arch/ia64/sn/kernel/xpc_partition.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * Copyright (c) 2004-2006 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   8
   9
  10 /*
  11  * Cross Partition Communication (XPC) partition support.
  12  *
  13  *      This is the part of XPC that detects the presence/absence of
  14  *      other partitions. It provides a heartbeat and monitors the
  15  *      heartbeats of other partitions.
  16  *
  17  */
  18
  19
  20 #include <linux/kernel.h>
  21 #include <linux/sysctl.h>
  22 #include <linux/cache.h>
  23 #include <linux/mmzone.h>
  24 #include <linux/nodemask.h>
  25 #include <asm/uncached.h>
  26 #include <asm/sn/bte.h>
  27 #include <asm/sn/intr.h>
  28 #include <asm/sn/sn_sal.h>
  29 #include <asm/sn/nodepda.h>
  30 #include <asm/sn/addrs.h>
  31 #include <asm/sn/xpc.h>
  32
  33
  34 /* XPC is exiting flag */
  35 int xpc_exiting;
  36
  37
  38 /* SH_IPI_ACCESS shub register value on startup */
  39 static u64 xpc_sh1_IPI_access;
  40 static u64 xpc_sh2_IPI_access0;
  41 static u64 xpc_sh2_IPI_access1;
  42 static u64 xpc_sh2_IPI_access2;
  43 static u64 xpc_sh2_IPI_access3;
  44
  45
  46 /* original protection values for each node */
  47 u64 xpc_prot_vec[MAX_NUMNODES];
  48
  49
  50 /* this partition's reserved page pointers */
  51 struct xpc_rsvd_page *xpc_rsvd_page;
  52 static u64 *xpc_part_nasids;
  53 static u64 *xpc_mach_nasids;
  54 struct xpc_vars *xpc_vars;
  55 struct xpc_vars_part *xpc_vars_part;
  56
  57 static int xp_nasid_mask_bytes; /* actual size in bytes of nasid mask */
  58 static int xp_nasid_mask_words; /* actual size in words of nasid mask */
  59
  60
  61 /*
  62  * For performance reasons, each entry of xpc_partitions[] is cacheline
  63  * aligned. And xpc_partitions[] is padded with an additional entry at the
  64  * end so that the last legitimate entry doesn't share its cacheline with
  65  * another variable.
  66  */
  67 struct xpc_partition xpc_partitions[XP_MAX_PARTITIONS + 1];
  68
  69
  70 /*
  71  * Generic buffer used to store a local copy of portions of a remote
  72  * partition's reserved page (either its header and part_nasids mask,
  73  * or its vars).
  74  */
  75 char *xpc_remote_copy_buffer;
  76 void *xpc_remote_copy_buffer_base;
  77
  78
  79 /*
  80  * Given a nasid, get the physical address of the  partition's reserved page
  81  * for that nasid. This function returns 0 on any error.
  82  */
  83 static u64
  84 xpc_get_rsvd_page_pa(int nasid)
  85 {
  86         bte_result_t bte_res;
  87         s64 status;
  88         u64 cookie = 0;
  89         u64 rp_pa = nasid;      /* seed with nasid */
  90         u64 len = 0;
  91         u64 buf = buf;
  92         u64 buf_len = 0;
  93         void *buf_base = NULL;
  94
  95
  96         while (1) {
  97
  98                 status = sn_partition_reserved_page_pa(buf, &cookie, &rp_pa,
  99                                                                 &len);
 100
 101                 dev_dbg(xpc_part, "SAL returned with status=%li, cookie="
 102                         "0x%016lx, address=0x%016lx, len=0x%016lx\n",
 103                         status, cookie, rp_pa, len);
 104
 105                 if (status != SALRET_MORE_PASSES) {
 106                         break;
 107                 }
 108
 109                 if (L1_CACHE_ALIGN(len) > buf_len) {
 110                         if (buf_base != NULL) {
 111                                 kfree(buf_base);
 112                         }
 113                         buf_len = L1_CACHE_ALIGN(len);
 114                         buf = (u64) xpc_kmalloc_cacheline_aligned(buf_len,
 115                                                         GFP_KERNEL, &buf_base);
 116                         if (buf_base == NULL) {
 117                                 dev_err(xpc_part, "unable to kmalloc "
 118                                         "len=0x%016lx\n", buf_len);
 119                                 status = SALRET_ERROR;
 120                                 break;
 121                         }
 122                 }
 123
 124                 bte_res = xp_bte_copy(rp_pa, buf, buf_len,
 125                                         (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 126                 if (bte_res != BTE_SUCCESS) {
 127                         dev_dbg(xpc_part, "xp_bte_copy failed %i\n", bte_res);
 128                         status = SALRET_ERROR;
 129                         break;
 130                 }
 131         }
 132
 133         if (buf_base != NULL) {
 134                 kfree(buf_base);
 135         }
 136
 137         if (status != SALRET_OK) {
 138                 rp_pa = 0;
 139         }
 140         dev_dbg(xpc_part, "reserved page at phys address 0x%016lx\n", rp_pa);
 141         return rp_pa;
 142 }
 143
 144
 145 /*
 146  * Fill the partition reserved page with the information needed by
 147  * other partitions to discover we are alive and establish initial
 148  * communications.
 149  */
 150 struct xpc_rsvd_page *
 151 xpc_rsvd_page_init(void)
 152 {
 153         struct xpc_rsvd_page *rp;
 154         AMO_t *amos_page;
 155         u64 rp_pa, nasid_array = 0;
 156         int i, ret;
 157
 158
 159         /* get the local reserved page's address */
 160
 161         preempt_disable();
 162         rp_pa = xpc_get_rsvd_page_pa(cpuid_to_nasid(smp_processor_id()));
 163         preempt_enable();
 164         if (rp_pa == 0) {
 165                 dev_err(xpc_part, "SAL failed to locate the reserved page\n");
 166                 return NULL;
 167         }
 168         rp = (struct xpc_rsvd_page *) __va(rp_pa);
 169
 170         if (rp->partid != sn_partition_id) {
 171                 dev_err(xpc_part, "the reserved page's partid of %d should be "
 172                         "%d\n", rp->partid, sn_partition_id);
 173                 return NULL;
 174         }
 175
 176         rp->version = XPC_RP_VERSION;
 177
 178         /* establish the actual sizes of the nasid masks */
 179         if (rp->SAL_version == 1) {
 180                 /* SAL_version 1 didn't set the nasids_size field */
 181                 rp->nasids_size = 128;
 182         }
 183         xp_nasid_mask_bytes = rp->nasids_size;
 184         xp_nasid_mask_words = xp_nasid_mask_bytes / 8;
 185
 186         /* setup the pointers to the various items in the reserved page */
 187         xpc_part_nasids = XPC_RP_PART_NASIDS(rp);
 188         xpc_mach_nasids = XPC_RP_MACH_NASIDS(rp);
 189         xpc_vars = XPC_RP_VARS(rp);
 190         xpc_vars_part = XPC_RP_VARS_PART(rp);
 191
 192         /*
 193          * Before clearing xpc_vars, see if a page of AMOs had been previously
 194          * allocated. If not we'll need to allocate one and set permissions
 195          * so that cross-partition AMOs are allowed.
 196          *
 197          * The allocated AMO page needs MCA reporting to remain disabled after
 198          * XPC has unloaded.  To make this work, we keep a copy of the pointer
 199          * to this page (i.e., amos_page) in the struct xpc_vars structure,
 200          * which is pointed to by the reserved page, and re-use that saved copy
 201          * on subsequent loads of XPC. This AMO page is never freed, and its
 202          * memory protections are never restricted.
 203          */
 204         if ((amos_page = xpc_vars->amos_page) == NULL) {
 205                 amos_page = (AMO_t *) TO_AMO(uncached_alloc_page(0));
 206                 if (amos_page == NULL) {
 207                         dev_err(xpc_part, "can't allocate page of AMOs\n");
 208                         return NULL;
 209                 }
 210
 211                 /*
 212                  * Open up AMO-R/W to cpu.  This is done for Shub 1.1 systems
 213                  * when xpc_allow_IPI_ops() is called via xpc_hb_init().
 214                  */
 215                 if (!enable_shub_wars_1_1()) {
 216                         ret = sn_change_memprotect(ia64_tpa((u64) amos_page),
 217                                         PAGE_SIZE, SN_MEMPROT_ACCESS_CLASS_1,
 218                                         &nasid_array);
 219                         if (ret != 0) {
 220                                 dev_err(xpc_part, "can't change memory "
 221                                         "protections\n");
 222                                 uncached_free_page(__IA64_UNCACHED_OFFSET |
 223                                                    TO_PHYS((u64) amos_page));
 224                                 return NULL;
 225                         }
 226                 }
 227         } else if (!IS_AMO_ADDRESS((u64) amos_page)) {
 228                 /*
 229                  * EFI's XPBOOT can also set amos_page in the reserved page,
 230                  * but it happens to leave it as an uncached physical address
 231                  * and we need it to be an uncached virtual, so we'll have to
 232                  * convert it.
 233                  */
 234                 if (!IS_AMO_PHYS_ADDRESS((u64) amos_page)) {
 235                         dev_err(xpc_part, "previously used amos_page address "
 236                                 "is bad = 0x%p\n", (void *) amos_page);
 237                         return NULL;
 238                 }
 239                 amos_page = (AMO_t *) TO_AMO((u64) amos_page);
 240         }
 241
 242         /* clear xpc_vars */
 243         memset(xpc_vars, 0, sizeof(struct xpc_vars));
 244
 245         xpc_vars->version = XPC_V_VERSION;
 246         xpc_vars->act_nasid = cpuid_to_nasid(0);
 247         xpc_vars->act_phys_cpuid = cpu_physical_id(0);
 248         xpc_vars->vars_part_pa = __pa(xpc_vars_part);
 249         xpc_vars->amos_page_pa = ia64_tpa((u64) amos_page);
 250         xpc_vars->amos_page = amos_page;  /* save for next load of XPC */
 251
 252
 253         /* clear xpc_vars_part */
 254         memset((u64 *) xpc_vars_part, 0, sizeof(struct xpc_vars_part) *
 255                                                         XP_MAX_PARTITIONS);
 256
 257         /* initialize the activate IRQ related AMO variables */
 258         for (i = 0; i < xp_nasid_mask_words; i++) {
 259                 (void) xpc_IPI_init(XPC_ACTIVATE_IRQ_AMOS + i);
 260         }
 261
 262         /* initialize the engaged remote partitions related AMO variables */
 263         (void) xpc_IPI_init(XPC_ENGAGED_PARTITIONS_AMO);
 264         (void) xpc_IPI_init(XPC_DISENGAGE_REQUEST_AMO);
 265
 266         /* timestamp of when reserved page was setup by XPC */
 267         rp->stamp = CURRENT_TIME;
 268
 269         /*
 270          * This signifies to the remote partition that our reserved
 271          * page is initialized.
 272          */
 273         rp->vars_pa = __pa(xpc_vars);
 274
 275         return rp;
 276 }
 277
 278
 279 /*
 280  * Change protections to allow IPI operations (and AMO operations on
 281  * Shub 1.1 systems).
 282  */
 283 void
 284 xpc_allow_IPI_ops(void)
 285 {
 286         int node;
 287         int nasid;
 288
 289
 290         // >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
 291
 292         if (is_shub2()) {
 293                 xpc_sh2_IPI_access0 =
 294                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS0));
 295                 xpc_sh2_IPI_access1 =
 296                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS1));
 297                 xpc_sh2_IPI_access2 =
 298                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS2));
 299                 xpc_sh2_IPI_access3 =
 300                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS3));
 301
 302                 for_each_online_node(node) {
 303                         nasid = cnodeid_to_nasid(node);
 304                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
 305                                                                 -1UL);
 306                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
 307                                                                 -1UL);
 308                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
 309                                                                 -1UL);
 310                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
 311                                                                 -1UL);
 312                 }
 313
 314         } else {
 315                 xpc_sh1_IPI_access =
 316                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH1_IPI_ACCESS));
 317
 318                 for_each_online_node(node) {
 319                         nasid = cnodeid_to_nasid(node);
 320                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
 321                                                                 -1UL);
 322
 323                         /*
 324                          * Since the BIST collides with memory operations on
 325                          * SHUB 1.1 sn_change_memprotect() cannot be used.
 326                          */
 327                         if (enable_shub_wars_1_1()) {
 328                                 /* open up everything */
 329                                 xpc_prot_vec[node] = (u64) HUB_L((u64 *)
 330                                                 GLOBAL_MMR_ADDR(nasid,
 331                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0));
 332                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 333                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
 334                                                                 -1UL);
 335                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 336                                                 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
 337                                                                 -1UL);
 338                         }
 339                 }
 340         }
 341 }
 342
 343
 344 /*
 345  * Restrict protections to disallow IPI operations (and AMO operations on
 346  * Shub 1.1 systems).
 347  */
 348 void
 349 xpc_restrict_IPI_ops(void)
 350 {
 351         int node;
 352         int nasid;
 353
 354
 355         // >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
 356
 357         if (is_shub2()) {
 358
 359                 for_each_online_node(node) {
 360                         nasid = cnodeid_to_nasid(node);
 361                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
 362                                                         xpc_sh2_IPI_access0);
 363                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
 364                                                         xpc_sh2_IPI_access1);
 365                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
 366                                                         xpc_sh2_IPI_access2);
 367                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
 368                                                         xpc_sh2_IPI_access3);
 369                 }
 370
 371         } else {
 372
 373                 for_each_online_node(node) {
 374                         nasid = cnodeid_to_nasid(node);
 375                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
 376                                                         xpc_sh1_IPI_access);
 377
 378                         if (enable_shub_wars_1_1()) {
 379                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 380                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
 381                                                         xpc_prot_vec[node]);
 382                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 383                                                 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
 384                                                         xpc_prot_vec[node]);
 385                         }
 386                 }
 387         }
 388 }
 389
 390
 391 /*
 392  * At periodic intervals, scan through all active partitions and ensure
 393  * their heartbeat is still active.  If not, the partition is deactivated.
 394  */
 395 void
 396 xpc_check_remote_hb(void)
 397 {
 398         struct xpc_vars *remote_vars;
 399         struct xpc_partition *part;
 400         partid_t partid;
 401         bte_result_t bres;
 402
 403
 404         remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
 405
 406         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 407
 408                 if (xpc_exiting) {
 409                         break;
 410                 }
 411
 412                 if (partid == sn_partition_id) {
 413                         continue;
 414                 }
 415
 416                 part = &xpc_partitions[partid];
 417
 418                 if (part->act_state == XPC_P_INACTIVE ||
 419                                 part->act_state == XPC_P_DEACTIVATING) {
 420                         continue;
 421                 }
 422
 423                 /* pull the remote_hb cache line */
 424                 bres = xp_bte_copy(part->remote_vars_pa,
 425                                         (u64) remote_vars,
 426                                         XPC_RP_VARS_SIZE,
 427                                         (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 428                 if (bres != BTE_SUCCESS) {
 429                         XPC_DEACTIVATE_PARTITION(part,
 430                                                 xpc_map_bte_errors(bres));
 431                         continue;
 432                 }
 433
 434                 dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
 435                         " = %ld, heartbeat_offline = %ld, HB_mask = 0x%lx\n",
 436                         partid, remote_vars->heartbeat, part->last_heartbeat,
 437                         remote_vars->heartbeat_offline,
 438                         remote_vars->heartbeating_to_mask);
 439
 440                 if (((remote_vars->heartbeat == part->last_heartbeat) &&
 441                         (remote_vars->heartbeat_offline == 0)) ||
 442                              !xpc_hb_allowed(sn_partition_id, remote_vars)) {
 443
 444                         XPC_DEACTIVATE_PARTITION(part, xpcNoHeartbeat);
 445                         continue;
 446                 }
 447
 448                 part->last_heartbeat = remote_vars->heartbeat;
 449         }
 450 }
 451
 452
 453 /*
 454  * Get a copy of a portion of the remote partition's rsvd page.
 455  *
 456  * remote_rp points to a buffer that is cacheline aligned for BTE copies and
 457  * is large enough to contain a copy of their reserved page header and
 458  * part_nasids mask.
 459  */
 460 static enum xpc_retval
 461 xpc_get_remote_rp(int nasid, u64 *discovered_nasids,
 462                 struct xpc_rsvd_page *remote_rp, u64 *remote_rp_pa)
 463 {
 464         int bres, i;
 465
 466
 467         /* get the reserved page's physical address */
 468
 469         *remote_rp_pa = xpc_get_rsvd_page_pa(nasid);
 470         if (*remote_rp_pa == 0) {
 471                 return xpcNoRsvdPageAddr;
 472         }
 473
 474
 475         /* pull over the reserved page header and part_nasids mask */
 476         bres = xp_bte_copy(*remote_rp_pa, (u64) remote_rp,
 477                                 XPC_RP_HEADER_SIZE + xp_nasid_mask_bytes,
 478                                 (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 479         if (bres != BTE_SUCCESS) {
 480                 return xpc_map_bte_errors(bres);
 481         }
 482
 483
 484         if (discovered_nasids != NULL) {
 485                 u64 *remote_part_nasids = XPC_RP_PART_NASIDS(remote_rp);
 486
 487
 488                 for (i = 0; i < xp_nasid_mask_words; i++) {
 489                         discovered_nasids[i] |= remote_part_nasids[i];
 490                 }
 491         }
 492
 493
 494         /* check that the partid is for another partition */
 495
 496         if (remote_rp->partid < 1 ||
 497                                 remote_rp->partid > (XP_MAX_PARTITIONS - 1)) {
 498                 return xpcInvalidPartid;
 499         }
 500
 501         if (remote_rp->partid == sn_partition_id) {
 502                 return xpcLocalPartid;
 503         }
 504
 505
 506         if (XPC_VERSION_MAJOR(remote_rp->version) !=
 507                                         XPC_VERSION_MAJOR(XPC_RP_VERSION)) {
 508                 return xpcBadVersion;
 509         }
 510
 511         return xpcSuccess;
 512 }
 513
 514
 515 /*
 516  * Get a copy of the remote partition's XPC variables from the reserved page.
 517  *
 518  * remote_vars points to a buffer that is cacheline aligned for BTE copies and
 519  * assumed to be of size XPC_RP_VARS_SIZE.
 520  */
 521 static enum xpc_retval
 522 xpc_get_remote_vars(u64 remote_vars_pa, struct xpc_vars *remote_vars)
 523 {
 524         int bres;
 525
 526
 527         if (remote_vars_pa == 0) {
 528                 return xpcVarsNotSet;
 529         }
 530
 531         /* pull over the cross partition variables */
 532         bres = xp_bte_copy(remote_vars_pa, (u64) remote_vars, XPC_RP_VARS_SIZE,
 533                                 (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 534         if (bres != BTE_SUCCESS) {
 535                 return xpc_map_bte_errors(bres);
 536         }
 537
 538         if (XPC_VERSION_MAJOR(remote_vars->version) !=
 539                                         XPC_VERSION_MAJOR(XPC_V_VERSION)) {
 540                 return xpcBadVersion;
 541         }
 542
 543         return xpcSuccess;
 544 }
 545
 546
 547 /*
 548  * Update the remote partition's info.
 549  */
 550 static void
 551 xpc_update_partition_info(struct xpc_partition *part, u8 remote_rp_version,
 552                 struct timespec *remote_rp_stamp, u64 remote_rp_pa,
 553                 u64 remote_vars_pa, struct xpc_vars *remote_vars)
 554 {
 555         part->remote_rp_version = remote_rp_version;
 556         dev_dbg(xpc_part, "  remote_rp_version = 0x%016lx\n",
 557                 part->remote_rp_version);
 558
 559         part->remote_rp_stamp = *remote_rp_stamp;
 560         dev_dbg(xpc_part, "  remote_rp_stamp (tv_sec = 0x%lx tv_nsec = 0x%lx\n",
 561                 part->remote_rp_stamp.tv_sec, part->remote_rp_stamp.tv_nsec);
 562
 563         part->remote_rp_pa = remote_rp_pa;
 564         dev_dbg(xpc_part, "  remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
 565
 566         part->remote_vars_pa = remote_vars_pa;
 567         dev_dbg(xpc_part, "  remote_vars_pa = 0x%016lx\n",
 568                 part->remote_vars_pa);
 569
 570         part->last_heartbeat = remote_vars->heartbeat;
 571         dev_dbg(xpc_part, "  last_heartbeat = 0x%016lx\n",
 572                 part->last_heartbeat);
 573
 574         part->remote_vars_part_pa = remote_vars->vars_part_pa;
 575         dev_dbg(xpc_part, "  remote_vars_part_pa = 0x%016lx\n",
 576                 part->remote_vars_part_pa);
 577
 578         part->remote_act_nasid = remote_vars->act_nasid;
 579         dev_dbg(xpc_part, "  remote_act_nasid = 0x%x\n",
 580                 part->remote_act_nasid);
 581
 582         part->remote_act_phys_cpuid = remote_vars->act_phys_cpuid;
 583         dev_dbg(xpc_part, "  remote_act_phys_cpuid = 0x%x\n",
 584                 part->remote_act_phys_cpuid);
 585
 586         part->remote_amos_page_pa = remote_vars->amos_page_pa;
 587         dev_dbg(xpc_part, "  remote_amos_page_pa = 0x%lx\n",
 588                 part->remote_amos_page_pa);
 589
 590         part->remote_vars_version = remote_vars->version;
 591         dev_dbg(xpc_part, "  remote_vars_version = 0x%x\n",
 592                 part->remote_vars_version);
 593 }
 594
 595
 596 /*
 597  * Prior code has determined the nasid which generated an IPI.  Inspect
 598  * that nasid to determine if its partition needs to be activated or
 599  * deactivated.
 600  *
 601  * A partition is consider "awaiting activation" if our partition
 602  * flags indicate it is not active and it has a heartbeat.  A
 603  * partition is considered "awaiting deactivation" if our partition
 604  * flags indicate it is active but it has no heartbeat or it is not
 605  * sending its heartbeat to us.
 606  *
 607  * To determine the heartbeat, the remote nasid must have a properly
 608  * initialized reserved page.
 609  */
 610 static void
 611 xpc_identify_act_IRQ_req(int nasid)
 612 {
 613         struct xpc_rsvd_page *remote_rp;
 614         struct xpc_vars *remote_vars;
 615         u64 remote_rp_pa;
 616         u64 remote_vars_pa;
 617         int remote_rp_version;
 618         int reactivate = 0;
 619         int stamp_diff;
 620         struct timespec remote_rp_stamp = { 0, 0 };
 621         partid_t partid;
 622         struct xpc_partition *part;
 623         enum xpc_retval ret;
 624
 625
 626         /* pull over the reserved page structure */
 627
 628         remote_rp = (struct xpc_rsvd_page *) xpc_remote_copy_buffer;
 629
 630         ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
 631         if (ret != xpcSuccess) {
 632                 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
 633                         "which sent interrupt, reason=%d\n", nasid, ret);
 634                 return;
 635         }
 636
 637         remote_vars_pa = remote_rp->vars_pa;
 638         remote_rp_version = remote_rp->version;
 639         if (XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 640                 remote_rp_stamp = remote_rp->stamp;
 641         }
 642         partid = remote_rp->partid;
 643         part = &xpc_partitions[partid];
 644
 645
 646         /* pull over the cross partition variables */
 647
 648         remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
 649
 650         ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
 651         if (ret != xpcSuccess) {
 652
 653                 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
 654                         "which sent interrupt, reason=%d\n", nasid, ret);
 655
 656                 XPC_DEACTIVATE_PARTITION(part, ret);
 657                 return;
 658         }
 659
 660
 661         part->act_IRQ_rcvd++;
 662
 663         dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
 664                 "%ld:0x%lx\n", (int) nasid, (int) partid, part->act_IRQ_rcvd,
 665                 remote_vars->heartbeat, remote_vars->heartbeating_to_mask);
 666
 667         if (xpc_partition_disengaged(part) &&
 668                                         part->act_state == XPC_P_INACTIVE) {
 669
 670                 xpc_update_partition_info(part, remote_rp_version,
 671                                         &remote_rp_stamp, remote_rp_pa,
 672                                         remote_vars_pa, remote_vars);
 673
 674                 if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 675                         if (xpc_partition_disengage_requested(1UL << partid)) {
 676                                 /*
 677                                  * Other side is waiting on us to disengage,
 678                                  * even though we already have.
 679                                  */
 680                                 return;
 681                         }
 682                 } else {
 683                         /* other side doesn't support disengage requests */
 684                         xpc_clear_partition_disengage_request(1UL << partid);
 685                 }
 686
 687                 xpc_activate_partition(part);
 688                 return;
 689         }
 690
 691         DBUG_ON(part->remote_rp_version == 0);
 692         DBUG_ON(part->remote_vars_version == 0);
 693
 694         if (!XPC_SUPPORTS_RP_STAMP(part->remote_rp_version)) {
 695                 DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(part->
 696                                                         remote_vars_version));
 697
 698                 if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 699                         DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->
 700                                                                 version));
 701                         /* see if the other side rebooted */
 702                         if (part->remote_amos_page_pa ==
 703                                 remote_vars->amos_page_pa &&
 704                                         xpc_hb_allowed(sn_partition_id,
 705                                                                 remote_vars)) {
 706                                 /* doesn't look that way, so ignore the IPI */
 707                                 return;
 708                         }
 709                 }
 710
 711                 /*
 712                  * Other side rebooted and previous XPC didn't support the
 713                  * disengage request, so we don't need to do anything special.
 714                  */
 715
 716                 xpc_update_partition_info(part, remote_rp_version,
 717                                                 &remote_rp_stamp, remote_rp_pa,
 718                                                 remote_vars_pa, remote_vars);
 719                 part->reactivate_nasid = nasid;
 720                 XPC_DEACTIVATE_PARTITION(part, xpcReactivating);
 721                 return;
 722         }
 723
 724         DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version));
 725
 726         if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 727                 DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version));
 728
 729                 /*
 730                  * Other side rebooted and previous XPC did support the
 731                  * disengage request, but the new one doesn't.
 732                  */
 733
 734                 xpc_clear_partition_engaged(1UL << partid);
 735                 xpc_clear_partition_disengage_request(1UL << partid);
 736
 737                 xpc_update_partition_info(part, remote_rp_version,
 738                                                 &remote_rp_stamp, remote_rp_pa,
 739                                                 remote_vars_pa, remote_vars);
 740                 reactivate = 1;
 741
 742         } else {
 743                 DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version));
 744
 745                 stamp_diff = xpc_compare_stamps(&part->remote_rp_stamp,
 746                                                         &remote_rp_stamp);
 747                 if (stamp_diff != 0) {
 748                         DBUG_ON(stamp_diff >= 0);
 749
 750                         /*
 751                          * Other side rebooted and the previous XPC did support
 752                          * the disengage request, as does the new one.
 753                          */
 754
 755                         DBUG_ON(xpc_partition_engaged(1UL << partid));
 756                         DBUG_ON(xpc_partition_disengage_requested(1UL <<
 757                                                                 partid));
 758
 759                         xpc_update_partition_info(part, remote_rp_version,
 760                                                 &remote_rp_stamp, remote_rp_pa,
 761                                                 remote_vars_pa, remote_vars);
 762                         reactivate = 1;
 763                 }
 764         }
 765
 766         if (part->disengage_request_timeout > 0 &&
 767                                         !xpc_partition_disengaged(part)) {
 768                 /* still waiting on other side to disengage from us */
 769                 return;
 770         }
 771
 772         if (reactivate) {
 773                 part->reactivate_nasid = nasid;
 774                 XPC_DEACTIVATE_PARTITION(part, xpcReactivating);
 775
 776         } else if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version) &&
 777                         xpc_partition_disengage_requested(1UL << partid)) {
 778                 XPC_DEACTIVATE_PARTITION(part, xpcOtherGoingDown);
 779         }
 780 }
 781
 782
 783 /*
 784  * Loop through the activation AMO variables and process any bits
 785  * which are set.  Each bit indicates a nasid sending a partition
 786  * activation or deactivation request.
 787  *
 788  * Return #of IRQs detected.
 789  */
 790 int
 791 xpc_identify_act_IRQ_sender(void)
 792 {
 793         int word, bit;
 794         u64 nasid_mask;
 795         u64 nasid;                      /* remote nasid */
 796         int n_IRQs_detected = 0;
 797         AMO_t *act_amos;
 798
 799
 800         act_amos = xpc_vars->amos_page + XPC_ACTIVATE_IRQ_AMOS;
 801
 802
 803         /* scan through act AMO variable looking for non-zero entries */
 804         for (word = 0; word < xp_nasid_mask_words; word++) {
 805
 806                 if (xpc_exiting) {
 807                         break;
 808                 }
 809
 810                 nasid_mask = xpc_IPI_receive(&act_amos[word]);
 811                 if (nasid_mask == 0) {
 812                         /* no IRQs from nasids in this variable */
 813                         continue;
 814                 }
 815
 816                 dev_dbg(xpc_part, "AMO[%d] gave back 0x%lx\n", word,
 817                         nasid_mask);
 818
 819
 820                 /*
 821                  * If this nasid has been added to the machine since
 822                  * our partition was reset, this will retain the
 823                  * remote nasid in our reserved pages machine mask.
 824                  * This is used in the event of module reload.
 825                  */
 826                 xpc_mach_nasids[word] |= nasid_mask;
 827
 828
 829                 /* locate the nasid(s) which sent interrupts */
 830
 831                 for (bit = 0; bit < (8 * sizeof(u64)); bit++) {
 832                         if (nasid_mask & (1UL << bit)) {
 833                                 n_IRQs_detected++;
 834                                 nasid = XPC_NASID_FROM_W_B(word, bit);
 835                                 dev_dbg(xpc_part, "interrupt from nasid %ld\n",
 836                                         nasid);
 837                                 xpc_identify_act_IRQ_req(nasid);
 838                         }
 839                 }
 840         }
 841         return n_IRQs_detected;
 842 }
 843
 844
 845 /*
 846  * See if the other side has responded to a partition disengage request
 847  * from us.
 848  */
 849 int
 850 xpc_partition_disengaged(struct xpc_partition *part)
 851 {
 852         partid_t partid = XPC_PARTID(part);
 853         int disengaged;
 854
 855
 856         disengaged = (xpc_partition_engaged(1UL << partid) == 0);
 857         if (part->disengage_request_timeout) {
 858                 if (!disengaged) {
 859                         if (jiffies < part->disengage_request_timeout) {
 860                                 /* timelimit hasn't been reached yet */
 861                                 return 0;
 862                         }
 863
 864                         /*
 865                          * Other side hasn't responded to our disengage
 866                          * request in a timely fashion, so assume it's dead.
 867                          */
 868
 869                         dev_info(xpc_part, "disengage from remote partition %d "
 870                                 "timed out\n", partid);
 871                         xpc_disengage_request_timedout = 1;
 872                         xpc_clear_partition_engaged(1UL << partid);
 873                         disengaged = 1;
 874                 }
 875                 part->disengage_request_timeout = 0;
 876
 877                 /* cancel the timer function, provided it's not us */
 878                 if (!in_interrupt()) {
 879                         del_singleshot_timer_sync(&part->
 880                                                       disengage_request_timer);
 881                 }
 882
 883                 DBUG_ON(part->act_state != XPC_P_DEACTIVATING &&
 884                                         part->act_state != XPC_P_INACTIVE);
 885                 if (part->act_state != XPC_P_INACTIVE) {
 886                         xpc_wakeup_channel_mgr(part);
 887                 }
 888
 889                 if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 890                         xpc_cancel_partition_disengage_request(part);
 891                 }
 892         }
 893         return disengaged;
 894 }
 895
 896
 897 /*
 898  * Mark specified partition as active.
 899  */
 900 enum xpc_retval
 901 xpc_mark_partition_active(struct xpc_partition *part)
 902 {
 903         unsigned long irq_flags;
 904         enum xpc_retval ret;
 905
 906
 907         dev_dbg(xpc_part, "setting partition %d to ACTIVE\n", XPC_PARTID(part));
 908
 909         spin_lock_irqsave(&part->act_lock, irq_flags);
 910         if (part->act_state == XPC_P_ACTIVATING) {
 911                 part->act_state = XPC_P_ACTIVE;
 912                 ret = xpcSuccess;
 913         } else {
 914                 DBUG_ON(part->reason == xpcSuccess);
 915                 ret = part->reason;
 916         }
 917         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 918
 919         return ret;
 920 }
 921
 922
 923 /*
 924  * Notify XPC that the partition is down.
 925  */
 926 void
 927 xpc_deactivate_partition(const int line, struct xpc_partition *part,
 928                                 enum xpc_retval reason)
 929 {
 930         unsigned long irq_flags;
 931
 932
 933         spin_lock_irqsave(&part->act_lock, irq_flags);
 934
 935         if (part->act_state == XPC_P_INACTIVE) {
 936                 XPC_SET_REASON(part, reason, line);
 937                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 938                 if (reason == xpcReactivating) {
 939                         /* we interrupt ourselves to reactivate partition */
 940                         xpc_IPI_send_reactivate(part);
 941                 }
 942                 return;
 943         }
 944         if (part->act_state == XPC_P_DEACTIVATING) {
 945                 if ((part->reason == xpcUnloading && reason != xpcUnloading) ||
 946                                         reason == xpcReactivating) {
 947                         XPC_SET_REASON(part, reason, line);
 948                 }
 949                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 950                 return;
 951         }
 952
 953         part->act_state = XPC_P_DEACTIVATING;
 954         XPC_SET_REASON(part, reason, line);
 955
 956         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 957
 958         if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 959                 xpc_request_partition_disengage(part);
 960                 xpc_IPI_send_disengage(part);
 961
 962                 /* set a timelimit on the disengage request */
 963                 part->disengage_request_timeout = jiffies +
 964                                         (xpc_disengage_request_timelimit * HZ);
 965                 part->disengage_request_timer.expires =
 966                                         part->disengage_request_timeout;
 967                 add_timer(&part->disengage_request_timer);
 968         }
 969
 970         dev_dbg(xpc_part, "bringing partition %d down, reason = %d\n",
 971                 XPC_PARTID(part), reason);
 972
 973         xpc_partition_going_down(part, reason);
 974 }
 975
 976
 977 /*
 978  * Mark specified partition as inactive.
 979  */
 980 void
 981 xpc_mark_partition_inactive(struct xpc_partition *part)
 982 {
 983         unsigned long irq_flags;
 984
 985
 986         dev_dbg(xpc_part, "setting partition %d to INACTIVE\n",
 987                 XPC_PARTID(part));
 988
 989         spin_lock_irqsave(&part->act_lock, irq_flags);
 990         part->act_state = XPC_P_INACTIVE;
 991         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 992         part->remote_rp_pa = 0;
 993 }
 994
 995
 996 /*
 997  * SAL has provided a partition and machine mask.  The partition mask
 998  * contains a bit for each even nasid in our partition.  The machine
 999  * mask contains a bit for each even nasid in the entire machine.
1000  *
1001  * Using those two bit arrays, we can determine which nasids are
1002  * known in the machine.  Each should also have a reserved page
1003  * initialized if they are available for partitioning.
1004  */
1005 void
1006 xpc_discovery(void)
1007 {
1008         void *remote_rp_base;
1009         struct xpc_rsvd_page *remote_rp;
1010         struct xpc_vars *remote_vars;
1011         u64 remote_rp_pa;
1012         u64 remote_vars_pa;
1013         int region;
1014         int region_size;
1015         int max_regions;
1016         int nasid;
1017         struct xpc_rsvd_page *rp;
1018         partid_t partid;
1019         struct xpc_partition *part;
1020         u64 *discovered_nasids;
1021         enum xpc_retval ret;
1022
1023
1024         remote_rp = xpc_kmalloc_cacheline_aligned(XPC_RP_HEADER_SIZE +
1025                                                 xp_nasid_mask_bytes,
1026                                                 GFP_KERNEL, &remote_rp_base);
1027         if (remote_rp == NULL) {
1028                 return;
1029         }
1030         remote_vars = (struct xpc_vars *) remote_rp;
1031
1032
1033         discovered_nasids = kmalloc(sizeof(u64) * xp_nasid_mask_words,
1034                                                         GFP_KERNEL);
1035         if (discovered_nasids == NULL) {
1036                 kfree(remote_rp_base);
1037                 return;
1038         }
1039         memset(discovered_nasids, 0, sizeof(u64) * xp_nasid_mask_words);
1040
1041         rp = (struct xpc_rsvd_page *) xpc_rsvd_page;
1042
1043         /*
1044          * The term 'region' in this context refers to the minimum number of
1045          * nodes that can comprise an access protection grouping. The access
1046          * protection is in regards to memory, IOI and IPI.
1047          */
1048         max_regions = 64;
1049         region_size = sn_region_size;
1050
1051         switch (region_size) {
1052         case 128:
1053                 max_regions *= 2;
1054         case 64:
1055                 max_regions *= 2;
1056         case 32:
1057                 max_regions *= 2;
1058                 region_size = 16;
1059                 DBUG_ON(!is_shub2());
1060         }
1061
1062         for (region = 0; region < max_regions; region++) {
1063
1064                 if ((volatile int) xpc_exiting) {
1065                         break;
1066                 }
1067
1068                 dev_dbg(xpc_part, "searching region %d\n", region);
1069
1070                 for (nasid = (region * region_size * 2);
1071                      nasid < ((region + 1) * region_size * 2);
1072                      nasid += 2) {
1073
1074                         if ((volatile int) xpc_exiting) {
1075                                 break;
1076                         }
1077
1078                         dev_dbg(xpc_part, "checking nasid %d\n", nasid);
1079
1080
1081                         if (XPC_NASID_IN_ARRAY(nasid, xpc_part_nasids)) {
1082                                 dev_dbg(xpc_part, "PROM indicates Nasid %d is "
1083                                         "part of the local partition; skipping "
1084                                         "region\n", nasid);
1085                                 break;
1086                         }
1087
1088                         if (!(XPC_NASID_IN_ARRAY(nasid, xpc_mach_nasids))) {
1089                                 dev_dbg(xpc_part, "PROM indicates Nasid %d was "
1090                                         "not on Numa-Link network at reset\n",
1091                                         nasid);
1092                                 continue;
1093                         }
1094
1095                         if (XPC_NASID_IN_ARRAY(nasid, discovered_nasids)) {
1096                                 dev_dbg(xpc_part, "Nasid %d is part of a "
1097                                         "partition which was previously "
1098                                         "discovered\n", nasid);
1099                                 continue;
1100                         }
1101
1102
1103                         /* pull over the reserved page structure */
1104
1105                         ret = xpc_get_remote_rp(nasid, discovered_nasids,
1106                                               remote_rp, &remote_rp_pa);
1107                         if (ret != xpcSuccess) {
1108                                 dev_dbg(xpc_part, "unable to get reserved page "
1109                                         "from nasid %d, reason=%d\n", nasid,
1110                                         ret);
1111
1112                                 if (ret == xpcLocalPartid) {
1113                                         break;
1114                                 }
1115                                 continue;
1116                         }
1117
1118                         remote_vars_pa = remote_rp->vars_pa;
1119
1120                         partid = remote_rp->partid;
1121                         part = &xpc_partitions[partid];
1122
1123
1124                         /* pull over the cross partition variables */
1125
1126                         ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
1127                         if (ret != xpcSuccess) {
1128                                 dev_dbg(xpc_part, "unable to get XPC variables "
1129                                         "from nasid %d, reason=%d\n", nasid,
1130                                         ret);
1131
1132                                 XPC_DEACTIVATE_PARTITION(part, ret);
1133                                 continue;
1134                         }
1135
1136                         if (part->act_state != XPC_P_INACTIVE) {
1137                                 dev_dbg(xpc_part, "partition %d on nasid %d is "
1138                                         "already activating\n", partid, nasid);
1139                                 break;
1140                         }
1141
1142                         /*
1143                          * Register the remote partition's AMOs with SAL so it
1144                          * can handle and cleanup errors within that address
1145                          * range should the remote partition go down. We don't
1146                          * unregister this range because it is difficult to
1147                          * tell when outstanding writes to the remote partition
1148                          * are finished and thus when it is thus safe to
1149                          * unregister. This should not result in wasted space
1150                          * in the SAL xp_addr_region table because we should
1151                          * get the same page for remote_act_amos_pa after
1152                          * module reloads and system reboots.
1153                          */
1154                         if (sn_register_xp_addr_region(
1155                                             remote_vars->amos_page_pa,
1156                                                         PAGE_SIZE, 1) < 0) {
1157                                 dev_dbg(xpc_part, "partition %d failed to "
1158                                         "register xp_addr region 0x%016lx\n",
1159                                         partid, remote_vars->amos_page_pa);
1160
1161                                 XPC_SET_REASON(part, xpcPhysAddrRegFailed,
1162                                                 __LINE__);
1163                                 break;
1164                         }
1165
1166                         /*
1167                          * The remote nasid is valid and available.
1168                          * Send an interrupt to that nasid to notify
1169                          * it that we are ready to begin activation.
1170                          */
1171                         dev_dbg(xpc_part, "sending an interrupt to AMO 0x%lx, "
1172                                 "nasid %d, phys_cpuid 0x%x\n",
1173                                 remote_vars->amos_page_pa,
1174                                 remote_vars->act_nasid,
1175                                 remote_vars->act_phys_cpuid);
1176
1177                         if (XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->
1178                                                                 version)) {
1179                                 part->remote_amos_page_pa =
1180                                                 remote_vars->amos_page_pa;
1181                                 xpc_mark_partition_disengaged(part);
1182                                 xpc_cancel_partition_disengage_request(part);
1183                         }
1184                         xpc_IPI_send_activate(remote_vars);
1185                 }
1186         }
1187
1188         kfree(discovered_nasids);
1189         kfree(remote_rp_base);
1190 }
1191
1192
1193 /*
1194  * Given a partid, get the nasids owned by that partition from the
1195  * remote partition's reserved page.
1196  */
1197 enum xpc_retval
1198 xpc_initiate_partid_to_nasids(partid_t partid, void *nasid_mask)
1199 {
1200         struct xpc_partition *part;
1201         u64 part_nasid_pa;
1202         int bte_res;
1203
1204
1205         part = &xpc_partitions[partid];
1206         if (part->remote_rp_pa == 0) {
1207                 return xpcPartitionDown;
1208         }
1209
1210         memset(nasid_mask, 0, XP_NASID_MASK_BYTES);
1211
1212         part_nasid_pa = (u64) XPC_RP_PART_NASIDS(part->remote_rp_pa);
1213
1214         bte_res = xp_bte_copy(part_nasid_pa, (u64) nasid_mask,
1215                         xp_nasid_mask_bytes, (BTE_NOTIFY | BTE_WACQUIRE), NULL);
1216
1217         return xpc_map_bte_errors(bte_res);
1218 }
1219