1 /* 2 * SN Platform GRU Driver 3 * 4 * DRIVER TABLE MANAGER + GRU CONTEXT LOAD/UNLOAD 5 * 6 * This file is subject to the terms and conditions of the GNU General Public 7 * License. See the file "COPYING" in the main directory of this archive 8 * for more details. 9 * 10 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved. 11 */ 12 13 #include <linux/kernel.h> 14 #include <linux/slab.h> 15 #include <linux/mm.h> 16 #include <linux/spinlock.h> 17 #include <linux/sched.h> 18 #include <linux/device.h> 19 #include <linux/list.h> 20 #include <asm/uv/uv_hub.h> 21 #include "gru.h" 22 #include "grutables.h" 23 #include "gruhandles.h" 24 25 unsigned long gru_options __read_mostly; 26 27 static struct device_driver gru_driver = { 28 .name = "gru" 29 }; 30 31 static struct device gru_device = { 32 .init_name = "", 33 .driver = &gru_driver, 34 }; 35 36 struct device *grudev = &gru_device; 37 38 /* 39 * Select a gru fault map to be used by the current cpu. Note that 40 * multiple cpus may be using the same map. 41 * ZZZ should "shift" be used?? Depends on HT cpu numbering 42 * ZZZ should be inline but did not work on emulator 43 */ 44 int gru_cpu_fault_map_id(void) 45 { 46 return uv_blade_processor_id() % GRU_NUM_TFM; 47 } 48 49 /*--------- ASID Management ------------------------------------------- 50 * 51 * Initially, assign asids sequentially from MIN_ASID .. MAX_ASID. 52 * Once MAX is reached, flush the TLB & start over. However, 53 * some asids may still be in use. There won't be many (percentage wise) still 54 * in use. Search active contexts & determine the value of the first 55 * asid in use ("x"s below). Set "limit" to this value. 56 * This defines a block of assignable asids. 57 * 58 * When "limit" is reached, search forward from limit+1 and determine the 59 * next block of assignable asids. 60 * 61 * Repeat until MAX_ASID is reached, then start over again. 62 * 63 * Each time MAX_ASID is reached, increment the asid generation. Since 64 * the search for in-use asids only checks contexts with GRUs currently 65 * assigned, asids in some contexts will be missed. Prior to loading 66 * a context, the asid generation of the GTS asid is rechecked. If it 67 * doesn't match the current generation, a new asid will be assigned. 68 * 69 * 0---------------x------------x---------------------x----| 70 * ^-next ^-limit ^-MAX_ASID 71 * 72 * All asid manipulation & context loading/unloading is protected by the 73 * gs_lock. 74 */ 75 76 /* Hit the asid limit. Start over */ 77 static int gru_wrap_asid(struct gru_state *gru) 78 { 79 gru_dbg(grudev, "gid %d\n", gru->gs_gid); 80 STAT(asid_wrap); 81 gru->gs_asid_gen++; 82 return MIN_ASID; 83 } 84 85 /* Find the next chunk of unused asids */ 86 static int gru_reset_asid_limit(struct gru_state *gru, int asid) 87 { 88 int i, gid, inuse_asid, limit; 89 90 gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid); 91 STAT(asid_next); 92 limit = MAX_ASID; 93 if (asid >= limit) 94 asid = gru_wrap_asid(gru); 95 gru_flush_all_tlb(gru); 96 gid = gru->gs_gid; 97 again: 98 for (i = 0; i < GRU_NUM_CCH; i++) { 99 if (!gru->gs_gts[i]) 100 continue; 101 inuse_asid = gru->gs_gts[i]->ts_gms->ms_asids[gid].mt_asid; 102 gru_dbg(grudev, "gid %d, gts %p, gms %p, inuse 0x%x, cxt %d\n", 103 gru->gs_gid, gru->gs_gts[i], gru->gs_gts[i]->ts_gms, 104 inuse_asid, i); 105 if (inuse_asid == asid) { 106 asid += ASID_INC; 107 if (asid >= limit) { 108 /* 109 * empty range: reset the range limit and 110 * start over 111 */ 112 limit = MAX_ASID; 113 if (asid >= MAX_ASID) 114 asid = gru_wrap_asid(gru); 115 goto again; 116 } 117 } 118 119 if ((inuse_asid > asid) && (inuse_asid < limit)) 120 limit = inuse_asid; 121 } 122 gru->gs_asid_limit = limit; 123 gru->gs_asid = asid; 124 gru_dbg(grudev, "gid %d, new asid 0x%x, new_limit 0x%x\n", gru->gs_gid, 125 asid, limit); 126 return asid; 127 } 128 129 /* Assign a new ASID to a thread context. */ 130 static int gru_assign_asid(struct gru_state *gru) 131 { 132 int asid; 133 134 gru->gs_asid += ASID_INC; 135 asid = gru->gs_asid; 136 if (asid >= gru->gs_asid_limit) 137 asid = gru_reset_asid_limit(gru, asid); 138 139 gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid); 140 return asid; 141 } 142 143 /* 144 * Clear n bits in a word. Return a word indicating the bits that were cleared. 145 * Optionally, build an array of chars that contain the bit numbers allocated. 146 */ 147 static unsigned long reserve_resources(unsigned long *p, int n, int mmax, 148 char *idx) 149 { 150 unsigned long bits = 0; 151 int i; 152 153 do { 154 i = find_first_bit(p, mmax); 155 if (i == mmax) 156 BUG(); 157 __clear_bit(i, p); 158 __set_bit(i, &bits); 159 if (idx) 160 *idx++ = i; 161 } while (--n); 162 return bits; 163 } 164 165 unsigned long gru_reserve_cb_resources(struct gru_state *gru, int cbr_au_count, 166 char *cbmap) 167 { 168 return reserve_resources(&gru->gs_cbr_map, cbr_au_count, GRU_CBR_AU, 169 cbmap); 170 } 171 172 unsigned long gru_reserve_ds_resources(struct gru_state *gru, int dsr_au_count, 173 char *dsmap) 174 { 175 return reserve_resources(&gru->gs_dsr_map, dsr_au_count, GRU_DSR_AU, 176 dsmap); 177 } 178 179 static void reserve_gru_resources(struct gru_state *gru, 180 struct gru_thread_state *gts) 181 { 182 gru->gs_active_contexts++; 183 gts->ts_cbr_map = 184 gru_reserve_cb_resources(gru, gts->ts_cbr_au_count, 185 gts->ts_cbr_idx); 186 gts->ts_dsr_map = 187 gru_reserve_ds_resources(gru, gts->ts_dsr_au_count, NULL); 188 } 189 190 static void free_gru_resources(struct gru_state *gru, 191 struct gru_thread_state *gts) 192 { 193 gru->gs_active_contexts--; 194 gru->gs_cbr_map |= gts->ts_cbr_map; 195 gru->gs_dsr_map |= gts->ts_dsr_map; 196 } 197 198 /* 199 * Check if a GRU has sufficient free resources to satisfy an allocation 200 * request. Note: GRU locks may or may not be held when this is called. If 201 * not held, recheck after acquiring the appropriate locks. 202 * 203 * Returns 1 if sufficient resources, 0 if not 204 */ 205 static int check_gru_resources(struct gru_state *gru, int cbr_au_count, 206 int dsr_au_count, int max_active_contexts) 207 { 208 return hweight64(gru->gs_cbr_map) >= cbr_au_count 209 && hweight64(gru->gs_dsr_map) >= dsr_au_count 210 && gru->gs_active_contexts < max_active_contexts; 211 } 212 213 /* 214 * TLB manangment requires tracking all GRU chiplets that have loaded a GSEG 215 * context. 216 */ 217 static int gru_load_mm_tracker(struct gru_state *gru, 218 struct gru_thread_state *gts) 219 { 220 struct gru_mm_struct *gms = gts->ts_gms; 221 struct gru_mm_tracker *asids = &gms->ms_asids[gru->gs_gid]; 222 unsigned short ctxbitmap = (1 << gts->ts_ctxnum); 223 int asid; 224 225 spin_lock(&gms->ms_asid_lock); 226 asid = asids->mt_asid; 227 228 spin_lock(&gru->gs_asid_lock); 229 if (asid == 0 || (asids->mt_ctxbitmap == 0 && asids->mt_asid_gen != 230 gru->gs_asid_gen)) { 231 asid = gru_assign_asid(gru); 232 asids->mt_asid = asid; 233 asids->mt_asid_gen = gru->gs_asid_gen; 234 STAT(asid_new); 235 } else { 236 STAT(asid_reuse); 237 } 238 spin_unlock(&gru->gs_asid_lock); 239 240 BUG_ON(asids->mt_ctxbitmap & ctxbitmap); 241 asids->mt_ctxbitmap |= ctxbitmap; 242 if (!test_bit(gru->gs_gid, gms->ms_asidmap)) 243 __set_bit(gru->gs_gid, gms->ms_asidmap); 244 spin_unlock(&gms->ms_asid_lock); 245 246 gru_dbg(grudev, 247 "gid %d, gts %p, gms %p, ctxnum %d, asid 0x%x, asidmap 0x%lx\n", 248 gru->gs_gid, gts, gms, gts->ts_ctxnum, asid, 249 gms->ms_asidmap[0]); 250 return asid; 251 } 252 253 static void gru_unload_mm_tracker(struct gru_state *gru, 254 struct gru_thread_state *gts) 255 { 256 struct gru_mm_struct *gms = gts->ts_gms; 257 struct gru_mm_tracker *asids; 258 unsigned short ctxbitmap; 259 260 asids = &gms->ms_asids[gru->gs_gid]; 261 ctxbitmap = (1 << gts->ts_ctxnum); 262 spin_lock(&gms->ms_asid_lock); 263 spin_lock(&gru->gs_asid_lock); 264 BUG_ON((asids->mt_ctxbitmap & ctxbitmap) != ctxbitmap); 265 asids->mt_ctxbitmap ^= ctxbitmap; 266 gru_dbg(grudev, "gid %d, gts %p, gms %p, ctxnum 0x%d, asidmap 0x%lx\n", 267 gru->gs_gid, gts, gms, gts->ts_ctxnum, gms->ms_asidmap[0]); 268 spin_unlock(&gru->gs_asid_lock); 269 spin_unlock(&gms->ms_asid_lock); 270 } 271 272 /* 273 * Decrement the reference count on a GTS structure. Free the structure 274 * if the reference count goes to zero. 275 */ 276 void gts_drop(struct gru_thread_state *gts) 277 { 278 if (gts && atomic_dec_return(>s->ts_refcnt) == 0) { 279 gru_drop_mmu_notifier(gts->ts_gms); 280 kfree(gts); 281 STAT(gts_free); 282 } 283 } 284 285 /* 286 * Locate the GTS structure for the current thread. 287 */ 288 static struct gru_thread_state *gru_find_current_gts_nolock(struct gru_vma_data 289 *vdata, int tsid) 290 { 291 struct gru_thread_state *gts; 292 293 list_for_each_entry(gts, &vdata->vd_head, ts_next) 294 if (gts->ts_tsid == tsid) 295 return gts; 296 return NULL; 297 } 298 299 /* 300 * Allocate a thread state structure. 301 */ 302 struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma, 303 int cbr_au_count, int dsr_au_count, int options, int tsid) 304 { 305 struct gru_thread_state *gts; 306 int bytes; 307 308 bytes = DSR_BYTES(dsr_au_count) + CBR_BYTES(cbr_au_count); 309 bytes += sizeof(struct gru_thread_state); 310 gts = kzalloc(bytes, GFP_KERNEL); 311 if (!gts) 312 return NULL; 313 314 STAT(gts_alloc); 315 atomic_set(>s->ts_refcnt, 1); 316 mutex_init(>s->ts_ctxlock); 317 gts->ts_cbr_au_count = cbr_au_count; 318 gts->ts_dsr_au_count = dsr_au_count; 319 gts->ts_user_options = options; 320 gts->ts_tsid = tsid; 321 gts->ts_ctxnum = NULLCTX; 322 gts->ts_tlb_int_select = -1; 323 gts->ts_sizeavail = GRU_SIZEAVAIL(PAGE_SHIFT); 324 if (vma) { 325 gts->ts_mm = current->mm; 326 gts->ts_vma = vma; 327 gts->ts_gms = gru_register_mmu_notifier(); 328 if (!gts->ts_gms) 329 goto err; 330 } 331 332 gru_dbg(grudev, "alloc gts %p\n", gts); 333 return gts; 334 335 err: 336 gts_drop(gts); 337 return NULL; 338 } 339 340 /* 341 * Allocate a vma private data structure. 342 */ 343 struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma, int tsid) 344 { 345 struct gru_vma_data *vdata = NULL; 346 347 vdata = kmalloc(sizeof(*vdata), GFP_KERNEL); 348 if (!vdata) 349 return NULL; 350 351 INIT_LIST_HEAD(&vdata->vd_head); 352 spin_lock_init(&vdata->vd_lock); 353 gru_dbg(grudev, "alloc vdata %p\n", vdata); 354 return vdata; 355 } 356 357 /* 358 * Find the thread state structure for the current thread. 359 */ 360 struct gru_thread_state *gru_find_thread_state(struct vm_area_struct *vma, 361 int tsid) 362 { 363 struct gru_vma_data *vdata = vma->vm_private_data; 364 struct gru_thread_state *gts; 365 366 spin_lock(&vdata->vd_lock); 367 gts = gru_find_current_gts_nolock(vdata, tsid); 368 spin_unlock(&vdata->vd_lock); 369 gru_dbg(grudev, "vma %p, gts %p\n", vma, gts); 370 return gts; 371 } 372 373 /* 374 * Allocate a new thread state for a GSEG. Note that races may allow 375 * another thread to race to create a gts. 376 */ 377 struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct *vma, 378 int tsid) 379 { 380 struct gru_vma_data *vdata = vma->vm_private_data; 381 struct gru_thread_state *gts, *ngts; 382 383 gts = gru_alloc_gts(vma, vdata->vd_cbr_au_count, vdata->vd_dsr_au_count, 384 vdata->vd_user_options, tsid); 385 if (!gts) 386 return NULL; 387 388 spin_lock(&vdata->vd_lock); 389 ngts = gru_find_current_gts_nolock(vdata, tsid); 390 if (ngts) { 391 gts_drop(gts); 392 gts = ngts; 393 STAT(gts_double_allocate); 394 } else { 395 list_add(>s->ts_next, &vdata->vd_head); 396 } 397 spin_unlock(&vdata->vd_lock); 398 gru_dbg(grudev, "vma %p, gts %p\n", vma, gts); 399 return gts; 400 } 401 402 /* 403 * Free the GRU context assigned to the thread state. 404 */ 405 static void gru_free_gru_context(struct gru_thread_state *gts) 406 { 407 struct gru_state *gru; 408 409 gru = gts->ts_gru; 410 gru_dbg(grudev, "gts %p, gid %d\n", gts, gru->gs_gid); 411 412 spin_lock(&gru->gs_lock); 413 gru->gs_gts[gts->ts_ctxnum] = NULL; 414 free_gru_resources(gru, gts); 415 BUG_ON(test_bit(gts->ts_ctxnum, &gru->gs_context_map) == 0); 416 __clear_bit(gts->ts_ctxnum, &gru->gs_context_map); 417 gts->ts_ctxnum = NULLCTX; 418 gts->ts_gru = NULL; 419 gts->ts_blade = -1; 420 spin_unlock(&gru->gs_lock); 421 422 gts_drop(gts); 423 STAT(free_context); 424 } 425 426 /* 427 * Prefetching cachelines help hardware performance. 428 * (Strictly a performance enhancement. Not functionally required). 429 */ 430 static void prefetch_data(void *p, int num, int stride) 431 { 432 while (num-- > 0) { 433 prefetchw(p); 434 p += stride; 435 } 436 } 437 438 static inline long gru_copy_handle(void *d, void *s) 439 { 440 memcpy(d, s, GRU_HANDLE_BYTES); 441 return GRU_HANDLE_BYTES; 442 } 443 444 static void gru_prefetch_context(void *gseg, void *cb, void *cbe, 445 unsigned long cbrmap, unsigned long length) 446 { 447 int i, scr; 448 449 prefetch_data(gseg + GRU_DS_BASE, length / GRU_CACHE_LINE_BYTES, 450 GRU_CACHE_LINE_BYTES); 451 452 for_each_cbr_in_allocation_map(i, &cbrmap, scr) { 453 prefetch_data(cb, 1, GRU_CACHE_LINE_BYTES); 454 prefetch_data(cbe + i * GRU_HANDLE_STRIDE, 1, 455 GRU_CACHE_LINE_BYTES); 456 cb += GRU_HANDLE_STRIDE; 457 } 458 } 459 460 static void gru_load_context_data(void *save, void *grubase, int ctxnum, 461 unsigned long cbrmap, unsigned long dsrmap) 462 { 463 void *gseg, *cb, *cbe; 464 unsigned long length; 465 int i, scr; 466 467 gseg = grubase + ctxnum * GRU_GSEG_STRIDE; 468 cb = gseg + GRU_CB_BASE; 469 cbe = grubase + GRU_CBE_BASE; 470 length = hweight64(dsrmap) * GRU_DSR_AU_BYTES; 471 gru_prefetch_context(gseg, cb, cbe, cbrmap, length); 472 473 for_each_cbr_in_allocation_map(i, &cbrmap, scr) { 474 save += gru_copy_handle(cb, save); 475 save += gru_copy_handle(cbe + i * GRU_HANDLE_STRIDE, save); 476 cb += GRU_HANDLE_STRIDE; 477 } 478 479 memcpy(gseg + GRU_DS_BASE, save, length); 480 } 481 482 static void gru_unload_context_data(void *save, void *grubase, int ctxnum, 483 unsigned long cbrmap, unsigned long dsrmap) 484 { 485 void *gseg, *cb, *cbe; 486 unsigned long length; 487 int i, scr; 488 489 gseg = grubase + ctxnum * GRU_GSEG_STRIDE; 490 cb = gseg + GRU_CB_BASE; 491 cbe = grubase + GRU_CBE_BASE; 492 length = hweight64(dsrmap) * GRU_DSR_AU_BYTES; 493 gru_prefetch_context(gseg, cb, cbe, cbrmap, length); 494 495 for_each_cbr_in_allocation_map(i, &cbrmap, scr) { 496 save += gru_copy_handle(save, cb); 497 save += gru_copy_handle(save, cbe + i * GRU_HANDLE_STRIDE); 498 cb += GRU_HANDLE_STRIDE; 499 } 500 memcpy(save, gseg + GRU_DS_BASE, length); 501 } 502 503 void gru_unload_context(struct gru_thread_state *gts, int savestate) 504 { 505 struct gru_state *gru = gts->ts_gru; 506 struct gru_context_configuration_handle *cch; 507 int ctxnum = gts->ts_ctxnum; 508 509 zap_vma_ptes(gts->ts_vma, UGRUADDR(gts), GRU_GSEG_PAGESIZE); 510 cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); 511 512 gru_dbg(grudev, "gts %p\n", gts); 513 lock_cch_handle(cch); 514 if (cch_interrupt_sync(cch)) 515 BUG(); 516 517 gru_unload_mm_tracker(gru, gts); 518 if (savestate) 519 gru_unload_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, 520 ctxnum, gts->ts_cbr_map, 521 gts->ts_dsr_map); 522 523 if (cch_deallocate(cch)) 524 BUG(); 525 gts->ts_force_unload = 0; /* ts_force_unload locked by CCH lock */ 526 unlock_cch_handle(cch); 527 528 gru_free_gru_context(gts); 529 STAT(unload_context); 530 } 531 532 /* 533 * Load a GRU context by copying it from the thread data structure in memory 534 * to the GRU. 535 */ 536 void gru_load_context(struct gru_thread_state *gts) 537 { 538 struct gru_state *gru = gts->ts_gru; 539 struct gru_context_configuration_handle *cch; 540 int i, err, asid, ctxnum = gts->ts_ctxnum; 541 542 gru_dbg(grudev, "gts %p\n", gts); 543 cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); 544 545 lock_cch_handle(cch); 546 cch->tfm_fault_bit_enable = 547 (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL 548 || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); 549 cch->tlb_int_enable = (gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); 550 if (cch->tlb_int_enable) { 551 gts->ts_tlb_int_select = gru_cpu_fault_map_id(); 552 cch->tlb_int_select = gts->ts_tlb_int_select; 553 } 554 cch->tfm_done_bit_enable = 0; 555 cch->dsr_allocation_map = gts->ts_dsr_map; 556 cch->cbr_allocation_map = gts->ts_cbr_map; 557 asid = gru_load_mm_tracker(gru, gts); 558 cch->unmap_enable = 0; 559 for (i = 0; i < 8; i++) { 560 cch->asid[i] = asid + i; 561 cch->sizeavail[i] = gts->ts_sizeavail; 562 } 563 564 err = cch_allocate(cch); 565 if (err) { 566 gru_dbg(grudev, 567 "err %d: cch %p, gts %p, cbr 0x%lx, dsr 0x%lx\n", 568 err, cch, gts, gts->ts_cbr_map, gts->ts_dsr_map); 569 BUG(); 570 } 571 572 gru_load_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, ctxnum, 573 gts->ts_cbr_map, gts->ts_dsr_map); 574 575 if (cch_start(cch)) 576 BUG(); 577 unlock_cch_handle(cch); 578 579 STAT(load_context); 580 } 581 582 /* 583 * Update fields in an active CCH: 584 * - retarget interrupts on local blade 585 * - update sizeavail mask 586 * - force a delayed context unload by clearing the CCH asids. This 587 * forces TLB misses for new GRU instructions. The context is unloaded 588 * when the next TLB miss occurs. 589 */ 590 int gru_update_cch(struct gru_thread_state *gts, int force_unload) 591 { 592 struct gru_context_configuration_handle *cch; 593 struct gru_state *gru = gts->ts_gru; 594 int i, ctxnum = gts->ts_ctxnum, ret = 0; 595 596 cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); 597 598 lock_cch_handle(cch); 599 if (cch->state == CCHSTATE_ACTIVE) { 600 if (gru->gs_gts[gts->ts_ctxnum] != gts) 601 goto exit; 602 if (cch_interrupt(cch)) 603 BUG(); 604 if (!force_unload) { 605 for (i = 0; i < 8; i++) 606 cch->sizeavail[i] = gts->ts_sizeavail; 607 gts->ts_tlb_int_select = gru_cpu_fault_map_id(); 608 cch->tlb_int_select = gru_cpu_fault_map_id(); 609 cch->tfm_fault_bit_enable = 610 (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL 611 || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); 612 } else { 613 for (i = 0; i < 8; i++) 614 cch->asid[i] = 0; 615 cch->tfm_fault_bit_enable = 0; 616 cch->tlb_int_enable = 0; 617 gts->ts_force_unload = 1; 618 } 619 if (cch_start(cch)) 620 BUG(); 621 ret = 1; 622 } 623 exit: 624 unlock_cch_handle(cch); 625 return ret; 626 } 627 628 /* 629 * Update CCH tlb interrupt select. Required when all the following is true: 630 * - task's GRU context is loaded into a GRU 631 * - task is using interrupt notification for TLB faults 632 * - task has migrated to a different cpu on the same blade where 633 * it was previously running. 634 */ 635 static int gru_retarget_intr(struct gru_thread_state *gts) 636 { 637 if (gts->ts_tlb_int_select < 0 638 || gts->ts_tlb_int_select == gru_cpu_fault_map_id()) 639 return 0; 640 641 gru_dbg(grudev, "retarget from %d to %d\n", gts->ts_tlb_int_select, 642 gru_cpu_fault_map_id()); 643 return gru_update_cch(gts, 0); 644 } 645 646 647 /* 648 * Insufficient GRU resources available on the local blade. Steal a context from 649 * a process. This is a hack until a _real_ resource scheduler is written.... 650 */ 651 #define next_ctxnum(n) ((n) < GRU_NUM_CCH - 2 ? (n) + 1 : 0) 652 #define next_gru(b, g) (((g) < &(b)->bs_grus[GRU_CHIPLETS_PER_BLADE - 1]) ? \ 653 ((g)+1) : &(b)->bs_grus[0]) 654 655 void gru_steal_context(struct gru_thread_state *gts, int blade_id) 656 { 657 struct gru_blade_state *blade; 658 struct gru_state *gru, *gru0; 659 struct gru_thread_state *ngts = NULL; 660 int ctxnum, ctxnum0, flag = 0, cbr, dsr; 661 662 cbr = gts->ts_cbr_au_count; 663 dsr = gts->ts_dsr_au_count; 664 665 blade = gru_base[blade_id]; 666 spin_lock(&blade->bs_lock); 667 668 ctxnum = next_ctxnum(blade->bs_lru_ctxnum); 669 gru = blade->bs_lru_gru; 670 if (ctxnum == 0) 671 gru = next_gru(blade, gru); 672 ctxnum0 = ctxnum; 673 gru0 = gru; 674 while (1) { 675 if (check_gru_resources(gru, cbr, dsr, GRU_NUM_CCH)) 676 break; 677 spin_lock(&gru->gs_lock); 678 for (; ctxnum < GRU_NUM_CCH; ctxnum++) { 679 if (flag && gru == gru0 && ctxnum == ctxnum0) 680 break; 681 ngts = gru->gs_gts[ctxnum]; 682 /* 683 * We are grabbing locks out of order, so trylock is 684 * needed. GTSs are usually not locked, so the odds of 685 * success are high. If trylock fails, try to steal a 686 * different GSEG. 687 */ 688 if (ngts && mutex_trylock(&ngts->ts_ctxlock)) 689 break; 690 ngts = NULL; 691 flag = 1; 692 } 693 spin_unlock(&gru->gs_lock); 694 if (ngts || (flag && gru == gru0 && ctxnum == ctxnum0)) 695 break; 696 ctxnum = 0; 697 gru = next_gru(blade, gru); 698 } 699 blade->bs_lru_gru = gru; 700 blade->bs_lru_ctxnum = ctxnum; 701 spin_unlock(&blade->bs_lock); 702 703 if (ngts) { 704 STAT(steal_context); 705 ngts->ts_steal_jiffies = jiffies; 706 gru_unload_context(ngts, 1); 707 mutex_unlock(&ngts->ts_ctxlock); 708 } else { 709 STAT(steal_context_failed); 710 } 711 gru_dbg(grudev, 712 "stole gid %d, ctxnum %d from gts %p. Need cb %d, ds %d;" 713 " avail cb %ld, ds %ld\n", 714 gru->gs_gid, ctxnum, ngts, cbr, dsr, hweight64(gru->gs_cbr_map), 715 hweight64(gru->gs_dsr_map)); 716 } 717 718 /* 719 * Scan the GRUs on the local blade & assign a GRU context. 720 */ 721 struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts, 722 int blade) 723 { 724 struct gru_state *gru, *grux; 725 int i, max_active_contexts; 726 727 728 again: 729 gru = NULL; 730 max_active_contexts = GRU_NUM_CCH; 731 for_each_gru_on_blade(grux, blade, i) { 732 if (check_gru_resources(grux, gts->ts_cbr_au_count, 733 gts->ts_dsr_au_count, 734 max_active_contexts)) { 735 gru = grux; 736 max_active_contexts = grux->gs_active_contexts; 737 if (max_active_contexts == 0) 738 break; 739 } 740 } 741 742 if (gru) { 743 spin_lock(&gru->gs_lock); 744 if (!check_gru_resources(gru, gts->ts_cbr_au_count, 745 gts->ts_dsr_au_count, GRU_NUM_CCH)) { 746 spin_unlock(&gru->gs_lock); 747 goto again; 748 } 749 reserve_gru_resources(gru, gts); 750 gts->ts_gru = gru; 751 gts->ts_blade = gru->gs_blade_id; 752 gts->ts_ctxnum = 753 find_first_zero_bit(&gru->gs_context_map, GRU_NUM_CCH); 754 BUG_ON(gts->ts_ctxnum == GRU_NUM_CCH); 755 atomic_inc(>s->ts_refcnt); 756 gru->gs_gts[gts->ts_ctxnum] = gts; 757 __set_bit(gts->ts_ctxnum, &gru->gs_context_map); 758 spin_unlock(&gru->gs_lock); 759 760 STAT(assign_context); 761 gru_dbg(grudev, 762 "gseg %p, gts %p, gid %d, ctx %d, cbr %d, dsr %d\n", 763 gseg_virtual_address(gts->ts_gru, gts->ts_ctxnum), gts, 764 gts->ts_gru->gs_gid, gts->ts_ctxnum, 765 gts->ts_cbr_au_count, gts->ts_dsr_au_count); 766 } else { 767 gru_dbg(grudev, "failed to allocate a GTS %s\n", ""); 768 STAT(assign_context_failed); 769 } 770 771 return gru; 772 } 773 774 /* 775 * gru_nopage 776 * 777 * Map the user's GRU segment 778 * 779 * Note: gru segments alway mmaped on GRU_GSEG_PAGESIZE boundaries. 780 */ 781 int gru_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 782 { 783 struct gru_thread_state *gts; 784 unsigned long paddr, vaddr; 785 int blade_id; 786 787 vaddr = (unsigned long)vmf->virtual_address; 788 gru_dbg(grudev, "vma %p, vaddr 0x%lx (0x%lx)\n", 789 vma, vaddr, GSEG_BASE(vaddr)); 790 STAT(nopfn); 791 792 /* The following check ensures vaddr is a valid address in the VMA */ 793 gts = gru_find_thread_state(vma, TSID(vaddr, vma)); 794 if (!gts) 795 return VM_FAULT_SIGBUS; 796 797 again: 798 mutex_lock(>s->ts_ctxlock); 799 preempt_disable(); 800 blade_id = uv_numa_blade_id(); 801 802 if (gts->ts_gru) { 803 if (gts->ts_gru->gs_blade_id != blade_id) { 804 STAT(migrated_nopfn_unload); 805 gru_unload_context(gts, 1); 806 } else { 807 if (gru_retarget_intr(gts)) 808 STAT(migrated_nopfn_retarget); 809 } 810 } 811 812 if (!gts->ts_gru) { 813 if (!gru_assign_gru_context(gts, blade_id)) { 814 preempt_enable(); 815 mutex_unlock(>s->ts_ctxlock); 816 set_current_state(TASK_INTERRUPTIBLE); 817 schedule_timeout(GRU_ASSIGN_DELAY); /* true hack ZZZ */ 818 blade_id = uv_numa_blade_id(); 819 if (gts->ts_steal_jiffies + GRU_STEAL_DELAY < jiffies) 820 gru_steal_context(gts, blade_id); 821 goto again; 822 } 823 gru_load_context(gts); 824 paddr = gseg_physical_address(gts->ts_gru, gts->ts_ctxnum); 825 remap_pfn_range(vma, vaddr & ~(GRU_GSEG_PAGESIZE - 1), 826 paddr >> PAGE_SHIFT, GRU_GSEG_PAGESIZE, 827 vma->vm_page_prot); 828 } 829 830 preempt_enable(); 831 mutex_unlock(>s->ts_ctxlock); 832 833 return VM_FAULT_NOPAGE; 834 } 835 836