arch/powerpc/platforms/pseries/lpar.c

   1 /*
   2  * pSeries_lpar.c
   3  * Copyright (C) 2001 Todd Inglett, IBM Corporation
   4  *
   5  * pSeries LPAR support.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  20  */
  21
  22 /* Enables debugging of low-level hash table routines - careful! */
  23 #undef DEBUG
  24
  25 #include <linux/kernel.h>
  26 #include <linux/dma-mapping.h>
  27 #include <linux/console.h>
  28 #include <linux/export.h>
  29 #include <linux/jump_label.h>
  30 #include <linux/delay.h>
  31 #include <linux/stop_machine.h>
  32 #include <asm/processor.h>
  33 #include <asm/mmu.h>
  34 #include <asm/page.h>
  35 #include <asm/pgtable.h>
  36 #include <asm/machdep.h>
  37 #include <asm/mmu_context.h>
  38 #include <asm/iommu.h>
  39 #include <asm/tlbflush.h>
  40 #include <asm/tlb.h>
  41 #include <asm/prom.h>
  42 #include <asm/cputable.h>
  43 #include <asm/udbg.h>
  44 #include <asm/smp.h>
  45 #include <asm/trace.h>
  46 #include <asm/firmware.h>
  47 #include <asm/plpar_wrappers.h>
  48 #include <asm/kexec.h>
  49 #include <asm/fadump.h>
  50 #include <asm/asm-prototypes.h>
  51 #include <asm/debugfs.h>
  52
  53 #include "pseries.h"
  54
  55 /* Flag bits for H_BULK_REMOVE */
  56 #define HBR_REQUEST     0x4000000000000000UL
  57 #define HBR_RESPONSE    0x8000000000000000UL
  58 #define HBR_END         0xc000000000000000UL
  59 #define HBR_AVPN        0x0200000000000000UL
  60 #define HBR_ANDCOND     0x0100000000000000UL
  61
  62
  63 /* in hvCall.S */
  64 EXPORT_SYMBOL(plpar_hcall);
  65 EXPORT_SYMBOL(plpar_hcall9);
  66 EXPORT_SYMBOL(plpar_hcall_norets);
  67
  68 void vpa_init(int cpu)
  69 {
  70         int hwcpu = get_hard_smp_processor_id(cpu);
  71         unsigned long addr;
  72         long ret;
  73         struct paca_struct *pp;
  74         struct dtl_entry *dtl;
  75
  76         /*
  77          * The spec says it "may be problematic" if CPU x registers the VPA of
  78          * CPU y. We should never do that, but wail if we ever do.
  79          */
  80         WARN_ON(cpu != smp_processor_id());
  81
  82         if (cpu_has_feature(CPU_FTR_ALTIVEC))
  83                 lppaca_of(cpu).vmxregs_in_use = 1;
  84
  85         if (cpu_has_feature(CPU_FTR_ARCH_207S))
  86                 lppaca_of(cpu).ebb_regs_in_use = 1;
  87
  88         addr = __pa(&lppaca_of(cpu));
  89         ret = register_vpa(hwcpu, addr);
  90
  91         if (ret) {
  92                 pr_err("WARNING: VPA registration for cpu %d (hw %d) of area "
  93                        "%lx failed with %ld\n", cpu, hwcpu, addr, ret);
  94                 return;
  95         }
  96
  97 #ifdef CONFIG_PPC_STD_MMU_64
  98         /*
  99          * PAPR says this feature is SLB-Buffer but firmware never
 100          * reports that.  All SPLPAR support SLB shadow buffer.
 101          */
 102         if (!radix_enabled() && firmware_has_feature(FW_FEATURE_SPLPAR)) {
 103                 addr = __pa(paca[cpu].slb_shadow_ptr);
 104                 ret = register_slb_shadow(hwcpu, addr);
 105                 if (ret)
 106                         pr_err("WARNING: SLB shadow buffer registration for "
 107                                "cpu %d (hw %d) of area %lx failed with %ld\n",
 108                                cpu, hwcpu, addr, ret);
 109         }
 110 #endif /* CONFIG_PPC_STD_MMU_64 */
 111
 112         /*
 113          * Register dispatch trace log, if one has been allocated.
 114          */
 115         pp = &paca[cpu];
 116         dtl = pp->dispatch_log;
 117         if (dtl) {
 118                 pp->dtl_ridx = 0;
 119                 pp->dtl_curr = dtl;
 120                 lppaca_of(cpu).dtl_idx = 0;
 121
 122                 /* hypervisor reads buffer length from this field */
 123                 dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
 124                 ret = register_dtl(hwcpu, __pa(dtl));
 125                 if (ret)
 126                         pr_err("WARNING: DTL registration of cpu %d (hw %d) "
 127                                "failed with %ld\n", smp_processor_id(),
 128                                hwcpu, ret);
 129                 lppaca_of(cpu).dtl_enable_mask = 2;
 130         }
 131 }
 132
 133 #ifdef CONFIG_PPC_STD_MMU_64
 134
 135 static long pSeries_lpar_hpte_insert(unsigned long hpte_group,
 136                                      unsigned long vpn, unsigned long pa,
 137                                      unsigned long rflags, unsigned long vflags,
 138                                      int psize, int apsize, int ssize)
 139 {
 140         unsigned long lpar_rc;
 141         unsigned long flags;
 142         unsigned long slot;
 143         unsigned long hpte_v, hpte_r;
 144
 145         if (!(vflags & HPTE_V_BOLTED))
 146                 pr_devel("hpte_insert(group=%lx, vpn=%016lx, "
 147                          "pa=%016lx, rflags=%lx, vflags=%lx, psize=%d)\n",
 148                          hpte_group, vpn,  pa, rflags, vflags, psize);
 149
 150         hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
 151         hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
 152
 153         if (!(vflags & HPTE_V_BOLTED))
 154                 pr_devel(" hpte_v=%016lx, hpte_r=%016lx\n", hpte_v, hpte_r);
 155
 156         /* Now fill in the actual HPTE */
 157         /* Set CEC cookie to 0         */
 158         /* Zero page = 0               */
 159         /* I-cache Invalidate = 0      */
 160         /* I-cache synchronize = 0     */
 161         /* Exact = 0                   */
 162         flags = 0;
 163
 164         if (firmware_has_feature(FW_FEATURE_XCMO) && !(hpte_r & HPTE_R_N))
 165                 flags |= H_COALESCE_CAND;
 166
 167         lpar_rc = plpar_pte_enter(flags, hpte_group, hpte_v, hpte_r, &slot);
 168         if (unlikely(lpar_rc == H_PTEG_FULL)) {
 169                 if (!(vflags & HPTE_V_BOLTED))
 170                         pr_devel(" full\n");
 171                 return -1;
 172         }
 173
 174         /*
 175          * Since we try and ioremap PHBs we don't own, the pte insert
 176          * will fail. However we must catch the failure in hash_page
 177          * or we will loop forever, so return -2 in this case.
 178          */
 179         if (unlikely(lpar_rc != H_SUCCESS)) {
 180                 if (!(vflags & HPTE_V_BOLTED))
 181                         pr_devel(" lpar err %ld\n", lpar_rc);
 182                 return -2;
 183         }
 184         if (!(vflags & HPTE_V_BOLTED))
 185                 pr_devel(" -> slot: %lu\n", slot & 7);
 186
 187         /* Because of iSeries, we have to pass down the secondary
 188          * bucket bit here as well
 189          */
 190         return (slot & 7) | (!!(vflags & HPTE_V_SECONDARY) << 3);
 191 }
 192
 193 static DEFINE_SPINLOCK(pSeries_lpar_tlbie_lock);
 194
 195 static long pSeries_lpar_hpte_remove(unsigned long hpte_group)
 196 {
 197         unsigned long slot_offset;
 198         unsigned long lpar_rc;
 199         int i;
 200         unsigned long dummy1, dummy2;
 201
 202         /* pick a random slot to start at */
 203         slot_offset = mftb() & 0x7;
 204
 205         for (i = 0; i < HPTES_PER_GROUP; i++) {
 206
 207                 /* don't remove a bolted entry */
 208                 lpar_rc = plpar_pte_remove(H_ANDCOND, hpte_group + slot_offset,
 209                                            (0x1UL << 4), &dummy1, &dummy2);
 210                 if (lpar_rc == H_SUCCESS)
 211                         return i;
 212
 213                 /*
 214                  * The test for adjunct partition is performed before the
 215                  * ANDCOND test.  H_RESOURCE may be returned, so we need to
 216                  * check for that as well.
 217                  */
 218                 BUG_ON(lpar_rc != H_NOT_FOUND && lpar_rc != H_RESOURCE);
 219
 220                 slot_offset++;
 221                 slot_offset &= 0x7;
 222         }
 223
 224         return -1;
 225 }
 226
 227 static void manual_hpte_clear_all(void)
 228 {
 229         unsigned long size_bytes = 1UL << ppc64_pft_size;
 230         unsigned long hpte_count = size_bytes >> 4;
 231         struct {
 232                 unsigned long pteh;
 233                 unsigned long ptel;
 234         } ptes[4];
 235         long lpar_rc;
 236         unsigned long i, j;
 237
 238         /* Read in batches of 4,
 239          * invalidate only valid entries not in the VRMA
 240          * hpte_count will be a multiple of 4
 241          */
 242         for (i = 0; i < hpte_count; i += 4) {
 243                 lpar_rc = plpar_pte_read_4_raw(0, i, (void *)ptes);
 244                 if (lpar_rc != H_SUCCESS)
 245                         continue;
 246                 for (j = 0; j < 4; j++){
 247                         if ((ptes[j].pteh & HPTE_V_VRMA_MASK) ==
 248                                 HPTE_V_VRMA_MASK)
 249                                 continue;
 250                         if (ptes[j].pteh & HPTE_V_VALID)
 251                                 plpar_pte_remove_raw(0, i + j, 0,
 252                                         &(ptes[j].pteh), &(ptes[j].ptel));
 253                 }
 254         }
 255 }
 256
 257 static int hcall_hpte_clear_all(void)
 258 {
 259         int rc;
 260
 261         do {
 262                 rc = plpar_hcall_norets(H_CLEAR_HPT);
 263         } while (rc == H_CONTINUE);
 264
 265         return rc;
 266 }
 267
 268 static void pseries_hpte_clear_all(void)
 269 {
 270         int rc;
 271
 272         rc = hcall_hpte_clear_all();
 273         if (rc != H_SUCCESS)
 274                 manual_hpte_clear_all();
 275
 276 #ifdef __LITTLE_ENDIAN__
 277         /*
 278          * Reset exceptions to big endian.
 279          *
 280          * FIXME this is a hack for kexec, we need to reset the exception
 281          * endian before starting the new kernel and this is a convenient place
 282          * to do it.
 283          *
 284          * This is also called on boot when a fadump happens. In that case we
 285          * must not change the exception endian mode.
 286          */
 287         if (firmware_has_feature(FW_FEATURE_SET_MODE) && !is_fadump_active())
 288                 pseries_big_endian_exceptions();
 289 #endif
 290 }
 291
 292 /*
 293  * NOTE: for updatepp ops we are fortunate that the linux "newpp" bits and
 294  * the low 3 bits of flags happen to line up.  So no transform is needed.
 295  * We can probably optimize here and assume the high bits of newpp are
 296  * already zero.  For now I am paranoid.
 297  */
 298 static long pSeries_lpar_hpte_updatepp(unsigned long slot,
 299                                        unsigned long newpp,
 300                                        unsigned long vpn,
 301                                        int psize, int apsize,
 302                                        int ssize, unsigned long inv_flags)
 303 {
 304         unsigned long lpar_rc;
 305         unsigned long flags;
 306         unsigned long want_v;
 307
 308         want_v = hpte_encode_avpn(vpn, psize, ssize);
 309
 310         pr_devel("    update: avpnv=%016lx, hash=%016lx, f=%lx, psize: %d ...",
 311                  want_v, slot, flags, psize);
 312
 313         flags = (newpp & 7) | H_AVPN;
 314         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 315                 /* Move pp0 into bit 8 (IBM 55) */
 316                 flags |= (newpp & HPTE_R_PP0) >> 55;
 317
 318         lpar_rc = plpar_pte_protect(flags, slot, want_v);
 319
 320         if (lpar_rc == H_NOT_FOUND) {
 321                 pr_devel("not found !\n");
 322                 return -1;
 323         }
 324
 325         pr_devel("ok\n");
 326
 327         BUG_ON(lpar_rc != H_SUCCESS);
 328
 329         return 0;
 330 }
 331
 332 static long __pSeries_lpar_hpte_find(unsigned long want_v, unsigned long hpte_group)
 333 {
 334         long lpar_rc;
 335         unsigned long i, j;
 336         struct {
 337                 unsigned long pteh;
 338                 unsigned long ptel;
 339         } ptes[4];
 340
 341         for (i = 0; i < HPTES_PER_GROUP; i += 4, hpte_group += 4) {
 342
 343                 lpar_rc = plpar_pte_read_4(0, hpte_group, (void *)ptes);
 344                 if (lpar_rc != H_SUCCESS)
 345                         continue;
 346
 347                 for (j = 0; j < 4; j++) {
 348                         if (HPTE_V_COMPARE(ptes[j].pteh, want_v) &&
 349                             (ptes[j].pteh & HPTE_V_VALID))
 350                                 return i + j;
 351                 }
 352         }
 353
 354         return -1;
 355 }
 356
 357 static long pSeries_lpar_hpte_find(unsigned long vpn, int psize, int ssize)
 358 {
 359         long slot;
 360         unsigned long hash;
 361         unsigned long want_v;
 362         unsigned long hpte_group;
 363
 364         hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
 365         want_v = hpte_encode_avpn(vpn, psize, ssize);
 366
 367         /* Bolted entries are always in the primary group */
 368         hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 369         slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
 370         if (slot < 0)
 371                 return -1;
 372         return hpte_group + slot;
 373 }
 374
 375 static void pSeries_lpar_hpte_updateboltedpp(unsigned long newpp,
 376                                              unsigned long ea,
 377                                              int psize, int ssize)
 378 {
 379         unsigned long vpn;
 380         unsigned long lpar_rc, slot, vsid, flags;
 381
 382         vsid = get_kernel_vsid(ea, ssize);
 383         vpn = hpt_vpn(ea, vsid, ssize);
 384
 385         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
 386         BUG_ON(slot == -1);
 387
 388         flags = newpp & 7;
 389         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 390                 /* Move pp0 into bit 8 (IBM 55) */
 391                 flags |= (newpp & HPTE_R_PP0) >> 55;
 392
 393         lpar_rc = plpar_pte_protect(flags, slot, 0);
 394
 395         BUG_ON(lpar_rc != H_SUCCESS);
 396 }
 397
 398 static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
 399                                          int psize, int apsize,
 400                                          int ssize, int local)
 401 {
 402         unsigned long want_v;
 403         unsigned long lpar_rc;
 404         unsigned long dummy1, dummy2;
 405
 406         pr_devel("    inval : slot=%lx, vpn=%016lx, psize: %d, local: %d\n",
 407                  slot, vpn, psize, local);
 408
 409         want_v = hpte_encode_avpn(vpn, psize, ssize);
 410         lpar_rc = plpar_pte_remove(H_AVPN, slot, want_v, &dummy1, &dummy2);
 411         if (lpar_rc == H_NOT_FOUND)
 412                 return;
 413
 414         BUG_ON(lpar_rc != H_SUCCESS);
 415 }
 416
 417 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 418 /*
 419  * Limit iterations holding pSeries_lpar_tlbie_lock to 3. We also need
 420  * to make sure that we avoid bouncing the hypervisor tlbie lock.
 421  */
 422 #define PPC64_HUGE_HPTE_BATCH 12
 423
 424 static void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
 425                                              unsigned long *vpn, int count,
 426                                              int psize, int ssize)
 427 {
 428         unsigned long param[PLPAR_HCALL9_BUFSIZE];
 429         int i = 0, pix = 0, rc;
 430         unsigned long flags = 0;
 431         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
 432
 433         if (lock_tlbie)
 434                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
 435
 436         for (i = 0; i < count; i++) {
 437
 438                 if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
 439                         pSeries_lpar_hpte_invalidate(slot[i], vpn[i], psize, 0,
 440                                                      ssize, 0);
 441                 } else {
 442                         param[pix] = HBR_REQUEST | HBR_AVPN | slot[i];
 443                         param[pix+1] = hpte_encode_avpn(vpn[i], psize, ssize);
 444                         pix += 2;
 445                         if (pix == 8) {
 446                                 rc = plpar_hcall9(H_BULK_REMOVE, param,
 447                                                   param[0], param[1], param[2],
 448                                                   param[3], param[4], param[5],
 449                                                   param[6], param[7]);
 450                                 BUG_ON(rc != H_SUCCESS);
 451                                 pix = 0;
 452                         }
 453                 }
 454         }
 455         if (pix) {
 456                 param[pix] = HBR_END;
 457                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
 458                                   param[2], param[3], param[4], param[5],
 459                                   param[6], param[7]);
 460                 BUG_ON(rc != H_SUCCESS);
 461         }
 462
 463         if (lock_tlbie)
 464                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
 465 }
 466
 467 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
 468                                              unsigned long addr,
 469                                              unsigned char *hpte_slot_array,
 470                                              int psize, int ssize, int local)
 471 {
 472         int i, index = 0;
 473         unsigned long s_addr = addr;
 474         unsigned int max_hpte_count, valid;
 475         unsigned long vpn_array[PPC64_HUGE_HPTE_BATCH];
 476         unsigned long slot_array[PPC64_HUGE_HPTE_BATCH];
 477         unsigned long shift, hidx, vpn = 0, hash, slot;
 478
 479         shift = mmu_psize_defs[psize].shift;
 480         max_hpte_count = 1U << (PMD_SHIFT - shift);
 481
 482         for (i = 0; i < max_hpte_count; i++) {
 483                 valid = hpte_valid(hpte_slot_array, i);
 484                 if (!valid)
 485                         continue;
 486                 hidx =  hpte_hash_index(hpte_slot_array, i);
 487
 488                 /* get the vpn */
 489                 addr = s_addr + (i * (1ul << shift));
 490                 vpn = hpt_vpn(addr, vsid, ssize);
 491                 hash = hpt_hash(vpn, shift, ssize);
 492                 if (hidx & _PTEIDX_SECONDARY)
 493                         hash = ~hash;
 494
 495                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 496                 slot += hidx & _PTEIDX_GROUP_IX;
 497
 498                 slot_array[index] = slot;
 499                 vpn_array[index] = vpn;
 500                 if (index == PPC64_HUGE_HPTE_BATCH - 1) {
 501                         /*
 502                          * Now do a bluk invalidate
 503                          */
 504                         __pSeries_lpar_hugepage_invalidate(slot_array,
 505                                                            vpn_array,
 506                                                            PPC64_HUGE_HPTE_BATCH,
 507                                                            psize, ssize);
 508                         index = 0;
 509                 } else
 510                         index++;
 511         }
 512         if (index)
 513                 __pSeries_lpar_hugepage_invalidate(slot_array, vpn_array,
 514                                                    index, psize, ssize);
 515 }
 516 #else
 517 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
 518                                              unsigned long addr,
 519                                              unsigned char *hpte_slot_array,
 520                                              int psize, int ssize, int local)
 521 {
 522         WARN(1, "%s called without THP support\n", __func__);
 523 }
 524 #endif
 525
 526 static int pSeries_lpar_hpte_removebolted(unsigned long ea,
 527                                           int psize, int ssize)
 528 {
 529         unsigned long vpn;
 530         unsigned long slot, vsid;
 531
 532         vsid = get_kernel_vsid(ea, ssize);
 533         vpn = hpt_vpn(ea, vsid, ssize);
 534
 535         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
 536         if (slot == -1)
 537                 return -ENOENT;
 538
 539         /*
 540          * lpar doesn't use the passed actual page size
 541          */
 542         pSeries_lpar_hpte_invalidate(slot, vpn, psize, 0, ssize, 0);
 543         return 0;
 544 }
 545
 546 /*
 547  * Take a spinlock around flushes to avoid bouncing the hypervisor tlbie
 548  * lock.
 549  */
 550 static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
 551 {
 552         unsigned long vpn;
 553         unsigned long i, pix, rc;
 554         unsigned long flags = 0;
 555         struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
 556         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
 557         unsigned long param[PLPAR_HCALL9_BUFSIZE];
 558         unsigned long hash, index, shift, hidx, slot;
 559         real_pte_t pte;
 560         int psize, ssize;
 561
 562         if (lock_tlbie)
 563                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
 564
 565         psize = batch->psize;
 566         ssize = batch->ssize;
 567         pix = 0;
 568         for (i = 0; i < number; i++) {
 569                 vpn = batch->vpn[i];
 570                 pte = batch->pte[i];
 571                 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
 572                         hash = hpt_hash(vpn, shift, ssize);
 573                         hidx = __rpte_to_hidx(pte, index);
 574                         if (hidx & _PTEIDX_SECONDARY)
 575                                 hash = ~hash;
 576                         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 577                         slot += hidx & _PTEIDX_GROUP_IX;
 578                         if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
 579                                 /*
 580                                  * lpar doesn't use the passed actual page size
 581                                  */
 582                                 pSeries_lpar_hpte_invalidate(slot, vpn, psize,
 583                                                              0, ssize, local);
 584                         } else {
 585                                 param[pix] = HBR_REQUEST | HBR_AVPN | slot;
 586                                 param[pix+1] = hpte_encode_avpn(vpn, psize,
 587                                                                 ssize);
 588                                 pix += 2;
 589                                 if (pix == 8) {
 590                                         rc = plpar_hcall9(H_BULK_REMOVE, param,
 591                                                 param[0], param[1], param[2],
 592                                                 param[3], param[4], param[5],
 593                                                 param[6], param[7]);
 594                                         BUG_ON(rc != H_SUCCESS);
 595                                         pix = 0;
 596                                 }
 597                         }
 598                 } pte_iterate_hashed_end();
 599         }
 600         if (pix) {
 601                 param[pix] = HBR_END;
 602                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
 603                                   param[2], param[3], param[4], param[5],
 604                                   param[6], param[7]);
 605                 BUG_ON(rc != H_SUCCESS);
 606         }
 607
 608         if (lock_tlbie)
 609                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
 610 }
 611
 612 static int __init disable_bulk_remove(char *str)
 613 {
 614         if (strcmp(str, "off") == 0 &&
 615             firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
 616                         printk(KERN_INFO "Disabling BULK_REMOVE firmware feature");
 617                         powerpc_firmware_features &= ~FW_FEATURE_BULK_REMOVE;
 618         }
 619         return 1;
 620 }
 621
 622 __setup("bulk_remove=", disable_bulk_remove);
 623
 624 #define HPT_RESIZE_TIMEOUT      10000 /* ms */
 625
 626 struct hpt_resize_state {
 627         unsigned long shift;
 628         int commit_rc;
 629 };
 630
 631 static int pseries_lpar_resize_hpt_commit(void *data)
 632 {
 633         struct hpt_resize_state *state = data;
 634
 635         state->commit_rc = plpar_resize_hpt_commit(0, state->shift);
 636         if (state->commit_rc != H_SUCCESS)
 637                 return -EIO;
 638
 639         /* Hypervisor has transitioned the HTAB, update our globals */
 640         ppc64_pft_size = state->shift;
 641         htab_size_bytes = 1UL << ppc64_pft_size;
 642         htab_hash_mask = (htab_size_bytes >> 7) - 1;
 643
 644         return 0;
 645 }
 646
 647 /*
 648  * Must be called in process context. The caller must hold the
 649  * cpus_lock.
 650  */
 651 static int pseries_lpar_resize_hpt(unsigned long shift)
 652 {
 653         struct hpt_resize_state state = {
 654                 .shift = shift,
 655                 .commit_rc = H_FUNCTION,
 656         };
 657         unsigned int delay, total_delay = 0;
 658         int rc;
 659         ktime_t t0, t1, t2;
 660
 661         might_sleep();
 662
 663         if (!firmware_has_feature(FW_FEATURE_HPT_RESIZE))
 664                 return -ENODEV;
 665
 666         printk(KERN_INFO "lpar: Attempting to resize HPT to shift %lu\n",
 667                shift);
 668
 669         t0 = ktime_get();
 670
 671         rc = plpar_resize_hpt_prepare(0, shift);
 672         while (H_IS_LONG_BUSY(rc)) {
 673                 delay = get_longbusy_msecs(rc);
 674                 total_delay += delay;
 675                 if (total_delay > HPT_RESIZE_TIMEOUT) {
 676                         /* prepare with shift==0 cancels an in-progress resize */
 677                         rc = plpar_resize_hpt_prepare(0, 0);
 678                         if (rc != H_SUCCESS)
 679                                 printk(KERN_WARNING
 680                                        "lpar: Unexpected error %d cancelling timed out HPT resize\n",
 681                                        rc);
 682                         return -ETIMEDOUT;
 683                 }
 684                 msleep(delay);
 685                 rc = plpar_resize_hpt_prepare(0, shift);
 686         };
 687
 688         switch (rc) {
 689         case H_SUCCESS:
 690                 /* Continue on */
 691                 break;
 692
 693         case H_PARAMETER:
 694                 return -EINVAL;
 695         case H_RESOURCE:
 696                 return -EPERM;
 697         default:
 698                 printk(KERN_WARNING
 699                        "lpar: Unexpected error %d from H_RESIZE_HPT_PREPARE\n",
 700                        rc);
 701                 return -EIO;
 702         }
 703
 704         t1 = ktime_get();
 705
 706         rc = stop_machine_cpuslocked(pseries_lpar_resize_hpt_commit,
 707                                      &state, NULL);
 708
 709         t2 = ktime_get();
 710
 711         if (rc != 0) {
 712                 switch (state.commit_rc) {
 713                 case H_PTEG_FULL:
 714                         printk(KERN_WARNING
 715                                "lpar: Hash collision while resizing HPT\n");
 716                         return -ENOSPC;
 717
 718                 default:
 719                         printk(KERN_WARNING
 720                                "lpar: Unexpected error %d from H_RESIZE_HPT_COMMIT\n",
 721                                state.commit_rc);
 722                         return -EIO;
 723                 };
 724         }
 725
 726         printk(KERN_INFO
 727                "lpar: HPT resize to shift %lu complete (%lld ms / %lld ms)\n",
 728                shift, (long long) ktime_ms_delta(t1, t0),
 729                (long long) ktime_ms_delta(t2, t1));
 730
 731         return 0;
 732 }
 733
 734 static int pseries_lpar_register_process_table(unsigned long base,
 735                         unsigned long page_size, unsigned long table_size)
 736 {
 737         long rc;
 738         unsigned long flags = 0;
 739
 740         if (table_size)
 741                 flags |= PROC_TABLE_NEW;
 742         if (radix_enabled())
 743                 flags |= PROC_TABLE_RADIX | PROC_TABLE_GTSE;
 744         else
 745                 flags |= PROC_TABLE_HPT_SLB;
 746         for (;;) {
 747                 rc = plpar_hcall_norets(H_REGISTER_PROC_TBL, flags, base,
 748                                         page_size, table_size);
 749                 if (!H_IS_LONG_BUSY(rc))
 750                         break;
 751                 mdelay(get_longbusy_msecs(rc));
 752         }
 753         if (rc != H_SUCCESS) {
 754                 pr_err("Failed to register process table (rc=%ld)\n", rc);
 755                 BUG();
 756         }
 757         return rc;
 758 }
 759
 760 void __init hpte_init_pseries(void)
 761 {
 762         mmu_hash_ops.hpte_invalidate     = pSeries_lpar_hpte_invalidate;
 763         mmu_hash_ops.hpte_updatepp       = pSeries_lpar_hpte_updatepp;
 764         mmu_hash_ops.hpte_updateboltedpp = pSeries_lpar_hpte_updateboltedpp;
 765         mmu_hash_ops.hpte_insert         = pSeries_lpar_hpte_insert;
 766         mmu_hash_ops.hpte_remove         = pSeries_lpar_hpte_remove;
 767         mmu_hash_ops.hpte_removebolted   = pSeries_lpar_hpte_removebolted;
 768         mmu_hash_ops.flush_hash_range    = pSeries_lpar_flush_hash_range;
 769         mmu_hash_ops.hpte_clear_all      = pseries_hpte_clear_all;
 770         mmu_hash_ops.hugepage_invalidate = pSeries_lpar_hugepage_invalidate;
 771         register_process_table           = pseries_lpar_register_process_table;
 772
 773         if (firmware_has_feature(FW_FEATURE_HPT_RESIZE))
 774                 mmu_hash_ops.resize_hpt = pseries_lpar_resize_hpt;
 775 }
 776
 777 void radix_init_pseries(void)
 778 {
 779         pr_info("Using radix MMU under hypervisor\n");
 780         register_process_table = pseries_lpar_register_process_table;
 781 }
 782
 783 #ifdef CONFIG_PPC_SMLPAR
 784 #define CMO_FREE_HINT_DEFAULT 1
 785 static int cmo_free_hint_flag = CMO_FREE_HINT_DEFAULT;
 786
 787 static int __init cmo_free_hint(char *str)
 788 {
 789         char *parm;
 790         parm = strstrip(str);
 791
 792         if (strcasecmp(parm, "no") == 0 || strcasecmp(parm, "off") == 0) {
 793                 printk(KERN_INFO "cmo_free_hint: CMO free page hinting is not active.\n");
 794                 cmo_free_hint_flag = 0;
 795                 return 1;
 796         }
 797
 798         cmo_free_hint_flag = 1;
 799         printk(KERN_INFO "cmo_free_hint: CMO free page hinting is active.\n");
 800
 801         if (strcasecmp(parm, "yes") == 0 || strcasecmp(parm, "on") == 0)
 802                 return 1;
 803
 804         return 0;
 805 }
 806
 807 __setup("cmo_free_hint=", cmo_free_hint);
 808
 809 static void pSeries_set_page_state(struct page *page, int order,
 810                                    unsigned long state)
 811 {
 812         int i, j;
 813         unsigned long cmo_page_sz, addr;
 814
 815         cmo_page_sz = cmo_get_page_size();
 816         addr = __pa((unsigned long)page_address(page));
 817
 818         for (i = 0; i < (1 << order); i++, addr += PAGE_SIZE) {
 819                 for (j = 0; j < PAGE_SIZE; j += cmo_page_sz)
 820                         plpar_hcall_norets(H_PAGE_INIT, state, addr + j, 0);
 821         }
 822 }
 823
 824 void arch_free_page(struct page *page, int order)
 825 {
 826         if (radix_enabled())
 827                 return;
 828         if (!cmo_free_hint_flag || !firmware_has_feature(FW_FEATURE_CMO))
 829                 return;
 830
 831         pSeries_set_page_state(page, order, H_PAGE_SET_UNUSED);
 832 }
 833 EXPORT_SYMBOL(arch_free_page);
 834
 835 #endif /* CONFIG_PPC_SMLPAR */
 836 #endif /* CONFIG_PPC_STD_MMU_64 */
 837
 838 #ifdef CONFIG_TRACEPOINTS
 839 #ifdef HAVE_JUMP_LABEL
 840 struct static_key hcall_tracepoint_key = STATIC_KEY_INIT;
 841
 842 int hcall_tracepoint_regfunc(void)
 843 {
 844         static_key_slow_inc(&hcall_tracepoint_key);
 845         return 0;
 846 }
 847
 848 void hcall_tracepoint_unregfunc(void)
 849 {
 850         static_key_slow_dec(&hcall_tracepoint_key);
 851 }
 852 #else
 853 /*
 854  * We optimise our hcall path by placing hcall_tracepoint_refcount
 855  * directly in the TOC so we can check if the hcall tracepoints are
 856  * enabled via a single load.
 857  */
 858
 859 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
 860 extern long hcall_tracepoint_refcount;
 861
 862 int hcall_tracepoint_regfunc(void)
 863 {
 864         hcall_tracepoint_refcount++;
 865         return 0;
 866 }
 867
 868 void hcall_tracepoint_unregfunc(void)
 869 {
 870         hcall_tracepoint_refcount--;
 871 }
 872 #endif
 873
 874 /*
 875  * Since the tracing code might execute hcalls we need to guard against
 876  * recursion. One example of this are spinlocks calling H_YIELD on
 877  * shared processor partitions.
 878  */
 879 static DEFINE_PER_CPU(unsigned int, hcall_trace_depth);
 880
 881
 882 void __trace_hcall_entry(unsigned long opcode, unsigned long *args)
 883 {
 884         unsigned long flags;
 885         unsigned int *depth;
 886
 887         /*
 888          * We cannot call tracepoints inside RCU idle regions which
 889          * means we must not trace H_CEDE.
 890          */
 891         if (opcode == H_CEDE)
 892                 return;
 893
 894         local_irq_save(flags);
 895
 896         depth = this_cpu_ptr(&hcall_trace_depth);
 897
 898         if (*depth)
 899                 goto out;
 900
 901         (*depth)++;
 902         preempt_disable();
 903         trace_hcall_entry(opcode, args);
 904         (*depth)--;
 905
 906 out:
 907         local_irq_restore(flags);
 908 }
 909
 910 void __trace_hcall_exit(long opcode, unsigned long retval,
 911                         unsigned long *retbuf)
 912 {
 913         unsigned long flags;
 914         unsigned int *depth;
 915
 916         if (opcode == H_CEDE)
 917                 return;
 918
 919         local_irq_save(flags);
 920
 921         depth = this_cpu_ptr(&hcall_trace_depth);
 922
 923         if (*depth)
 924                 goto out;
 925
 926         (*depth)++;
 927         trace_hcall_exit(opcode, retval, retbuf);
 928         preempt_enable();
 929         (*depth)--;
 930
 931 out:
 932         local_irq_restore(flags);
 933 }
 934 #endif
 935
 936 /**
 937  * h_get_mpp
 938  * H_GET_MPP hcall returns info in 7 parms
 939  */
 940 int h_get_mpp(struct hvcall_mpp_data *mpp_data)
 941 {
 942         int rc;
 943         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
 944
 945         rc = plpar_hcall9(H_GET_MPP, retbuf);
 946
 947         mpp_data->entitled_mem = retbuf[0];
 948         mpp_data->mapped_mem = retbuf[1];
 949
 950         mpp_data->group_num = (retbuf[2] >> 2 * 8) & 0xffff;
 951         mpp_data->pool_num = retbuf[2] & 0xffff;
 952
 953         mpp_data->mem_weight = (retbuf[3] >> 7 * 8) & 0xff;
 954         mpp_data->unallocated_mem_weight = (retbuf[3] >> 6 * 8) & 0xff;
 955         mpp_data->unallocated_entitlement = retbuf[3] & 0xffffffffffffUL;
 956
 957         mpp_data->pool_size = retbuf[4];
 958         mpp_data->loan_request = retbuf[5];
 959         mpp_data->backing_mem = retbuf[6];
 960
 961         return rc;
 962 }
 963 EXPORT_SYMBOL(h_get_mpp);
 964
 965 int h_get_mpp_x(struct hvcall_mpp_x_data *mpp_x_data)
 966 {
 967         int rc;
 968         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE] = { 0 };
 969
 970         rc = plpar_hcall9(H_GET_MPP_X, retbuf);
 971
 972         mpp_x_data->coalesced_bytes = retbuf[0];
 973         mpp_x_data->pool_coalesced_bytes = retbuf[1];
 974         mpp_x_data->pool_purr_cycles = retbuf[2];
 975         mpp_x_data->pool_spurr_cycles = retbuf[3];
 976
 977         return rc;
 978 }
 979
 980 static unsigned long vsid_unscramble(unsigned long vsid, int ssize)
 981 {
 982         unsigned long protovsid;
 983         unsigned long va_bits = VA_BITS;
 984         unsigned long modinv, vsid_modulus;
 985         unsigned long max_mod_inv, tmp_modinv;
 986
 987         if (!mmu_has_feature(MMU_FTR_68_BIT_VA))
 988                 va_bits = 65;
 989
 990         if (ssize == MMU_SEGSIZE_256M) {
 991                 modinv = VSID_MULINV_256M;
 992                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT)) - 1);
 993         } else {
 994                 modinv = VSID_MULINV_1T;
 995                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT_1T)) - 1);
 996         }
 997
 998         /*
 999          * vsid outside our range.
1000          */
1001         if (vsid >= vsid_modulus)
1002                 return 0;
1003
1004         /*
1005          * If modinv is the modular multiplicate inverse of (x % vsid_modulus)
1006          * and vsid = (protovsid * x) % vsid_modulus, then we say:
1007          *   protovsid = (vsid * modinv) % vsid_modulus
1008          */
1009
1010         /* Check if (vsid * modinv) overflow (63 bits) */
1011         max_mod_inv = 0x7fffffffffffffffull / vsid;
1012         if (modinv < max_mod_inv)
1013                 return (vsid * modinv) % vsid_modulus;
1014
1015         tmp_modinv = modinv/max_mod_inv;
1016         modinv %= max_mod_inv;
1017
1018         protovsid = (((vsid * max_mod_inv) % vsid_modulus) * tmp_modinv) % vsid_modulus;
1019         protovsid = (protovsid + vsid * modinv) % vsid_modulus;
1020
1021         return protovsid;
1022 }
1023
1024 static int __init reserve_vrma_context_id(void)
1025 {
1026         unsigned long protovsid;
1027
1028         /*
1029          * Reserve context ids which map to reserved virtual addresses. For now
1030          * we only reserve the context id which maps to the VRMA VSID. We ignore
1031          * the addresses in "ibm,adjunct-virtual-addresses" because we don't
1032          * enable adjunct support via the "ibm,client-architecture-support"
1033          * interface.
1034          */
1035         protovsid = vsid_unscramble(VRMA_VSID, MMU_SEGSIZE_1T);
1036         hash__reserve_context_id(protovsid >> ESID_BITS_1T);
1037         return 0;
1038 }
1039 machine_device_initcall(pseries, reserve_vrma_context_id);
1040
1041 #ifdef CONFIG_DEBUG_FS
1042 /* debugfs file interface for vpa data */
1043 static ssize_t vpa_file_read(struct file *filp, char __user *buf, size_t len,
1044                               loff_t *pos)
1045 {
1046         int cpu = (long)filp->private_data;
1047         struct lppaca *lppaca = &lppaca_of(cpu);
1048
1049         return simple_read_from_buffer(buf, len, pos, lppaca,
1050                                 sizeof(struct lppaca));
1051 }
1052
1053 static const struct file_operations vpa_fops = {
1054         .open           = simple_open,
1055         .read           = vpa_file_read,
1056         .llseek         = default_llseek,
1057 };
1058
1059 static int __init vpa_debugfs_init(void)
1060 {
1061         char name[16];
1062         long i;
1063         struct dentry *vpa_dir;
1064
1065         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
1066                 return 0;
1067
1068         vpa_dir = debugfs_create_dir("vpa", powerpc_debugfs_root);
1069         if (!vpa_dir) {
1070                 pr_warn("%s: can't create vpa root dir\n", __func__);
1071                 return -ENOMEM;
1072         }
1073
1074         /* set up the per-cpu vpa file*/
1075         for_each_possible_cpu(i) {
1076                 struct dentry *d;
1077
1078                 sprintf(name, "cpu-%ld", i);
1079
1080                 d = debugfs_create_file(name, 0400, vpa_dir, (void *)i,
1081                                         &vpa_fops);
1082                 if (!d) {
1083                         pr_warn("%s: can't create per-cpu vpa file\n",
1084                                         __func__);
1085                         return -ENOMEM;
1086                 }
1087         }
1088
1089         return 0;
1090 }
1091 machine_arch_initcall(pseries, vpa_debugfs_init);
1092 #endif /* CONFIG_DEBUG_FS */