Commit | Line | Data |
---|---|---|
28bffaf0 JS |
1 | /* |
2 | * SN Platform GRU Driver | |
3 | * | |
4 | * KERNEL SERVICES THAT USE THE GRU | |
5 | * | |
6 | * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved. | |
7 | * | |
8 | * This program is free software; you can redistribute it and/or modify | |
9 | * it under the terms of the GNU General Public License as published by | |
10 | * the Free Software Foundation; either version 2 of the License, or | |
11 | * (at your option) any later version. | |
12 | * | |
13 | * This program is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | * GNU General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU General Public License | |
19 | * along with this program; if not, write to the Free Software | |
20 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
21 | */ | |
22 | ||
23 | #include <linux/kernel.h> | |
24 | #include <linux/errno.h> | |
25 | #include <linux/slab.h> | |
26 | #include <linux/mm.h> | |
27 | #include <linux/smp_lock.h> | |
28 | #include <linux/spinlock.h> | |
29 | #include <linux/device.h> | |
30 | #include <linux/miscdevice.h> | |
31 | #include <linux/proc_fs.h> | |
32 | #include <linux/interrupt.h> | |
33 | #include <linux/uaccess.h> | |
836ce679 | 34 | #include <linux/delay.h> |
28bffaf0 JS |
35 | #include "gru.h" |
36 | #include "grulib.h" | |
37 | #include "grutables.h" | |
38 | #include "grukservices.h" | |
39 | #include "gru_instructions.h" | |
40 | #include <asm/uv/uv_hub.h> | |
41 | ||
42 | /* | |
43 | * Kernel GRU Usage | |
44 | * | |
45 | * The following is an interim algorithm for management of kernel GRU | |
46 | * resources. This will likely be replaced when we better understand the | |
47 | * kernel/user requirements. | |
48 | * | |
836ce679 JS |
49 | * Blade percpu resources reserved for kernel use. These resources are |
50 | * reserved whenever the the kernel context for the blade is loaded. Note | |
51 | * that the kernel context is not guaranteed to be always available. It is | |
52 | * loaded on demand & can be stolen by a user if the user demand exceeds the | |
53 | * kernel demand. The kernel can always reload the kernel context but | |
54 | * a SLEEP may be required!!!. | |
9120dec4 JS |
55 | * |
56 | * Async Overview: | |
57 | * | |
58 | * Each blade has one "kernel context" that owns GRU kernel resources | |
59 | * located on the blade. Kernel drivers use GRU resources in this context | |
60 | * for sending messages, zeroing memory, etc. | |
61 | * | |
62 | * The kernel context is dynamically loaded on demand. If it is not in | |
63 | * use by the kernel, the kernel context can be unloaded & given to a user. | |
64 | * The kernel context will be reloaded when needed. This may require that | |
65 | * a context be stolen from a user. | |
66 | * NOTE: frequent unloading/reloading of the kernel context is | |
67 | * expensive. We are depending on batch schedulers, cpusets, sane | |
68 | * drivers or some other mechanism to prevent the need for frequent | |
69 | * stealing/reloading. | |
70 | * | |
71 | * The kernel context consists of two parts: | |
72 | * - 1 CB & a few DSRs that are reserved for each cpu on the blade. | |
73 | * Each cpu has it's own private resources & does not share them | |
74 | * with other cpus. These resources are used serially, ie, | |
75 | * locked, used & unlocked on each call to a function in | |
76 | * grukservices. | |
77 | * (Now that we have dynamic loading of kernel contexts, I | |
78 | * may rethink this & allow sharing between cpus....) | |
79 | * | |
80 | * - Additional resources can be reserved long term & used directly | |
81 | * by UV drivers located in the kernel. Drivers using these GRU | |
82 | * resources can use asynchronous GRU instructions that send | |
83 | * interrupts on completion. | |
84 | * - these resources must be explicitly locked/unlocked | |
85 | * - locked resources prevent (obviously) the kernel | |
86 | * context from being unloaded. | |
87 | * - drivers using these resource directly issue their own | |
88 | * GRU instruction and must wait/check completion. | |
89 | * | |
90 | * When these resources are reserved, the caller can optionally | |
91 | * associate a wait_queue with the resources and use asynchronous | |
92 | * GRU instructions. When an async GRU instruction completes, the | |
93 | * driver will do a wakeup on the event. | |
94 | * | |
28bffaf0 | 95 | */ |
9120dec4 JS |
96 | |
97 | ||
98 | #define ASYNC_HAN_TO_BID(h) ((h) - 1) | |
99 | #define ASYNC_BID_TO_HAN(b) ((b) + 1) | |
100 | #define ASYNC_HAN_TO_BS(h) gru_base[ASYNC_HAN_TO_BID(h)] | |
101 | ||
6f2584f4 | 102 | #define GRU_NUM_KERNEL_CBR 1 |
28bffaf0 | 103 | #define GRU_NUM_KERNEL_DSR_BYTES 256 |
6f2584f4 JS |
104 | #define GRU_NUM_KERNEL_DSR_CL (GRU_NUM_KERNEL_DSR_BYTES / \ |
105 | GRU_CACHE_LINE_BYTES) | |
28bffaf0 JS |
106 | |
107 | /* GRU instruction attributes for all instructions */ | |
108 | #define IMA IMA_CB_DELAY | |
109 | ||
110 | /* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */ | |
111 | #define __gru_cacheline_aligned__ \ | |
112 | __attribute__((__aligned__(GRU_CACHE_LINE_BYTES))) | |
113 | ||
114 | #define MAGIC 0x1234567887654321UL | |
115 | ||
116 | /* Default retry count for GRU errors on kernel instructions */ | |
117 | #define EXCEPTION_RETRY_LIMIT 3 | |
118 | ||
119 | /* Status of message queue sections */ | |
120 | #define MQS_EMPTY 0 | |
121 | #define MQS_FULL 1 | |
122 | #define MQS_NOOP 2 | |
123 | ||
124 | /*----------------- RESOURCE MANAGEMENT -------------------------------------*/ | |
125 | /* optimized for x86_64 */ | |
126 | struct message_queue { | |
127 | union gru_mesqhead head __gru_cacheline_aligned__; /* CL 0 */ | |
128 | int qlines; /* DW 1 */ | |
129 | long hstatus[2]; | |
130 | void *next __gru_cacheline_aligned__;/* CL 1 */ | |
131 | void *limit; | |
132 | void *start; | |
133 | void *start2; | |
134 | char data ____cacheline_aligned; /* CL 2 */ | |
135 | }; | |
136 | ||
137 | /* First word in every message - used by mesq interface */ | |
138 | struct message_header { | |
139 | char present; | |
140 | char present2; | |
141 | char lines; | |
142 | char fill; | |
143 | }; | |
144 | ||
28bffaf0 JS |
145 | #define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h])) |
146 | ||
836ce679 JS |
147 | /* |
148 | * Reload the blade's kernel context into a GRU chiplet. Called holding | |
149 | * the bs_kgts_sema for READ. Will steal user contexts if necessary. | |
150 | */ | |
151 | static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id) | |
152 | { | |
153 | struct gru_state *gru; | |
154 | struct gru_thread_state *kgts; | |
155 | void *vaddr; | |
9120dec4 | 156 | int ctxnum, ncpus; |
836ce679 JS |
157 | |
158 | up_read(&bs->bs_kgts_sema); | |
159 | down_write(&bs->bs_kgts_sema); | |
160 | ||
161 | if (!bs->bs_kgts) | |
9120dec4 | 162 | bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0); |
836ce679 JS |
163 | kgts = bs->bs_kgts; |
164 | ||
165 | if (!kgts->ts_gru) { | |
166 | STAT(load_kernel_context); | |
9120dec4 JS |
167 | ncpus = uv_blade_nr_possible_cpus(blade_id); |
168 | kgts->ts_cbr_au_count = GRU_CB_COUNT_TO_AU( | |
169 | GRU_NUM_KERNEL_CBR * ncpus + bs->bs_async_cbrs); | |
170 | kgts->ts_dsr_au_count = GRU_DS_BYTES_TO_AU( | |
171 | GRU_NUM_KERNEL_DSR_BYTES * ncpus + | |
172 | bs->bs_async_dsr_bytes); | |
836ce679 JS |
173 | while (!gru_assign_gru_context(kgts, blade_id)) { |
174 | msleep(1); | |
175 | gru_steal_context(kgts, blade_id); | |
176 | } | |
177 | gru_load_context(kgts); | |
178 | gru = bs->bs_kgts->ts_gru; | |
179 | vaddr = gru->gs_gru_base_vaddr; | |
180 | ctxnum = kgts->ts_ctxnum; | |
181 | bs->kernel_cb = get_gseg_base_address_cb(vaddr, ctxnum, 0); | |
182 | bs->kernel_dsr = get_gseg_base_address_ds(vaddr, ctxnum, 0); | |
183 | } | |
184 | downgrade_write(&bs->bs_kgts_sema); | |
185 | } | |
186 | ||
187 | /* | |
188 | * Lock & load the kernel context for the specified blade. | |
189 | */ | |
190 | static struct gru_blade_state *gru_lock_kernel_context(int blade_id) | |
191 | { | |
192 | struct gru_blade_state *bs; | |
193 | ||
194 | STAT(lock_kernel_context); | |
195 | bs = gru_base[blade_id]; | |
196 | ||
197 | down_read(&bs->bs_kgts_sema); | |
198 | if (!bs->bs_kgts || !bs->bs_kgts->ts_gru) | |
199 | gru_load_kernel_context(bs, blade_id); | |
200 | return bs; | |
201 | ||
202 | } | |
203 | ||
204 | /* | |
205 | * Unlock the kernel context for the specified blade. Context is not | |
206 | * unloaded but may be stolen before next use. | |
207 | */ | |
208 | static void gru_unlock_kernel_context(int blade_id) | |
209 | { | |
210 | struct gru_blade_state *bs; | |
211 | ||
212 | bs = gru_base[blade_id]; | |
213 | up_read(&bs->bs_kgts_sema); | |
214 | STAT(unlock_kernel_context); | |
215 | } | |
216 | ||
217 | /* | |
218 | * Reserve & get pointers to the DSR/CBRs reserved for the current cpu. | |
219 | * - returns with preemption disabled | |
220 | */ | |
28bffaf0 JS |
221 | static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr) |
222 | { | |
223 | struct gru_blade_state *bs; | |
224 | int lcpu; | |
225 | ||
226 | BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES); | |
227 | preempt_disable(); | |
836ce679 | 228 | bs = gru_lock_kernel_context(uv_numa_blade_id()); |
28bffaf0 JS |
229 | lcpu = uv_blade_processor_id(); |
230 | *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE; | |
231 | *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES; | |
232 | return 0; | |
233 | } | |
234 | ||
836ce679 JS |
235 | /* |
236 | * Free the current cpus reserved DSR/CBR resources. | |
237 | */ | |
28bffaf0 JS |
238 | static void gru_free_cpu_resources(void *cb, void *dsr) |
239 | { | |
836ce679 | 240 | gru_unlock_kernel_context(uv_numa_blade_id()); |
28bffaf0 JS |
241 | preempt_enable(); |
242 | } | |
243 | ||
9120dec4 JS |
244 | /* |
245 | * Reserve GRU resources to be used asynchronously. | |
246 | * Note: currently supports only 1 reservation per blade. | |
247 | * | |
248 | * input: | |
249 | * blade_id - blade on which resources should be reserved | |
250 | * cbrs - number of CBRs | |
251 | * dsr_bytes - number of DSR bytes needed | |
252 | * output: | |
253 | * handle to identify resource | |
254 | * (0 = async resources already reserved) | |
255 | */ | |
256 | unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes, | |
257 | struct completion *cmp) | |
258 | { | |
259 | struct gru_blade_state *bs; | |
260 | struct gru_thread_state *kgts; | |
261 | int ret = 0; | |
262 | ||
263 | bs = gru_base[blade_id]; | |
264 | ||
265 | down_write(&bs->bs_kgts_sema); | |
266 | ||
267 | /* Verify no resources already reserved */ | |
268 | if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs) | |
269 | goto done; | |
270 | bs->bs_async_dsr_bytes = dsr_bytes; | |
271 | bs->bs_async_cbrs = cbrs; | |
272 | bs->bs_async_wq = cmp; | |
273 | kgts = bs->bs_kgts; | |
274 | ||
275 | /* Resources changed. Unload context if already loaded */ | |
276 | if (kgts && kgts->ts_gru) | |
277 | gru_unload_context(kgts, 0); | |
278 | ret = ASYNC_BID_TO_HAN(blade_id); | |
279 | ||
280 | done: | |
281 | up_write(&bs->bs_kgts_sema); | |
282 | return ret; | |
283 | } | |
284 | ||
285 | /* | |
286 | * Release async resources previously reserved. | |
287 | * | |
288 | * input: | |
289 | * han - handle to identify resources | |
290 | */ | |
291 | void gru_release_async_resources(unsigned long han) | |
292 | { | |
293 | struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han); | |
294 | ||
295 | down_write(&bs->bs_kgts_sema); | |
296 | bs->bs_async_dsr_bytes = 0; | |
297 | bs->bs_async_cbrs = 0; | |
298 | bs->bs_async_wq = NULL; | |
299 | up_write(&bs->bs_kgts_sema); | |
300 | } | |
301 | ||
302 | /* | |
303 | * Wait for async GRU instructions to complete. | |
304 | * | |
305 | * input: | |
306 | * han - handle to identify resources | |
307 | */ | |
308 | void gru_wait_async_cbr(unsigned long han) | |
309 | { | |
310 | struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han); | |
311 | ||
312 | wait_for_completion(bs->bs_async_wq); | |
313 | mb(); | |
314 | } | |
315 | ||
316 | /* | |
317 | * Lock previous reserved async GRU resources | |
318 | * | |
319 | * input: | |
320 | * han - handle to identify resources | |
321 | * output: | |
322 | * cb - pointer to first CBR | |
323 | * dsr - pointer to first DSR | |
324 | */ | |
325 | void gru_lock_async_resource(unsigned long han, void **cb, void **dsr) | |
326 | { | |
327 | struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han); | |
328 | int blade_id = ASYNC_HAN_TO_BID(han); | |
329 | int ncpus; | |
330 | ||
331 | gru_lock_kernel_context(blade_id); | |
332 | ncpus = uv_blade_nr_possible_cpus(blade_id); | |
333 | if (cb) | |
334 | *cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE; | |
335 | if (dsr) | |
336 | *dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES; | |
337 | } | |
338 | ||
339 | /* | |
340 | * Unlock previous reserved async GRU resources | |
341 | * | |
342 | * input: | |
343 | * han - handle to identify resources | |
344 | */ | |
345 | void gru_unlock_async_resource(unsigned long han) | |
346 | { | |
347 | int blade_id = ASYNC_HAN_TO_BID(han); | |
348 | ||
349 | gru_unlock_kernel_context(blade_id); | |
350 | } | |
351 | ||
836ce679 | 352 | /*----------------------------------------------------------------------*/ |
28bffaf0 JS |
353 | int gru_get_cb_exception_detail(void *cb, |
354 | struct control_block_extended_exc_detail *excdet) | |
355 | { | |
356 | struct gru_control_block_extended *cbe; | |
357 | ||
358 | cbe = get_cbe(GRUBASE(cb), get_cb_number(cb)); | |
fe5bb6b0 | 359 | prefetchw(cbe); /* Harmless on hardware, required for emulator */ |
28bffaf0 JS |
360 | excdet->opc = cbe->opccpy; |
361 | excdet->exopc = cbe->exopccpy; | |
362 | excdet->ecause = cbe->ecause; | |
363 | excdet->exceptdet0 = cbe->idef1upd; | |
364 | excdet->exceptdet1 = cbe->idef3upd; | |
365 | return 0; | |
366 | } | |
367 | ||
368 | char *gru_get_cb_exception_detail_str(int ret, void *cb, | |
369 | char *buf, int size) | |
370 | { | |
371 | struct gru_control_block_status *gen = (void *)cb; | |
372 | struct control_block_extended_exc_detail excdet; | |
373 | ||
374 | if (ret > 0 && gen->istatus == CBS_EXCEPTION) { | |
375 | gru_get_cb_exception_detail(cb, &excdet); | |
376 | snprintf(buf, size, | |
377 | "GRU exception: cb %p, opc %d, exopc %d, ecause 0x%x," | |
378 | "excdet0 0x%lx, excdet1 0x%x", | |
379 | gen, excdet.opc, excdet.exopc, excdet.ecause, | |
380 | excdet.exceptdet0, excdet.exceptdet1); | |
381 | } else { | |
382 | snprintf(buf, size, "No exception"); | |
383 | } | |
384 | return buf; | |
385 | } | |
386 | ||
387 | static int gru_wait_idle_or_exception(struct gru_control_block_status *gen) | |
388 | { | |
389 | while (gen->istatus >= CBS_ACTIVE) { | |
390 | cpu_relax(); | |
391 | barrier(); | |
392 | } | |
393 | return gen->istatus; | |
394 | } | |
395 | ||
396 | static int gru_retry_exception(void *cb) | |
397 | { | |
398 | struct gru_control_block_status *gen = (void *)cb; | |
399 | struct control_block_extended_exc_detail excdet; | |
400 | int retry = EXCEPTION_RETRY_LIMIT; | |
401 | ||
402 | while (1) { | |
403 | if (gru_get_cb_message_queue_substatus(cb)) | |
404 | break; | |
405 | if (gru_wait_idle_or_exception(gen) == CBS_IDLE) | |
406 | return CBS_IDLE; | |
407 | ||
408 | gru_get_cb_exception_detail(cb, &excdet); | |
409 | if (excdet.ecause & ~EXCEPTION_RETRY_BITS) | |
410 | break; | |
411 | if (retry-- == 0) | |
412 | break; | |
413 | gen->icmd = 1; | |
414 | gru_flush_cache(gen); | |
415 | } | |
416 | return CBS_EXCEPTION; | |
417 | } | |
418 | ||
419 | int gru_check_status_proc(void *cb) | |
420 | { | |
421 | struct gru_control_block_status *gen = (void *)cb; | |
422 | int ret; | |
423 | ||
424 | ret = gen->istatus; | |
425 | if (ret != CBS_EXCEPTION) | |
426 | return ret; | |
427 | return gru_retry_exception(cb); | |
428 | ||
429 | } | |
430 | ||
431 | int gru_wait_proc(void *cb) | |
432 | { | |
433 | struct gru_control_block_status *gen = (void *)cb; | |
434 | int ret; | |
435 | ||
436 | ret = gru_wait_idle_or_exception(gen); | |
437 | if (ret == CBS_EXCEPTION) | |
438 | ret = gru_retry_exception(cb); | |
439 | ||
440 | return ret; | |
441 | } | |
442 | ||
443 | void gru_abort(int ret, void *cb, char *str) | |
444 | { | |
445 | char buf[GRU_EXC_STR_SIZE]; | |
446 | ||
447 | panic("GRU FATAL ERROR: %s - %s\n", str, | |
448 | gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf))); | |
449 | } | |
450 | ||
451 | void gru_wait_abort_proc(void *cb) | |
452 | { | |
453 | int ret; | |
454 | ||
455 | ret = gru_wait_proc(cb); | |
456 | if (ret) | |
457 | gru_abort(ret, cb, "gru_wait_abort"); | |
458 | } | |
459 | ||
460 | ||
461 | /*------------------------------ MESSAGE QUEUES -----------------------------*/ | |
462 | ||
463 | /* Internal status . These are NOT returned to the user. */ | |
464 | #define MQIE_AGAIN -1 /* try again */ | |
465 | ||
466 | ||
467 | /* | |
468 | * Save/restore the "present" flag that is in the second line of 2-line | |
469 | * messages | |
470 | */ | |
471 | static inline int get_present2(void *p) | |
472 | { | |
473 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; | |
474 | return mhdr->present; | |
475 | } | |
476 | ||
477 | static inline void restore_present2(void *p, int val) | |
478 | { | |
479 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; | |
480 | mhdr->present = val; | |
481 | } | |
482 | ||
483 | /* | |
484 | * Create a message queue. | |
485 | * qlines - message queue size in cache lines. Includes 2-line header. | |
486 | */ | |
6f2584f4 JS |
487 | int gru_create_message_queue(struct gru_message_queue_desc *mqd, |
488 | void *p, unsigned int bytes, int nasid, int vector, int apicid) | |
28bffaf0 JS |
489 | { |
490 | struct message_queue *mq = p; | |
491 | unsigned int qlines; | |
492 | ||
493 | qlines = bytes / GRU_CACHE_LINE_BYTES - 2; | |
494 | memset(mq, 0, bytes); | |
495 | mq->start = &mq->data; | |
496 | mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES; | |
497 | mq->next = &mq->data; | |
498 | mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES; | |
499 | mq->qlines = qlines; | |
500 | mq->hstatus[0] = 0; | |
501 | mq->hstatus[1] = 1; | |
502 | mq->head = gru_mesq_head(2, qlines / 2 + 1); | |
6f2584f4 JS |
503 | mqd->mq = mq; |
504 | mqd->mq_gpa = uv_gpa(mq); | |
505 | mqd->qlines = qlines; | |
506 | mqd->interrupt_pnode = UV_NASID_TO_PNODE(nasid); | |
507 | mqd->interrupt_vector = vector; | |
508 | mqd->interrupt_apicid = apicid; | |
28bffaf0 JS |
509 | return 0; |
510 | } | |
511 | EXPORT_SYMBOL_GPL(gru_create_message_queue); | |
512 | ||
513 | /* | |
514 | * Send a NOOP message to a message queue | |
515 | * Returns: | |
516 | * 0 - if queue is full after the send. This is the normal case | |
517 | * but various races can change this. | |
518 | * -1 - if mesq sent successfully but queue not full | |
519 | * >0 - unexpected error. MQE_xxx returned | |
520 | */ | |
6f2584f4 JS |
521 | static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd, |
522 | void *mesg) | |
28bffaf0 JS |
523 | { |
524 | const struct message_header noop_header = { | |
525 | .present = MQS_NOOP, .lines = 1}; | |
526 | unsigned long m; | |
527 | int substatus, ret; | |
528 | struct message_header save_mhdr, *mhdr = mesg; | |
529 | ||
530 | STAT(mesq_noop); | |
531 | save_mhdr = *mhdr; | |
532 | *mhdr = noop_header; | |
6f2584f4 | 533 | gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA); |
28bffaf0 JS |
534 | ret = gru_wait(cb); |
535 | ||
536 | if (ret) { | |
537 | substatus = gru_get_cb_message_queue_substatus(cb); | |
538 | switch (substatus) { | |
539 | case CBSS_NO_ERROR: | |
540 | STAT(mesq_noop_unexpected_error); | |
541 | ret = MQE_UNEXPECTED_CB_ERR; | |
542 | break; | |
543 | case CBSS_LB_OVERFLOWED: | |
544 | STAT(mesq_noop_lb_overflow); | |
545 | ret = MQE_CONGESTION; | |
546 | break; | |
547 | case CBSS_QLIMIT_REACHED: | |
548 | STAT(mesq_noop_qlimit_reached); | |
549 | ret = 0; | |
550 | break; | |
551 | case CBSS_AMO_NACKED: | |
552 | STAT(mesq_noop_amo_nacked); | |
553 | ret = MQE_CONGESTION; | |
554 | break; | |
555 | case CBSS_PUT_NACKED: | |
556 | STAT(mesq_noop_put_nacked); | |
6f2584f4 | 557 | m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6); |
28bffaf0 JS |
558 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1, |
559 | IMA); | |
560 | if (gru_wait(cb) == CBS_IDLE) | |
561 | ret = MQIE_AGAIN; | |
562 | else | |
563 | ret = MQE_UNEXPECTED_CB_ERR; | |
564 | break; | |
565 | case CBSS_PAGE_OVERFLOW: | |
566 | default: | |
567 | BUG(); | |
568 | } | |
569 | } | |
570 | *mhdr = save_mhdr; | |
571 | return ret; | |
572 | } | |
573 | ||
574 | /* | |
575 | * Handle a gru_mesq full. | |
576 | */ | |
6f2584f4 JS |
577 | static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd, |
578 | void *mesg, int lines) | |
28bffaf0 JS |
579 | { |
580 | union gru_mesqhead mqh; | |
581 | unsigned int limit, head; | |
582 | unsigned long avalue; | |
6f2584f4 | 583 | int half, qlines; |
28bffaf0 JS |
584 | |
585 | /* Determine if switching to first/second half of q */ | |
586 | avalue = gru_get_amo_value(cb); | |
587 | head = gru_get_amo_value_head(cb); | |
588 | limit = gru_get_amo_value_limit(cb); | |
589 | ||
6f2584f4 | 590 | qlines = mqd->qlines; |
28bffaf0 JS |
591 | half = (limit != qlines); |
592 | ||
593 | if (half) | |
594 | mqh = gru_mesq_head(qlines / 2 + 1, qlines); | |
595 | else | |
596 | mqh = gru_mesq_head(2, qlines / 2 + 1); | |
597 | ||
598 | /* Try to get lock for switching head pointer */ | |
6f2584f4 | 599 | gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA); |
28bffaf0 JS |
600 | if (gru_wait(cb) != CBS_IDLE) |
601 | goto cberr; | |
602 | if (!gru_get_amo_value(cb)) { | |
603 | STAT(mesq_qf_locked); | |
604 | return MQE_QUEUE_FULL; | |
605 | } | |
606 | ||
607 | /* Got the lock. Send optional NOP if queue not full, */ | |
608 | if (head != limit) { | |
6f2584f4 JS |
609 | if (send_noop_message(cb, mqd, mesg)) { |
610 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), | |
28bffaf0 JS |
611 | XTYPE_DW, IMA); |
612 | if (gru_wait(cb) != CBS_IDLE) | |
613 | goto cberr; | |
614 | STAT(mesq_qf_noop_not_full); | |
615 | return MQIE_AGAIN; | |
616 | } | |
617 | avalue++; | |
618 | } | |
619 | ||
620 | /* Then flip queuehead to other half of queue. */ | |
6f2584f4 JS |
621 | gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue, |
622 | IMA); | |
28bffaf0 JS |
623 | if (gru_wait(cb) != CBS_IDLE) |
624 | goto cberr; | |
625 | ||
626 | /* If not successfully in swapping queue head, clear the hstatus lock */ | |
627 | if (gru_get_amo_value(cb) != avalue) { | |
628 | STAT(mesq_qf_switch_head_failed); | |
6f2584f4 JS |
629 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, |
630 | IMA); | |
28bffaf0 JS |
631 | if (gru_wait(cb) != CBS_IDLE) |
632 | goto cberr; | |
633 | } | |
634 | return MQIE_AGAIN; | |
635 | cberr: | |
636 | STAT(mesq_qf_unexpected_error); | |
637 | return MQE_UNEXPECTED_CB_ERR; | |
638 | } | |
639 | ||
6f2584f4 JS |
640 | /* |
641 | * Send a cross-partition interrupt to the SSI that contains the target | |
642 | * message queue. Normally, the interrupt is automatically delivered by hardware | |
643 | * but some error conditions require explicit delivery. | |
644 | */ | |
645 | static void send_message_queue_interrupt(struct gru_message_queue_desc *mqd) | |
646 | { | |
647 | if (mqd->interrupt_vector) | |
648 | uv_hub_send_ipi(mqd->interrupt_pnode, mqd->interrupt_apicid, | |
649 | mqd->interrupt_vector); | |
650 | } | |
651 | ||
17b49a67 JS |
652 | /* |
653 | * Handle a PUT failure. Note: if message was a 2-line message, one of the | |
654 | * lines might have successfully have been written. Before sending the | |
655 | * message, "present" must be cleared in BOTH lines to prevent the receiver | |
656 | * from prematurely seeing the full message. | |
657 | */ | |
658 | static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd, | |
659 | void *mesg, int lines) | |
660 | { | |
661 | unsigned long m; | |
662 | ||
663 | m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6); | |
664 | if (lines == 2) { | |
665 | gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA); | |
666 | if (gru_wait(cb) != CBS_IDLE) | |
667 | return MQE_UNEXPECTED_CB_ERR; | |
668 | } | |
669 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA); | |
670 | if (gru_wait(cb) != CBS_IDLE) | |
671 | return MQE_UNEXPECTED_CB_ERR; | |
672 | send_message_queue_interrupt(mqd); | |
673 | return MQE_OK; | |
674 | } | |
28bffaf0 JS |
675 | |
676 | /* | |
677 | * Handle a gru_mesq failure. Some of these failures are software recoverable | |
678 | * or retryable. | |
679 | */ | |
6f2584f4 JS |
680 | static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd, |
681 | void *mesg, int lines) | |
28bffaf0 JS |
682 | { |
683 | int substatus, ret = 0; | |
28bffaf0 JS |
684 | |
685 | substatus = gru_get_cb_message_queue_substatus(cb); | |
686 | switch (substatus) { | |
687 | case CBSS_NO_ERROR: | |
688 | STAT(mesq_send_unexpected_error); | |
689 | ret = MQE_UNEXPECTED_CB_ERR; | |
690 | break; | |
691 | case CBSS_LB_OVERFLOWED: | |
692 | STAT(mesq_send_lb_overflow); | |
693 | ret = MQE_CONGESTION; | |
694 | break; | |
695 | case CBSS_QLIMIT_REACHED: | |
696 | STAT(mesq_send_qlimit_reached); | |
6f2584f4 | 697 | ret = send_message_queue_full(cb, mqd, mesg, lines); |
28bffaf0 JS |
698 | break; |
699 | case CBSS_AMO_NACKED: | |
700 | STAT(mesq_send_amo_nacked); | |
701 | ret = MQE_CONGESTION; | |
702 | break; | |
703 | case CBSS_PUT_NACKED: | |
704 | STAT(mesq_send_put_nacked); | |
17b49a67 | 705 | ret = send_message_put_nacked(cb, mqd, mesg, lines); |
28bffaf0 JS |
706 | break; |
707 | default: | |
708 | BUG(); | |
709 | } | |
710 | return ret; | |
711 | } | |
712 | ||
713 | /* | |
714 | * Send a message to a message queue | |
6f2584f4 | 715 | * mqd message queue descriptor |
28bffaf0 JS |
716 | * mesg message. ust be vaddr within a GSEG |
717 | * bytes message size (<= 2 CL) | |
718 | */ | |
6f2584f4 JS |
719 | int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg, |
720 | unsigned int bytes) | |
28bffaf0 JS |
721 | { |
722 | struct message_header *mhdr; | |
723 | void *cb; | |
724 | void *dsr; | |
725 | int istatus, clines, ret; | |
726 | ||
727 | STAT(mesq_send); | |
728 | BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES); | |
729 | ||
cbf330b9 | 730 | clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES); |
28bffaf0 JS |
731 | if (gru_get_cpu_resources(bytes, &cb, &dsr)) |
732 | return MQE_BUG_NO_RESOURCES; | |
733 | memcpy(dsr, mesg, bytes); | |
734 | mhdr = dsr; | |
735 | mhdr->present = MQS_FULL; | |
736 | mhdr->lines = clines; | |
737 | if (clines == 2) { | |
738 | mhdr->present2 = get_present2(mhdr); | |
739 | restore_present2(mhdr, MQS_FULL); | |
740 | } | |
741 | ||
742 | do { | |
743 | ret = MQE_OK; | |
6f2584f4 | 744 | gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA); |
28bffaf0 JS |
745 | istatus = gru_wait(cb); |
746 | if (istatus != CBS_IDLE) | |
6f2584f4 | 747 | ret = send_message_failure(cb, mqd, dsr, clines); |
28bffaf0 JS |
748 | } while (ret == MQIE_AGAIN); |
749 | gru_free_cpu_resources(cb, dsr); | |
750 | ||
751 | if (ret) | |
752 | STAT(mesq_send_failed); | |
753 | return ret; | |
754 | } | |
755 | EXPORT_SYMBOL_GPL(gru_send_message_gpa); | |
756 | ||
757 | /* | |
758 | * Advance the receive pointer for the queue to the next message. | |
759 | */ | |
6f2584f4 | 760 | void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg) |
28bffaf0 | 761 | { |
6f2584f4 | 762 | struct message_queue *mq = mqd->mq; |
28bffaf0 JS |
763 | struct message_header *mhdr = mq->next; |
764 | void *next, *pnext; | |
765 | int half = -1; | |
766 | int lines = mhdr->lines; | |
767 | ||
768 | if (lines == 2) | |
769 | restore_present2(mhdr, MQS_EMPTY); | |
770 | mhdr->present = MQS_EMPTY; | |
771 | ||
772 | pnext = mq->next; | |
773 | next = pnext + GRU_CACHE_LINE_BYTES * lines; | |
774 | if (next == mq->limit) { | |
775 | next = mq->start; | |
776 | half = 1; | |
777 | } else if (pnext < mq->start2 && next >= mq->start2) { | |
778 | half = 0; | |
779 | } | |
780 | ||
781 | if (half >= 0) | |
782 | mq->hstatus[half] = 1; | |
783 | mq->next = next; | |
784 | } | |
785 | EXPORT_SYMBOL_GPL(gru_free_message); | |
786 | ||
787 | /* | |
788 | * Get next message from message queue. Return NULL if no message | |
789 | * present. User must call next_message() to move to next message. | |
790 | * rmq message queue | |
791 | */ | |
6f2584f4 | 792 | void *gru_get_next_message(struct gru_message_queue_desc *mqd) |
28bffaf0 | 793 | { |
6f2584f4 | 794 | struct message_queue *mq = mqd->mq; |
28bffaf0 JS |
795 | struct message_header *mhdr = mq->next; |
796 | int present = mhdr->present; | |
797 | ||
798 | /* skip NOOP messages */ | |
799 | STAT(mesq_receive); | |
800 | while (present == MQS_NOOP) { | |
6f2584f4 | 801 | gru_free_message(mqd, mhdr); |
28bffaf0 JS |
802 | mhdr = mq->next; |
803 | present = mhdr->present; | |
804 | } | |
805 | ||
806 | /* Wait for both halves of 2 line messages */ | |
807 | if (present == MQS_FULL && mhdr->lines == 2 && | |
808 | get_present2(mhdr) == MQS_EMPTY) | |
809 | present = MQS_EMPTY; | |
810 | ||
811 | if (!present) { | |
812 | STAT(mesq_receive_none); | |
813 | return NULL; | |
814 | } | |
815 | ||
816 | if (mhdr->lines == 2) | |
817 | restore_present2(mhdr, mhdr->present2); | |
818 | ||
819 | return mhdr; | |
820 | } | |
821 | EXPORT_SYMBOL_GPL(gru_get_next_message); | |
822 | ||
823 | /* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/ | |
824 | ||
825 | /* | |
826 | * Copy a block of data using the GRU resources | |
827 | */ | |
828 | int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa, | |
829 | unsigned int bytes) | |
830 | { | |
831 | void *cb; | |
832 | void *dsr; | |
833 | int ret; | |
834 | ||
835 | STAT(copy_gpa); | |
836 | if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr)) | |
837 | return MQE_BUG_NO_RESOURCES; | |
838 | gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr), | |
6f2584f4 | 839 | XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA); |
28bffaf0 JS |
840 | ret = gru_wait(cb); |
841 | gru_free_cpu_resources(cb, dsr); | |
842 | return ret; | |
843 | } | |
844 | EXPORT_SYMBOL_GPL(gru_copy_gpa); | |
845 | ||
846 | /* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/ | |
847 | /* Temp - will delete after we gain confidence in the GRU */ | |
28bffaf0 | 848 | |
eb5bd5e5 | 849 | static int quicktest0(unsigned long arg) |
28bffaf0 | 850 | { |
836ce679 JS |
851 | unsigned long word0; |
852 | unsigned long word1; | |
28bffaf0 | 853 | void *cb; |
836ce679 | 854 | void *dsr; |
28bffaf0 | 855 | unsigned long *p; |
eb5bd5e5 | 856 | int ret = -EIO; |
28bffaf0 | 857 | |
836ce679 JS |
858 | if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr)) |
859 | return MQE_BUG_NO_RESOURCES; | |
860 | p = dsr; | |
28bffaf0 | 861 | word0 = MAGIC; |
836ce679 | 862 | word1 = 0; |
28bffaf0 | 863 | |
836ce679 | 864 | gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA); |
eb5bd5e5 JS |
865 | if (gru_wait(cb) != CBS_IDLE) { |
866 | printk(KERN_DEBUG "GRU quicktest0: CBR failure 1\n"); | |
867 | goto done; | |
868 | } | |
28bffaf0 | 869 | |
eb5bd5e5 JS |
870 | if (*p != MAGIC) { |
871 | printk(KERN_DEBUG "GRU: quicktest0 bad magic 0x%lx\n", *p); | |
872 | goto done; | |
873 | } | |
836ce679 | 874 | gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA); |
eb5bd5e5 JS |
875 | if (gru_wait(cb) != CBS_IDLE) { |
876 | printk(KERN_DEBUG "GRU quicktest0: CBR failure 2\n"); | |
877 | goto done; | |
878 | } | |
28bffaf0 | 879 | |
836ce679 | 880 | if (word0 != word1 || word1 != MAGIC) { |
eb5bd5e5 JS |
881 | printk(KERN_DEBUG |
882 | "GRU quicktest0 err: found 0x%lx, expected 0x%lx\n", | |
836ce679 | 883 | word1, MAGIC); |
eb5bd5e5 | 884 | goto done; |
28bffaf0 | 885 | } |
eb5bd5e5 | 886 | ret = 0; |
28bffaf0 | 887 | |
eb5bd5e5 JS |
888 | done: |
889 | gru_free_cpu_resources(cb, dsr); | |
890 | return ret; | |
28bffaf0 JS |
891 | } |
892 | ||
eb5bd5e5 JS |
893 | #define ALIGNUP(p, q) ((void *)(((unsigned long)(p) + (q) - 1) & ~(q - 1))) |
894 | ||
895 | static int quicktest1(unsigned long arg) | |
896 | { | |
897 | struct gru_message_queue_desc mqd; | |
898 | void *p, *mq; | |
899 | unsigned long *dw; | |
900 | int i, ret = -EIO; | |
901 | char mes[GRU_CACHE_LINE_BYTES], *m; | |
902 | ||
903 | /* Need 1K cacheline aligned that does not cross page boundary */ | |
904 | p = kmalloc(4096, 0); | |
905 | mq = ALIGNUP(p, 1024); | |
906 | memset(mes, 0xee, sizeof(mes)); | |
907 | dw = mq; | |
908 | ||
909 | gru_create_message_queue(&mqd, mq, 8 * GRU_CACHE_LINE_BYTES, 0, 0, 0); | |
910 | for (i = 0; i < 6; i++) { | |
911 | mes[8] = i; | |
912 | do { | |
913 | ret = gru_send_message_gpa(&mqd, mes, sizeof(mes)); | |
914 | } while (ret == MQE_CONGESTION); | |
915 | if (ret) | |
916 | break; | |
917 | } | |
918 | if (ret != MQE_QUEUE_FULL || i != 4) | |
919 | goto done; | |
920 | ||
921 | for (i = 0; i < 6; i++) { | |
922 | m = gru_get_next_message(&mqd); | |
923 | if (!m || m[8] != i) | |
924 | break; | |
925 | gru_free_message(&mqd, m); | |
926 | } | |
927 | ret = (i == 4) ? 0 : -EIO; | |
928 | ||
929 | done: | |
930 | kfree(p); | |
931 | return ret; | |
932 | } | |
933 | ||
934 | static int quicktest2(unsigned long arg) | |
935 | { | |
936 | static DECLARE_COMPLETION(cmp); | |
937 | unsigned long han; | |
938 | int blade_id = 0; | |
939 | int numcb = 4; | |
940 | int ret = 0; | |
941 | unsigned long *buf; | |
942 | void *cb0, *cb; | |
943 | int i, k, istatus, bytes; | |
944 | ||
945 | bytes = numcb * 4 * 8; | |
946 | buf = kmalloc(bytes, GFP_KERNEL); | |
947 | if (!buf) | |
948 | return -ENOMEM; | |
949 | ||
950 | ret = -EBUSY; | |
951 | han = gru_reserve_async_resources(blade_id, numcb, 0, &cmp); | |
952 | if (!han) | |
953 | goto done; | |
954 | ||
955 | gru_lock_async_resource(han, &cb0, NULL); | |
956 | memset(buf, 0xee, bytes); | |
957 | for (i = 0; i < numcb; i++) | |
958 | gru_vset(cb0 + i * GRU_HANDLE_STRIDE, uv_gpa(&buf[i * 4]), 0, | |
959 | XTYPE_DW, 4, 1, IMA_INTERRUPT); | |
960 | ||
961 | ret = 0; | |
962 | for (k = 0; k < numcb; k++) { | |
963 | gru_wait_async_cbr(han); | |
964 | for (i = 0; i < numcb; i++) { | |
965 | cb = cb0 + i * GRU_HANDLE_STRIDE; | |
966 | istatus = gru_check_status(cb); | |
967 | if (istatus == CBS_ACTIVE) | |
968 | continue; | |
969 | if (istatus == CBS_EXCEPTION) | |
970 | ret = -EFAULT; | |
971 | else if (buf[i] || buf[i + 1] || buf[i + 2] || | |
972 | buf[i + 3]) | |
973 | ret = -EIO; | |
974 | } | |
975 | } | |
976 | BUG_ON(cmp.done); | |
977 | ||
978 | gru_unlock_async_resource(han); | |
979 | gru_release_async_resources(han); | |
980 | done: | |
981 | kfree(buf); | |
982 | return ret; | |
983 | } | |
984 | ||
985 | /* | |
986 | * Debugging only. User hook for various kernel tests | |
987 | * of driver & gru. | |
988 | */ | |
989 | int gru_ktest(unsigned long arg) | |
990 | { | |
991 | int ret = -EINVAL; | |
992 | ||
993 | switch (arg & 0xff) { | |
994 | case 0: | |
995 | ret = quicktest0(arg); | |
996 | break; | |
997 | case 1: | |
998 | ret = quicktest1(arg); | |
999 | break; | |
1000 | case 2: | |
1001 | ret = quicktest2(arg); | |
1002 | break; | |
1003 | } | |
1004 | return ret; | |
1005 | ||
1006 | } | |
28bffaf0 JS |
1007 | |
1008 | int gru_kservices_init(struct gru_state *gru) | |
1009 | { | |
1010 | struct gru_blade_state *bs; | |
836ce679 | 1011 | |
28bffaf0 | 1012 | bs = gru->gs_blade; |
836ce679 | 1013 | if (gru != &bs->bs_grus[0]) |
28bffaf0 JS |
1014 | return 0; |
1015 | ||
836ce679 | 1016 | init_rwsem(&bs->bs_kgts_sema); |
28bffaf0 JS |
1017 | return 0; |
1018 | } | |
27ca8a7b JS |
1019 | |
1020 | void gru_kservices_exit(struct gru_state *gru) | |
1021 | { | |
27ca8a7b | 1022 | struct gru_blade_state *bs; |
836ce679 | 1023 | struct gru_thread_state *kgts; |
27ca8a7b JS |
1024 | |
1025 | bs = gru->gs_blade; | |
836ce679 | 1026 | if (gru != &bs->bs_grus[0]) |
27ca8a7b JS |
1027 | return; |
1028 | ||
836ce679 JS |
1029 | kgts = bs->bs_kgts; |
1030 | if (kgts && kgts->ts_gru) | |
1031 | gru_unload_context(kgts, 0); | |
1032 | kfree(kgts); | |
27ca8a7b JS |
1033 | } |
1034 |