Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame^] | 1 | /* |
| 2 | * SN Platform GRU Driver |
| 3 | * |
| 4 | * KERNEL SERVICES THAT USE THE GRU |
| 5 | * |
| 6 | * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved. |
| 7 | * |
| 8 | * This program is free software; you can redistribute it and/or modify |
| 9 | * it under the terms of the GNU General Public License as published by |
| 10 | * the Free Software Foundation; either version 2 of the License, or |
| 11 | * (at your option) any later version. |
| 12 | * |
| 13 | * This program is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | * GNU General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU General Public License |
| 19 | * along with this program; if not, write to the Free Software |
| 20 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 21 | */ |
| 22 | |
| 23 | #include <linux/kernel.h> |
| 24 | #include <linux/errno.h> |
| 25 | #include <linux/slab.h> |
| 26 | #include <linux/mm.h> |
| 27 | #include <linux/smp_lock.h> |
| 28 | #include <linux/spinlock.h> |
| 29 | #include <linux/device.h> |
| 30 | #include <linux/miscdevice.h> |
| 31 | #include <linux/proc_fs.h> |
| 32 | #include <linux/interrupt.h> |
| 33 | #include <linux/uaccess.h> |
| 34 | #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 | * |
| 48 | * At boot time, the kernel permanently reserves a fixed number of |
| 49 | * CBRs/DSRs for each cpu to use. The resources are all taken from |
| 50 | * the GRU chiplet 1 on the blade. This leaves the full set of resources |
| 51 | * of chiplet 0 available to be allocated to a single user. |
| 52 | */ |
| 53 | |
| 54 | /* Blade percpu resources PERMANENTLY reserved for kernel use */ |
| 55 | #define GRU_NUM_KERNEL_CBR 1 |
| 56 | #define GRU_NUM_KERNEL_DSR_BYTES 256 |
| 57 | #define KERNEL_CTXNUM 15 |
| 58 | |
| 59 | /* GRU instruction attributes for all instructions */ |
| 60 | #define IMA IMA_CB_DELAY |
| 61 | |
| 62 | /* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */ |
| 63 | #define __gru_cacheline_aligned__ \ |
| 64 | __attribute__((__aligned__(GRU_CACHE_LINE_BYTES))) |
| 65 | |
| 66 | #define MAGIC 0x1234567887654321UL |
| 67 | |
| 68 | /* Default retry count for GRU errors on kernel instructions */ |
| 69 | #define EXCEPTION_RETRY_LIMIT 3 |
| 70 | |
| 71 | /* Status of message queue sections */ |
| 72 | #define MQS_EMPTY 0 |
| 73 | #define MQS_FULL 1 |
| 74 | #define MQS_NOOP 2 |
| 75 | |
| 76 | /*----------------- RESOURCE MANAGEMENT -------------------------------------*/ |
| 77 | /* optimized for x86_64 */ |
| 78 | struct message_queue { |
| 79 | union gru_mesqhead head __gru_cacheline_aligned__; /* CL 0 */ |
| 80 | int qlines; /* DW 1 */ |
| 81 | long hstatus[2]; |
| 82 | void *next __gru_cacheline_aligned__;/* CL 1 */ |
| 83 | void *limit; |
| 84 | void *start; |
| 85 | void *start2; |
| 86 | char data ____cacheline_aligned; /* CL 2 */ |
| 87 | }; |
| 88 | |
| 89 | /* First word in every message - used by mesq interface */ |
| 90 | struct message_header { |
| 91 | char present; |
| 92 | char present2; |
| 93 | char lines; |
| 94 | char fill; |
| 95 | }; |
| 96 | |
| 97 | #define QLINES(mq) ((mq) + offsetof(struct message_queue, qlines)) |
| 98 | #define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h])) |
| 99 | |
| 100 | static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr) |
| 101 | { |
| 102 | struct gru_blade_state *bs; |
| 103 | int lcpu; |
| 104 | |
| 105 | BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES); |
| 106 | preempt_disable(); |
| 107 | bs = gru_base[uv_numa_blade_id()]; |
| 108 | lcpu = uv_blade_processor_id(); |
| 109 | *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE; |
| 110 | *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES; |
| 111 | return 0; |
| 112 | } |
| 113 | |
| 114 | static void gru_free_cpu_resources(void *cb, void *dsr) |
| 115 | { |
| 116 | preempt_enable(); |
| 117 | } |
| 118 | |
| 119 | int gru_get_cb_exception_detail(void *cb, |
| 120 | struct control_block_extended_exc_detail *excdet) |
| 121 | { |
| 122 | struct gru_control_block_extended *cbe; |
| 123 | |
| 124 | cbe = get_cbe(GRUBASE(cb), get_cb_number(cb)); |
| 125 | excdet->opc = cbe->opccpy; |
| 126 | excdet->exopc = cbe->exopccpy; |
| 127 | excdet->ecause = cbe->ecause; |
| 128 | excdet->exceptdet0 = cbe->idef1upd; |
| 129 | excdet->exceptdet1 = cbe->idef3upd; |
| 130 | return 0; |
| 131 | } |
| 132 | |
| 133 | char *gru_get_cb_exception_detail_str(int ret, void *cb, |
| 134 | char *buf, int size) |
| 135 | { |
| 136 | struct gru_control_block_status *gen = (void *)cb; |
| 137 | struct control_block_extended_exc_detail excdet; |
| 138 | |
| 139 | if (ret > 0 && gen->istatus == CBS_EXCEPTION) { |
| 140 | gru_get_cb_exception_detail(cb, &excdet); |
| 141 | snprintf(buf, size, |
| 142 | "GRU exception: cb %p, opc %d, exopc %d, ecause 0x%x," |
| 143 | "excdet0 0x%lx, excdet1 0x%x", |
| 144 | gen, excdet.opc, excdet.exopc, excdet.ecause, |
| 145 | excdet.exceptdet0, excdet.exceptdet1); |
| 146 | } else { |
| 147 | snprintf(buf, size, "No exception"); |
| 148 | } |
| 149 | return buf; |
| 150 | } |
| 151 | |
| 152 | static int gru_wait_idle_or_exception(struct gru_control_block_status *gen) |
| 153 | { |
| 154 | while (gen->istatus >= CBS_ACTIVE) { |
| 155 | cpu_relax(); |
| 156 | barrier(); |
| 157 | } |
| 158 | return gen->istatus; |
| 159 | } |
| 160 | |
| 161 | static int gru_retry_exception(void *cb) |
| 162 | { |
| 163 | struct gru_control_block_status *gen = (void *)cb; |
| 164 | struct control_block_extended_exc_detail excdet; |
| 165 | int retry = EXCEPTION_RETRY_LIMIT; |
| 166 | |
| 167 | while (1) { |
| 168 | if (gru_get_cb_message_queue_substatus(cb)) |
| 169 | break; |
| 170 | if (gru_wait_idle_or_exception(gen) == CBS_IDLE) |
| 171 | return CBS_IDLE; |
| 172 | |
| 173 | gru_get_cb_exception_detail(cb, &excdet); |
| 174 | if (excdet.ecause & ~EXCEPTION_RETRY_BITS) |
| 175 | break; |
| 176 | if (retry-- == 0) |
| 177 | break; |
| 178 | gen->icmd = 1; |
| 179 | gru_flush_cache(gen); |
| 180 | } |
| 181 | return CBS_EXCEPTION; |
| 182 | } |
| 183 | |
| 184 | int gru_check_status_proc(void *cb) |
| 185 | { |
| 186 | struct gru_control_block_status *gen = (void *)cb; |
| 187 | int ret; |
| 188 | |
| 189 | ret = gen->istatus; |
| 190 | if (ret != CBS_EXCEPTION) |
| 191 | return ret; |
| 192 | return gru_retry_exception(cb); |
| 193 | |
| 194 | } |
| 195 | |
| 196 | int gru_wait_proc(void *cb) |
| 197 | { |
| 198 | struct gru_control_block_status *gen = (void *)cb; |
| 199 | int ret; |
| 200 | |
| 201 | ret = gru_wait_idle_or_exception(gen); |
| 202 | if (ret == CBS_EXCEPTION) |
| 203 | ret = gru_retry_exception(cb); |
| 204 | |
| 205 | return ret; |
| 206 | } |
| 207 | |
| 208 | void gru_abort(int ret, void *cb, char *str) |
| 209 | { |
| 210 | char buf[GRU_EXC_STR_SIZE]; |
| 211 | |
| 212 | panic("GRU FATAL ERROR: %s - %s\n", str, |
| 213 | gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf))); |
| 214 | } |
| 215 | |
| 216 | void gru_wait_abort_proc(void *cb) |
| 217 | { |
| 218 | int ret; |
| 219 | |
| 220 | ret = gru_wait_proc(cb); |
| 221 | if (ret) |
| 222 | gru_abort(ret, cb, "gru_wait_abort"); |
| 223 | } |
| 224 | |
| 225 | |
| 226 | /*------------------------------ MESSAGE QUEUES -----------------------------*/ |
| 227 | |
| 228 | /* Internal status . These are NOT returned to the user. */ |
| 229 | #define MQIE_AGAIN -1 /* try again */ |
| 230 | |
| 231 | |
| 232 | /* |
| 233 | * Save/restore the "present" flag that is in the second line of 2-line |
| 234 | * messages |
| 235 | */ |
| 236 | static inline int get_present2(void *p) |
| 237 | { |
| 238 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; |
| 239 | return mhdr->present; |
| 240 | } |
| 241 | |
| 242 | static inline void restore_present2(void *p, int val) |
| 243 | { |
| 244 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; |
| 245 | mhdr->present = val; |
| 246 | } |
| 247 | |
| 248 | /* |
| 249 | * Create a message queue. |
| 250 | * qlines - message queue size in cache lines. Includes 2-line header. |
| 251 | */ |
| 252 | int gru_create_message_queue(void *p, unsigned int bytes) |
| 253 | { |
| 254 | struct message_queue *mq = p; |
| 255 | unsigned int qlines; |
| 256 | |
| 257 | qlines = bytes / GRU_CACHE_LINE_BYTES - 2; |
| 258 | memset(mq, 0, bytes); |
| 259 | mq->start = &mq->data; |
| 260 | mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES; |
| 261 | mq->next = &mq->data; |
| 262 | mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES; |
| 263 | mq->qlines = qlines; |
| 264 | mq->hstatus[0] = 0; |
| 265 | mq->hstatus[1] = 1; |
| 266 | mq->head = gru_mesq_head(2, qlines / 2 + 1); |
| 267 | return 0; |
| 268 | } |
| 269 | EXPORT_SYMBOL_GPL(gru_create_message_queue); |
| 270 | |
| 271 | /* |
| 272 | * Send a NOOP message to a message queue |
| 273 | * Returns: |
| 274 | * 0 - if queue is full after the send. This is the normal case |
| 275 | * but various races can change this. |
| 276 | * -1 - if mesq sent successfully but queue not full |
| 277 | * >0 - unexpected error. MQE_xxx returned |
| 278 | */ |
| 279 | static int send_noop_message(void *cb, |
| 280 | unsigned long mq, void *mesg) |
| 281 | { |
| 282 | const struct message_header noop_header = { |
| 283 | .present = MQS_NOOP, .lines = 1}; |
| 284 | unsigned long m; |
| 285 | int substatus, ret; |
| 286 | struct message_header save_mhdr, *mhdr = mesg; |
| 287 | |
| 288 | STAT(mesq_noop); |
| 289 | save_mhdr = *mhdr; |
| 290 | *mhdr = noop_header; |
| 291 | gru_mesq(cb, mq, gru_get_tri(mhdr), 1, IMA); |
| 292 | ret = gru_wait(cb); |
| 293 | |
| 294 | if (ret) { |
| 295 | substatus = gru_get_cb_message_queue_substatus(cb); |
| 296 | switch (substatus) { |
| 297 | case CBSS_NO_ERROR: |
| 298 | STAT(mesq_noop_unexpected_error); |
| 299 | ret = MQE_UNEXPECTED_CB_ERR; |
| 300 | break; |
| 301 | case CBSS_LB_OVERFLOWED: |
| 302 | STAT(mesq_noop_lb_overflow); |
| 303 | ret = MQE_CONGESTION; |
| 304 | break; |
| 305 | case CBSS_QLIMIT_REACHED: |
| 306 | STAT(mesq_noop_qlimit_reached); |
| 307 | ret = 0; |
| 308 | break; |
| 309 | case CBSS_AMO_NACKED: |
| 310 | STAT(mesq_noop_amo_nacked); |
| 311 | ret = MQE_CONGESTION; |
| 312 | break; |
| 313 | case CBSS_PUT_NACKED: |
| 314 | STAT(mesq_noop_put_nacked); |
| 315 | m = mq + (gru_get_amo_value_head(cb) << 6); |
| 316 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1, |
| 317 | IMA); |
| 318 | if (gru_wait(cb) == CBS_IDLE) |
| 319 | ret = MQIE_AGAIN; |
| 320 | else |
| 321 | ret = MQE_UNEXPECTED_CB_ERR; |
| 322 | break; |
| 323 | case CBSS_PAGE_OVERFLOW: |
| 324 | default: |
| 325 | BUG(); |
| 326 | } |
| 327 | } |
| 328 | *mhdr = save_mhdr; |
| 329 | return ret; |
| 330 | } |
| 331 | |
| 332 | /* |
| 333 | * Handle a gru_mesq full. |
| 334 | */ |
| 335 | static int send_message_queue_full(void *cb, |
| 336 | unsigned long mq, void *mesg, int lines) |
| 337 | { |
| 338 | union gru_mesqhead mqh; |
| 339 | unsigned int limit, head; |
| 340 | unsigned long avalue; |
| 341 | int half, qlines, save; |
| 342 | |
| 343 | /* Determine if switching to first/second half of q */ |
| 344 | avalue = gru_get_amo_value(cb); |
| 345 | head = gru_get_amo_value_head(cb); |
| 346 | limit = gru_get_amo_value_limit(cb); |
| 347 | |
| 348 | /* |
| 349 | * Fetch "qlines" from the queue header. Since the queue may be |
| 350 | * in memory that can't be accessed using socket addresses, use |
| 351 | * the GRU to access the data. Use DSR space from the message. |
| 352 | */ |
| 353 | save = *(int *)mesg; |
| 354 | gru_vload(cb, QLINES(mq), gru_get_tri(mesg), XTYPE_W, 1, 1, IMA); |
| 355 | if (gru_wait(cb) != CBS_IDLE) |
| 356 | goto cberr; |
| 357 | qlines = *(int *)mesg; |
| 358 | *(int *)mesg = save; |
| 359 | half = (limit != qlines); |
| 360 | |
| 361 | if (half) |
| 362 | mqh = gru_mesq_head(qlines / 2 + 1, qlines); |
| 363 | else |
| 364 | mqh = gru_mesq_head(2, qlines / 2 + 1); |
| 365 | |
| 366 | /* Try to get lock for switching head pointer */ |
| 367 | gru_gamir(cb, EOP_IR_CLR, HSTATUS(mq, half), XTYPE_DW, IMA); |
| 368 | if (gru_wait(cb) != CBS_IDLE) |
| 369 | goto cberr; |
| 370 | if (!gru_get_amo_value(cb)) { |
| 371 | STAT(mesq_qf_locked); |
| 372 | return MQE_QUEUE_FULL; |
| 373 | } |
| 374 | |
| 375 | /* Got the lock. Send optional NOP if queue not full, */ |
| 376 | if (head != limit) { |
| 377 | if (send_noop_message(cb, mq, mesg)) { |
| 378 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mq, half), |
| 379 | XTYPE_DW, IMA); |
| 380 | if (gru_wait(cb) != CBS_IDLE) |
| 381 | goto cberr; |
| 382 | STAT(mesq_qf_noop_not_full); |
| 383 | return MQIE_AGAIN; |
| 384 | } |
| 385 | avalue++; |
| 386 | } |
| 387 | |
| 388 | /* Then flip queuehead to other half of queue. */ |
| 389 | gru_gamer(cb, EOP_ERR_CSWAP, mq, XTYPE_DW, mqh.val, avalue, IMA); |
| 390 | if (gru_wait(cb) != CBS_IDLE) |
| 391 | goto cberr; |
| 392 | |
| 393 | /* If not successfully in swapping queue head, clear the hstatus lock */ |
| 394 | if (gru_get_amo_value(cb) != avalue) { |
| 395 | STAT(mesq_qf_switch_head_failed); |
| 396 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mq, half), XTYPE_DW, IMA); |
| 397 | if (gru_wait(cb) != CBS_IDLE) |
| 398 | goto cberr; |
| 399 | } |
| 400 | return MQIE_AGAIN; |
| 401 | cberr: |
| 402 | STAT(mesq_qf_unexpected_error); |
| 403 | return MQE_UNEXPECTED_CB_ERR; |
| 404 | } |
| 405 | |
| 406 | |
| 407 | /* |
| 408 | * Handle a gru_mesq failure. Some of these failures are software recoverable |
| 409 | * or retryable. |
| 410 | */ |
| 411 | static int send_message_failure(void *cb, |
| 412 | unsigned long mq, |
| 413 | void *mesg, |
| 414 | int lines) |
| 415 | { |
| 416 | int substatus, ret = 0; |
| 417 | unsigned long m; |
| 418 | |
| 419 | substatus = gru_get_cb_message_queue_substatus(cb); |
| 420 | switch (substatus) { |
| 421 | case CBSS_NO_ERROR: |
| 422 | STAT(mesq_send_unexpected_error); |
| 423 | ret = MQE_UNEXPECTED_CB_ERR; |
| 424 | break; |
| 425 | case CBSS_LB_OVERFLOWED: |
| 426 | STAT(mesq_send_lb_overflow); |
| 427 | ret = MQE_CONGESTION; |
| 428 | break; |
| 429 | case CBSS_QLIMIT_REACHED: |
| 430 | STAT(mesq_send_qlimit_reached); |
| 431 | ret = send_message_queue_full(cb, mq, mesg, lines); |
| 432 | break; |
| 433 | case CBSS_AMO_NACKED: |
| 434 | STAT(mesq_send_amo_nacked); |
| 435 | ret = MQE_CONGESTION; |
| 436 | break; |
| 437 | case CBSS_PUT_NACKED: |
| 438 | STAT(mesq_send_put_nacked); |
| 439 | m =mq + (gru_get_amo_value_head(cb) << 6); |
| 440 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA); |
| 441 | if (gru_wait(cb) == CBS_IDLE) |
| 442 | ret = MQE_OK; |
| 443 | else |
| 444 | ret = MQE_UNEXPECTED_CB_ERR; |
| 445 | break; |
| 446 | default: |
| 447 | BUG(); |
| 448 | } |
| 449 | return ret; |
| 450 | } |
| 451 | |
| 452 | /* |
| 453 | * Send a message to a message queue |
| 454 | * cb GRU control block to use to send message |
| 455 | * mq message queue |
| 456 | * mesg message. ust be vaddr within a GSEG |
| 457 | * bytes message size (<= 2 CL) |
| 458 | */ |
| 459 | int gru_send_message_gpa(unsigned long mq, void *mesg, unsigned int bytes) |
| 460 | { |
| 461 | struct message_header *mhdr; |
| 462 | void *cb; |
| 463 | void *dsr; |
| 464 | int istatus, clines, ret; |
| 465 | |
| 466 | STAT(mesq_send); |
| 467 | BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES); |
| 468 | |
| 469 | clines = (bytes + GRU_CACHE_LINE_BYTES - 1) / GRU_CACHE_LINE_BYTES; |
| 470 | if (gru_get_cpu_resources(bytes, &cb, &dsr)) |
| 471 | return MQE_BUG_NO_RESOURCES; |
| 472 | memcpy(dsr, mesg, bytes); |
| 473 | mhdr = dsr; |
| 474 | mhdr->present = MQS_FULL; |
| 475 | mhdr->lines = clines; |
| 476 | if (clines == 2) { |
| 477 | mhdr->present2 = get_present2(mhdr); |
| 478 | restore_present2(mhdr, MQS_FULL); |
| 479 | } |
| 480 | |
| 481 | do { |
| 482 | ret = MQE_OK; |
| 483 | gru_mesq(cb, mq, gru_get_tri(mhdr), clines, IMA); |
| 484 | istatus = gru_wait(cb); |
| 485 | if (istatus != CBS_IDLE) |
| 486 | ret = send_message_failure(cb, mq, dsr, clines); |
| 487 | } while (ret == MQIE_AGAIN); |
| 488 | gru_free_cpu_resources(cb, dsr); |
| 489 | |
| 490 | if (ret) |
| 491 | STAT(mesq_send_failed); |
| 492 | return ret; |
| 493 | } |
| 494 | EXPORT_SYMBOL_GPL(gru_send_message_gpa); |
| 495 | |
| 496 | /* |
| 497 | * Advance the receive pointer for the queue to the next message. |
| 498 | */ |
| 499 | void gru_free_message(void *rmq, void *mesg) |
| 500 | { |
| 501 | struct message_queue *mq = rmq; |
| 502 | struct message_header *mhdr = mq->next; |
| 503 | void *next, *pnext; |
| 504 | int half = -1; |
| 505 | int lines = mhdr->lines; |
| 506 | |
| 507 | if (lines == 2) |
| 508 | restore_present2(mhdr, MQS_EMPTY); |
| 509 | mhdr->present = MQS_EMPTY; |
| 510 | |
| 511 | pnext = mq->next; |
| 512 | next = pnext + GRU_CACHE_LINE_BYTES * lines; |
| 513 | if (next == mq->limit) { |
| 514 | next = mq->start; |
| 515 | half = 1; |
| 516 | } else if (pnext < mq->start2 && next >= mq->start2) { |
| 517 | half = 0; |
| 518 | } |
| 519 | |
| 520 | if (half >= 0) |
| 521 | mq->hstatus[half] = 1; |
| 522 | mq->next = next; |
| 523 | } |
| 524 | EXPORT_SYMBOL_GPL(gru_free_message); |
| 525 | |
| 526 | /* |
| 527 | * Get next message from message queue. Return NULL if no message |
| 528 | * present. User must call next_message() to move to next message. |
| 529 | * rmq message queue |
| 530 | */ |
| 531 | void *gru_get_next_message(void *rmq) |
| 532 | { |
| 533 | struct message_queue *mq = rmq; |
| 534 | struct message_header *mhdr = mq->next; |
| 535 | int present = mhdr->present; |
| 536 | |
| 537 | /* skip NOOP messages */ |
| 538 | STAT(mesq_receive); |
| 539 | while (present == MQS_NOOP) { |
| 540 | gru_free_message(rmq, mhdr); |
| 541 | mhdr = mq->next; |
| 542 | present = mhdr->present; |
| 543 | } |
| 544 | |
| 545 | /* Wait for both halves of 2 line messages */ |
| 546 | if (present == MQS_FULL && mhdr->lines == 2 && |
| 547 | get_present2(mhdr) == MQS_EMPTY) |
| 548 | present = MQS_EMPTY; |
| 549 | |
| 550 | if (!present) { |
| 551 | STAT(mesq_receive_none); |
| 552 | return NULL; |
| 553 | } |
| 554 | |
| 555 | if (mhdr->lines == 2) |
| 556 | restore_present2(mhdr, mhdr->present2); |
| 557 | |
| 558 | return mhdr; |
| 559 | } |
| 560 | EXPORT_SYMBOL_GPL(gru_get_next_message); |
| 561 | |
| 562 | /* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/ |
| 563 | |
| 564 | /* |
| 565 | * Copy a block of data using the GRU resources |
| 566 | */ |
| 567 | int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa, |
| 568 | unsigned int bytes) |
| 569 | { |
| 570 | void *cb; |
| 571 | void *dsr; |
| 572 | int ret; |
| 573 | |
| 574 | STAT(copy_gpa); |
| 575 | if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr)) |
| 576 | return MQE_BUG_NO_RESOURCES; |
| 577 | gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr), |
| 578 | XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_BYTES, IMA); |
| 579 | ret = gru_wait(cb); |
| 580 | gru_free_cpu_resources(cb, dsr); |
| 581 | return ret; |
| 582 | } |
| 583 | EXPORT_SYMBOL_GPL(gru_copy_gpa); |
| 584 | |
| 585 | /* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/ |
| 586 | /* Temp - will delete after we gain confidence in the GRU */ |
| 587 | static __cacheline_aligned unsigned long word0; |
| 588 | static __cacheline_aligned unsigned long word1; |
| 589 | |
| 590 | static int quicktest(struct gru_state *gru) |
| 591 | { |
| 592 | void *cb; |
| 593 | void *ds; |
| 594 | unsigned long *p; |
| 595 | |
| 596 | cb = get_gseg_base_address_cb(gru->gs_gru_base_vaddr, KERNEL_CTXNUM, 0); |
| 597 | ds = get_gseg_base_address_ds(gru->gs_gru_base_vaddr, KERNEL_CTXNUM, 0); |
| 598 | p = ds; |
| 599 | word0 = MAGIC; |
| 600 | |
| 601 | gru_vload(cb, uv_gpa(&word0), 0, XTYPE_DW, 1, 1, IMA); |
| 602 | if (gru_wait(cb) != CBS_IDLE) |
| 603 | BUG(); |
| 604 | |
| 605 | if (*(unsigned long *)ds != MAGIC) |
| 606 | BUG(); |
| 607 | gru_vstore(cb, uv_gpa(&word1), 0, XTYPE_DW, 1, 1, IMA); |
| 608 | if (gru_wait(cb) != CBS_IDLE) |
| 609 | BUG(); |
| 610 | |
| 611 | if (word0 != word1 || word0 != MAGIC) { |
| 612 | printk |
| 613 | ("GRU quicktest err: gru %d, found 0x%lx, expected 0x%lx\n", |
| 614 | gru->gs_gid, word1, MAGIC); |
| 615 | BUG(); /* ZZZ should not be fatal */ |
| 616 | } |
| 617 | |
| 618 | return 0; |
| 619 | } |
| 620 | |
| 621 | |
| 622 | int gru_kservices_init(struct gru_state *gru) |
| 623 | { |
| 624 | struct gru_blade_state *bs; |
| 625 | struct gru_context_configuration_handle *cch; |
| 626 | unsigned long cbr_map, dsr_map; |
| 627 | int err, num, cpus_possible; |
| 628 | |
| 629 | /* |
| 630 | * Currently, resources are reserved ONLY on the second chiplet |
| 631 | * on each blade. This leaves ALL resources on chiplet 0 available |
| 632 | * for user code. |
| 633 | */ |
| 634 | bs = gru->gs_blade; |
| 635 | if (gru != &bs->bs_grus[1]) |
| 636 | return 0; |
| 637 | |
| 638 | cpus_possible = uv_blade_nr_possible_cpus(gru->gs_blade_id); |
| 639 | |
| 640 | num = GRU_NUM_KERNEL_CBR * cpus_possible; |
| 641 | cbr_map = reserve_gru_cb_resources(gru, GRU_CB_COUNT_TO_AU(num), NULL); |
| 642 | gru->gs_reserved_cbrs += num; |
| 643 | |
| 644 | num = GRU_NUM_KERNEL_DSR_BYTES * cpus_possible; |
| 645 | dsr_map = reserve_gru_ds_resources(gru, GRU_DS_BYTES_TO_AU(num), NULL); |
| 646 | gru->gs_reserved_dsr_bytes += num; |
| 647 | |
| 648 | gru->gs_active_contexts++; |
| 649 | __set_bit(KERNEL_CTXNUM, &gru->gs_context_map); |
| 650 | cch = get_cch(gru->gs_gru_base_vaddr, KERNEL_CTXNUM); |
| 651 | |
| 652 | bs->kernel_cb = get_gseg_base_address_cb(gru->gs_gru_base_vaddr, |
| 653 | KERNEL_CTXNUM, 0); |
| 654 | bs->kernel_dsr = get_gseg_base_address_ds(gru->gs_gru_base_vaddr, |
| 655 | KERNEL_CTXNUM, 0); |
| 656 | |
| 657 | lock_cch_handle(cch); |
| 658 | cch->tfm_fault_bit_enable = 0; |
| 659 | cch->tlb_int_enable = 0; |
| 660 | cch->tfm_done_bit_enable = 0; |
| 661 | cch->unmap_enable = 1; |
| 662 | err = cch_allocate(cch, 0, cbr_map, dsr_map); |
| 663 | if (err) { |
| 664 | gru_dbg(grudev, |
| 665 | "Unable to allocate kernel CCH: gru %d, err %d\n", |
| 666 | gru->gs_gid, err); |
| 667 | BUG(); |
| 668 | } |
| 669 | if (cch_start(cch)) { |
| 670 | gru_dbg(grudev, "Unable to start kernel CCH: gru %d, err %d\n", |
| 671 | gru->gs_gid, err); |
| 672 | BUG(); |
| 673 | } |
| 674 | unlock_cch_handle(cch); |
| 675 | |
| 676 | if (options & GRU_QUICKLOOK) |
| 677 | quicktest(gru); |
| 678 | return 0; |
| 679 | } |