPC_VARS_ALIGNED_SIZE > XPC_RSVD_PAGE_ALIGNED_SIZE) { | |
899 | dev_err(xpc_part, "xpc_remote_copy_buffer is not big enough\n"); | |
900 | return -EPERM; | |
901 | } | |
902 | DBUG_ON((u64) xpc_remote_copy_buffer != | |
903 | L1_CACHE_ALIGN((u64) xpc_remote_copy_buffer)); | |
904 | ||
905 | snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part"); | |
906 | snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan"); | |
907 | ||
908 | xpc_sysctl = register_sysctl_table(xpc_sys_dir, 1); | |
909 | ||
910 | /* | |
911 | * The first few fields of each entry of xpc_partitions[] need to | |
912 | * be initialized now so that calls to xpc_connect() and | |
913 | * xpc_disconnect() can be made prior to the activation of any remote | |
914 | * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE | |
915 | * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING | |
916 | * PARTITION HAS BEEN ACTIVATED. | |
917 | */ | |
918 | for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) { | |
919 | part = &xpc_partitions[partid]; | |
920 | ||
921 | DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part)); | |
922 | ||
923 | part->act_IRQ_rcvd = 0; | |
924 | spin_lock_init(&part->act_lock); | |
925 | part->act_state = XPC_P_INACTIVE; | |
926 | XPC_SET_REASON(part, 0, 0); | |
927 | part->setup_state = XPC_P_UNSET; | |
928 | init_waitqueue_head(&part->teardown_wq); | |
929 | atomic_set(&part->references, 0); | |
930 | } | |
931 | ||
932 | /* | |
933 | * Open up protections for IPI operations (and AMO operations on | |
934 | * Shub 1.1 systems). | |
935 | */ | |
936 | xpc_allow_IPI_ops(); | |
937 | ||
938 | /* | |
939 | * Interrupts being processed will increment this atomic variable and | |
940 | * awaken the heartbeat thread which will process the interrupts. | |
941 | */ | |
942 | atomic_set(&xpc_act_IRQ_rcvd, 0); | |
943 | ||
944 | /* | |
945 | * This is safe to do before the xpc_hb_checker thread has started | |
946 | * because the handler releases a wait queue. If an interrupt is | |
947 | * received before the thread is waiting, it will not go to sleep, | |
948 | * but rather immediately process the interrupt. | |
949 | */ | |
950 | ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0, | |
951 | "xpc hb", NULL); | |
952 | if (ret != 0) { | |
953 | dev_err(xpc_part, "can't register ACTIVATE IRQ handler, " | |
954 | "errno=%d\n", -ret); | |
955 | ||
956 | xpc_restrict_IPI_ops(); | |
957 | ||
958 | if (xpc_sysctl) { | |
959 | unregister_sysctl_table(xpc_sysctl); | |
960 | } | |
961 | return -EBUSY; | |
962 | } | |
963 | ||
964 | /* | |
965 | * Fill the partition reserved page with the information needed by | |
966 | * other partitions to discover we are alive and establish initial | |
967 | * communications. | |
968 | */ | |
969 | xpc_rsvd_page = xpc_rsvd_page_init(); | |
970 | if (xpc_rsvd_page == NULL) { | |
971 | dev_err(xpc_part, "could not setup our reserved page\n"); | |
972 | ||
973 | free_irq(SGI_XPC_ACTIVATE, NULL); | |
974 | xpc_restrict_IPI_ops(); | |
975 | ||
976 | if (xpc_sysctl) { | |
977 | unregister_sysctl_table(xpc_sysctl); | |
978 | } | |
979 | return -EBUSY; | |
980 | } | |
981 | ||
982 | ||
983 | /* | |
984 | * Set the beating to other partitions into motion. This is | |
985 | * the last requirement for other partitions' discovery to | |
986 | * initiate communications with us. | |
987 | */ | |
988 | init_timer(&xpc_hb_timer); | |
989 | xpc_hb_timer.function = xpc_hb_beater; | |
990 | xpc_hb_beater(0); | |
991 | ||
992 | ||
993 | /* | |
994 | * The real work-horse behind xpc. This processes incoming | |
995 | * interrupts and monitors remote heartbeats. | |
996 | */ | |
997 | pid = kernel_thread(xpc_hb_checker, NULL, 0); | |
998 | if (pid < 0) { | |
999 | dev_err(xpc_part, "failed while forking hb check thread\n"); | |
1000 | ||
1001 | /* indicate to others that our reserved page is uninitialized */ | |
1002 | xpc_rsvd_page->vars_pa = 0; | |
1003 | ||
1004 | del_timer_sync(&xpc_hb_timer); | |
1005 | free_irq(SGI_XPC_ACTIVATE, NULL); | |
1006 | xpc_restrict_IPI_ops(); | |
1007 | ||
1008 | if (xpc_sysctl) { | |
1009 | unregister_sysctl_table(xpc_sysctl); | |
1010 | } | |
1011 | return -EBUSY; | |
1012 | } | |
1013 | ||
1014 | ||
1015 | /* | |
1016 | * Startup a thread that will attempt to discover other partitions to | |
1017 | * activate based on info provided by SAL. This new thread is short | |
1018 | * lived and will exit once discovery is complete. | |
1019 | */ | |
1020 | pid = kernel_thread(xpc_initiate_discovery, NULL, 0); | |
1021 | if (pid < 0) { | |
1022 | dev_err(xpc_part, "failed while forking discovery thread\n"); | |
1023 | ||
1024 | /* mark this new thread as a non-starter */ | |
1025 | up(&xpc_discovery_exited); | |
1026 | ||
1027 | xpc_do_exit(); | |
1028 | return -EBUSY; | |
1029 | } | |
1030 | ||
1031 | ||
1032 | /* set the interface to point at XPC's functions */ | |
1033 | xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect, | |
1034 | xpc_initiate_allocate, xpc_initiate_send, | |
1035 | xpc_initiate_send_notify, xpc_initiate_received, | |
1036 | xpc_initiate_partid_to_nasids); | |
1037 | ||
1038 | return 0; | |
1039 | } | |
1040 | module_init(xpc_init); | |
1041 | ||
1042 | ||
1043 | void __exit | |
1044 | xpc_exit(void) | |
1045 | { | |
1046 | xpc_do_exit(); | |
1047 | } | |
1048 | module_exit(xpc_exit); | |
1049 | ||
1050 | ||
1051 | MODULE_AUTHOR("Silicon Graphics, Inc."); | |
1052 | MODULE_DESCRIPTION("Cross Partition Communication (XPC) support"); | |
1053 | MODULE_LICENSE("GPL"); | |
1054 | ||
1055 | module_param(xpc_hb_interval, int, 0); | |
1056 | MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between " | |
1057 | "heartbeat increments."); | |
1058 | ||
1059 | module_param(xpc_hb_check_interval, int, 0); | |
1060 | MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between " | |
1061 | "heartbeat checks."); | |
1062 |