grukservices.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * SN Platform GRU Driver
   4 *
   5 *              KERNEL SERVICES THAT USE THE GRU
   6 *
   7 *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
   8 */
   9
  10#include <linux/kernel.h>
  11#include <linux/errno.h>
  12#include <linux/slab.h>
  13#include <linux/mm.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>
  24#include "gru.h"
  25#include "grulib.h"
  26#include "grutables.h"
  27#include "grukservices.h"
  28#include "gru_instructions.h"
  29#include <asm/uv/uv_hub.h>
  30
  31/*
  32 * Kernel GRU Usage
  33 *
  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.
  37 *
  38 * Blade percpu resources reserved for kernel use. These resources are
  39 * reserved whenever 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!!!.
  44 *
  45 * Async Overview:
  46 *
  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.
  50 *
  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
  58 *	 	stealing/reloading.
  59 *
  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
  65 * 		  grukservices.
  66 * 		  	(Now that we have dynamic loading of kernel contexts, I
  67 * 		  	 may rethink this & allow sharing between cpus....)
  68 *
  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.
  78 *
  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.
  83 *
  84 */
  85
  86
  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)]
  90
  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 /		\
  94					GRU_CACHE_LINE_BYTES)
  95
  96/* GRU instruction attributes for all instructions */
  97#define IMA			IMA_CB_DELAY
  98
  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)))
 102
 103#define MAGIC	0x1234567887654321UL
 104
 105/* Default retry count for GRU errors on kernel instructions */
 106#define EXCEPTION_RETRY_LIMIT	3
 107
 108/* Status of message queue sections */
 109#define MQS_EMPTY		0
 110#define MQS_FULL		1
 111#define MQS_NOOP		2
 112
 113/*----------------- RESOURCE MANAGEMENT -------------------------------------*/
 114/* optimized for x86_64 */
 115struct message_queue {
 116	union gru_mesqhead	head __gru_cacheline_aligned__;	/* CL 0 */
 117	int			qlines;				/* DW 1 */
 118	long 			hstatus[2];
 119	void 			*next __gru_cacheline_aligned__;/* CL 1 */
 120	void 			*limit;
 121	void 			*start;
 122	void 			*start2;
 123	char			data ____cacheline_aligned;	/* CL 2 */
 124};
 125
 126/* First word in every message - used by mesq interface */
 127struct message_header {
 128	char	present;
 129	char	present2;
 130	char 	lines;
 131	char	fill;
 132};
 133
 134#define HSTATUS(mq, h)	((mq) + offsetof(struct message_queue, hstatus[h]))
 135
 136/*
 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.
 139 */
 140static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id)
 141{
 142	struct gru_state *gru;
 143	struct gru_thread_state *kgts;
 144	void *vaddr;
 145	int ctxnum, ncpus;
 146
 147	up_read(&bs->bs_kgts_sema);
 148	down_write(&bs->bs_kgts_sema);
 149
 150	if (!bs->bs_kgts) {
 151		do {
 152			bs->bs_kgts = gru_alloc_gts(NULL, 0, 0, 0, 0, 0);
 153			if (!IS_ERR(bs->bs_kgts))
 154				break;
 155			msleep(1);
 156		} while (true);
 157		bs->bs_kgts->ts_user_blade_id = blade_id;
 158	}
 159	kgts = bs->bs_kgts;
 160
 161	if (!kgts->ts_gru) {
 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)) {
 170			msleep(1);
 171			gru_steal_context(kgts);
 172		}
 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);
 179	}
 180	downgrade_write(&bs->bs_kgts_sema);
 181}
 182
 183/*
 184 * Free all kernel contexts that are not currently in use.
 185 *   Returns 0 if all freed, else number of inuse context.
 186 */
 187static int gru_free_kernel_contexts(void)
 188{
 189	struct gru_blade_state *bs;
 190	struct gru_thread_state *kgts;
 191	int bid, ret = 0;
 192
 193	for (bid = 0; bid < GRU_MAX_BLADES; bid++) {
 194		bs = gru_base[bid];
 195		if (!bs)
 196			continue;
 197
 198		/* Ignore busy contexts. Don't want to block here.  */
 199		if (down_write_trylock(&bs->bs_kgts_sema)) {
 200			kgts = bs->bs_kgts;
 201			if (kgts && kgts->ts_gru)
 202				gru_unload_context(kgts, 0);
 203			bs->bs_kgts = NULL;
 204			up_write(&bs->bs_kgts_sema);
 205			kfree(kgts);
 206		} else {
 207			ret++;
 208		}
 209	}
 210	return ret;
 211}
 212
 213/*
 214 * Lock & load the kernel context for the specified blade.
 215 */
 216static struct gru_blade_state *gru_lock_kernel_context(int blade_id)
 217{
 218	struct gru_blade_state *bs;
 219	int bid;
 220
 221	STAT(lock_kernel_context);
 222again:
 223	bid = blade_id < 0 ? uv_numa_blade_id() : blade_id;
 224	bs = gru_base[bid];
 225
 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);
 230		goto again;
 231	}
 232	if (!bs->bs_kgts || !bs->bs_kgts->ts_gru)
 233		gru_load_kernel_context(bs, bid);
 234	return bs;
 235
 236}
 237
 238/*
 239 * Unlock the kernel context for the specified blade. Context is not
 240 * unloaded but may be stolen before next use.
 241 */
 242static void gru_unlock_kernel_context(int blade_id)
 243{
 244	struct gru_blade_state *bs;
 245
 246	bs = gru_base[blade_id];
 247	up_read(&bs->bs_kgts_sema);
 248	STAT(unlock_kernel_context);
 249}
 250
 251/*
 252 * Reserve & get pointers to the DSR/CBRs reserved for the current cpu.
 253 * 	- returns with preemption disabled
 254 */
 255static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr)
 256{
 257	struct gru_blade_state *bs;
 258	int lcpu;
 259
 260	BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES);
 261	bs = gru_lock_kernel_context(-1);
 262	lcpu = uv_blade_processor_id();
 263	*cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE;
 264	*dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES;
 265	return 0;
 266}
 267
 268/*
 269 * Free the current cpus reserved DSR/CBR resources.
 270 */
 271static void gru_free_cpu_resources(void *cb, void *dsr)
 272{
 273	gru_unlock_kernel_context(uv_numa_blade_id());
 274}
 275
 276/*
 277 * Reserve GRU resources to be used asynchronously.
 278 *   Note: currently supports only 1 reservation per blade.
 279 *
 280 * 	input:
 281 * 		blade_id  - blade on which resources should be reserved
 282 * 		cbrs	  - number of CBRs
 283 * 		dsr_bytes - number of DSR bytes needed
 284 *	output:
 285 *		handle to identify resource
 286 *		(0 = async resources already reserved)
 287 */
 288unsigned long gru_reserve_async_resources(int blade_id, int cbrs, int dsr_bytes,
 289			struct completion *cmp)
 290{
 291	struct gru_blade_state *bs;
 292	struct gru_thread_state *kgts;
 293	int ret = 0;
 294
 295	bs = gru_base[blade_id];
 296
 297	down_write(&bs->bs_kgts_sema);
 298
 299	/* Verify no resources already reserved */
 300	if (bs->bs_async_dsr_bytes + bs->bs_async_cbrs)
 301		goto done;
 302	bs->bs_async_dsr_bytes = dsr_bytes;
 303	bs->bs_async_cbrs = cbrs;
 304	bs->bs_async_wq = cmp;
 305	kgts = bs->bs_kgts;
 306
 307	/* Resources changed. Unload context if already loaded */
 308	if (kgts && kgts->ts_gru)
 309		gru_unload_context(kgts, 0);
 310	ret = ASYNC_BID_TO_HAN(blade_id);
 311
 312done:
 313	up_write(&bs->bs_kgts_sema);
 314	return ret;
 315}
 316
 317/*
 318 * Release async resources previously reserved.
 319 *
 320 *	input:
 321 *		han - handle to identify resources
 322 */
 323void gru_release_async_resources(unsigned long han)
 324{
 325	struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 326
 327	down_write(&bs->bs_kgts_sema);
 328	bs->bs_async_dsr_bytes = 0;
 329	bs->bs_async_cbrs = 0;
 330	bs->bs_async_wq = NULL;
 331	up_write(&bs->bs_kgts_sema);
 332}
 333
 334/*
 335 * Wait for async GRU instructions to complete.
 336 *
 337 *	input:
 338 *		han - handle to identify resources
 339 */
 340void gru_wait_async_cbr(unsigned long han)
 341{
 342	struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 343
 344	wait_for_completion(bs->bs_async_wq);
 345	mb();
 346}
 347
 348/*
 349 * Lock previous reserved async GRU resources
 350 *
 351 *	input:
 352 *		han - handle to identify resources
 353 *	output:
 354 *		cb  - pointer to first CBR
 355 *		dsr - pointer to first DSR
 356 */
 357void gru_lock_async_resource(unsigned long han,  void **cb, void **dsr)
 358{
 359	struct gru_blade_state *bs = ASYNC_HAN_TO_BS(han);
 360	int blade_id = ASYNC_HAN_TO_BID(han);
 361	int ncpus;
 362
 363	gru_lock_kernel_context(blade_id);
 364	ncpus = uv_blade_nr_possible_cpus(blade_id);
 365	if (cb)
 366		*cb = bs->kernel_cb + ncpus * GRU_HANDLE_STRIDE;
 367	if (dsr)
 368		*dsr = bs->kernel_dsr + ncpus * GRU_NUM_KERNEL_DSR_BYTES;
 369}
 370
 371/*
 372 * Unlock previous reserved async GRU resources
 373 *
 374 *	input:
 375 *		han - handle to identify resources
 376 */
 377void gru_unlock_async_resource(unsigned long han)
 378{
 379	int blade_id = ASYNC_HAN_TO_BID(han);
 380
 381	gru_unlock_kernel_context(blade_id);
 382}
 383
 384/*----------------------------------------------------------------------*/
 385int gru_get_cb_exception_detail(void *cb,
 386		struct control_block_extended_exc_detail *excdet)
 387{
 388	struct gru_control_block_extended *cbe;
 389	struct gru_thread_state *kgts = NULL;
 390	unsigned long off;
 391	int cbrnum, bid;
 392
 393	/*
 394	 * Locate kgts for cb. This algorithm is SLOW but
 395	 * this function is rarely called (ie., almost never).
 396	 * Performance does not matter.
 397	 */
 398	for_each_possible_blade(bid) {
 399		if (!gru_base[bid])
 400			break;
 401		kgts = gru_base[bid]->bs_kgts;
 402		if (!kgts || !kgts->ts_gru)
 403			continue;
 404		off = cb - kgts->ts_gru->gs_gru_base_vaddr;
 405		if (off < GRU_SIZE)
 406			break;
 407		kgts = NULL;
 408	}
 409	BUG_ON(!kgts);
 410	cbrnum = thread_cbr_number(kgts, get_cb_number(cb));
 411	cbe = get_cbe(GRUBASE(cb), cbrnum);
 412	gru_flush_cache(cbe);	/* CBE not coherent */
 413	sync_core();
 414	excdet->opc = cbe->opccpy;
 415	excdet->exopc = cbe->exopccpy;
 416	excdet->ecause = cbe->ecause;
 417	excdet->exceptdet0 = cbe->idef1upd;
 418	excdet->exceptdet1 = cbe->idef3upd;
 419	gru_flush_cache(cbe);
 420	return 0;
 421}
 422
 423static char *gru_get_cb_exception_detail_str(int ret, void *cb,
 424					     char *buf, int size)
 425{
 426	struct gru_control_block_status *gen = cb;
 427	struct control_block_extended_exc_detail excdet;
 428
 429	if (ret > 0 && gen->istatus == CBS_EXCEPTION) {
 430		gru_get_cb_exception_detail(cb, &excdet);
 431		snprintf(buf, size,
 432			"GRU:%d exception: cb %p, opc %d, exopc %d, ecause 0x%x,"
 433			"excdet0 0x%lx, excdet1 0x%x", smp_processor_id(),
 434			gen, excdet.opc, excdet.exopc, excdet.ecause,
 435			excdet.exceptdet0, excdet.exceptdet1);
 436	} else {
 437		snprintf(buf, size, "No exception");
 438	}
 439	return buf;
 440}
 441
 442static int gru_wait_idle_or_exception(struct gru_control_block_status *gen)
 443{
 444	while (gen->istatus >= CBS_ACTIVE) {
 445		cpu_relax();
 446		barrier();
 447	}
 448	return gen->istatus;
 449}
 450
 451static int gru_retry_exception(void *cb)
 452{
 453	struct gru_control_block_status *gen = cb;
 454	struct control_block_extended_exc_detail excdet;
 455	int retry = EXCEPTION_RETRY_LIMIT;
 456
 457	while (1)  {
 458		if (gru_wait_idle_or_exception(gen) == CBS_IDLE)
 459			return CBS_IDLE;
 460		if (gru_get_cb_message_queue_substatus(cb))
 461			return CBS_EXCEPTION;
 462		gru_get_cb_exception_detail(cb, &excdet);
 463		if ((excdet.ecause & ~EXCEPTION_RETRY_BITS) ||
 464				(excdet.cbrexecstatus & CBR_EXS_ABORT_OCC))
 465			break;
 466		if (retry-- == 0)
 467			break;
 468		gen->icmd = 1;
 469		gru_flush_cache(gen);
 470	}
 471	return CBS_EXCEPTION;
 472}
 473
 474int gru_check_status_proc(void *cb)
 475{
 476	struct gru_control_block_status *gen = cb;
 477	int ret;
 478
 479	ret = gen->istatus;
 480	if (ret == CBS_EXCEPTION)
 481		ret = gru_retry_exception(cb);
 482	rmb();
 483	return ret;
 484
 485}
 486
 487int gru_wait_proc(void *cb)
 488{
 489	struct gru_control_block_status *gen = cb;
 490	int ret;
 491
 492	ret = gru_wait_idle_or_exception(gen);
 493	if (ret == CBS_EXCEPTION)
 494		ret = gru_retry_exception(cb);
 495	rmb();
 496	return ret;
 497}
 498
 499static void gru_abort(int ret, void *cb, char *str)
 500{
 501	char buf[GRU_EXC_STR_SIZE];
 502
 503	panic("GRU FATAL ERROR: %s - %s\n", str,
 504	      gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf)));
 505}
 506
 507void gru_wait_abort_proc(void *cb)
 508{
 509	int ret;
 510
 511	ret = gru_wait_proc(cb);
 512	if (ret)
 513		gru_abort(ret, cb, "gru_wait_abort");
 514}
 515
 516
 517/*------------------------------ MESSAGE QUEUES -----------------------------*/
 518
 519/* Internal status . These are NOT returned to the user. */
 520#define MQIE_AGAIN		-1	/* try again */
 521
 522
 523/*
 524 * Save/restore the "present" flag that is in the second line of 2-line
 525 * messages
 526 */
 527static inline int get_present2(void *p)
 528{
 529	struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
 530	return mhdr->present;
 531}
 532
 533static inline void restore_present2(void *p, int val)
 534{
 535	struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES;
 536	mhdr->present = val;
 537}
 538
 539/*
 540 * Create a message queue.
 541 * 	qlines - message queue size in cache lines. Includes 2-line header.
 542 */
 543int gru_create_message_queue(struct gru_message_queue_desc *mqd,
 544		void *p, unsigned int bytes, int nasid, int vector, int apicid)
 545{
 546	struct message_queue *mq = p;
 547	unsigned int qlines;
 548
 549	qlines = bytes / GRU_CACHE_LINE_BYTES - 2;
 550	memset(mq, 0, bytes);
 551	mq->start = &mq->data;
 552	mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES;
 553	mq->next = &mq->data;
 554	mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES;
 555	mq->qlines = qlines;
 556	mq->hstatus[0] = 0;
 557	mq->hstatus[1] = 1;
 558	mq->head = gru_mesq_head(2, qlines / 2 + 1);
 559	mqd->mq = mq;
 560	mqd->mq_gpa = uv_gpa(mq);
 561	mqd->qlines = qlines;
 562	mqd->interrupt_pnode = nasid >> 1;
 563	mqd->interrupt_vector = vector;
 564	mqd->interrupt_apicid = apicid;
 565	return 0;
 566}
 567EXPORT_SYMBOL_GPL(gru_create_message_queue);
 568
 569/*
 570 * Send a NOOP message to a message queue
 571 * 	Returns:
 572 * 		 0 - if queue is full after the send. This is the normal case
 573 * 		     but various races can change this.
 574 *		-1 - if mesq sent successfully but queue not full
 575 *		>0 - unexpected error. MQE_xxx returned
 576 */
 577static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd,
 578				void *mesg)
 579{
 580	const struct message_header noop_header = {
 581					.present = MQS_NOOP, .lines = 1};
 582	unsigned long m;
 583	int substatus, ret;
 584	struct message_header save_mhdr, *mhdr = mesg;
 585
 586	STAT(mesq_noop);
 587	save_mhdr = *mhdr;
 588	*mhdr = noop_header;
 589	gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA);
 590	ret = gru_wait(cb);
 591
 592	if (ret) {
 593		substatus = gru_get_cb_message_queue_substatus(cb);
 594		switch (substatus) {
 595		case CBSS_NO_ERROR:
 596			STAT(mesq_noop_unexpected_error);
 597			ret = MQE_UNEXPECTED_CB_ERR;
 598			break;
 599		case CBSS_LB_OVERFLOWED:
 600			STAT(mesq_noop_lb_overflow);
 601			ret = MQE_CONGESTION;
 602			break;
 603		case CBSS_QLIMIT_REACHED:
 604			STAT(mesq_noop_qlimit_reached);
 605			ret = 0;
 606			break;
 607		case CBSS_AMO_NACKED:
 608			STAT(mesq_noop_amo_nacked);
 609			ret = MQE_CONGESTION;
 610			break;
 611		case CBSS_PUT_NACKED:
 612			STAT(mesq_noop_put_nacked);
 613			m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
 614			gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1,
 615						IMA);
 616			if (gru_wait(cb) == CBS_IDLE)
 617				ret = MQIE_AGAIN;
 618			else
 619				ret = MQE_UNEXPECTED_CB_ERR;
 620			break;
 621		case CBSS_PAGE_OVERFLOW:
 622			STAT(mesq_noop_page_overflow);
 623			fallthrough;
 624		default:
 625			BUG();
 626		}
 627	}
 628	*mhdr = save_mhdr;
 629	return ret;
 630}
 631
 632/*
 633 * Handle a gru_mesq full.
 634 */
 635static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd,
 636				void *mesg, int lines)
 637{
 638	union gru_mesqhead mqh;
 639	unsigned int limit, head;
 640	unsigned long avalue;
 641	int half, qlines;
 642
 643	/* Determine if switching to first/second half of q */
 644	avalue = gru_get_amo_value(cb);
 645	head = gru_get_amo_value_head(cb);
 646	limit = gru_get_amo_value_limit(cb);
 647
 648	qlines = mqd->qlines;
 649	half = (limit != qlines);
 650
 651	if (half)
 652		mqh = gru_mesq_head(qlines / 2 + 1, qlines);
 653	else
 654		mqh = gru_mesq_head(2, qlines / 2 + 1);
 655
 656	/* Try to get lock for switching head pointer */
 657	gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA);
 658	if (gru_wait(cb) != CBS_IDLE)
 659		goto cberr;
 660	if (!gru_get_amo_value(cb)) {
 661		STAT(mesq_qf_locked);
 662		return MQE_QUEUE_FULL;
 663	}
 664
 665	/* Got the lock. Send optional NOP if queue not full, */
 666	if (head != limit) {
 667		if (send_noop_message(cb, mqd, mesg)) {
 668			gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half),
 669					XTYPE_DW, IMA);
 670			if (gru_wait(cb) != CBS_IDLE)
 671				goto cberr;
 672			STAT(mesq_qf_noop_not_full);
 673			return MQIE_AGAIN;
 674		}
 675		avalue++;
 676	}
 677
 678	/* Then flip queuehead to other half of queue. */
 679	gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue,
 680							IMA);
 681	if (gru_wait(cb) != CBS_IDLE)
 682		goto cberr;
 683
 684	/* If not successfully in swapping queue head, clear the hstatus lock */
 685	if (gru_get_amo_value(cb) != avalue) {
 686		STAT(mesq_qf_switch_head_failed);
 687		gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW,
 688							IMA);
 689		if (gru_wait(cb) != CBS_IDLE)
 690			goto cberr;
 691	}
 692	return MQIE_AGAIN;
 693cberr:
 694	STAT(mesq_qf_unexpected_error);
 695	return MQE_UNEXPECTED_CB_ERR;
 696}
 697
 698/*
 699 * Handle a PUT failure. Note: if message was a 2-line message, one of the
 700 * lines might have successfully have been written. Before sending the
 701 * message, "present" must be cleared in BOTH lines to prevent the receiver
 702 * from prematurely seeing the full message.
 703 */
 704static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd,
 705			void *mesg, int lines)
 706{
 707	unsigned long m;
 708	int ret, loops = 200;	/* experimentally determined */
 709
 710	m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6);
 711	if (lines == 2) {
 712		gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA);
 713		if (gru_wait(cb) != CBS_IDLE)
 714			return MQE_UNEXPECTED_CB_ERR;
 715	}
 716	gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA);
 717	if (gru_wait(cb) != CBS_IDLE)
 718		return MQE_UNEXPECTED_CB_ERR;
 719
 720	if (!mqd->interrupt_vector)
 721		return MQE_OK;
 722
 723	/*
 724	 * Send a noop message in order to deliver a cross-partition interrupt
 725	 * to the SSI that contains the target message queue. Normally, the
 726	 * interrupt is automatically delivered by hardware following mesq
 727	 * operations, but some error conditions require explicit delivery.
 728	 * The noop message will trigger delivery. Otherwise partition failures
 729	 * could cause unrecovered errors.
 730	 */
 731	do {
 732		ret = send_noop_message(cb, mqd, mesg);
 733	} while ((ret == MQIE_AGAIN || ret == MQE_CONGESTION) && (loops-- > 0));
 734
 735	if (ret == MQIE_AGAIN || ret == MQE_CONGESTION) {
 736		/*
 737		 * Don't indicate to the app to resend the message, as it's
 738		 * already been successfully sent.  We simply send an OK
 739		 * (rather than fail the send with MQE_UNEXPECTED_CB_ERR),
 740		 * assuming that the other side is receiving enough
 741		 * interrupts to get this message processed anyway.
 742		 */
 743		ret = MQE_OK;
 744	}
 745	return ret;
 746}
 747
 748/*
 749 * Handle a gru_mesq failure. Some of these failures are software recoverable
 750 * or retryable.
 751 */
 752static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd,
 753				void *mesg, int lines)
 754{
 755	int substatus, ret = 0;
 756
 757	substatus = gru_get_cb_message_queue_substatus(cb);
 758	switch (substatus) {
 759	case CBSS_NO_ERROR:
 760		STAT(mesq_send_unexpected_error);
 761		ret = MQE_UNEXPECTED_CB_ERR;
 762		break;
 763	case CBSS_LB_OVERFLOWED:
 764		STAT(mesq_send_lb_overflow);
 765		ret = MQE_CONGESTION;
 766		break;
 767	case CBSS_QLIMIT_REACHED:
 768		STAT(mesq_send_qlimit_reached);
 769		ret = send_message_queue_full(cb, mqd, mesg, lines);
 770		break;
 771	case CBSS_AMO_NACKED:
 772		STAT(mesq_send_amo_nacked);
 773		ret = MQE_CONGESTION;
 774		break;
 775	case CBSS_PUT_NACKED:
 776		STAT(mesq_send_put_nacked);
 777		ret = send_message_put_nacked(cb, mqd, mesg, lines);
 778		break;
 779	case CBSS_PAGE_OVERFLOW:
 780		STAT(mesq_page_overflow);
 781		fallthrough;
 782	default:
 783		BUG();
 784	}
 785	return ret;
 786}
 787
 788/*
 789 * Send a message to a message queue
 790 * 	mqd	message queue descriptor
 791 * 	mesg	message. ust be vaddr within a GSEG
 792 * 	bytes	message size (<= 2 CL)
 793 */
 794int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg,
 795				unsigned int bytes)
 796{
 797	struct message_header *mhdr;
 798	void *cb;
 799	void *dsr;
 800	int istatus, clines, ret;
 801
 802	STAT(mesq_send);
 803	BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES);
 804
 805	clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES);
 806	if (gru_get_cpu_resources(bytes, &cb, &dsr))
 807		return MQE_BUG_NO_RESOURCES;
 808	memcpy(dsr, mesg, bytes);
 809	mhdr = dsr;
 810	mhdr->present = MQS_FULL;
 811	mhdr->lines = clines;
 812	if (clines == 2) {
 813		mhdr->present2 = get_present2(mhdr);
 814		restore_present2(mhdr, MQS_FULL);
 815	}
 816
 817	do {
 818		ret = MQE_OK;
 819		gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA);
 820		istatus = gru_wait(cb);
 821		if (istatus != CBS_IDLE)
 822			ret = send_message_failure(cb, mqd, dsr, clines);
 823	} while (ret == MQIE_AGAIN);
 824	gru_free_cpu_resources(cb, dsr);
 825
 826	if (ret)
 827		STAT(mesq_send_failed);
 828	return ret;
 829}
 830EXPORT_SYMBOL_GPL(gru_send_message_gpa);
 831
 832/*
 833 * Advance the receive pointer for the queue to the next message.
 834 */
 835void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg)
 836{
 837	struct message_queue *mq = mqd->mq;
 838	struct message_header *mhdr = mq->next;
 839	void *next, *pnext;
 840	int half = -1;
 841	int lines = mhdr->lines;
 842
 843	if (lines == 2)
 844		restore_present2(mhdr, MQS_EMPTY);
 845	mhdr->present = MQS_EMPTY;
 846
 847	pnext = mq->next;
 848	next = pnext + GRU_CACHE_LINE_BYTES * lines;
 849	if (next == mq->limit) {
 850		next = mq->start;
 851		half = 1;
 852	} else if (pnext < mq->start2 && next >= mq->start2) {
 853		half = 0;
 854	}
 855
 856	if (half >= 0)
 857		mq->hstatus[half] = 1;
 858	mq->next = next;
 859}
 860EXPORT_SYMBOL_GPL(gru_free_message);
 861
 862/*
 863 * Get next message from message queue. Return NULL if no message
 864 * present. User must call next_message() to move to next message.
 865 * 	rmq	message queue
 866 */
 867void *gru_get_next_message(struct gru_message_queue_desc *mqd)
 868{
 869	struct message_queue *mq = mqd->mq;
 870	struct message_header *mhdr = mq->next;
 871	int present = mhdr->present;
 872
 873	/* skip NOOP messages */
 874	while (present == MQS_NOOP) {
 875		gru_free_message(mqd, mhdr);
 876		mhdr = mq->next;
 877		present = mhdr->present;
 878	}
 879
 880	/* Wait for both halves of 2 line messages */
 881	if (present == MQS_FULL && mhdr->lines == 2 &&
 882				get_present2(mhdr) == MQS_EMPTY)
 883		present = MQS_EMPTY;
 884
 885	if (!present) {
 886		STAT(mesq_receive_none);
 887		return NULL;
 888	}
 889
 890	if (mhdr->lines == 2)
 891		restore_present2(mhdr, mhdr->present2);
 892
 893	STAT(mesq_receive);
 894	return mhdr;
 895}
 896EXPORT_SYMBOL_GPL(gru_get_next_message);
 897
 898/* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/
 899
 900/*
 901 * Load a DW from a global GPA. The GPA can be a memory or MMR address.
 902 */
 903int gru_read_gpa(unsigned long *value, unsigned long gpa)
 904{
 905	void *cb;
 906	void *dsr;
 907	int ret, iaa;
 908
 909	STAT(read_gpa);
 910	if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
 911		return MQE_BUG_NO_RESOURCES;
 912	iaa = gpa >> 62;
 913	gru_vload_phys(cb, gpa, gru_get_tri(dsr), iaa, IMA);
 914	ret = gru_wait(cb);
 915	if (ret == CBS_IDLE)
 916		*value = *(unsigned long *)dsr;
 917	gru_free_cpu_resources(cb, dsr);
 918	return ret;
 919}
 920EXPORT_SYMBOL_GPL(gru_read_gpa);
 921
 922
 923/*
 924 * Copy a block of data using the GRU resources
 925 */
 926int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa,
 927				unsigned int bytes)
 928{
 929	void *cb;
 930	void *dsr;
 931	int ret;
 932
 933	STAT(copy_gpa);
 934	if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr))
 935		return MQE_BUG_NO_RESOURCES;
 936	gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr),
 937		  XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA);
 938	ret = gru_wait(cb);
 939	gru_free_cpu_resources(cb, dsr);
 940	return ret;
 941}
 942EXPORT_SYMBOL_GPL(gru_copy_gpa);
 943
 944/* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/
 945/* 	Temp - will delete after we gain confidence in the GRU		*/
 946
 947static int quicktest0(unsigned long arg)
 948{
 949	unsigned long word0;
 950	unsigned long word1;
 951	void *cb;
 952	void *dsr;
 953	unsigned long *p;
 954	int ret = -EIO;
 955
 956	if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr))
 957		return MQE_BUG_NO_RESOURCES;
 958	p = dsr;
 959	word0 = MAGIC;
 960	word1 = 0;
 961
 962	gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
 963	if (gru_wait(cb) != CBS_IDLE) {
 964		printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 1\n", smp_processor_id());
 965		goto done;
 966	}
 967
 968	if (*p != MAGIC) {
 969		printk(KERN_DEBUG "GRU:%d quicktest0 bad magic 0x%lx\n", smp_processor_id(), *p);
 970		goto done;
 971	}
 972	gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA);
 973	if (gru_wait(cb) != CBS_IDLE) {
 974		printk(KERN_DEBUG "GRU:%d quicktest0: CBR failure 2\n", smp_processor_id());
 975		goto done;
 976	}
 977
 978	if (word0 != word1 || word1 != MAGIC) {
 979		printk(KERN_DEBUG
 980		       "GRU:%d quicktest0 err: found 0x%lx, expected 0x%lx\n",
 981		     smp_processor_id(), word1, MAGIC);
 982		goto done;
 983	}
 984	ret = 0;
 985
 986done:
 987	gru_free_cpu_resources(cb, dsr);
 988	return ret;
 989}
 990
 991#define ALIGNUP(p, q)	((void *)(((unsigned long)(p) + (q) - 1) & ~(q - 1)))
 992
 993static int quicktest1(unsigned long arg)
 994{
 995	struct gru_message_queue_desc mqd;
 996	void *p, *mq;
 997	int i, ret = -EIO;
 998	char mes[GRU_CACHE_LINE_BYTES], *m;
 999
1000	/* Need  1K cacheline aligned that does not cross page boundary */
1001	p = kmalloc(4096, 0);
1002	if (p == NULL)
1003		return -ENOMEM;
1004	mq = ALIGNUP(p, 1024);
1005	memset(mes, 0xee, sizeof(mes));
1006
1007	gru_create_message_queue(&mqd, mq, 8 * GRU_CACHE_LINE_BYTES, 0, 0, 0);
1008	for (i = 0; i < 6; i++) {
1009		mes[8] = i;
1010		do {
1011			ret = gru_send_message_gpa(&mqd, mes, sizeof(mes));
1012		} while (ret == MQE_CONGESTION);
1013		if (ret)
1014			break;
1015	}
1016	if (ret != MQE_QUEUE_FULL || i != 4) {
1017		printk(KERN_DEBUG "GRU:%d quicktest1: unexpected status %d, i %d\n",
1018		       smp_processor_id(), ret, i);
1019		goto done;
1020	}
1021
1022	for (i = 0; i < 6; i++) {
1023		m = gru_get_next_message(&mqd);
1024		if (!m || m[8] != i)
1025			break;
1026		gru_free_message(&mqd, m);
1027	}
1028	if (i != 4) {
1029		printk(KERN_DEBUG "GRU:%d quicktest2: bad message, i %d, m %p, m8 %d\n",
1030			smp_processor_id(), i, m, m ? m[8] : -1);
1031		goto done;
1032	}
1033	ret = 0;
1034
1035done:
1036	kfree(p);
1037	return ret;
1038}
1039
1040static int quicktest2(unsigned long arg)
1041{
1042	static DECLARE_COMPLETION(cmp);
1043	unsigned long han;
1044	int blade_id = 0;
1045	int numcb = 4;
1046	int ret = 0;
1047	unsigned long *buf;
1048	void *cb0, *cb;
1049	struct gru_control_block_status *gen;
1050	int i, k, istatus, bytes;
1051
1052	bytes = numcb * 4 * 8;
1053	buf = kmalloc(bytes, GFP_KERNEL);
1054	if (!buf)
1055		return -ENOMEM;
1056
1057	ret = -EBUSY;
1058	han = gru_reserve_async_resources(blade_id, numcb, 0, &cmp);
1059	if (!han)
1060		goto done;
1061
1062	gru_lock_async_resource(han, &cb0, NULL);
1063	memset(buf, 0xee, bytes);
1064	for (i = 0; i < numcb; i++)
1065		gru_vset(cb0 + i * GRU_HANDLE_STRIDE, uv_gpa(&buf[i * 4]), 0,
1066				XTYPE_DW, 4, 1, IMA_INTERRUPT);
1067
1068	ret = 0;
1069	k = numcb;
1070	do {
1071		gru_wait_async_cbr(han);
1072		for (i = 0; i < numcb; i++) {
1073			cb = cb0 + i * GRU_HANDLE_STRIDE;
1074			istatus = gru_check_status(cb);
1075			if (istatus != CBS_ACTIVE && istatus != CBS_CALL_OS)
1076				break;
1077		}
1078		if (i == numcb)
1079			continue;
1080		if (istatus != CBS_IDLE) {
1081			printk(KERN_DEBUG "GRU:%d quicktest2: cb %d, exception\n", smp_processor_id(), i);
1082			ret = -EFAULT;
1083		} else if (buf[4 * i] || buf[4 * i + 1] || buf[4 * i + 2] ||
1084				buf[4 * i + 3]) {
1085			printk(KERN_DEBUG "GRU:%d quicktest2:cb %d,  buf 0x%lx, 0x%lx, 0x%lx, 0x%lx\n",
1086			       smp_processor_id(), i, buf[4 * i], buf[4 * i + 1], buf[4 * i + 2], buf[4 * i + 3]);
1087			ret = -EIO;
1088		}
1089		k--;
1090		gen = cb;
1091		gen->istatus = CBS_CALL_OS; /* don't handle this CBR again */
1092	} while (k);
1093	BUG_ON(cmp.done);
1094
1095	gru_unlock_async_resource(han);
1096	gru_release_async_resources(han);
1097done:
1098	kfree(buf);
1099	return ret;
1100}
1101
1102#define BUFSIZE 200
1103static int quicktest3(unsigned long arg)
1104{
1105	char buf1[BUFSIZE], buf2[BUFSIZE];
1106	int ret = 0;
1107
1108	memset(buf2, 0, sizeof(buf2));
1109	memset(buf1, get_cycles() & 255, sizeof(buf1));
1110	gru_copy_gpa(uv_gpa(buf2), uv_gpa(buf1), BUFSIZE);
1111	if (memcmp(buf1, buf2, BUFSIZE)) {
1112		printk(KERN_DEBUG "GRU:%d quicktest3 error\n", smp_processor_id());
1113		ret = -EIO;
1114	}
1115	return ret;
1116}
1117
1118/*
1119 * Debugging only. User hook for various kernel tests
1120 * of driver & gru.
1121 */
1122int gru_ktest(unsigned long arg)
1123{
1124	int ret = -EINVAL;
1125
1126	switch (arg & 0xff) {
1127	case 0:
1128		ret = quicktest0(arg);
1129		break;
1130	case 1:
1131		ret = quicktest1(arg);
1132		break;
1133	case 2:
1134		ret = quicktest2(arg);
1135		break;
1136	case 3:
1137		ret = quicktest3(arg);
1138		break;
1139	case 99:
1140		ret = gru_free_kernel_contexts();
1141		break;
1142	}
1143	return ret;
1144
1145}
1146
1147int gru_kservices_init(void)
1148{
1149	return 0;
1150}
1151
1152void gru_kservices_exit(void)
1153{
1154	if (gru_free_kernel_contexts())
1155		BUG();
1156}
1157