1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * SN Platform GRU Driver
5 * KERNEL SERVICES THAT USE THE GRU
7 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
10 #include <linux/kernel.h>
11 #include <linux/errno.h>
12 #include <linux/slab.h>
14 #include <linux/spinlock.h>
15 #include <linux/device.h>
16 #include <linux/miscdevice.h>
17 #include <linux/proc_fs.h>
18 #include <linux/interrupt.h>
19 #include <linux/sync_core.h>
20 #include <linux/uaccess.h>
21 #include <linux/delay.h>
22 #include <linux/export.h>
23 #include <asm/io_apic.h>
26 #include "grutables.h"
27 #include "grukservices.h"
28 #include "gru_instructions.h"
29 #include <asm/uv/uv_hub.h>
34 * The following is an interim algorithm for management of kernel GRU
35 * resources. This will likely be replaced when we better understand the
36 * kernel/user requirements.
38 * Blade percpu resources reserved for kernel use. These resources are
39 * reserved whenever the the kernel context for the blade is loaded. Note
40 * that the kernel context is not guaranteed to be always available. It is
41 * loaded on demand & can be stolen by a user if the user demand exceeds the
42 * kernel demand. The kernel can always reload the kernel context but
43 * a SLEEP may be required!!!.
47 * Each blade has one "kernel context" that owns GRU kernel resources
48 * located on the blade. Kernel drivers use GRU resources in this context
49 * for sending messages, zeroing memory, etc.
51 * The kernel context is dynamically loaded on demand. If it is not in
52 * use by the kernel, the kernel context can be unloaded & given to a user.
53 * The kernel context will be reloaded when needed. This may require that
54 * a context be stolen from a user.
55 * NOTE: frequent unloading/reloading of the kernel context is
56 * expensive. We are depending on batch schedulers, cpusets, sane
57 * drivers or some other mechanism to prevent the need for frequent
60 * The kernel context consists of two parts:
61 * - 1 CB & a few DSRs that are reserved for each cpu on the blade.
62 * Each cpu has it's own private resources & does not share them
63 * with other cpus. These resources are used serially, ie,
64 * locked, used & unlocked on each call to a function in
66 * (Now that we have dynamic loading of kernel contexts, I
67 * may rethink this & allow sharing between cpus....)
69 * - Additional resources can be reserved long term & used directly
70 * by UV drivers located in the kernel. Drivers using these GRU
71 * resources can use asynchronous GRU instructions that send
72 * interrupts on completion.
73 * - these resources must be explicitly locked/unlocked
74 * - locked resources prevent (obviously) the kernel
75 * context from being unloaded.
76 * - drivers using these resource directly issue their own
77 * GRU instruction and must wait/check completion.
79 * When these resources are reserved, the caller can optionally
80 * associate a wait_queue with the resources and use asynchronous
81 * GRU instructions. When an async GRU instruction completes, the
82 * driver will do a wakeup on the event.
87 #define ASYNC_HAN_TO_BID(h) ((h) - 1)
88 #define ASYNC_BID_TO_HAN(b) ((b) + 1)
89 #define ASYNC_HAN_TO_BS(h) gru_base[ASYNC_HAN_TO_BID(h)]
91 #define GRU_NUM_KERNEL_CBR 1
92 #define GRU_NUM_KERNEL_DSR_BYTES 256
93 #define GRU_NUM_KERNEL_DSR_CL (GRU_NUM_KERNEL_DSR_BYTES / \
96 /* GRU instruction attributes for all instructions */
97 #define IMA IMA_CB_DELAY
99 /* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */
100 #define __gru_cacheline_aligned__ \
101 __attribute__((__aligned__(GRU_CACHE_LINE_BYTES)))
103 #define MAGIC 0x1234567887654321UL
105 /* Default retry count for GRU errors on kernel instructions */
106 #define EXCEPTION_RETRY_LIMIT 3
108 /* Status of message queue sections */
113 /*----------------- RESOURCE MANAGEMENT -------------------------------------*/
114 /* optimized for x86_64 */
115 struct message_queue {
116 union gru_mesqhead head __gru_cacheline_aligned__; /* CL 0 */
117 int qlines; /* DW 1 */
119 void *next __gru_cacheline_aligned__;/* CL 1 */
123 char data ____cacheline_aligned; /* CL 2 */
126 /* First word in every message - used by mesq interface */
127 struct message_header {
134 #define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h]))
137 * Reload the blade's kernel context into a GRU chiplet. Called holding
138 * the bs_kgts_sema for READ. Will steal user contexts if necessary.
140 static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id)
142 struct gru_state *gru;
143 struct gru_thread_state *kgts;
147 up_read(&bs->bs_kgts_sema);
148 down_write(&bs->bs_kgts_sema);
152 bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0, 0);
153 if (!IS_ERR(bs->bs_kgts))
157 bs->bs_kgts->ts_user_blade_id = blade_id;
162 STAT(load_kernel_context);
163 ncpus = uv_blade_nr_possible_cpus(blade_id);
164 kgts->ts_cbr_au_count = GRU_CB_COUNT_TO_AU(
165 GRU_NUM_KERNEL_CBR * ncpus + bs->bs_async_cbrs);
166 kgts->ts_dsr_au_count = GRU_DS_BYTES_TO_AU(
167 GRU_NUM_KERNEL_DSR_BYTES * ncpus +
168 bs->bs_async_dsr_bytes);
169 while (!gru_assign_gru_context(kgts)) {
171 gru_steal_context(kgts);
173 gru_load_context(kgts);
174 gru = bs->bs_kgts->ts_gru;
175 vaddr = gru->gs_gru_base_vaddr;
176 ctxnum = kgts->ts_ctxnum;
177 bs->kernel_cb = get_gseg_base_address_cb(vaddr, ctxnum, 0);
178 bs->kernel_dsr = get_gseg_base_address_ds(vaddr, ctxnum, 0);
180 downgrade_write(&bs->bs_kgts_sema);
184 * Free all kernel contexts that are not currently in use.
185 * Returns 0 if all freed, else number of inuse context.
187 static int gru_free_kernel_contexts(void)
189 struct gru_blade_state *bs;
190 struct gru_thread_state *kgts;
193 for (bid = 0; bid < GRU_MAX_BLADES; bid++) {
198 /* Ignore busy contexts. Don't want to block here. */
199 if (down_write_trylock(&bs->bs_kgts_sema)) {
201 if (kgts && kgts->ts_gru)
202 gru_unload_context(kgts, 0);
204 up_write(&bs->bs_kgts_sema);
214 * Lock & load the kernel context for the specified blade.
216 static struct gru_blade_state *gru_lock_kernel_context(int blade_id)
218 struct gru_blade_state *bs;
221 STAT(lock_kernel_context);
223 bid = blade_id < 0 ? uv_numa_blade_id() : blade_id;
226 /* Handle the case where migration occurred while waiting for the sema */
227 down_read(&bs->bs_kgts_sema);
228 if (blade_id < 0 && bid != uv_numa_blade_id()) {
229 up_read(&bs->bs_kgts_sema);
232 if (!bs->bs_kgts || !bs->bs_kgts->ts_gru)
233 gru_load_kernel_context(bs, bid);
239 * Unlock the kernel context for the specified blade. Context is not
240 * unloaded but may be stolen before next use.
242 static void gru_unlock_kernel_context(int blade_id)
244 struct gru_blade_state *bs;
246 bs = gru_base[blade_id];
247 up_read(&bs->bs_kgts_sema);
248 STAT(unlock_kernel_context);
252 * Reserve & get pointers to the DSR/CBRs reserved for the current cpu.
253 * - returns with preemption disabled
255 static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr)
257 struct gru_blade_state *bs;
260 BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES);
262 bs = gru_lock_kernel_context(-1);
263 lcpu = uv_blade_processor_id();
264 *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE;
265 *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES;
270 * Free the current cpus reserved DSR/CBR resources.
272 static void gru_free_cpu_resources(void *cb, void *dsr)
274 gru_unlock_kernel_context(uv_numa_blade_id());
279 * Reserve GRU resources to be used asynchronously.
280 * Note: currently supports only 1 reservation per blade.
283 * blade_id - blade on which resources should be reserved
284 * cbrs - number of CBRs
285 * dsr_bytes - number of DSR bytes needed
287 * handle to identify resource
288 * (0 = async resources already reserved)
290 unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
291 struct completion *cmp)
293 struct gru_blade_state *bs;
294 struct gru_thread_state *kgts;
297 bs = gru_base[blade_id];
299 down_write(&bs->bs_kgts_sema);
301 /* Verify no resources already reserved */
302 if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs)
304 bs->bs_async_dsr_bytes = dsr_bytes;
305 bs->bs_async_cbrs = cbrs;
306 bs->bs_async_wq = cmp;
309 /* Resources changed. Unload context if already loaded */
310 if (kgts && kgts->ts_gru)
311 gru_unload_context(kgts, 0);
312 ret = ASYNC_BID_TO_HAN(blade_id);
315 up_write(&bs->bs_kgts_sema);
320 * Release async resources previously reserved.
323 * han - handle to identify resources
325 void gru_release_async_resources(unsigned long han)
327 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
329 down_write(&bs->bs_kgts_sema);
330 bs->bs_async_dsr_bytes = 0;
331 bs->bs_async_cbrs = 0;
332 bs->bs_async_wq = NULL;
333 up_write(&bs->bs_kgts_sema);
337 * Wait for async GRU instructions to complete.
340 * han - handle to identify resources
342 void gru_wait_async_cbr(unsigned long han)
344 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
346 wait_for_completion(bs->bs_async_wq);
351 * Lock previous reserved async GRU resources
354 * han - handle to identify resources
356 * cb - pointer to first CBR
357 * dsr - pointer to first DSR
359 void gru_lock_async_resource(unsigned long han, void **cb, void **dsr)
361 struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
362 int blade_id = ASYNC_HAN_TO_BID(han);
365 gru_lock_kernel_context(blade_id);
366 ncpus = uv_blade_nr_possible_cpus(blade_id);
368 *cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE;
370 *dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES;
374 * Unlock previous reserved async GRU resources
377 * han - handle to identify resources
379 void gru_unlock_async_resource(unsigned long han)
381 int blade_id = ASYNC_HAN_TO_BID(han);
383 gru_unlock_kernel_context(blade_id);
386 /*----------------------------------------------------------------------*/
387 int gru_get_cb_exception_detail(void *cb,
388 struct control_block_extended_exc_detail *excdet)
390 struct gru_control_block_extended *cbe;
391 struct gru_thread_state *kgts = NULL;
396 * Locate kgts for cb. This algorithm is SLOW but
397 * this function is rarely called (ie., almost never).
398 * Performance does not matter.
400 for_each_possible_blade(bid) {
403 kgts = gru_base[bid]->bs_kgts;
404 if (!kgts || !kgts->ts_gru)
406 off = cb - kgts->ts_gru->gs_gru_base_vaddr;
412 cbrnum = thread_cbr_number(kgts, get_cb_number(cb));
413 cbe = get_cbe(GRUBASE(cb), cbrnum);
414 gru_flush_cache(cbe); /* CBE not coherent */
416 excdet->opc = cbe->opccpy;
417 excdet->exopc = cbe->exopccpy;
418 excdet->ecause = cbe->ecause;
419 excdet->exceptdet0 = cbe->idef1upd;
420 excdet->exceptdet1 = cbe->idef3upd;
421 gru_flush_cache(cbe);
425 static char *gru_get_cb_exception_detail_str(int ret, void *cb,
428 struct gru_control_block_status *gen = (void *)cb;
429 struct control_block_extended_exc_detail excdet;
431 if (ret > 0 && gen->istatus == CBS_EXCEPTION) {
432 gru_get_cb_exception_detail(cb, &excdet);
434 "GRU:%d exception: cb %p, opc %d, exopc %d, ecause 0x%x,"
435 "excdet0 0x%lx, excdet1 0x%x", smp_processor_id(),
436 gen, excdet.opc, excdet.exopc, excdet.ecause,
437 excdet.exceptdet0, excdet.exceptdet1);
439 snprintf(buf, size, "No exception");
444 static int gru_wait_idle_or_exception(struct gru_control_block_status *gen)
446 while (gen->istatus >= CBS_ACTIVE) {
453 static int gru_retry_exception(void *cb)
455 struct gru_control_block_status *gen = (void *)cb;
456 struct control_block_extended_exc_detail excdet;
457 int retry = EXCEPTION_RETRY_LIMIT;
460 if (gru_wait_idle_or_exception(gen) == CBS_IDLE)
462 if (gru_get_cb_message_queue_substatus(cb))
463 return CBS_EXCEPTION;
464 gru_get_cb_exception_detail(cb, &excdet);
465 if ((excdet.ecause & ~EXCEPTION_RETRY_BITS) ||
466 (excdet.cbrexecstatus & CBR_EXS_ABORT_OCC))
471 gru_flush_cache(gen);
473 return CBS_EXCEPTION;
476 int gru_check_status_proc(void *cb)
478 struct gru_control_block_status *gen = (void *)cb;
482 if (ret == CBS_EXCEPTION)
483 ret = gru_retry_exception(cb);
489 int gru_wait_proc(void *cb)
491 struct gru_control_block_status *gen = (void *)cb;
494 ret = gru_wait_idle_or_exception(gen);
495 if (ret == CBS_EXCEPTION)
496 ret = gru_retry_exception(cb);
501 static void gru_abort(int ret, void *cb, char *str)
503 char buf[GRU_EXC_STR_SIZE];
505 panic("GRU FATAL ERROR: %s - %s\n", str,
506 gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf)));
509 void gru_wait_abort_proc(void *cb)
513 ret = gru_wait_proc(cb);
515 gru_abort(ret, cb, "gru_wait_abort");
519 /*------------------------------ MESSAGE QUEUES -----------------------------*/
521 /* Internal status . These are NOT returned to the user. */
522 #define MQIE_AGAIN -1 /* try again */
526 * Save/restore the "present" flag that is in the second line of 2-line
529 static inline int get_present2(void *p)
531 struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
532 return mhdr->present;
535 static inline void restore_present2(void *p, int val)
537 struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
542 * Create a message queue.
543 * qlines - message queue size in cache lines. Includes 2-line header.
545 int gru_create_message_queue(struct gru_message_queue_desc *mqd,
546 void *p, unsigned int bytes, int nasid, int vector, int apicid)
548 struct message_queue *mq = p;
551 qlines = bytes / GRU_CACHE_LINE_BYTES - 2;
552 memset(mq, 0, bytes);
553 mq->start = &mq->data;
554 mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES;
555 mq->next = &mq->data;
556 mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES;
560 mq->head = gru_mesq_head(2, qlines / 2 + 1);
562 mqd->mq_gpa = uv_gpa(mq);
563 mqd->qlines = qlines;
564 mqd->interrupt_pnode = nasid >> 1;
565 mqd->interrupt_vector = vector;
566 mqd->interrupt_apicid = apicid;
569 EXPORT_SYMBOL_GPL(gru_create_message_queue);
572 * Send a NOOP message to a message queue
574 * 0 - if queue is full after the send. This is the normal case
575 * but various races can change this.
576 * -1 - if mesq sent successfully but queue not full
577 * >0 - unexpected error. MQE_xxx returned
579 static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd,
582 const struct message_header noop_header = {
583 .present = MQS_NOOP, .lines = 1};
586 struct message_header save_mhdr, *mhdr = mesg;
591 gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA);
595 substatus = gru_get_cb_message_queue_substatus(cb);
598 STAT(mesq_noop_unexpected_error);
599 ret = MQE_UNEXPECTED_CB_ERR;
601 case CBSS_LB_OVERFLOWED:
602 STAT(mesq_noop_lb_overflow);
603 ret = MQE_CONGESTION;
605 case CBSS_QLIMIT_REACHED:
606 STAT(mesq_noop_qlimit_reached);
609 case CBSS_AMO_NACKED:
610 STAT(mesq_noop_amo_nacked);
611 ret = MQE_CONGESTION;
613 case CBSS_PUT_NACKED:
614 STAT(mesq_noop_put_nacked);
615 m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
616 gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1,
618 if (gru_wait(cb) == CBS_IDLE)
621 ret = MQE_UNEXPECTED_CB_ERR;
623 case CBSS_PAGE_OVERFLOW:
624 STAT(mesq_noop_page_overflow);
635 * Handle a gru_mesq full.
637 static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd,
638 void *mesg, int lines)
640 union gru_mesqhead mqh;
641 unsigned int limit, head;
642 unsigned long avalue;
645 /* Determine if switching to first/second half of q */
646 avalue = gru_get_amo_value(cb);
647 head = gru_get_amo_value_head(cb);
648 limit = gru_get_amo_value_limit(cb);
650 qlines = mqd->qlines;
651 half = (limit != qlines);
654 mqh = gru_mesq_head(qlines / 2 + 1, qlines);
656 mqh = gru_mesq_head(2, qlines / 2 + 1);
658 /* Try to get lock for switching head pointer */
659 gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA);
660 if (gru_wait(cb) != CBS_IDLE)
662 if (!gru_get_amo_value(cb)) {
663 STAT(mesq_qf_locked);
664 return MQE_QUEUE_FULL;
667 /* Got the lock. Send optional NOP if queue not full, */
669 if (send_noop_message(cb, mqd, mesg)) {
670 gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half),
672 if (gru_wait(cb) != CBS_IDLE)
674 STAT(mesq_qf_noop_not_full);
680 /* Then flip queuehead to other half of queue. */
681 gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue,
683 if (gru_wait(cb) != CBS_IDLE)
686 /* If not successfully in swapping queue head, clear the hstatus lock */
687 if (gru_get_amo_value(cb) != avalue) {
688 STAT(mesq_qf_switch_head_failed);
689 gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW,
691 if (gru_wait(cb) != CBS_IDLE)
696 STAT(mesq_qf_unexpected_error);
697 return MQE_UNEXPECTED_CB_ERR;
701 * Handle a PUT failure. Note: if message was a 2-line message, one of the
702 * lines might have successfully have been written. Before sending the
703 * message, "present" must be cleared in BOTH lines to prevent the receiver
704 * from prematurely seeing the full message.
706 static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd,
707 void *mesg, int lines)
710 int ret, loops = 200; /* experimentally determined */
712 m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
714 gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA);
715 if (gru_wait(cb) != CBS_IDLE)
716 return MQE_UNEXPECTED_CB_ERR;
718 gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA);
719 if (gru_wait(cb) != CBS_IDLE)
720 return MQE_UNEXPECTED_CB_ERR;
722 if (!mqd->interrupt_vector)
726 * Send a noop message in order to deliver a cross-partition interrupt
727 * to the SSI that contains the target message queue. Normally, the
728 * interrupt is automatically delivered by hardware following mesq
729 * operations, but some error conditions require explicit delivery.
730 * The noop message will trigger delivery. Otherwise partition failures
731 * could cause unrecovered errors.
734 ret = send_noop_message(cb, mqd, mesg);
735 } while ((ret == MQIE_AGAIN || ret == MQE_CONGESTION) && (loops-- > 0));
737 if (ret == MQIE_AGAIN || ret == MQE_CONGESTION) {
739 * Don't indicate to the app to resend the message, as it's
740 * already been successfully sent. We simply send an OK
741 * (rather than fail the send with MQE_UNEXPECTED_CB_ERR),
742 * assuming that the other side is receiving enough
743 * interrupts to get this message processed anyway.
751 * Handle a gru_mesq failure. Some of these failures are software recoverable
754 static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd,
755 void *mesg, int lines)
757 int substatus, ret = 0;
759 substatus = gru_get_cb_message_queue_substatus(cb);
762 STAT(mesq_send_unexpected_error);
763 ret = MQE_UNEXPECTED_CB_ERR;
765 case CBSS_LB_OVERFLOWED:
766 STAT(mesq_send_lb_overflow);
767 ret = MQE_CONGESTION;
769 case CBSS_QLIMIT_REACHED:
770 STAT(mesq_send_qlimit_reached);
771 ret = send_message_queue_full(cb, mqd, mesg, lines);
773 case CBSS_AMO_NACKED:
774 STAT(mesq_send_amo_nacked);
775 ret = MQE_CONGESTION;
777 case CBSS_PUT_NACKED:
778 STAT(mesq_send_put_nacked);
779 ret = send_message_put_nacked(cb, mqd, mesg, lines);
781 case CBSS_PAGE_OVERFLOW:
782 STAT(mesq_page_overflow);
791 * Send a message to a message queue
792 * mqd message queue descriptor
793 * mesg message. ust be vaddr within a GSEG
794 * bytes message size (<= 2 CL)
796 int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg,
799 struct message_header *mhdr;
802 int istatus, clines, ret;
805 BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES);
807 clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES);
808 if (gru_get_cpu_resources(bytes, &cb, &dsr))
809 return MQE_BUG_NO_RESOURCES;
810 memcpy(dsr, mesg, bytes);
812 mhdr->present = MQS_FULL;
813 mhdr->lines = clines;
815 mhdr->present2 = get_present2(mhdr);
816 restore_present2(mhdr, MQS_FULL);
821 gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA);
822 istatus = gru_wait(cb);
823 if (istatus != CBS_IDLE)
824 ret = send_message_failure(cb, mqd, dsr, clines);
825 } while (ret == MQIE_AGAIN);
826 gru_free_cpu_resources(cb, dsr);
829 STAT(mesq_send_failed);
832 EXPORT_SYMBOL_GPL(gru_send_message_gpa);
835 * Advance the receive pointer for the queue to the next message.
837 void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg)
839 struct message_queue *mq = mqd->mq;
840 struct message_header *mhdr = mq->next;
843 int lines = mhdr->lines;
846 restore_present2(mhdr, MQS_EMPTY);
847 mhdr->present = MQS_EMPTY;
850 next = pnext + GRU_CACHE_LINE_BYTES * lines;
851 if (next == mq->limit) {
854 } else if (pnext < mq->start2 && next >= mq->start2) {
859 mq->hstatus[half] = 1;
862 EXPORT_SYMBOL_GPL(gru_free_message);
865 * Get next message from message queue. Return NULL if no message
866 * present. User must call next_message() to move to next message.
869 void *gru_get_next_message(struct gru_message_queue_desc *mqd)
871 struct message_queue *mq = mqd->mq;
872 struct message_header *mhdr = mq->next;
873 int present = mhdr->present;
875 /* skip NOOP messages */
876 while (present == MQS_NOOP) {
877 gru_free_message(mqd, mhdr);
879 present = mhdr->present;
882 /* Wait for both halves of 2 line messages */
883 if (present == MQS_FULL && mhdr->lines == 2 &&
884 get_present2(mhdr) == MQS_EMPTY)
888 STAT(mesq_receive_none);
892 if (mhdr->lines == 2)
893 restore_present2(mhdr, mhdr->present2);
898 EXPORT_SYMBOL_GPL(gru_get_next_message);
900 /* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/
903 * Load a DW from a global GPA. The GPA can be a memory or MMR address.
905 int gru_read_gpa(unsigned long *value, unsigned long gpa)
912 if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
913 return MQE_BUG_NO_RESOURCES;
915 gru_vload_phys(cb, gpa, gru_get_tri(dsr), iaa, IMA);
918 *value = *(unsigned long *)dsr;
919 gru_free_cpu_resources(cb, dsr);
922 EXPORT_SYMBOL_GPL(gru_read_gpa);
926 * Copy a block of data using the GRU resources
928 int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa,
936 if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
937 return MQE_BUG_NO_RESOURCES;
938 gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr),
939 XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA);
941 gru_free_cpu_resources(cb, dsr);
944 EXPORT_SYMBOL_GPL(gru_copy_gpa);
946 /* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/
947 /* Temp - will delete after we gain confidence in the GRU */
949 static int quicktest0(unsigned long arg)
958 if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr))
959 return MQE_BUG_NO_RESOURCES;
964 gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
965 if (gru_wait(cb) != CBS_IDLE) {
966 printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 1\n", smp_processor_id());
971 printk(KERN_DEBUG "GRU:%d quicktest0 bad magic 0x%lx\n", smp_processor_id(), *p);
974 gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
975 if (gru_wait(cb) != CBS_IDLE) {
976 printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 2\n", smp_processor_id());
980 if (word0 != word1 || word1 != MAGIC) {
982 "GRU:%d quicktest0 err: found 0x%lx, expected 0x%lx\n",
983 smp_processor_id(), word1, MAGIC);
989 gru_free_cpu_resources(cb, dsr);
993 #define ALIGNUP(p, q) ((void *)(((unsigned long)(p) + (q) - 1) & ~(q - 1)))
995 static int quicktest1(unsigned long arg)
997 struct gru_message_queue_desc mqd;
1000 char mes[GRU_CACHE_LINE_BYTES], *m;
1002 /* Need 1K cacheline aligned that does not cross page boundary */
1003 p = kmalloc(4096, 0);
1006 mq = ALIGNUP(p, 1024);
1007 memset(mes, 0xee, sizeof(mes));
1009 gru_create_message_queue(&mqd, mq, 8 * GRU_CACHE_LINE_BYTES, 0, 0, 0);
1010 for (i = 0; i < 6; i++) {
1013 ret = gru_send_message_gpa(&mqd, mes, sizeof(mes));
1014 } while (ret == MQE_CONGESTION);
1018 if (ret != MQE_QUEUE_FULL || i != 4) {
1019 printk(KERN_DEBUG "GRU:%d quicktest1: unexpected status %d, i %d\n",
1020 smp_processor_id(), ret, i);
1024 for (i = 0; i < 6; i++) {
1025 m = gru_get_next_message(&mqd);
1026 if (!m || m[8] != i)
1028 gru_free_message(&mqd, m);
1031 printk(KERN_DEBUG "GRU:%d quicktest2: bad message, i %d, m %p, m8 %d\n",
1032 smp_processor_id(), i, m, m ? m[8] : -1);
1042 static int quicktest2(unsigned long arg)
1044 static DECLARE_COMPLETION(cmp);
1051 struct gru_control_block_status *gen;
1052 int i, k, istatus, bytes;
1054 bytes = numcb * 4 * 8;
1055 buf = kmalloc(bytes, GFP_KERNEL);
1060 han = gru_reserve_async_resources(blade_id, numcb, 0, &cmp);
1064 gru_lock_async_resource(han, &cb0, NULL);
1065 memset(buf, 0xee, bytes);
1066 for (i = 0; i < numcb; i++)
1067 gru_vset(cb0 + i * GRU_HANDLE_STRIDE, uv_gpa(&buf[i * 4]), 0,
1068 XTYPE_DW, 4, 1, IMA_INTERRUPT);
1073 gru_wait_async_cbr(han);
1074 for (i = 0; i < numcb; i++) {
1075 cb = cb0 + i * GRU_HANDLE_STRIDE;
1076 istatus = gru_check_status(cb);
1077 if (istatus != CBS_ACTIVE && istatus != CBS_CALL_OS)
1082 if (istatus != CBS_IDLE) {
1083 printk(KERN_DEBUG "GRU:%d quicktest2: cb %d, exception\n", smp_processor_id(), i);
1085 } else if (buf[4 * i] || buf[4 * i + 1] || buf[4 * i + 2] ||
1087 printk(KERN_DEBUG "GRU:%d quicktest2:cb %d, buf 0x%lx, 0x%lx, 0x%lx, 0x%lx\n",
1088 smp_processor_id(), i, buf[4 * i], buf[4 * i + 1], buf[4 * i + 2], buf[4 * i + 3]);
1093 gen->istatus = CBS_CALL_OS; /* don't handle this CBR again */
1097 gru_unlock_async_resource(han);
1098 gru_release_async_resources(han);
1105 static int quicktest3(unsigned long arg)
1107 char buf1[BUFSIZE], buf2[BUFSIZE];
1110 memset(buf2, 0, sizeof(buf2));
1111 memset(buf1, get_cycles() & 255, sizeof(buf1));
1112 gru_copy_gpa(uv_gpa(buf2), uv_gpa(buf1), BUFSIZE);
1113 if (memcmp(buf1, buf2, BUFSIZE)) {
1114 printk(KERN_DEBUG "GRU:%d quicktest3 error\n", smp_processor_id());
1121 * Debugging only. User hook for various kernel tests
1124 int gru_ktest(unsigned long arg)
1128 switch (arg & 0xff) {
1130 ret = quicktest0(arg);
1133 ret = quicktest1(arg);
1136 ret = quicktest2(arg);
1139 ret = quicktest3(arg);
1142 ret = gru_free_kernel_contexts();
1149 int gru_kservices_init(void)
1154 void gru_kservices_exit(void)
1156 if (gru_free_kernel_contexts())