blob: 766e21e15574c75ceba18828a4d492ac6a90a189 [file] [log] [blame]
Jack Steiner28bffaf2008-07-29 22:33:57 -07001/*
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>
Jack Steiner28bffaf2008-07-29 22:33:57 -070027#include <linux/spinlock.h>
28#include <linux/device.h>
29#include <linux/miscdevice.h>
30#include <linux/proc_fs.h>
31#include <linux/interrupt.h>
32#include <linux/uaccess.h>
Jack Steiner836ce672009-06-17 16:28:22 -070033#include <linux/delay.h>
Jack Steiner28bffaf2008-07-29 22:33:57 -070034#include "gru.h"
35#include "grulib.h"
36#include "grutables.h"
37#include "grukservices.h"
38#include "gru_instructions.h"
39#include <asm/uv/uv_hub.h>
40
41/*
42 * Kernel GRU Usage
43 *
44 * The following is an interim algorithm for management of kernel GRU
45 * resources. This will likely be replaced when we better understand the
46 * kernel/user requirements.
47 *
Jack Steiner836ce672009-06-17 16:28:22 -070048 * Blade percpu resources reserved for kernel use. These resources are
49 * reserved whenever the the kernel context for the blade is loaded. Note
50 * that the kernel context is not guaranteed to be always available. It is
51 * loaded on demand & can be stolen by a user if the user demand exceeds the
52 * kernel demand. The kernel can always reload the kernel context but
53 * a SLEEP may be required!!!.
Jack Steiner9120dec2009-06-17 16:28:25 -070054 *
55 * Async Overview:
56 *
57 * Each blade has one "kernel context" that owns GRU kernel resources
58 * located on the blade. Kernel drivers use GRU resources in this context
59 * for sending messages, zeroing memory, etc.
60 *
61 * The kernel context is dynamically loaded on demand. If it is not in
62 * use by the kernel, the kernel context can be unloaded & given to a user.
63 * The kernel context will be reloaded when needed. This may require that
64 * a context be stolen from a user.
65 * NOTE: frequent unloading/reloading of the kernel context is
66 * expensive. We are depending on batch schedulers, cpusets, sane
67 * drivers or some other mechanism to prevent the need for frequent
68 * stealing/reloading.
69 *
70 * The kernel context consists of two parts:
71 * - 1 CB & a few DSRs that are reserved for each cpu on the blade.
72 * Each cpu has it's own private resources & does not share them
73 * with other cpus. These resources are used serially, ie,
74 * locked, used & unlocked on each call to a function in
75 * grukservices.
76 * (Now that we have dynamic loading of kernel contexts, I
77 * may rethink this & allow sharing between cpus....)
78 *
79 * - Additional resources can be reserved long term & used directly
80 * by UV drivers located in the kernel. Drivers using these GRU
81 * resources can use asynchronous GRU instructions that send
82 * interrupts on completion.
83 * - these resources must be explicitly locked/unlocked
84 * - locked resources prevent (obviously) the kernel
85 * context from being unloaded.
86 * - drivers using these resource directly issue their own
87 * GRU instruction and must wait/check completion.
88 *
89 * When these resources are reserved, the caller can optionally
90 * associate a wait_queue with the resources and use asynchronous
91 * GRU instructions. When an async GRU instruction completes, the
92 * driver will do a wakeup on the event.
93 *
Jack Steiner28bffaf2008-07-29 22:33:57 -070094 */
Jack Steiner9120dec2009-06-17 16:28:25 -070095
96
97#define ASYNC_HAN_TO_BID(h) ((h) - 1)
98#define ASYNC_BID_TO_HAN(b) ((b) + 1)
99#define ASYNC_HAN_TO_BS(h) gru_base[ASYNC_HAN_TO_BID(h)]
Jack Steiner1a2c09e2009-06-17 16:28:28 -0700100#define KCB_TO_GID(cb) ((cb - gru_start_vaddr) / \
101 (GRU_SIZE * GRU_CHIPLETS_PER_BLADE))
102#define KCB_TO_BS(cb) gru_base[KCB_TO_GID(cb)]
Jack Steiner9120dec2009-06-17 16:28:25 -0700103
Jack Steiner6f2584f2009-04-02 16:59:10 -0700104#define GRU_NUM_KERNEL_CBR 1
Jack Steiner28bffaf2008-07-29 22:33:57 -0700105#define GRU_NUM_KERNEL_DSR_BYTES 256
Jack Steiner6f2584f2009-04-02 16:59:10 -0700106#define GRU_NUM_KERNEL_DSR_CL (GRU_NUM_KERNEL_DSR_BYTES / \
107 GRU_CACHE_LINE_BYTES)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700108
109/* GRU instruction attributes for all instructions */
110#define IMA IMA_CB_DELAY
111
112/* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */
113#define __gru_cacheline_aligned__ \
114 __attribute__((__aligned__(GRU_CACHE_LINE_BYTES)))
115
116#define MAGIC 0x1234567887654321UL
117
118/* Default retry count for GRU errors on kernel instructions */
119#define EXCEPTION_RETRY_LIMIT 3
120
121/* Status of message queue sections */
122#define MQS_EMPTY 0
123#define MQS_FULL 1
124#define MQS_NOOP 2
125
126/*----------------- RESOURCE MANAGEMENT -------------------------------------*/
127/* optimized for x86_64 */
128struct message_queue {
129 union gru_mesqhead head __gru_cacheline_aligned__; /* CL 0 */
130 int qlines; /* DW 1 */
131 long hstatus[2];
132 void *next __gru_cacheline_aligned__;/* CL 1 */
133 void *limit;
134 void *start;
135 void *start2;
136 char data ____cacheline_aligned; /* CL 2 */
137};
138
139/* First word in every message - used by mesq interface */
140struct message_header {
141 char present;
142 char present2;
143 char lines;
144 char fill;
145};
146
Jack Steiner28bffaf2008-07-29 22:33:57 -0700147#define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h]))
148
Jack Steiner836ce672009-06-17 16:28:22 -0700149/*
Jack Steiner836ce672009-06-17 16:28:22 -0700150 * Reload the blade's kernel context into a GRU chiplet. Called holding
151 * the bs_kgts_sema for READ. Will steal user contexts if necessary.
152 */
153static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id)
154{
155 struct gru_state *gru;
156 struct gru_thread_state *kgts;
157 void *vaddr;
Jack Steiner9120dec2009-06-17 16:28:25 -0700158 int ctxnum, ncpus;
Jack Steiner836ce672009-06-17 16:28:22 -0700159
160 up_read(&bs->bs_kgts_sema);
161 down_write(&bs->bs_kgts_sema);
162
163 if (!bs->bs_kgts)
Jack Steiner9120dec2009-06-17 16:28:25 -0700164 bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0);
Jack Steiner836ce672009-06-17 16:28:22 -0700165 kgts = bs->bs_kgts;
166
167 if (!kgts->ts_gru) {
168 STAT(load_kernel_context);
Jack Steiner9120dec2009-06-17 16:28:25 -0700169 ncpus = uv_blade_nr_possible_cpus(blade_id);
170 kgts->ts_cbr_au_count = GRU_CB_COUNT_TO_AU(
171 GRU_NUM_KERNEL_CBR * ncpus + bs->bs_async_cbrs);
172 kgts->ts_dsr_au_count = GRU_DS_BYTES_TO_AU(
173 GRU_NUM_KERNEL_DSR_BYTES * ncpus +
174 bs->bs_async_dsr_bytes);
Jack Steiner836ce672009-06-17 16:28:22 -0700175 while (!gru_assign_gru_context(kgts, blade_id)) {
176 msleep(1);
177 gru_steal_context(kgts, blade_id);
178 }
179 gru_load_context(kgts);
180 gru = bs->bs_kgts->ts_gru;
181 vaddr = gru->gs_gru_base_vaddr;
182 ctxnum = kgts->ts_ctxnum;
183 bs->kernel_cb = get_gseg_base_address_cb(vaddr, ctxnum, 0);
184 bs->kernel_dsr = get_gseg_base_address_ds(vaddr, ctxnum, 0);
185 }
186 downgrade_write(&bs->bs_kgts_sema);
187}
188
189/*
Jack Steinerd5826dd2009-06-17 16:28:28 -0700190 * Free all kernel contexts that are not currently in use.
191 * Returns 0 if all freed, else number of inuse context.
192 */
193static int gru_free_kernel_contexts(void)
194{
195 struct gru_blade_state *bs;
196 struct gru_thread_state *kgts;
197 int bid, ret = 0;
198
199 for (bid = 0; bid < GRU_MAX_BLADES; bid++) {
200 bs = gru_base[bid];
201 if (!bs)
202 continue;
203 if (down_write_trylock(&bs->bs_kgts_sema)) {
204 kgts = bs->bs_kgts;
205 if (kgts && kgts->ts_gru)
206 gru_unload_context(kgts, 0);
207 kfree(kgts);
208 bs->bs_kgts = NULL;
209 up_write(&bs->bs_kgts_sema);
210 } else {
211 ret++;
212 }
213 }
214 return ret;
215}
216
217/*
Jack Steiner836ce672009-06-17 16:28:22 -0700218 * Lock & load the kernel context for the specified blade.
219 */
220static struct gru_blade_state *gru_lock_kernel_context(int blade_id)
221{
222 struct gru_blade_state *bs;
223
224 STAT(lock_kernel_context);
225 bs = gru_base[blade_id];
226
227 down_read(&bs->bs_kgts_sema);
228 if (!bs->bs_kgts || !bs->bs_kgts->ts_gru)
229 gru_load_kernel_context(bs, blade_id);
230 return bs;
231
232}
233
234/*
235 * Unlock the kernel context for the specified blade. Context is not
236 * unloaded but may be stolen before next use.
237 */
238static void gru_unlock_kernel_context(int blade_id)
239{
240 struct gru_blade_state *bs;
241
242 bs = gru_base[blade_id];
243 up_read(&bs->bs_kgts_sema);
244 STAT(unlock_kernel_context);
245}
246
247/*
248 * Reserve & get pointers to the DSR/CBRs reserved for the current cpu.
249 * - returns with preemption disabled
250 */
Jack Steiner28bffaf2008-07-29 22:33:57 -0700251static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr)
252{
253 struct gru_blade_state *bs;
254 int lcpu;
255
256 BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES);
257 preempt_disable();
Jack Steiner836ce672009-06-17 16:28:22 -0700258 bs = gru_lock_kernel_context(uv_numa_blade_id());
Jack Steiner28bffaf2008-07-29 22:33:57 -0700259 lcpu = uv_blade_processor_id();
260 *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE;
261 *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES;
262 return 0;
263}
264
Jack Steiner836ce672009-06-17 16:28:22 -0700265/*
266 * Free the current cpus reserved DSR/CBR resources.
267 */
Jack Steiner28bffaf2008-07-29 22:33:57 -0700268static void gru_free_cpu_resources(void *cb, void *dsr)
269{
Jack Steiner836ce672009-06-17 16:28:22 -0700270 gru_unlock_kernel_context(uv_numa_blade_id());
Jack Steiner28bffaf2008-07-29 22:33:57 -0700271 preempt_enable();
272}
273
Jack Steiner9120dec2009-06-17 16:28:25 -0700274/*
275 * Reserve GRU resources to be used asynchronously.
276 * Note: currently supports only 1 reservation per blade.
277 *
278 * input:
279 * blade_id - blade on which resources should be reserved
280 * cbrs - number of CBRs
281 * dsr_bytes - number of DSR bytes needed
282 * output:
283 * handle to identify resource
284 * (0 = async resources already reserved)
285 */
286unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
287 struct completion *cmp)
288{
289 struct gru_blade_state *bs;
290 struct gru_thread_state *kgts;
291 int ret = 0;
292
293 bs = gru_base[blade_id];
294
295 down_write(&bs->bs_kgts_sema);
296
297 /* Verify no resources already reserved */
298 if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs)
299 goto done;
300 bs->bs_async_dsr_bytes = dsr_bytes;
301 bs->bs_async_cbrs = cbrs;
302 bs->bs_async_wq = cmp;
303 kgts = bs->bs_kgts;
304
305 /* Resources changed. Unload context if already loaded */
306 if (kgts && kgts->ts_gru)
307 gru_unload_context(kgts, 0);
308 ret = ASYNC_BID_TO_HAN(blade_id);
309
310done:
311 up_write(&bs->bs_kgts_sema);
312 return ret;
313}
314
315/*
316 * Release async resources previously reserved.
317 *
318 * input:
319 * han - handle to identify resources
320 */
321void gru_release_async_resources(unsigned long han)
322{
323 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
324
325 down_write(&bs->bs_kgts_sema);
326 bs->bs_async_dsr_bytes = 0;
327 bs->bs_async_cbrs = 0;
328 bs->bs_async_wq = NULL;
329 up_write(&bs->bs_kgts_sema);
330}
331
332/*
333 * Wait for async GRU instructions to complete.
334 *
335 * input:
336 * han - handle to identify resources
337 */
338void gru_wait_async_cbr(unsigned long han)
339{
340 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
341
342 wait_for_completion(bs->bs_async_wq);
343 mb();
344}
345
346/*
347 * Lock previous reserved async GRU resources
348 *
349 * input:
350 * han - handle to identify resources
351 * output:
352 * cb - pointer to first CBR
353 * dsr - pointer to first DSR
354 */
355void gru_lock_async_resource(unsigned long han, void **cb, void **dsr)
356{
357 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
358 int blade_id = ASYNC_HAN_TO_BID(han);
359 int ncpus;
360
361 gru_lock_kernel_context(blade_id);
362 ncpus = uv_blade_nr_possible_cpus(blade_id);
363 if (cb)
364 *cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE;
365 if (dsr)
366 *dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES;
367}
368
369/*
370 * Unlock previous reserved async GRU resources
371 *
372 * input:
373 * han - handle to identify resources
374 */
375void gru_unlock_async_resource(unsigned long han)
376{
377 int blade_id = ASYNC_HAN_TO_BID(han);
378
379 gru_unlock_kernel_context(blade_id);
380}
381
Jack Steiner836ce672009-06-17 16:28:22 -0700382/*----------------------------------------------------------------------*/
Jack Steiner28bffaf2008-07-29 22:33:57 -0700383int gru_get_cb_exception_detail(void *cb,
384 struct control_block_extended_exc_detail *excdet)
385{
386 struct gru_control_block_extended *cbe;
Jack Steiner1a2c09e2009-06-17 16:28:28 -0700387 struct gru_blade_state *bs;
388 int cbrnum;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700389
Jack Steiner1a2c09e2009-06-17 16:28:28 -0700390 bs = KCB_TO_BS(cb);
391 cbrnum = thread_cbr_number(bs->bs_kgts, get_cb_number(cb));
392 cbe = get_cbe(GRUBASE(cb), cbrnum);
393 gru_flush_cache(cbe); /* CBE not coherent */
Jack Steiner28bffaf2008-07-29 22:33:57 -0700394 excdet->opc = cbe->opccpy;
395 excdet->exopc = cbe->exopccpy;
396 excdet->ecause = cbe->ecause;
397 excdet->exceptdet0 = cbe->idef1upd;
398 excdet->exceptdet1 = cbe->idef3upd;
Jack Steiner1a2c09e2009-06-17 16:28:28 -0700399 gru_flush_cache(cbe);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700400 return 0;
401}
402
403char *gru_get_cb_exception_detail_str(int ret, void *cb,
404 char *buf, int size)
405{
406 struct gru_control_block_status *gen = (void *)cb;
407 struct control_block_extended_exc_detail excdet;
408
409 if (ret > 0 && gen->istatus == CBS_EXCEPTION) {
410 gru_get_cb_exception_detail(cb, &excdet);
411 snprintf(buf, size,
412 "GRU exception: cb %p, opc %d, exopc %d, ecause 0x%x,"
413 "excdet0 0x%lx, excdet1 0x%x",
414 gen, excdet.opc, excdet.exopc, excdet.ecause,
415 excdet.exceptdet0, excdet.exceptdet1);
416 } else {
417 snprintf(buf, size, "No exception");
418 }
419 return buf;
420}
421
422static int gru_wait_idle_or_exception(struct gru_control_block_status *gen)
423{
424 while (gen->istatus >= CBS_ACTIVE) {
425 cpu_relax();
426 barrier();
427 }
428 return gen->istatus;
429}
430
431static int gru_retry_exception(void *cb)
432{
433 struct gru_control_block_status *gen = (void *)cb;
434 struct control_block_extended_exc_detail excdet;
435 int retry = EXCEPTION_RETRY_LIMIT;
436
437 while (1) {
Jack Steiner28bffaf2008-07-29 22:33:57 -0700438 if (gru_wait_idle_or_exception(gen) == CBS_IDLE)
439 return CBS_IDLE;
Jack Steinerd6e2fbc2009-06-17 16:28:29 -0700440 if (gru_get_cb_message_queue_substatus(cb))
441 return CBS_EXCEPTION;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700442 gru_get_cb_exception_detail(cb, &excdet);
Jack Steiner270952a2009-06-17 16:28:27 -0700443 if ((excdet.ecause & ~EXCEPTION_RETRY_BITS) ||
444 (excdet.cbrexecstatus & CBR_EXS_ABORT_OCC))
Jack Steiner28bffaf2008-07-29 22:33:57 -0700445 break;
446 if (retry-- == 0)
447 break;
448 gen->icmd = 1;
449 gru_flush_cache(gen);
450 }
451 return CBS_EXCEPTION;
452}
453
454int gru_check_status_proc(void *cb)
455{
456 struct gru_control_block_status *gen = (void *)cb;
457 int ret;
458
459 ret = gen->istatus;
460 if (ret != CBS_EXCEPTION)
461 return ret;
462 return gru_retry_exception(cb);
463
464}
465
466int gru_wait_proc(void *cb)
467{
468 struct gru_control_block_status *gen = (void *)cb;
469 int ret;
470
471 ret = gru_wait_idle_or_exception(gen);
472 if (ret == CBS_EXCEPTION)
473 ret = gru_retry_exception(cb);
474
475 return ret;
476}
477
478void gru_abort(int ret, void *cb, char *str)
479{
480 char buf[GRU_EXC_STR_SIZE];
481
482 panic("GRU FATAL ERROR: %s - %s\n", str,
483 gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf)));
484}
485
486void gru_wait_abort_proc(void *cb)
487{
488 int ret;
489
490 ret = gru_wait_proc(cb);
491 if (ret)
492 gru_abort(ret, cb, "gru_wait_abort");
493}
494
495
496/*------------------------------ MESSAGE QUEUES -----------------------------*/
497
498/* Internal status . These are NOT returned to the user. */
499#define MQIE_AGAIN -1 /* try again */
500
501
502/*
503 * Save/restore the "present" flag that is in the second line of 2-line
504 * messages
505 */
506static inline int get_present2(void *p)
507{
508 struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
509 return mhdr->present;
510}
511
512static inline void restore_present2(void *p, int val)
513{
514 struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
515 mhdr->present = val;
516}
517
518/*
519 * Create a message queue.
520 * qlines - message queue size in cache lines. Includes 2-line header.
521 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700522int gru_create_message_queue(struct gru_message_queue_desc *mqd,
523 void *p, unsigned int bytes, int nasid, int vector, int apicid)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700524{
525 struct message_queue *mq = p;
526 unsigned int qlines;
527
528 qlines = bytes / GRU_CACHE_LINE_BYTES - 2;
529 memset(mq, 0, bytes);
530 mq->start = &mq->data;
531 mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES;
532 mq->next = &mq->data;
533 mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES;
534 mq->qlines = qlines;
535 mq->hstatus[0] = 0;
536 mq->hstatus[1] = 1;
537 mq->head = gru_mesq_head(2, qlines / 2 + 1);
Jack Steiner6f2584f2009-04-02 16:59:10 -0700538 mqd->mq = mq;
539 mqd->mq_gpa = uv_gpa(mq);
540 mqd->qlines = qlines;
541 mqd->interrupt_pnode = UV_NASID_TO_PNODE(nasid);
542 mqd->interrupt_vector = vector;
543 mqd->interrupt_apicid = apicid;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700544 return 0;
545}
546EXPORT_SYMBOL_GPL(gru_create_message_queue);
547
548/*
549 * Send a NOOP message to a message queue
550 * Returns:
551 * 0 - if queue is full after the send. This is the normal case
552 * but various races can change this.
553 * -1 - if mesq sent successfully but queue not full
554 * >0 - unexpected error. MQE_xxx returned
555 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700556static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd,
557 void *mesg)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700558{
559 const struct message_header noop_header = {
560 .present = MQS_NOOP, .lines = 1};
561 unsigned long m;
562 int substatus, ret;
563 struct message_header save_mhdr, *mhdr = mesg;
564
565 STAT(mesq_noop);
566 save_mhdr = *mhdr;
567 *mhdr = noop_header;
Jack Steiner6f2584f2009-04-02 16:59:10 -0700568 gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700569 ret = gru_wait(cb);
570
571 if (ret) {
572 substatus = gru_get_cb_message_queue_substatus(cb);
573 switch (substatus) {
574 case CBSS_NO_ERROR:
575 STAT(mesq_noop_unexpected_error);
576 ret = MQE_UNEXPECTED_CB_ERR;
577 break;
578 case CBSS_LB_OVERFLOWED:
579 STAT(mesq_noop_lb_overflow);
580 ret = MQE_CONGESTION;
581 break;
582 case CBSS_QLIMIT_REACHED:
583 STAT(mesq_noop_qlimit_reached);
584 ret = 0;
585 break;
586 case CBSS_AMO_NACKED:
587 STAT(mesq_noop_amo_nacked);
588 ret = MQE_CONGESTION;
589 break;
590 case CBSS_PUT_NACKED:
591 STAT(mesq_noop_put_nacked);
Jack Steiner6f2584f2009-04-02 16:59:10 -0700592 m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700593 gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1,
594 IMA);
595 if (gru_wait(cb) == CBS_IDLE)
596 ret = MQIE_AGAIN;
597 else
598 ret = MQE_UNEXPECTED_CB_ERR;
599 break;
600 case CBSS_PAGE_OVERFLOW:
601 default:
602 BUG();
603 }
604 }
605 *mhdr = save_mhdr;
606 return ret;
607}
608
609/*
610 * Handle a gru_mesq full.
611 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700612static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd,
613 void *mesg, int lines)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700614{
615 union gru_mesqhead mqh;
616 unsigned int limit, head;
617 unsigned long avalue;
Jack Steiner6f2584f2009-04-02 16:59:10 -0700618 int half, qlines;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700619
620 /* Determine if switching to first/second half of q */
621 avalue = gru_get_amo_value(cb);
622 head = gru_get_amo_value_head(cb);
623 limit = gru_get_amo_value_limit(cb);
624
Jack Steiner6f2584f2009-04-02 16:59:10 -0700625 qlines = mqd->qlines;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700626 half = (limit != qlines);
627
628 if (half)
629 mqh = gru_mesq_head(qlines / 2 + 1, qlines);
630 else
631 mqh = gru_mesq_head(2, qlines / 2 + 1);
632
633 /* Try to get lock for switching head pointer */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700634 gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700635 if (gru_wait(cb) != CBS_IDLE)
636 goto cberr;
637 if (!gru_get_amo_value(cb)) {
638 STAT(mesq_qf_locked);
639 return MQE_QUEUE_FULL;
640 }
641
642 /* Got the lock. Send optional NOP if queue not full, */
643 if (head != limit) {
Jack Steiner6f2584f2009-04-02 16:59:10 -0700644 if (send_noop_message(cb, mqd, mesg)) {
645 gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half),
Jack Steiner28bffaf2008-07-29 22:33:57 -0700646 XTYPE_DW, IMA);
647 if (gru_wait(cb) != CBS_IDLE)
648 goto cberr;
649 STAT(mesq_qf_noop_not_full);
650 return MQIE_AGAIN;
651 }
652 avalue++;
653 }
654
655 /* Then flip queuehead to other half of queue. */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700656 gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue,
657 IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700658 if (gru_wait(cb) != CBS_IDLE)
659 goto cberr;
660
661 /* If not successfully in swapping queue head, clear the hstatus lock */
662 if (gru_get_amo_value(cb) != avalue) {
663 STAT(mesq_qf_switch_head_failed);
Jack Steiner6f2584f2009-04-02 16:59:10 -0700664 gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW,
665 IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700666 if (gru_wait(cb) != CBS_IDLE)
667 goto cberr;
668 }
669 return MQIE_AGAIN;
670cberr:
671 STAT(mesq_qf_unexpected_error);
672 return MQE_UNEXPECTED_CB_ERR;
673}
674
Jack Steiner6f2584f2009-04-02 16:59:10 -0700675/*
676 * Send a cross-partition interrupt to the SSI that contains the target
677 * message queue. Normally, the interrupt is automatically delivered by hardware
678 * but some error conditions require explicit delivery.
679 */
680static void send_message_queue_interrupt(struct gru_message_queue_desc *mqd)
681{
682 if (mqd->interrupt_vector)
683 uv_hub_send_ipi(mqd->interrupt_pnode, mqd->interrupt_apicid,
684 mqd->interrupt_vector);
685}
686
Jack Steiner17b49a62009-06-17 16:28:23 -0700687/*
688 * Handle a PUT failure. Note: if message was a 2-line message, one of the
689 * lines might have successfully have been written. Before sending the
690 * message, "present" must be cleared in BOTH lines to prevent the receiver
691 * from prematurely seeing the full message.
692 */
693static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd,
694 void *mesg, int lines)
695{
696 unsigned long m;
697
698 m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
699 if (lines == 2) {
700 gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA);
701 if (gru_wait(cb) != CBS_IDLE)
702 return MQE_UNEXPECTED_CB_ERR;
703 }
704 gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA);
705 if (gru_wait(cb) != CBS_IDLE)
706 return MQE_UNEXPECTED_CB_ERR;
707 send_message_queue_interrupt(mqd);
708 return MQE_OK;
709}
Jack Steiner28bffaf2008-07-29 22:33:57 -0700710
711/*
712 * Handle a gru_mesq failure. Some of these failures are software recoverable
713 * or retryable.
714 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700715static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd,
716 void *mesg, int lines)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700717{
718 int substatus, ret = 0;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700719
720 substatus = gru_get_cb_message_queue_substatus(cb);
721 switch (substatus) {
722 case CBSS_NO_ERROR:
723 STAT(mesq_send_unexpected_error);
724 ret = MQE_UNEXPECTED_CB_ERR;
725 break;
726 case CBSS_LB_OVERFLOWED:
727 STAT(mesq_send_lb_overflow);
728 ret = MQE_CONGESTION;
729 break;
730 case CBSS_QLIMIT_REACHED:
731 STAT(mesq_send_qlimit_reached);
Jack Steiner6f2584f2009-04-02 16:59:10 -0700732 ret = send_message_queue_full(cb, mqd, mesg, lines);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700733 break;
734 case CBSS_AMO_NACKED:
735 STAT(mesq_send_amo_nacked);
736 ret = MQE_CONGESTION;
737 break;
738 case CBSS_PUT_NACKED:
739 STAT(mesq_send_put_nacked);
Jack Steiner17b49a62009-06-17 16:28:23 -0700740 ret = send_message_put_nacked(cb, mqd, mesg, lines);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700741 break;
742 default:
743 BUG();
744 }
745 return ret;
746}
747
748/*
749 * Send a message to a message queue
Jack Steiner6f2584f2009-04-02 16:59:10 -0700750 * mqd message queue descriptor
Jack Steiner28bffaf2008-07-29 22:33:57 -0700751 * mesg message. ust be vaddr within a GSEG
752 * bytes message size (<= 2 CL)
753 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700754int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg,
755 unsigned int bytes)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700756{
757 struct message_header *mhdr;
758 void *cb;
759 void *dsr;
760 int istatus, clines, ret;
761
762 STAT(mesq_send);
763 BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES);
764
Julia Lawallcbf330b2008-10-15 22:01:27 -0700765 clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700766 if (gru_get_cpu_resources(bytes, &cb, &dsr))
767 return MQE_BUG_NO_RESOURCES;
768 memcpy(dsr, mesg, bytes);
769 mhdr = dsr;
770 mhdr->present = MQS_FULL;
771 mhdr->lines = clines;
772 if (clines == 2) {
773 mhdr->present2 = get_present2(mhdr);
774 restore_present2(mhdr, MQS_FULL);
775 }
776
777 do {
778 ret = MQE_OK;
Jack Steiner6f2584f2009-04-02 16:59:10 -0700779 gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700780 istatus = gru_wait(cb);
781 if (istatus != CBS_IDLE)
Jack Steiner6f2584f2009-04-02 16:59:10 -0700782 ret = send_message_failure(cb, mqd, dsr, clines);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700783 } while (ret == MQIE_AGAIN);
784 gru_free_cpu_resources(cb, dsr);
785
786 if (ret)
787 STAT(mesq_send_failed);
788 return ret;
789}
790EXPORT_SYMBOL_GPL(gru_send_message_gpa);
791
792/*
793 * Advance the receive pointer for the queue to the next message.
794 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700795void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700796{
Jack Steiner6f2584f2009-04-02 16:59:10 -0700797 struct message_queue *mq = mqd->mq;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700798 struct message_header *mhdr = mq->next;
799 void *next, *pnext;
800 int half = -1;
801 int lines = mhdr->lines;
802
803 if (lines == 2)
804 restore_present2(mhdr, MQS_EMPTY);
805 mhdr->present = MQS_EMPTY;
806
807 pnext = mq->next;
808 next = pnext + GRU_CACHE_LINE_BYTES * lines;
809 if (next == mq->limit) {
810 next = mq->start;
811 half = 1;
812 } else if (pnext < mq->start2 && next >= mq->start2) {
813 half = 0;
814 }
815
816 if (half >= 0)
817 mq->hstatus[half] = 1;
818 mq->next = next;
819}
820EXPORT_SYMBOL_GPL(gru_free_message);
821
822/*
823 * Get next message from message queue. Return NULL if no message
824 * present. User must call next_message() to move to next message.
825 * rmq message queue
826 */
Jack Steiner6f2584f2009-04-02 16:59:10 -0700827void *gru_get_next_message(struct gru_message_queue_desc *mqd)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700828{
Jack Steiner6f2584f2009-04-02 16:59:10 -0700829 struct message_queue *mq = mqd->mq;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700830 struct message_header *mhdr = mq->next;
831 int present = mhdr->present;
832
833 /* skip NOOP messages */
834 STAT(mesq_receive);
835 while (present == MQS_NOOP) {
Jack Steiner6f2584f2009-04-02 16:59:10 -0700836 gru_free_message(mqd, mhdr);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700837 mhdr = mq->next;
838 present = mhdr->present;
839 }
840
841 /* Wait for both halves of 2 line messages */
842 if (present == MQS_FULL && mhdr->lines == 2 &&
843 get_present2(mhdr) == MQS_EMPTY)
844 present = MQS_EMPTY;
845
846 if (!present) {
847 STAT(mesq_receive_none);
848 return NULL;
849 }
850
851 if (mhdr->lines == 2)
852 restore_present2(mhdr, mhdr->present2);
853
854 return mhdr;
855}
856EXPORT_SYMBOL_GPL(gru_get_next_message);
857
858/* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/
859
860/*
861 * Copy a block of data using the GRU resources
862 */
863int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa,
864 unsigned int bytes)
865{
866 void *cb;
867 void *dsr;
868 int ret;
869
870 STAT(copy_gpa);
871 if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
872 return MQE_BUG_NO_RESOURCES;
873 gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr),
Jack Steiner6f2584f2009-04-02 16:59:10 -0700874 XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA);
Jack Steiner28bffaf2008-07-29 22:33:57 -0700875 ret = gru_wait(cb);
876 gru_free_cpu_resources(cb, dsr);
877 return ret;
878}
879EXPORT_SYMBOL_GPL(gru_copy_gpa);
880
881/* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/
882/* Temp - will delete after we gain confidence in the GRU */
Jack Steiner28bffaf2008-07-29 22:33:57 -0700883
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700884static int quicktest0(unsigned long arg)
Jack Steiner28bffaf2008-07-29 22:33:57 -0700885{
Jack Steiner836ce672009-06-17 16:28:22 -0700886 unsigned long word0;
887 unsigned long word1;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700888 void *cb;
Jack Steiner836ce672009-06-17 16:28:22 -0700889 void *dsr;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700890 unsigned long *p;
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700891 int ret = -EIO;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700892
Jack Steiner836ce672009-06-17 16:28:22 -0700893 if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr))
894 return MQE_BUG_NO_RESOURCES;
895 p = dsr;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700896 word0 = MAGIC;
Jack Steiner836ce672009-06-17 16:28:22 -0700897 word1 = 0;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700898
Jack Steiner836ce672009-06-17 16:28:22 -0700899 gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700900 if (gru_wait(cb) != CBS_IDLE) {
901 printk(KERN_DEBUG "GRU quicktest0: CBR failure 1\n");
902 goto done;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700903 }
904
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700905 if (*p != MAGIC) {
906 printk(KERN_DEBUG "GRU: quicktest0 bad magic 0x%lx\n", *p);
907 goto done;
908 }
909 gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
910 if (gru_wait(cb) != CBS_IDLE) {
911 printk(KERN_DEBUG "GRU quicktest0: CBR failure 2\n");
912 goto done;
913 }
914
915 if (word0 != word1 || word1 != MAGIC) {
916 printk(KERN_DEBUG
917 "GRU quicktest0 err: found 0x%lx, expected 0x%lx\n",
918 word1, MAGIC);
919 goto done;
920 }
921 ret = 0;
922
923done:
924 gru_free_cpu_resources(cb, dsr);
925 return ret;
Jack Steiner28bffaf2008-07-29 22:33:57 -0700926}
927
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700928#define ALIGNUP(p, q) ((void *)(((unsigned long)(p) + (q) - 1) & ~(q - 1)))
929
930static int quicktest1(unsigned long arg)
931{
932 struct gru_message_queue_desc mqd;
933 void *p, *mq;
934 unsigned long *dw;
935 int i, ret = -EIO;
936 char mes[GRU_CACHE_LINE_BYTES], *m;
937
938 /* Need 1K cacheline aligned that does not cross page boundary */
939 p = kmalloc(4096, 0);
Roel Kluin9e5f1132009-09-23 15:57:33 -0700940 if (p == NULL)
941 return -ENOMEM;
Jack Steinereb5bd5e2009-06-17 16:28:26 -0700942 mq = ALIGNUP(p, 1024);
943 memset(mes, 0xee, sizeof(mes));
944 dw = mq;
945
946 gru_create_message_queue(&mqd, mq, 8 * GRU_CACHE_LINE_BYTES, 0, 0, 0);
947 for (i = 0; i < 6; i++) {
948 mes[8] = i;
949 do {
950 ret = gru_send_message_gpa(&mqd, mes, sizeof(mes));
951 } while (ret == MQE_CONGESTION);
952 if (ret)
953 break;
954 }
955 if (ret != MQE_QUEUE_FULL || i != 4)
956 goto done;
957
958 for (i = 0; i < 6; i++) {
959 m = gru_get_next_message(&mqd);
960 if (!m || m[8] != i)
961 break;
962 gru_free_message(&mqd, m);
963 }
964 ret = (i == 4) ? 0 : -EIO;
965
966done:
967 kfree(p);
968 return ret;
969}
970
971static int quicktest2(unsigned long arg)
972{
973 static DECLARE_COMPLETION(cmp);
974 unsigned long han;
975 int blade_id = 0;
976 int numcb = 4;
977 int ret = 0;
978 unsigned long *buf;
979 void *cb0, *cb;
980 int i, k, istatus, bytes;
981
982 bytes = numcb * 4 * 8;
983 buf = kmalloc(bytes, GFP_KERNEL);
984 if (!buf)
985 return -ENOMEM;
986
987 ret = -EBUSY;
988 han = gru_reserve_async_resources(blade_id, numcb, 0, &cmp);
989 if (!han)
990 goto done;
991
992 gru_lock_async_resource(han, &cb0, NULL);
993 memset(buf, 0xee, bytes);
994 for (i = 0; i < numcb; i++)
995 gru_vset(cb0 + i * GRU_HANDLE_STRIDE, uv_gpa(&buf[i * 4]), 0,
996 XTYPE_DW, 4, 1, IMA_INTERRUPT);
997
998 ret = 0;
999 for (k = 0; k < numcb; k++) {
1000 gru_wait_async_cbr(han);
1001 for (i = 0; i < numcb; i++) {
1002 cb = cb0 + i * GRU_HANDLE_STRIDE;
1003 istatus = gru_check_status(cb);
1004 if (istatus == CBS_ACTIVE)
1005 continue;
1006 if (istatus == CBS_EXCEPTION)
1007 ret = -EFAULT;
1008 else if (buf[i] || buf[i + 1] || buf[i + 2] ||
1009 buf[i + 3])
1010 ret = -EIO;
1011 }
1012 }
1013 BUG_ON(cmp.done);
1014
1015 gru_unlock_async_resource(han);
1016 gru_release_async_resources(han);
1017done:
1018 kfree(buf);
1019 return ret;
1020}
1021
1022/*
1023 * Debugging only. User hook for various kernel tests
1024 * of driver & gru.
1025 */
1026int gru_ktest(unsigned long arg)
1027{
1028 int ret = -EINVAL;
1029
1030 switch (arg & 0xff) {
1031 case 0:
1032 ret = quicktest0(arg);
1033 break;
1034 case 1:
1035 ret = quicktest1(arg);
1036 break;
1037 case 2:
1038 ret = quicktest2(arg);
1039 break;
Jack Steinerd5826dd2009-06-17 16:28:28 -07001040 case 99:
1041 ret = gru_free_kernel_contexts();
1042 break;
Jack Steinereb5bd5e2009-06-17 16:28:26 -07001043 }
1044 return ret;
1045
1046}
Jack Steiner28bffaf2008-07-29 22:33:57 -07001047
Jack Steinerd5826dd2009-06-17 16:28:28 -07001048int gru_kservices_init(void)
Jack Steiner28bffaf2008-07-29 22:33:57 -07001049{
Jack Steiner28bffaf2008-07-29 22:33:57 -07001050 return 0;
1051}
Jack Steiner27ca8a72009-04-02 16:59:11 -07001052
Jack Steinerd5826dd2009-06-17 16:28:28 -07001053void gru_kservices_exit(void)
Jack Steiner27ca8a72009-04-02 16:59:11 -07001054{
Jack Steinerd5826dd2009-06-17 16:28:28 -07001055 if (gru_free_kernel_contexts())
1056 BUG();
Jack Steiner27ca8a72009-04-02 16:59:11 -07001057}
1058