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> |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 34 | #include <linux/delay.h> |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 35 | #include "gru.h" |
| 36 | #include "grulib.h" |
| 37 | #include "grutables.h" |
| 38 | #include "grukservices.h" |
| 39 | #include "gru_instructions.h" |
| 40 | #include <asm/uv/uv_hub.h> |
| 41 | |
| 42 | /* |
| 43 | * Kernel GRU Usage |
| 44 | * |
| 45 | * The following is an interim algorithm for management of kernel GRU |
| 46 | * resources. This will likely be replaced when we better understand the |
| 47 | * kernel/user requirements. |
| 48 | * |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 49 | * Blade percpu resources reserved for kernel use. These resources are |
| 50 | * reserved whenever the the kernel context for the blade is loaded. Note |
| 51 | * that the kernel context is not guaranteed to be always available. It is |
| 52 | * loaded on demand & can be stolen by a user if the user demand exceeds the |
| 53 | * kernel demand. The kernel can always reload the kernel context but |
| 54 | * a SLEEP may be required!!!. |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 55 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 56 | #define GRU_NUM_KERNEL_CBR 1 |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 57 | #define GRU_NUM_KERNEL_DSR_BYTES 256 |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 58 | #define GRU_NUM_KERNEL_DSR_CL (GRU_NUM_KERNEL_DSR_BYTES / \ |
| 59 | GRU_CACHE_LINE_BYTES) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 60 | |
| 61 | /* GRU instruction attributes for all instructions */ |
| 62 | #define IMA IMA_CB_DELAY |
| 63 | |
| 64 | /* GRU cacheline size is always 64 bytes - even on arches with 128 byte lines */ |
| 65 | #define __gru_cacheline_aligned__ \ |
| 66 | __attribute__((__aligned__(GRU_CACHE_LINE_BYTES))) |
| 67 | |
| 68 | #define MAGIC 0x1234567887654321UL |
| 69 | |
| 70 | /* Default retry count for GRU errors on kernel instructions */ |
| 71 | #define EXCEPTION_RETRY_LIMIT 3 |
| 72 | |
| 73 | /* Status of message queue sections */ |
| 74 | #define MQS_EMPTY 0 |
| 75 | #define MQS_FULL 1 |
| 76 | #define MQS_NOOP 2 |
| 77 | |
| 78 | /*----------------- RESOURCE MANAGEMENT -------------------------------------*/ |
| 79 | /* optimized for x86_64 */ |
| 80 | struct message_queue { |
| 81 | union gru_mesqhead head __gru_cacheline_aligned__; /* CL 0 */ |
| 82 | int qlines; /* DW 1 */ |
| 83 | long hstatus[2]; |
| 84 | void *next __gru_cacheline_aligned__;/* CL 1 */ |
| 85 | void *limit; |
| 86 | void *start; |
| 87 | void *start2; |
| 88 | char data ____cacheline_aligned; /* CL 2 */ |
| 89 | }; |
| 90 | |
| 91 | /* First word in every message - used by mesq interface */ |
| 92 | struct message_header { |
| 93 | char present; |
| 94 | char present2; |
| 95 | char lines; |
| 96 | char fill; |
| 97 | }; |
| 98 | |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 99 | #define HSTATUS(mq, h) ((mq) + offsetof(struct message_queue, hstatus[h])) |
| 100 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 101 | /* |
| 102 | * Allocate a kernel context (GTS) for the specified blade. |
| 103 | * - protected by writelock on bs_kgts_sema. |
| 104 | */ |
| 105 | static void gru_alloc_kernel_context(struct gru_blade_state *bs, int blade_id) |
| 106 | { |
| 107 | int cbr_au_count, dsr_au_count, ncpus; |
| 108 | |
| 109 | ncpus = uv_blade_nr_possible_cpus(blade_id); |
| 110 | cbr_au_count = GRU_CB_COUNT_TO_AU(GRU_NUM_KERNEL_CBR * ncpus); |
| 111 | dsr_au_count = GRU_DS_BYTES_TO_AU(GRU_NUM_KERNEL_DSR_BYTES * ncpus); |
| 112 | bs->bs_kgts = gru_alloc_gts(NULL, cbr_au_count, dsr_au_count, 0, 0); |
| 113 | } |
| 114 | |
| 115 | /* |
| 116 | * Reload the blade's kernel context into a GRU chiplet. Called holding |
| 117 | * the bs_kgts_sema for READ. Will steal user contexts if necessary. |
| 118 | */ |
| 119 | static void gru_load_kernel_context(struct gru_blade_state *bs, int blade_id) |
| 120 | { |
| 121 | struct gru_state *gru; |
| 122 | struct gru_thread_state *kgts; |
| 123 | void *vaddr; |
| 124 | int ctxnum; |
| 125 | |
| 126 | up_read(&bs->bs_kgts_sema); |
| 127 | down_write(&bs->bs_kgts_sema); |
| 128 | |
| 129 | if (!bs->bs_kgts) |
| 130 | gru_alloc_kernel_context(bs, blade_id); |
| 131 | kgts = bs->bs_kgts; |
| 132 | |
| 133 | if (!kgts->ts_gru) { |
| 134 | STAT(load_kernel_context); |
| 135 | while (!gru_assign_gru_context(kgts, blade_id)) { |
| 136 | msleep(1); |
| 137 | gru_steal_context(kgts, blade_id); |
| 138 | } |
| 139 | gru_load_context(kgts); |
| 140 | gru = bs->bs_kgts->ts_gru; |
| 141 | vaddr = gru->gs_gru_base_vaddr; |
| 142 | ctxnum = kgts->ts_ctxnum; |
| 143 | bs->kernel_cb = get_gseg_base_address_cb(vaddr, ctxnum, 0); |
| 144 | bs->kernel_dsr = get_gseg_base_address_ds(vaddr, ctxnum, 0); |
| 145 | } |
| 146 | downgrade_write(&bs->bs_kgts_sema); |
| 147 | } |
| 148 | |
| 149 | /* |
| 150 | * Lock & load the kernel context for the specified blade. |
| 151 | */ |
| 152 | static struct gru_blade_state *gru_lock_kernel_context(int blade_id) |
| 153 | { |
| 154 | struct gru_blade_state *bs; |
| 155 | |
| 156 | STAT(lock_kernel_context); |
| 157 | bs = gru_base[blade_id]; |
| 158 | |
| 159 | down_read(&bs->bs_kgts_sema); |
| 160 | if (!bs->bs_kgts || !bs->bs_kgts->ts_gru) |
| 161 | gru_load_kernel_context(bs, blade_id); |
| 162 | return bs; |
| 163 | |
| 164 | } |
| 165 | |
| 166 | /* |
| 167 | * Unlock the kernel context for the specified blade. Context is not |
| 168 | * unloaded but may be stolen before next use. |
| 169 | */ |
| 170 | static void gru_unlock_kernel_context(int blade_id) |
| 171 | { |
| 172 | struct gru_blade_state *bs; |
| 173 | |
| 174 | bs = gru_base[blade_id]; |
| 175 | up_read(&bs->bs_kgts_sema); |
| 176 | STAT(unlock_kernel_context); |
| 177 | } |
| 178 | |
| 179 | /* |
| 180 | * Reserve & get pointers to the DSR/CBRs reserved for the current cpu. |
| 181 | * - returns with preemption disabled |
| 182 | */ |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 183 | static int gru_get_cpu_resources(int dsr_bytes, void **cb, void **dsr) |
| 184 | { |
| 185 | struct gru_blade_state *bs; |
| 186 | int lcpu; |
| 187 | |
| 188 | BUG_ON(dsr_bytes > GRU_NUM_KERNEL_DSR_BYTES); |
| 189 | preempt_disable(); |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 190 | bs = gru_lock_kernel_context(uv_numa_blade_id()); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 191 | lcpu = uv_blade_processor_id(); |
| 192 | *cb = bs->kernel_cb + lcpu * GRU_HANDLE_STRIDE; |
| 193 | *dsr = bs->kernel_dsr + lcpu * GRU_NUM_KERNEL_DSR_BYTES; |
| 194 | return 0; |
| 195 | } |
| 196 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 197 | /* |
| 198 | * Free the current cpus reserved DSR/CBR resources. |
| 199 | */ |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 200 | static void gru_free_cpu_resources(void *cb, void *dsr) |
| 201 | { |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 202 | gru_unlock_kernel_context(uv_numa_blade_id()); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 203 | preempt_enable(); |
| 204 | } |
| 205 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 206 | /*----------------------------------------------------------------------*/ |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 207 | int gru_get_cb_exception_detail(void *cb, |
| 208 | struct control_block_extended_exc_detail *excdet) |
| 209 | { |
| 210 | struct gru_control_block_extended *cbe; |
| 211 | |
| 212 | cbe = get_cbe(GRUBASE(cb), get_cb_number(cb)); |
Jack Steiner | fe5bb6b | 2009-04-02 16:59:04 -0700 | [diff] [blame] | 213 | prefetchw(cbe); /* Harmless on hardware, required for emulator */ |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 214 | excdet->opc = cbe->opccpy; |
| 215 | excdet->exopc = cbe->exopccpy; |
| 216 | excdet->ecause = cbe->ecause; |
| 217 | excdet->exceptdet0 = cbe->idef1upd; |
| 218 | excdet->exceptdet1 = cbe->idef3upd; |
| 219 | return 0; |
| 220 | } |
| 221 | |
| 222 | char *gru_get_cb_exception_detail_str(int ret, void *cb, |
| 223 | char *buf, int size) |
| 224 | { |
| 225 | struct gru_control_block_status *gen = (void *)cb; |
| 226 | struct control_block_extended_exc_detail excdet; |
| 227 | |
| 228 | if (ret > 0 && gen->istatus == CBS_EXCEPTION) { |
| 229 | gru_get_cb_exception_detail(cb, &excdet); |
| 230 | snprintf(buf, size, |
| 231 | "GRU exception: cb %p, opc %d, exopc %d, ecause 0x%x," |
| 232 | "excdet0 0x%lx, excdet1 0x%x", |
| 233 | gen, excdet.opc, excdet.exopc, excdet.ecause, |
| 234 | excdet.exceptdet0, excdet.exceptdet1); |
| 235 | } else { |
| 236 | snprintf(buf, size, "No exception"); |
| 237 | } |
| 238 | return buf; |
| 239 | } |
| 240 | |
| 241 | static int gru_wait_idle_or_exception(struct gru_control_block_status *gen) |
| 242 | { |
| 243 | while (gen->istatus >= CBS_ACTIVE) { |
| 244 | cpu_relax(); |
| 245 | barrier(); |
| 246 | } |
| 247 | return gen->istatus; |
| 248 | } |
| 249 | |
| 250 | static int gru_retry_exception(void *cb) |
| 251 | { |
| 252 | struct gru_control_block_status *gen = (void *)cb; |
| 253 | struct control_block_extended_exc_detail excdet; |
| 254 | int retry = EXCEPTION_RETRY_LIMIT; |
| 255 | |
| 256 | while (1) { |
| 257 | if (gru_get_cb_message_queue_substatus(cb)) |
| 258 | break; |
| 259 | if (gru_wait_idle_or_exception(gen) == CBS_IDLE) |
| 260 | return CBS_IDLE; |
| 261 | |
| 262 | gru_get_cb_exception_detail(cb, &excdet); |
| 263 | if (excdet.ecause & ~EXCEPTION_RETRY_BITS) |
| 264 | break; |
| 265 | if (retry-- == 0) |
| 266 | break; |
| 267 | gen->icmd = 1; |
| 268 | gru_flush_cache(gen); |
| 269 | } |
| 270 | return CBS_EXCEPTION; |
| 271 | } |
| 272 | |
| 273 | int gru_check_status_proc(void *cb) |
| 274 | { |
| 275 | struct gru_control_block_status *gen = (void *)cb; |
| 276 | int ret; |
| 277 | |
| 278 | ret = gen->istatus; |
| 279 | if (ret != CBS_EXCEPTION) |
| 280 | return ret; |
| 281 | return gru_retry_exception(cb); |
| 282 | |
| 283 | } |
| 284 | |
| 285 | int gru_wait_proc(void *cb) |
| 286 | { |
| 287 | struct gru_control_block_status *gen = (void *)cb; |
| 288 | int ret; |
| 289 | |
| 290 | ret = gru_wait_idle_or_exception(gen); |
| 291 | if (ret == CBS_EXCEPTION) |
| 292 | ret = gru_retry_exception(cb); |
| 293 | |
| 294 | return ret; |
| 295 | } |
| 296 | |
| 297 | void gru_abort(int ret, void *cb, char *str) |
| 298 | { |
| 299 | char buf[GRU_EXC_STR_SIZE]; |
| 300 | |
| 301 | panic("GRU FATAL ERROR: %s - %s\n", str, |
| 302 | gru_get_cb_exception_detail_str(ret, cb, buf, sizeof(buf))); |
| 303 | } |
| 304 | |
| 305 | void gru_wait_abort_proc(void *cb) |
| 306 | { |
| 307 | int ret; |
| 308 | |
| 309 | ret = gru_wait_proc(cb); |
| 310 | if (ret) |
| 311 | gru_abort(ret, cb, "gru_wait_abort"); |
| 312 | } |
| 313 | |
| 314 | |
| 315 | /*------------------------------ MESSAGE QUEUES -----------------------------*/ |
| 316 | |
| 317 | /* Internal status . These are NOT returned to the user. */ |
| 318 | #define MQIE_AGAIN -1 /* try again */ |
| 319 | |
| 320 | |
| 321 | /* |
| 322 | * Save/restore the "present" flag that is in the second line of 2-line |
| 323 | * messages |
| 324 | */ |
| 325 | static inline int get_present2(void *p) |
| 326 | { |
| 327 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; |
| 328 | return mhdr->present; |
| 329 | } |
| 330 | |
| 331 | static inline void restore_present2(void *p, int val) |
| 332 | { |
| 333 | struct message_header *mhdr = p + GRU_CACHE_LINE_BYTES; |
| 334 | mhdr->present = val; |
| 335 | } |
| 336 | |
| 337 | /* |
| 338 | * Create a message queue. |
| 339 | * qlines - message queue size in cache lines. Includes 2-line header. |
| 340 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 341 | int gru_create_message_queue(struct gru_message_queue_desc *mqd, |
| 342 | void *p, unsigned int bytes, int nasid, int vector, int apicid) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 343 | { |
| 344 | struct message_queue *mq = p; |
| 345 | unsigned int qlines; |
| 346 | |
| 347 | qlines = bytes / GRU_CACHE_LINE_BYTES - 2; |
| 348 | memset(mq, 0, bytes); |
| 349 | mq->start = &mq->data; |
| 350 | mq->start2 = &mq->data + (qlines / 2 - 1) * GRU_CACHE_LINE_BYTES; |
| 351 | mq->next = &mq->data; |
| 352 | mq->limit = &mq->data + (qlines - 2) * GRU_CACHE_LINE_BYTES; |
| 353 | mq->qlines = qlines; |
| 354 | mq->hstatus[0] = 0; |
| 355 | mq->hstatus[1] = 1; |
| 356 | mq->head = gru_mesq_head(2, qlines / 2 + 1); |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 357 | mqd->mq = mq; |
| 358 | mqd->mq_gpa = uv_gpa(mq); |
| 359 | mqd->qlines = qlines; |
| 360 | mqd->interrupt_pnode = UV_NASID_TO_PNODE(nasid); |
| 361 | mqd->interrupt_vector = vector; |
| 362 | mqd->interrupt_apicid = apicid; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 363 | return 0; |
| 364 | } |
| 365 | EXPORT_SYMBOL_GPL(gru_create_message_queue); |
| 366 | |
| 367 | /* |
| 368 | * Send a NOOP message to a message queue |
| 369 | * Returns: |
| 370 | * 0 - if queue is full after the send. This is the normal case |
| 371 | * but various races can change this. |
| 372 | * -1 - if mesq sent successfully but queue not full |
| 373 | * >0 - unexpected error. MQE_xxx returned |
| 374 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 375 | static int send_noop_message(void *cb, struct gru_message_queue_desc *mqd, |
| 376 | void *mesg) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 377 | { |
| 378 | const struct message_header noop_header = { |
| 379 | .present = MQS_NOOP, .lines = 1}; |
| 380 | unsigned long m; |
| 381 | int substatus, ret; |
| 382 | struct message_header save_mhdr, *mhdr = mesg; |
| 383 | |
| 384 | STAT(mesq_noop); |
| 385 | save_mhdr = *mhdr; |
| 386 | *mhdr = noop_header; |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 387 | gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), 1, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 388 | ret = gru_wait(cb); |
| 389 | |
| 390 | if (ret) { |
| 391 | substatus = gru_get_cb_message_queue_substatus(cb); |
| 392 | switch (substatus) { |
| 393 | case CBSS_NO_ERROR: |
| 394 | STAT(mesq_noop_unexpected_error); |
| 395 | ret = MQE_UNEXPECTED_CB_ERR; |
| 396 | break; |
| 397 | case CBSS_LB_OVERFLOWED: |
| 398 | STAT(mesq_noop_lb_overflow); |
| 399 | ret = MQE_CONGESTION; |
| 400 | break; |
| 401 | case CBSS_QLIMIT_REACHED: |
| 402 | STAT(mesq_noop_qlimit_reached); |
| 403 | ret = 0; |
| 404 | break; |
| 405 | case CBSS_AMO_NACKED: |
| 406 | STAT(mesq_noop_amo_nacked); |
| 407 | ret = MQE_CONGESTION; |
| 408 | break; |
| 409 | case CBSS_PUT_NACKED: |
| 410 | STAT(mesq_noop_put_nacked); |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 411 | m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 412 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, 1, 1, |
| 413 | IMA); |
| 414 | if (gru_wait(cb) == CBS_IDLE) |
| 415 | ret = MQIE_AGAIN; |
| 416 | else |
| 417 | ret = MQE_UNEXPECTED_CB_ERR; |
| 418 | break; |
| 419 | case CBSS_PAGE_OVERFLOW: |
| 420 | default: |
| 421 | BUG(); |
| 422 | } |
| 423 | } |
| 424 | *mhdr = save_mhdr; |
| 425 | return ret; |
| 426 | } |
| 427 | |
| 428 | /* |
| 429 | * Handle a gru_mesq full. |
| 430 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 431 | static int send_message_queue_full(void *cb, struct gru_message_queue_desc *mqd, |
| 432 | void *mesg, int lines) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 433 | { |
| 434 | union gru_mesqhead mqh; |
| 435 | unsigned int limit, head; |
| 436 | unsigned long avalue; |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 437 | int half, qlines; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 438 | |
| 439 | /* Determine if switching to first/second half of q */ |
| 440 | avalue = gru_get_amo_value(cb); |
| 441 | head = gru_get_amo_value_head(cb); |
| 442 | limit = gru_get_amo_value_limit(cb); |
| 443 | |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 444 | qlines = mqd->qlines; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 445 | half = (limit != qlines); |
| 446 | |
| 447 | if (half) |
| 448 | mqh = gru_mesq_head(qlines / 2 + 1, qlines); |
| 449 | else |
| 450 | mqh = gru_mesq_head(2, qlines / 2 + 1); |
| 451 | |
| 452 | /* Try to get lock for switching head pointer */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 453 | gru_gamir(cb, EOP_IR_CLR, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 454 | if (gru_wait(cb) != CBS_IDLE) |
| 455 | goto cberr; |
| 456 | if (!gru_get_amo_value(cb)) { |
| 457 | STAT(mesq_qf_locked); |
| 458 | return MQE_QUEUE_FULL; |
| 459 | } |
| 460 | |
| 461 | /* Got the lock. Send optional NOP if queue not full, */ |
| 462 | if (head != limit) { |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 463 | if (send_noop_message(cb, mqd, mesg)) { |
| 464 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 465 | XTYPE_DW, IMA); |
| 466 | if (gru_wait(cb) != CBS_IDLE) |
| 467 | goto cberr; |
| 468 | STAT(mesq_qf_noop_not_full); |
| 469 | return MQIE_AGAIN; |
| 470 | } |
| 471 | avalue++; |
| 472 | } |
| 473 | |
| 474 | /* Then flip queuehead to other half of queue. */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 475 | gru_gamer(cb, EOP_ERR_CSWAP, mqd->mq_gpa, XTYPE_DW, mqh.val, avalue, |
| 476 | IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 477 | if (gru_wait(cb) != CBS_IDLE) |
| 478 | goto cberr; |
| 479 | |
| 480 | /* If not successfully in swapping queue head, clear the hstatus lock */ |
| 481 | if (gru_get_amo_value(cb) != avalue) { |
| 482 | STAT(mesq_qf_switch_head_failed); |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 483 | gru_gamir(cb, EOP_IR_INC, HSTATUS(mqd->mq_gpa, half), XTYPE_DW, |
| 484 | IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 485 | if (gru_wait(cb) != CBS_IDLE) |
| 486 | goto cberr; |
| 487 | } |
| 488 | return MQIE_AGAIN; |
| 489 | cberr: |
| 490 | STAT(mesq_qf_unexpected_error); |
| 491 | return MQE_UNEXPECTED_CB_ERR; |
| 492 | } |
| 493 | |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 494 | /* |
| 495 | * Send a cross-partition interrupt to the SSI that contains the target |
| 496 | * message queue. Normally, the interrupt is automatically delivered by hardware |
| 497 | * but some error conditions require explicit delivery. |
| 498 | */ |
| 499 | static void send_message_queue_interrupt(struct gru_message_queue_desc *mqd) |
| 500 | { |
| 501 | if (mqd->interrupt_vector) |
| 502 | uv_hub_send_ipi(mqd->interrupt_pnode, mqd->interrupt_apicid, |
| 503 | mqd->interrupt_vector); |
| 504 | } |
| 505 | |
Jack Steiner | 17b49a6 | 2009-06-17 16:28:23 -0700 | [diff] [blame^] | 506 | /* |
| 507 | * Handle a PUT failure. Note: if message was a 2-line message, one of the |
| 508 | * lines might have successfully have been written. Before sending the |
| 509 | * message, "present" must be cleared in BOTH lines to prevent the receiver |
| 510 | * from prematurely seeing the full message. |
| 511 | */ |
| 512 | static int send_message_put_nacked(void *cb, struct gru_message_queue_desc *mqd, |
| 513 | void *mesg, int lines) |
| 514 | { |
| 515 | unsigned long m; |
| 516 | |
| 517 | m = mqd->mq_gpa + (gru_get_amo_value_head(cb) << 6); |
| 518 | if (lines == 2) { |
| 519 | gru_vset(cb, m, 0, XTYPE_CL, lines, 1, IMA); |
| 520 | if (gru_wait(cb) != CBS_IDLE) |
| 521 | return MQE_UNEXPECTED_CB_ERR; |
| 522 | } |
| 523 | gru_vstore(cb, m, gru_get_tri(mesg), XTYPE_CL, lines, 1, IMA); |
| 524 | if (gru_wait(cb) != CBS_IDLE) |
| 525 | return MQE_UNEXPECTED_CB_ERR; |
| 526 | send_message_queue_interrupt(mqd); |
| 527 | return MQE_OK; |
| 528 | } |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 529 | |
| 530 | /* |
| 531 | * Handle a gru_mesq failure. Some of these failures are software recoverable |
| 532 | * or retryable. |
| 533 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 534 | static int send_message_failure(void *cb, struct gru_message_queue_desc *mqd, |
| 535 | void *mesg, int lines) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 536 | { |
| 537 | int substatus, ret = 0; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 538 | |
| 539 | substatus = gru_get_cb_message_queue_substatus(cb); |
| 540 | switch (substatus) { |
| 541 | case CBSS_NO_ERROR: |
| 542 | STAT(mesq_send_unexpected_error); |
| 543 | ret = MQE_UNEXPECTED_CB_ERR; |
| 544 | break; |
| 545 | case CBSS_LB_OVERFLOWED: |
| 546 | STAT(mesq_send_lb_overflow); |
| 547 | ret = MQE_CONGESTION; |
| 548 | break; |
| 549 | case CBSS_QLIMIT_REACHED: |
| 550 | STAT(mesq_send_qlimit_reached); |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 551 | ret = send_message_queue_full(cb, mqd, mesg, lines); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 552 | break; |
| 553 | case CBSS_AMO_NACKED: |
| 554 | STAT(mesq_send_amo_nacked); |
| 555 | ret = MQE_CONGESTION; |
| 556 | break; |
| 557 | case CBSS_PUT_NACKED: |
| 558 | STAT(mesq_send_put_nacked); |
Jack Steiner | 17b49a6 | 2009-06-17 16:28:23 -0700 | [diff] [blame^] | 559 | ret = send_message_put_nacked(cb, mqd, mesg, lines); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 560 | break; |
| 561 | default: |
| 562 | BUG(); |
| 563 | } |
| 564 | return ret; |
| 565 | } |
| 566 | |
| 567 | /* |
| 568 | * Send a message to a message queue |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 569 | * mqd message queue descriptor |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 570 | * mesg message. ust be vaddr within a GSEG |
| 571 | * bytes message size (<= 2 CL) |
| 572 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 573 | int gru_send_message_gpa(struct gru_message_queue_desc *mqd, void *mesg, |
| 574 | unsigned int bytes) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 575 | { |
| 576 | struct message_header *mhdr; |
| 577 | void *cb; |
| 578 | void *dsr; |
| 579 | int istatus, clines, ret; |
| 580 | |
| 581 | STAT(mesq_send); |
| 582 | BUG_ON(bytes < sizeof(int) || bytes > 2 * GRU_CACHE_LINE_BYTES); |
| 583 | |
Julia Lawall | cbf330b | 2008-10-15 22:01:27 -0700 | [diff] [blame] | 584 | clines = DIV_ROUND_UP(bytes, GRU_CACHE_LINE_BYTES); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 585 | if (gru_get_cpu_resources(bytes, &cb, &dsr)) |
| 586 | return MQE_BUG_NO_RESOURCES; |
| 587 | memcpy(dsr, mesg, bytes); |
| 588 | mhdr = dsr; |
| 589 | mhdr->present = MQS_FULL; |
| 590 | mhdr->lines = clines; |
| 591 | if (clines == 2) { |
| 592 | mhdr->present2 = get_present2(mhdr); |
| 593 | restore_present2(mhdr, MQS_FULL); |
| 594 | } |
| 595 | |
| 596 | do { |
| 597 | ret = MQE_OK; |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 598 | gru_mesq(cb, mqd->mq_gpa, gru_get_tri(mhdr), clines, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 599 | istatus = gru_wait(cb); |
| 600 | if (istatus != CBS_IDLE) |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 601 | ret = send_message_failure(cb, mqd, dsr, clines); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 602 | } while (ret == MQIE_AGAIN); |
| 603 | gru_free_cpu_resources(cb, dsr); |
| 604 | |
| 605 | if (ret) |
| 606 | STAT(mesq_send_failed); |
| 607 | return ret; |
| 608 | } |
| 609 | EXPORT_SYMBOL_GPL(gru_send_message_gpa); |
| 610 | |
| 611 | /* |
| 612 | * Advance the receive pointer for the queue to the next message. |
| 613 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 614 | void gru_free_message(struct gru_message_queue_desc *mqd, void *mesg) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 615 | { |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 616 | struct message_queue *mq = mqd->mq; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 617 | struct message_header *mhdr = mq->next; |
| 618 | void *next, *pnext; |
| 619 | int half = -1; |
| 620 | int lines = mhdr->lines; |
| 621 | |
| 622 | if (lines == 2) |
| 623 | restore_present2(mhdr, MQS_EMPTY); |
| 624 | mhdr->present = MQS_EMPTY; |
| 625 | |
| 626 | pnext = mq->next; |
| 627 | next = pnext + GRU_CACHE_LINE_BYTES * lines; |
| 628 | if (next == mq->limit) { |
| 629 | next = mq->start; |
| 630 | half = 1; |
| 631 | } else if (pnext < mq->start2 && next >= mq->start2) { |
| 632 | half = 0; |
| 633 | } |
| 634 | |
| 635 | if (half >= 0) |
| 636 | mq->hstatus[half] = 1; |
| 637 | mq->next = next; |
| 638 | } |
| 639 | EXPORT_SYMBOL_GPL(gru_free_message); |
| 640 | |
| 641 | /* |
| 642 | * Get next message from message queue. Return NULL if no message |
| 643 | * present. User must call next_message() to move to next message. |
| 644 | * rmq message queue |
| 645 | */ |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 646 | void *gru_get_next_message(struct gru_message_queue_desc *mqd) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 647 | { |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 648 | struct message_queue *mq = mqd->mq; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 649 | struct message_header *mhdr = mq->next; |
| 650 | int present = mhdr->present; |
| 651 | |
| 652 | /* skip NOOP messages */ |
| 653 | STAT(mesq_receive); |
| 654 | while (present == MQS_NOOP) { |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 655 | gru_free_message(mqd, mhdr); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 656 | mhdr = mq->next; |
| 657 | present = mhdr->present; |
| 658 | } |
| 659 | |
| 660 | /* Wait for both halves of 2 line messages */ |
| 661 | if (present == MQS_FULL && mhdr->lines == 2 && |
| 662 | get_present2(mhdr) == MQS_EMPTY) |
| 663 | present = MQS_EMPTY; |
| 664 | |
| 665 | if (!present) { |
| 666 | STAT(mesq_receive_none); |
| 667 | return NULL; |
| 668 | } |
| 669 | |
| 670 | if (mhdr->lines == 2) |
| 671 | restore_present2(mhdr, mhdr->present2); |
| 672 | |
| 673 | return mhdr; |
| 674 | } |
| 675 | EXPORT_SYMBOL_GPL(gru_get_next_message); |
| 676 | |
| 677 | /* ---------------------- GRU DATA COPY FUNCTIONS ---------------------------*/ |
| 678 | |
| 679 | /* |
| 680 | * Copy a block of data using the GRU resources |
| 681 | */ |
| 682 | int gru_copy_gpa(unsigned long dest_gpa, unsigned long src_gpa, |
| 683 | unsigned int bytes) |
| 684 | { |
| 685 | void *cb; |
| 686 | void *dsr; |
| 687 | int ret; |
| 688 | |
| 689 | STAT(copy_gpa); |
| 690 | if (gru_get_cpu_resources(GRU_NUM_KERNEL_DSR_BYTES, &cb, &dsr)) |
| 691 | return MQE_BUG_NO_RESOURCES; |
| 692 | gru_bcopy(cb, src_gpa, dest_gpa, gru_get_tri(dsr), |
Jack Steiner | 6f2584f | 2009-04-02 16:59:10 -0700 | [diff] [blame] | 693 | XTYPE_B, bytes, GRU_NUM_KERNEL_DSR_CL, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 694 | ret = gru_wait(cb); |
| 695 | gru_free_cpu_resources(cb, dsr); |
| 696 | return ret; |
| 697 | } |
| 698 | EXPORT_SYMBOL_GPL(gru_copy_gpa); |
| 699 | |
| 700 | /* ------------------- KERNEL QUICKTESTS RUN AT STARTUP ----------------*/ |
| 701 | /* Temp - will delete after we gain confidence in the GRU */ |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 702 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 703 | int quicktest(void) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 704 | { |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 705 | unsigned long word0; |
| 706 | unsigned long word1; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 707 | void *cb; |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 708 | void *dsr; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 709 | unsigned long *p; |
| 710 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 711 | if (gru_get_cpu_resources(GRU_CACHE_LINE_BYTES, &cb, &dsr)) |
| 712 | return MQE_BUG_NO_RESOURCES; |
| 713 | p = dsr; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 714 | word0 = MAGIC; |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 715 | word1 = 0; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 716 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 717 | gru_vload(cb, uv_gpa(&word0), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 718 | if (gru_wait(cb) != CBS_IDLE) |
| 719 | BUG(); |
| 720 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 721 | if (*p != MAGIC) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 722 | BUG(); |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 723 | gru_vstore(cb, uv_gpa(&word1), gru_get_tri(dsr), XTYPE_DW, 1, 1, IMA); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 724 | if (gru_wait(cb) != CBS_IDLE) |
| 725 | BUG(); |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 726 | gru_free_cpu_resources(cb, dsr); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 727 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 728 | if (word0 != word1 || word1 != MAGIC) { |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 729 | printk |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 730 | ("GRU quicktest err: found 0x%lx, expected 0x%lx\n", |
| 731 | word1, MAGIC); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 732 | BUG(); /* ZZZ should not be fatal */ |
| 733 | } |
| 734 | |
| 735 | return 0; |
| 736 | } |
| 737 | |
| 738 | |
| 739 | int gru_kservices_init(struct gru_state *gru) |
| 740 | { |
| 741 | struct gru_blade_state *bs; |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 742 | |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 743 | bs = gru->gs_blade; |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 744 | if (gru != &bs->bs_grus[0]) |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 745 | return 0; |
| 746 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 747 | init_rwsem(&bs->bs_kgts_sema); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 748 | |
Jack Steiner | 9ca8e40c1 | 2008-07-29 22:34:02 -0700 | [diff] [blame] | 749 | if (gru_options & GRU_QUICKLOOK) |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 750 | quicktest(); |
Jack Steiner | 28bffaf | 2008-07-29 22:33:57 -0700 | [diff] [blame] | 751 | return 0; |
| 752 | } |
Jack Steiner | 27ca8a7 | 2009-04-02 16:59:11 -0700 | [diff] [blame] | 753 | |
| 754 | void gru_kservices_exit(struct gru_state *gru) |
| 755 | { |
Jack Steiner | 27ca8a7 | 2009-04-02 16:59:11 -0700 | [diff] [blame] | 756 | struct gru_blade_state *bs; |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 757 | struct gru_thread_state *kgts; |
Jack Steiner | 27ca8a7 | 2009-04-02 16:59:11 -0700 | [diff] [blame] | 758 | |
| 759 | bs = gru->gs_blade; |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 760 | if (gru != &bs->bs_grus[0]) |
Jack Steiner | 27ca8a7 | 2009-04-02 16:59:11 -0700 | [diff] [blame] | 761 | return; |
| 762 | |
Jack Steiner | 836ce67 | 2009-06-17 16:28:22 -0700 | [diff] [blame] | 763 | kgts = bs->bs_kgts; |
| 764 | if (kgts && kgts->ts_gru) |
| 765 | gru_unload_context(kgts, 0); |
| 766 | kfree(kgts); |
Jack Steiner | 27ca8a7 | 2009-04-02 16:59:11 -0700 | [diff] [blame] | 767 | } |
| 768 | |