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  * Guarantee that the kmalloc'd memory is cacheline aligned.
  81  */
  82 void *
  83 xpc_kmalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
  84 {
  85         /* see if kmalloc will give us cachline aligned memory by default */
  86         *base = kmalloc(size, flags);
  87         if (*base == NULL) {
  88                 return NULL;
  89         }
  90         if ((u64) *base == L1_CACHE_ALIGN((u64) *base)) {
  91                 return *base;
  92         }
  93         kfree(*base);
  94
  95         /* nope, we'll have to do it ourselves */
  96         *base = kmalloc(size + L1_CACHE_BYTES, flags);
  97         if (*base == NULL) {
  98                 return NULL;
  99         }
 100         return (void *) L1_CACHE_ALIGN((u64) *base);
 101 }
 102
 103
 104 /*
 105  * Given a nasid, get the physical address of the  partition's reserved page
 106  * for that nasid. This function returns 0 on any error.
 107  */
 108 static u64
 109 xpc_get_rsvd_page_pa(int nasid)
 110 {
 111         bte_result_t bte_res;
 112         s64 status;
 113         u64 cookie = 0;
 114         u64 rp_pa = nasid;      /* seed with nasid */
 115         u64 len = 0;
 116         u64 buf = buf;
 117         u64 buf_len = 0;
 118         void *buf_base = NULL;
 119
 120
 121         while (1) {
 122
 123                 status = sn_partition_reserved_page_pa(buf, &cookie, &rp_pa,
 124                                                                 &len);
 125
 126                 dev_dbg(xpc_part, "SAL returned with status=%li, cookie="
 127                         "0x%016lx, address=0x%016lx, len=0x%016lx\n",
 128                         status, cookie, rp_pa, len);
 129
 130                 if (status != SALRET_MORE_PASSES) {
 131                         break;
 132                 }
 133
 134                 if (L1_CACHE_ALIGN(len) > buf_len) {
 135                         kfree(buf_base);
 136                         buf_len = L1_CACHE_ALIGN(len);
 137                         buf = (u64) xpc_kmalloc_cacheline_aligned(buf_len,
 138                                                         GFP_KERNEL, &buf_base);
 139                         if (buf_base == NULL) {
 140                                 dev_err(xpc_part, "unable to kmalloc "
 141                                         "len=0x%016lx\n", buf_len);
 142                                 status = SALRET_ERROR;
 143                                 break;
 144                         }
 145                 }
 146
 147                 bte_res = xp_bte_copy(rp_pa, buf, buf_len,
 148                                         (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 149                 if (bte_res != BTE_SUCCESS) {
 150                         dev_dbg(xpc_part, "xp_bte_copy failed %i\n", bte_res);
 151                         status = SALRET_ERROR;
 152                         break;
 153                 }
 154         }
 155
 156         kfree(buf_base);
 157
 158         if (status != SALRET_OK) {
 159                 rp_pa = 0;
 160         }
 161         dev_dbg(xpc_part, "reserved page at phys address 0x%016lx\n", rp_pa);
 162         return rp_pa;
 163 }
 164
 165
 166 /*
 167  * Fill the partition reserved page with the information needed by
 168  * other partitions to discover we are alive and establish initial
 169  * communications.
 170  */
 171 struct xpc_rsvd_page *
 172 xpc_rsvd_page_init(void)
 173 {
 174         struct xpc_rsvd_page *rp;
 175         AMO_t *amos_page;
 176         u64 rp_pa, nasid_array = 0;
 177         int i, ret;
 178
 179
 180         /* get the local reserved page's address */
 181
 182         preempt_disable();
 183         rp_pa = xpc_get_rsvd_page_pa(cpuid_to_nasid(smp_processor_id()));
 184         preempt_enable();
 185         if (rp_pa == 0) {
 186                 dev_err(xpc_part, "SAL failed to locate the reserved page\n");
 187                 return NULL;
 188         }
 189         rp = (struct xpc_rsvd_page *) __va(rp_pa);
 190
 191         if (rp->partid != sn_partition_id) {
 192                 dev_err(xpc_part, "the reserved page's partid of %d should be "
 193                         "%d\n", rp->partid, sn_partition_id);
 194                 return NULL;
 195         }
 196
 197         rp->version = XPC_RP_VERSION;
 198
 199         /* establish the actual sizes of the nasid masks */
 200         if (rp->SAL_version == 1) {
 201                 /* SAL_version 1 didn't set the nasids_size field */
 202                 rp->nasids_size = 128;
 203         }
 204         xp_nasid_mask_bytes = rp->nasids_size;
 205         xp_nasid_mask_words = xp_nasid_mask_bytes / 8;
 206
 207         /* setup the pointers to the various items in the reserved page */
 208         xpc_part_nasids = XPC_RP_PART_NASIDS(rp);
 209         xpc_mach_nasids = XPC_RP_MACH_NASIDS(rp);
 210         xpc_vars = XPC_RP_VARS(rp);
 211         xpc_vars_part = XPC_RP_VARS_PART(rp);
 212
 213         /*
 214          * Before clearing xpc_vars, see if a page of AMOs had been previously
 215          * allocated. If not we'll need to allocate one and set permissions
 216          * so that cross-partition AMOs are allowed.
 217          *
 218          * The allocated AMO page needs MCA reporting to remain disabled after
 219          * XPC has unloaded.  To make this work, we keep a copy of the pointer
 220          * to this page (i.e., amos_page) in the struct xpc_vars structure,
 221          * which is pointed to by the reserved page, and re-use that saved copy
 222          * on subsequent loads of XPC. This AMO page is never freed, and its
 223          * memory protections are never restricted.
 224          */
 225         if ((amos_page = xpc_vars->amos_page) == NULL) {
 226                 amos_page = (AMO_t *) TO_AMO(uncached_alloc_page(0));
 227                 if (amos_page == NULL) {
 228                         dev_err(xpc_part, "can't allocate page of AMOs\n");
 229                         return NULL;
 230                 }
 231
 232                 /*
 233                  * Open up AMO-R/W to cpu.  This is done for Shub 1.1 systems
 234                  * when xpc_allow_IPI_ops() is called via xpc_hb_init().
 235                  */
 236                 if (!enable_shub_wars_1_1()) {
 237                         ret = sn_change_memprotect(ia64_tpa((u64) amos_page),
 238                                         PAGE_SIZE, SN_MEMPROT_ACCESS_CLASS_1,
 239                                         &nasid_array);
 240                         if (ret != 0) {
 241                                 dev_err(xpc_part, "can't change memory "
 242                                         "protections\n");
 243                                 uncached_free_page(__IA64_UNCACHED_OFFSET |
 244                                                    TO_PHYS((u64) amos_page));
 245                                 return NULL;
 246                         }
 247                 }
 248         } else if (!IS_AMO_ADDRESS((u64) amos_page)) {
 249                 /*
 250                  * EFI's XPBOOT can also set amos_page in the reserved page,
 251                  * but it happens to leave it as an uncached physical address
 252                  * and we need it to be an uncached virtual, so we'll have to
 253                  * convert it.
 254                  */
 255                 if (!IS_AMO_PHYS_ADDRESS((u64) amos_page)) {
 256                         dev_err(xpc_part, "previously used amos_page address "
 257                                 "is bad = 0x%p\n", (void *) amos_page);
 258                         return NULL;
 259                 }
 260                 amos_page = (AMO_t *) TO_AMO((u64) amos_page);
 261         }
 262
 263         /* clear xpc_vars */
 264         memset(xpc_vars, 0, sizeof(struct xpc_vars));
 265
 266         xpc_vars->version = XPC_V_VERSION;
 267         xpc_vars->act_nasid = cpuid_to_nasid(0);
 268         xpc_vars->act_phys_cpuid = cpu_physical_id(0);
 269         xpc_vars->vars_part_pa = __pa(xpc_vars_part);
 270         xpc_vars->amos_page_pa = ia64_tpa((u64) amos_page);
 271         xpc_vars->amos_page = amos_page;  /* save for next load of XPC */
 272
 273
 274         /* clear xpc_vars_part */
 275         memset((u64 *) xpc_vars_part, 0, sizeof(struct xpc_vars_part) *
 276                                                         XP_MAX_PARTITIONS);
 277
 278         /* initialize the activate IRQ related AMO variables */
 279         for (i = 0; i < xp_nasid_mask_words; i++) {
 280                 (void) xpc_IPI_init(XPC_ACTIVATE_IRQ_AMOS + i);
 281         }
 282
 283         /* initialize the engaged remote partitions related AMO variables */
 284         (void) xpc_IPI_init(XPC_ENGAGED_PARTITIONS_AMO);
 285         (void) xpc_IPI_init(XPC_DISENGAGE_REQUEST_AMO);
 286
 287         /* timestamp of when reserved page was setup by XPC */
 288         rp->stamp = CURRENT_TIME;
 289
 290         /*
 291          * This signifies to the remote partition that our reserved
 292          * page is initialized.
 293          */
 294         rp->vars_pa = __pa(xpc_vars);
 295
 296         return rp;
 297 }
 298
 299
 300 /*
 301  * Change protections to allow IPI operations (and AMO operations on
 302  * Shub 1.1 systems).
 303  */
 304 void
 305 xpc_allow_IPI_ops(void)
 306 {
 307         int node;
 308         int nasid;
 309
 310
 311         // >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
 312
 313         if (is_shub2()) {
 314                 xpc_sh2_IPI_access0 =
 315                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS0));
 316                 xpc_sh2_IPI_access1 =
 317                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS1));
 318                 xpc_sh2_IPI_access2 =
 319                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS2));
 320                 xpc_sh2_IPI_access3 =
 321                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS3));
 322
 323                 for_each_online_node(node) {
 324                         nasid = cnodeid_to_nasid(node);
 325                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
 326                                                                 -1UL);
 327                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
 328                                                                 -1UL);
 329                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
 330                                                                 -1UL);
 331                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
 332                                                                 -1UL);
 333                 }
 334
 335         } else {
 336                 xpc_sh1_IPI_access =
 337                         (u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH1_IPI_ACCESS));
 338
 339                 for_each_online_node(node) {
 340                         nasid = cnodeid_to_nasid(node);
 341                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
 342                                                                 -1UL);
 343
 344                         /*
 345                          * Since the BIST collides with memory operations on
 346                          * SHUB 1.1 sn_change_memprotect() cannot be used.
 347                          */
 348                         if (enable_shub_wars_1_1()) {
 349                                 /* open up everything */
 350                                 xpc_prot_vec[node] = (u64) HUB_L((u64 *)
 351                                                 GLOBAL_MMR_ADDR(nasid,
 352                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0));
 353                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 354                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
 355                                                                 -1UL);
 356                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 357                                                 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
 358                                                                 -1UL);
 359                         }
 360                 }
 361         }
 362 }
 363
 364
 365 /*
 366  * Restrict protections to disallow IPI operations (and AMO operations on
 367  * Shub 1.1 systems).
 368  */
 369 void
 370 xpc_restrict_IPI_ops(void)
 371 {
 372         int node;
 373         int nasid;
 374
 375
 376         // >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
 377
 378         if (is_shub2()) {
 379
 380                 for_each_online_node(node) {
 381                         nasid = cnodeid_to_nasid(node);
 382                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
 383                                                         xpc_sh2_IPI_access0);
 384                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
 385                                                         xpc_sh2_IPI_access1);
 386                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
 387                                                         xpc_sh2_IPI_access2);
 388                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
 389                                                         xpc_sh2_IPI_access3);
 390                 }
 391
 392         } else {
 393
 394                 for_each_online_node(node) {
 395                         nasid = cnodeid_to_nasid(node);
 396                         HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
 397                                                         xpc_sh1_IPI_access);
 398
 399                         if (enable_shub_wars_1_1()) {
 400                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 401                                                 SH1_MD_DQLP_MMR_DIR_PRIVEC0),
 402                                                         xpc_prot_vec[node]);
 403                                 HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
 404                                                 SH1_MD_DQRP_MMR_DIR_PRIVEC0),
 405                                                         xpc_prot_vec[node]);
 406                         }
 407                 }
 408         }
 409 }
 410
 411
 412 /*
 413  * At periodic intervals, scan through all active partitions and ensure
 414  * their heartbeat is still active.  If not, the partition is deactivated.
 415  */
 416 void
 417 xpc_check_remote_hb(void)
 418 {
 419         struct xpc_vars *remote_vars;
 420         struct xpc_partition *part;
 421         partid_t partid;
 422         bte_result_t bres;
 423
 424
 425         remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
 426
 427         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 428
 429                 if (xpc_exiting) {
 430                         break;
 431                 }
 432
 433                 if (partid == sn_partition_id) {
 434                         continue;
 435                 }
 436
 437                 part = &xpc_partitions[partid];
 438
 439                 if (part->act_state == XPC_P_INACTIVE ||
 440                                 part->act_state == XPC_P_DEACTIVATING) {
 441                         continue;
 442                 }
 443
 444                 /* pull the remote_hb cache line */
 445                 bres = xp_bte_copy(part->remote_vars_pa,
 446                                         (u64) remote_vars,
 447                                         XPC_RP_VARS_SIZE,
 448                                         (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 449                 if (bres != BTE_SUCCESS) {
 450                         XPC_DEACTIVATE_PARTITION(part,
 451                                                 xpc_map_bte_errors(bres));
 452                         continue;
 453                 }
 454
 455                 dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
 456                         " = %ld, heartbeat_offline = %ld, HB_mask = 0x%lx\n",
 457                         partid, remote_vars->heartbeat, part->last_heartbeat,
 458                         remote_vars->heartbeat_offline,
 459                         remote_vars->heartbeating_to_mask);
 460
 461                 if (((remote_vars->heartbeat == part->last_heartbeat) &&
 462                         (remote_vars->heartbeat_offline == 0)) ||
 463                              !xpc_hb_allowed(sn_partition_id, remote_vars)) {
 464
 465                         XPC_DEACTIVATE_PARTITION(part, xpcNoHeartbeat);
 466                         continue;
 467                 }
 468
 469                 part->last_heartbeat = remote_vars->heartbeat;
 470         }
 471 }
 472
 473
 474 /*
 475  * Get a copy of a portion of the remote partition's rsvd page.
 476  *
 477  * remote_rp points to a buffer that is cacheline aligned for BTE copies and
 478  * is large enough to contain a copy of their reserved page header and
 479  * part_nasids mask.
 480  */
 481 static enum xpc_retval
 482 xpc_get_remote_rp(int nasid, u64 *discovered_nasids,
 483                 struct xpc_rsvd_page *remote_rp, u64 *remote_rp_pa)
 484 {
 485         int bres, i;
 486
 487
 488         /* get the reserved page's physical address */
 489
 490         *remote_rp_pa = xpc_get_rsvd_page_pa(nasid);
 491         if (*remote_rp_pa == 0) {
 492                 return xpcNoRsvdPageAddr;
 493         }
 494
 495
 496         /* pull over the reserved page header and part_nasids mask */
 497         bres = xp_bte_copy(*remote_rp_pa, (u64) remote_rp,
 498                                 XPC_RP_HEADER_SIZE + xp_nasid_mask_bytes,
 499                                 (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 500         if (bres != BTE_SUCCESS) {
 501                 return xpc_map_bte_errors(bres);
 502         }
 503
 504
 505         if (discovered_nasids != NULL) {
 506                 u64 *remote_part_nasids = XPC_RP_PART_NASIDS(remote_rp);
 507
 508
 509                 for (i = 0; i < xp_nasid_mask_words; i++) {
 510                         discovered_nasids[i] |= remote_part_nasids[i];
 511                 }
 512         }
 513
 514
 515         /* check that the partid is for another partition */
 516
 517         if (remote_rp->partid < 1 ||
 518                                 remote_rp->partid > (XP_MAX_PARTITIONS - 1)) {
 519                 return xpcInvalidPartid;
 520         }
 521
 522         if (remote_rp->partid == sn_partition_id) {
 523                 return xpcLocalPartid;
 524         }
 525
 526
 527         if (XPC_VERSION_MAJOR(remote_rp->version) !=
 528                                         XPC_VERSION_MAJOR(XPC_RP_VERSION)) {
 529                 return xpcBadVersion;
 530         }
 531
 532         return xpcSuccess;
 533 }
 534
 535
 536 /*
 537  * Get a copy of the remote partition's XPC variables from the reserved page.
 538  *
 539  * remote_vars points to a buffer that is cacheline aligned for BTE copies and
 540  * assumed to be of size XPC_RP_VARS_SIZE.
 541  */
 542 static enum xpc_retval
 543 xpc_get_remote_vars(u64 remote_vars_pa, struct xpc_vars *remote_vars)
 544 {
 545         int bres;
 546
 547
 548         if (remote_vars_pa == 0) {
 549                 return xpcVarsNotSet;
 550         }
 551
 552         /* pull over the cross partition variables */
 553         bres = xp_bte_copy(remote_vars_pa, (u64) remote_vars, XPC_RP_VARS_SIZE,
 554                                 (BTE_NOTIFY | BTE_WACQUIRE), NULL);
 555         if (bres != BTE_SUCCESS) {
 556                 return xpc_map_bte_errors(bres);
 557         }
 558
 559         if (XPC_VERSION_MAJOR(remote_vars->version) !=
 560                                         XPC_VERSION_MAJOR(XPC_V_VERSION)) {
 561                 return xpcBadVersion;
 562         }
 563
 564         return xpcSuccess;
 565 }
 566
 567
 568 /*
 569  * Update the remote partition's info.
 570  */
 571 static void
 572 xpc_update_partition_info(struct xpc_partition *part, u8 remote_rp_version,
 573                 struct timespec *remote_rp_stamp, u64 remote_rp_pa,
 574                 u64 remote_vars_pa, struct xpc_vars *remote_vars)
 575 {
 576         part->remote_rp_version = remote_rp_version;
 577         dev_dbg(xpc_part, "  remote_rp_version = 0x%016lx\n",
 578                 part->remote_rp_version);
 579
 580         part->remote_rp_stamp = *remote_rp_stamp;
 581         dev_dbg(xpc_part, "  remote_rp_stamp (tv_sec = 0x%lx tv_nsec = 0x%lx\n",
 582                 part->remote_rp_stamp.tv_sec, part->remote_rp_stamp.tv_nsec);
 583
 584         part->remote_rp_pa = remote_rp_pa;
 585         dev_dbg(xpc_part, "  remote_rp_pa = 0x%016lx\n", part->remote_rp_pa);
 586
 587         part->remote_vars_pa = remote_vars_pa;
 588         dev_dbg(xpc_part, "  remote_vars_pa = 0x%016lx\n",
 589                 part->remote_vars_pa);
 590
 591         part->last_heartbeat = remote_vars->heartbeat;
 592         dev_dbg(xpc_part, "  last_heartbeat = 0x%016lx\n",
 593                 part->last_heartbeat);
 594
 595         part->remote_vars_part_pa = remote_vars->vars_part_pa;
 596         dev_dbg(xpc_part, "  remote_vars_part_pa = 0x%016lx\n",
 597                 part->remote_vars_part_pa);
 598
 599         part->remote_act_nasid = remote_vars->act_nasid;
 600         dev_dbg(xpc_part, "  remote_act_nasid = 0x%x\n",
 601                 part->remote_act_nasid);
 602
 603         part->remote_act_phys_cpuid = remote_vars->act_phys_cpuid;
 604         dev_dbg(xpc_part, "  remote_act_phys_cpuid = 0x%x\n",
 605                 part->remote_act_phys_cpuid);
 606
 607         part->remote_amos_page_pa = remote_vars->amos_page_pa;
 608         dev_dbg(xpc_part, "  remote_amos_page_pa = 0x%lx\n",
 609                 part->remote_amos_page_pa);
 610
 611         part->remote_vars_version = remote_vars->version;
 612         dev_dbg(xpc_part, "  remote_vars_version = 0x%x\n",
 613                 part->remote_vars_version);
 614 }
 615
 616
 617 /*
 618  * Prior code has determined the nasid which generated an IPI.  Inspect
 619  * that nasid to determine if its partition needs to be activated or
 620  * deactivated.
 621  *
 622  * A partition is consider "awaiting activation" if our partition
 623  * flags indicate it is not active and it has a heartbeat.  A
 624  * partition is considered "awaiting deactivation" if our partition
 625  * flags indicate it is active but it has no heartbeat or it is not
 626  * sending its heartbeat to us.
 627  *
 628  * To determine the heartbeat, the remote nasid must have a properly
 629  * initialized reserved page.
 630  */
 631 static void
 632 xpc_identify_act_IRQ_req(int nasid)
 633 {
 634         struct xpc_rsvd_page *remote_rp;
 635         struct xpc_vars *remote_vars;
 636         u64 remote_rp_pa;
 637         u64 remote_vars_pa;
 638         int remote_rp_version;
 639         int reactivate = 0;
 640         int stamp_diff;
 641         struct timespec remote_rp_stamp = { 0, 0 };
 642         partid_t partid;
 643         struct xpc_partition *part;
 644         enum xpc_retval ret;
 645
 646
 647         /* pull over the reserved page structure */
 648
 649         remote_rp = (struct xpc_rsvd_page *) xpc_remote_copy_buffer;
 650
 651         ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rp_pa);
 652         if (ret != xpcSuccess) {
 653                 dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
 654                         "which sent interrupt, reason=%d\n", nasid, ret);
 655                 return;
 656         }
 657
 658         remote_vars_pa = remote_rp->vars_pa;
 659         remote_rp_version = remote_rp->version;
 660         if (XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 661                 remote_rp_stamp = remote_rp->stamp;
 662         }
 663         partid = remote_rp->partid;
 664         part = &xpc_partitions[partid];
 665
 666
 667         /* pull over the cross partition variables */
 668
 669         remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
 670
 671         ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
 672         if (ret != xpcSuccess) {
 673
 674                 dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
 675                         "which sent interrupt, reason=%d\n", nasid, ret);
 676
 677                 XPC_DEACTIVATE_PARTITION(part, ret);
 678                 return;
 679         }
 680
 681
 682         part->act_IRQ_rcvd++;
 683
 684         dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
 685                 "%ld:0x%lx\n", (int) nasid, (int) partid, part->act_IRQ_rcvd,
 686                 remote_vars->heartbeat, remote_vars->heartbeating_to_mask);
 687
 688         if (xpc_partition_disengaged(part) &&
 689                                         part->act_state == XPC_P_INACTIVE) {
 690
 691                 xpc_update_partition_info(part, remote_rp_version,
 692                                         &remote_rp_stamp, remote_rp_pa,
 693                                         remote_vars_pa, remote_vars);
 694
 695                 if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 696                         if (xpc_partition_disengage_requested(1UL << partid)) {
 697                                 /*
 698                                  * Other side is waiting on us to disengage,
 699                                  * even though we already have.
 700                                  */
 701                                 return;
 702                         }
 703                 } else {
 704                         /* other side doesn't support disengage requests */
 705                         xpc_clear_partition_disengage_request(1UL << partid);
 706                 }
 707
 708                 xpc_activate_partition(part);
 709                 return;
 710         }
 711
 712         DBUG_ON(part->remote_rp_version == 0);
 713         DBUG_ON(part->remote_vars_version == 0);
 714
 715         if (!XPC_SUPPORTS_RP_STAMP(part->remote_rp_version)) {
 716                 DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(part->
 717                                                         remote_vars_version));
 718
 719                 if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 720                         DBUG_ON(XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->
 721                                                                 version));
 722                         /* see if the other side rebooted */
 723                         if (part->remote_amos_page_pa ==
 724                                 remote_vars->amos_page_pa &&
 725                                         xpc_hb_allowed(sn_partition_id,
 726                                                                 remote_vars)) {
 727                                 /* doesn't look that way, so ignore the IPI */
 728                                 return;
 729                         }
 730                 }
 731
 732                 /*
 733                  * Other side rebooted and previous XPC didn't support the
 734                  * disengage request, so we don't need to do anything special.
 735                  */
 736
 737                 xpc_update_partition_info(part, remote_rp_version,
 738                                                 &remote_rp_stamp, remote_rp_pa,
 739                                                 remote_vars_pa, remote_vars);
 740                 part->reactivate_nasid = nasid;
 741                 XPC_DEACTIVATE_PARTITION(part, xpcReactivating);
 742                 return;
 743         }
 744
 745         DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version));
 746
 747         if (!XPC_SUPPORTS_RP_STAMP(remote_rp_version)) {
 748                 DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version));
 749
 750                 /*
 751                  * Other side rebooted and previous XPC did support the
 752                  * disengage request, but the new one doesn't.
 753                  */
 754
 755                 xpc_clear_partition_engaged(1UL << partid);
 756                 xpc_clear_partition_disengage_request(1UL << partid);
 757
 758                 xpc_update_partition_info(part, remote_rp_version,
 759                                                 &remote_rp_stamp, remote_rp_pa,
 760                                                 remote_vars_pa, remote_vars);
 761                 reactivate = 1;
 762
 763         } else {
 764                 DBUG_ON(!XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->version));
 765
 766                 stamp_diff = xpc_compare_stamps(&part->remote_rp_stamp,
 767                                                         &remote_rp_stamp);
 768                 if (stamp_diff != 0) {
 769                         DBUG_ON(stamp_diff >= 0);
 770
 771                         /*
 772                          * Other side rebooted and the previous XPC did support
 773                          * the disengage request, as does the new one.
 774                          */
 775
 776                         DBUG_ON(xpc_partition_engaged(1UL << partid));
 777                         DBUG_ON(xpc_partition_disengage_requested(1UL <<
 778                                                                 partid));
 779
 780                         xpc_update_partition_info(part, remote_rp_version,
 781                                                 &remote_rp_stamp, remote_rp_pa,
 782                                                 remote_vars_pa, remote_vars);
 783                         reactivate = 1;
 784                 }
 785         }
 786
 787         if (part->disengage_request_timeout > 0 &&
 788                                         !xpc_partition_disengaged(part)) {
 789                 /* still waiting on other side to disengage from us */
 790                 return;
 791         }
 792
 793         if (reactivate) {
 794                 part->reactivate_nasid = nasid;
 795                 XPC_DEACTIVATE_PARTITION(part, xpcReactivating);
 796
 797         } else if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version) &&
 798                         xpc_partition_disengage_requested(1UL << partid)) {
 799                 XPC_DEACTIVATE_PARTITION(part, xpcOtherGoingDown);
 800         }
 801 }
 802
 803
 804 /*
 805  * Loop through the activation AMO variables and process any bits
 806  * which are set.  Each bit indicates a nasid sending a partition
 807  * activation or deactivation request.
 808  *
 809  * Return #of IRQs detected.
 810  */
 811 int
 812 xpc_identify_act_IRQ_sender(void)
 813 {
 814         int word, bit;
 815         u64 nasid_mask;
 816         u64 nasid;                      /* remote nasid */
 817         int n_IRQs_detected = 0;
 818         AMO_t *act_amos;
 819
 820
 821         act_amos = xpc_vars->amos_page + XPC_ACTIVATE_IRQ_AMOS;
 822
 823
 824         /* scan through act AMO variable looking for non-zero entries */
 825         for (word = 0; word < xp_nasid_mask_words; word++) {
 826
 827                 if (xpc_exiting) {
 828                         break;
 829                 }
 830
 831                 nasid_mask = xpc_IPI_receive(&act_amos[word]);
 832                 if (nasid_mask == 0) {
 833                         /* no IRQs from nasids in this variable */
 834                         continue;
 835                 }
 836
 837                 dev_dbg(xpc_part, "AMO[%d] gave back 0x%lx\n", word,
 838                         nasid_mask);
 839
 840
 841                 /*
 842                  * If this nasid has been added to the machine since
 843                  * our partition was reset, this will retain the
 844                  * remote nasid in our reserved pages machine mask.
 845                  * This is used in the event of module reload.
 846                  */
 847                 xpc_mach_nasids[word] |= nasid_mask;
 848
 849
 850                 /* locate the nasid(s) which sent interrupts */
 851
 852                 for (bit = 0; bit < (8 * sizeof(u64)); bit++) {
 853                         if (nasid_mask & (1UL << bit)) {
 854                                 n_IRQs_detected++;
 855                                 nasid = XPC_NASID_FROM_W_B(word, bit);
 856                                 dev_dbg(xpc_part, "interrupt from nasid %ld\n",
 857                                         nasid);
 858                                 xpc_identify_act_IRQ_req(nasid);
 859                         }
 860                 }
 861         }
 862         return n_IRQs_detected;
 863 }
 864
 865
 866 /*
 867  * See if the other side has responded to a partition disengage request
 868  * from us.
 869  */
 870 int
 871 xpc_partition_disengaged(struct xpc_partition *part)
 872 {
 873         partid_t partid = XPC_PARTID(part);
 874         int disengaged;
 875
 876
 877         disengaged = (xpc_partition_engaged(1UL << partid) == 0);
 878         if (part->disengage_request_timeout) {
 879                 if (!disengaged) {
 880                         if (jiffies < part->disengage_request_timeout) {
 881                                 /* timelimit hasn't been reached yet */
 882                                 return 0;
 883                         }
 884
 885                         /*
 886                          * Other side hasn't responded to our disengage
 887                          * request in a timely fashion, so assume it's dead.
 888                          */
 889
 890                         dev_info(xpc_part, "disengage from remote partition %d "
 891                                 "timed out\n", partid);
 892                         xpc_disengage_request_timedout = 1;
 893                         xpc_clear_partition_engaged(1UL << partid);
 894                         disengaged = 1;
 895                 }
 896                 part->disengage_request_timeout = 0;
 897
 898                 /* cancel the timer function, provided it's not us */
 899                 if (!in_interrupt()) {
 900                         del_singleshot_timer_sync(&part->
 901                                                       disengage_request_timer);
 902                 }
 903
 904                 DBUG_ON(part->act_state != XPC_P_DEACTIVATING &&
 905                                         part->act_state != XPC_P_INACTIVE);
 906                 if (part->act_state != XPC_P_INACTIVE) {
 907                         xpc_wakeup_channel_mgr(part);
 908                 }
 909
 910                 if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 911                         xpc_cancel_partition_disengage_request(part);
 912                 }
 913         }
 914         return disengaged;
 915 }
 916
 917
 918 /*
 919  * Mark specified partition as active.
 920  */
 921 enum xpc_retval
 922 xpc_mark_partition_active(struct xpc_partition *part)
 923 {
 924         unsigned long irq_flags;
 925         enum xpc_retval ret;
 926
 927
 928         dev_dbg(xpc_part, "setting partition %d to ACTIVE\n", XPC_PARTID(part));
 929
 930         spin_lock_irqsave(&part->act_lock, irq_flags);
 931         if (part->act_state == XPC_P_ACTIVATING) {
 932                 part->act_state = XPC_P_ACTIVE;
 933                 ret = xpcSuccess;
 934         } else {
 935                 DBUG_ON(part->reason == xpcSuccess);
 936                 ret = part->reason;
 937         }
 938         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 939
 940         return ret;
 941 }
 942
 943
 944 /*
 945  * Notify XPC that the partition is down.
 946  */
 947 void
 948 xpc_deactivate_partition(const int line, struct xpc_partition *part,
 949                                 enum xpc_retval reason)
 950 {
 951         unsigned long irq_flags;
 952
 953
 954         spin_lock_irqsave(&part->act_lock, irq_flags);
 955
 956         if (part->act_state == XPC_P_INACTIVE) {
 957                 XPC_SET_REASON(part, reason, line);
 958                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 959                 if (reason == xpcReactivating) {
 960                         /* we interrupt ourselves to reactivate partition */
 961                         xpc_IPI_send_reactivate(part);
 962                 }
 963                 return;
 964         }
 965         if (part->act_state == XPC_P_DEACTIVATING) {
 966                 if ((part->reason == xpcUnloading && reason != xpcUnloading) ||
 967                                         reason == xpcReactivating) {
 968                         XPC_SET_REASON(part, reason, line);
 969                 }
 970                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 971                 return;
 972         }
 973
 974         part->act_state = XPC_P_DEACTIVATING;
 975         XPC_SET_REASON(part, reason, line);
 976
 977         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 978
 979         if (XPC_SUPPORTS_DISENGAGE_REQUEST(part->remote_vars_version)) {
 980                 xpc_request_partition_disengage(part);
 981                 xpc_IPI_send_disengage(part);
 982
 983                 /* set a timelimit on the disengage request */
 984                 part->disengage_request_timeout = jiffies +
 985                                         (xpc_disengage_request_timelimit * HZ);
 986                 part->disengage_request_timer.expires =
 987                                         part->disengage_request_timeout;
 988                 add_timer(&part->disengage_request_timer);
 989         }
 990
 991         dev_dbg(xpc_part, "bringing partition %d down, reason = %d\n",
 992                 XPC_PARTID(part), reason);
 993
 994         xpc_partition_going_down(part, reason);
 995 }
 996
 997
 998 /*
 999  * Mark specified partition as inactive.
1000  */
1001 void
1002 xpc_mark_partition_inactive(struct xpc_partition *part)
1003 {
1004         unsigned long irq_flags;
1005
1006
1007         dev_dbg(xpc_part, "setting partition %d to INACTIVE\n",
1008                 XPC_PARTID(part));
1009
1010         spin_lock_irqsave(&part->act_lock, irq_flags);
1011         part->act_state = XPC_P_INACTIVE;
1012         spin_unlock_irqrestore(&part->act_lock, irq_flags);
1013         part->remote_rp_pa = 0;
1014 }
1015
1016
1017 /*
1018  * SAL has provided a partition and machine mask.  The partition mask
1019  * contains a bit for each even nasid in our partition.  The machine
1020  * mask contains a bit for each even nasid in the entire machine.
1021  *
1022  * Using those two bit arrays, we can determine which nasids are
1023  * known in the machine.  Each should also have a reserved page
1024  * initialized if they are available for partitioning.
1025  */
1026 void
1027 xpc_discovery(void)
1028 {
1029         void *remote_rp_base;
1030         struct xpc_rsvd_page *remote_rp;
1031         struct xpc_vars *remote_vars;
1032         u64 remote_rp_pa;
1033         u64 remote_vars_pa;
1034         int region;
1035         int region_size;
1036         int max_regions;
1037         int nasid;
1038         struct xpc_rsvd_page *rp;
1039         partid_t partid;
1040         struct xpc_partition *part;
1041         u64 *discovered_nasids;
1042         enum xpc_retval ret;
1043
1044
1045         remote_rp = xpc_kmalloc_cacheline_aligned(XPC_RP_HEADER_SIZE +
1046                                                 xp_nasid_mask_bytes,
1047                                                 GFP_KERNEL, &remote_rp_base);
1048         if (remote_rp == NULL) {
1049                 return;
1050         }
1051         remote_vars = (struct xpc_vars *) remote_rp;
1052
1053
1054         discovered_nasids = kzalloc(sizeof(u64) * xp_nasid_mask_words,
1055                                                         GFP_KERNEL);
1056         if (discovered_nasids == NULL) {
1057                 kfree(remote_rp_base);
1058                 return;
1059         }
1060
1061         rp = (struct xpc_rsvd_page *) xpc_rsvd_page;
1062
1063         /*
1064          * The term 'region' in this context refers to the minimum number of
1065          * nodes that can comprise an access protection grouping. The access
1066          * protection is in regards to memory, IOI and IPI.
1067          */
1068         max_regions = 64;
1069         region_size = sn_region_size;
1070
1071         switch (region_size) {
1072         case 128:
1073                 max_regions *= 2;
1074         case 64:
1075                 max_regions *= 2;
1076         case 32:
1077                 max_regions *= 2;
1078                 region_size = 16;
1079                 DBUG_ON(!is_shub2());
1080         }
1081
1082         for (region = 0; region < max_regions; region++) {
1083
1084                 if ((volatile int) xpc_exiting) {
1085                         break;
1086                 }
1087
1088                 dev_dbg(xpc_part, "searching region %d\n", region);
1089
1090                 for (nasid = (region * region_size * 2);
1091                      nasid < ((region + 1) * region_size * 2);
1092                      nasid += 2) {
1093
1094                         if ((volatile int) xpc_exiting) {
1095                                 break;
1096                         }
1097
1098                         dev_dbg(xpc_part, "checking nasid %d\n", nasid);
1099
1100
1101                         if (XPC_NASID_IN_ARRAY(nasid, xpc_part_nasids)) {
1102                                 dev_dbg(xpc_part, "PROM indicates Nasid %d is "
1103                                         "part of the local partition; skipping "
1104                                         "region\n", nasid);
1105                                 break;
1106                         }
1107
1108                         if (!(XPC_NASID_IN_ARRAY(nasid, xpc_mach_nasids))) {
1109                                 dev_dbg(xpc_part, "PROM indicates Nasid %d was "
1110                                         "not on Numa-Link network at reset\n",
1111                                         nasid);
1112                                 continue;
1113                         }
1114
1115                         if (XPC_NASID_IN_ARRAY(nasid, discovered_nasids)) {
1116                                 dev_dbg(xpc_part, "Nasid %d is part of a "
1117                                         "partition which was previously "
1118                                         "discovered\n", nasid);
1119                                 continue;
1120                         }
1121
1122
1123                         /* pull over the reserved page structure */
1124
1125                         ret = xpc_get_remote_rp(nasid, discovered_nasids,
1126                                               remote_rp, &remote_rp_pa);
1127                         if (ret != xpcSuccess) {
1128                                 dev_dbg(xpc_part, "unable to get reserved page "
1129                                         "from nasid %d, reason=%d\n", nasid,
1130                                         ret);
1131
1132                                 if (ret == xpcLocalPartid) {
1133                                         break;
1134                                 }
1135                                 continue;
1136                         }
1137
1138                         remote_vars_pa = remote_rp->vars_pa;
1139
1140                         partid = remote_rp->partid;
1141                         part = &xpc_partitions[partid];
1142
1143
1144                         /* pull over the cross partition variables */
1145
1146                         ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
1147                         if (ret != xpcSuccess) {
1148                                 dev_dbg(xpc_part, "unable to get XPC variables "
1149                                         "from nasid %d, reason=%d\n", nasid,
1150                                         ret);
1151
1152                                 XPC_DEACTIVATE_PARTITION(part, ret);
1153                                 continue;
1154                         }
1155
1156                         if (part->act_state != XPC_P_INACTIVE) {
1157                                 dev_dbg(xpc_part, "partition %d on nasid %d is "
1158                                         "already activating\n", partid, nasid);
1159                                 break;
1160                         }
1161
1162                         /*
1163                          * Register the remote partition's AMOs with SAL so it
1164                          * can handle and cleanup errors within that address
1165                          * range should the remote partition go down. We don't
1166                          * unregister this range because it is difficult to
1167                          * tell when outstanding writes to the remote partition
1168                          * are finished and thus when it is thus safe to
1169                          * unregister. This should not result in wasted space
1170                          * in the SAL xp_addr_region table because we should
1171                          * get the same page for remote_act_amos_pa after
1172                          * module reloads and system reboots.
1173                          */
1174                         if (sn_register_xp_addr_region(
1175                                             remote_vars->amos_page_pa,
1176                                                         PAGE_SIZE, 1) < 0) {
1177                                 dev_dbg(xpc_part, "partition %d failed to "
1178                                         "register xp_addr region 0x%016lx\n",
1179                                         partid, remote_vars->amos_page_pa);
1180
1181                                 XPC_SET_REASON(part, xpcPhysAddrRegFailed,
1182                                                 __LINE__);
1183                                 break;
1184                         }
1185
1186                         /*
1187                          * The remote nasid is valid and available.
1188                          * Send an interrupt to that nasid to notify
1189                          * it that we are ready to begin activation.
1190                          */
1191                         dev_dbg(xpc_part, "sending an interrupt to AMO 0x%lx, "
1192                                 "nasid %d, phys_cpuid 0x%x\n",
1193                                 remote_vars->amos_page_pa,
1194                                 remote_vars->act_nasid,
1195                                 remote_vars->act_phys_cpuid);
1196
1197                         if (XPC_SUPPORTS_DISENGAGE_REQUEST(remote_vars->
1198                                                                 version)) {
1199                                 part->remote_amos_page_pa =
1200                                                 remote_vars->amos_page_pa;
1201                                 xpc_mark_partition_disengaged(part);
1202                                 xpc_cancel_partition_disengage_request(part);
1203                         }
1204                         xpc_IPI_send_activate(remote_vars);
1205                 }
1206         }
1207
1208         kfree(discovered_nasids);
1209         kfree(remote_rp_base);
1210 }
1211
1212
1213 /*
1214  * Given a partid, get the nasids owned by that partition from the
1215  * remote partition's reserved page.
1216  */
1217 enum xpc_retval
1218 xpc_initiate_partid_to_nasids(partid_t partid, void *nasid_mask)
1219 {
1220         struct xpc_partition *part;
1221         u64 part_nasid_pa;
1222         int bte_res;
1223
1224
1225         part = &xpc_partitions[partid];
1226         if (part->remote_rp_pa == 0) {
1227                 return xpcPartitionDown;
1228         }
1229
1230         memset(nasid_mask, 0, XP_NASID_MASK_BYTES);
1231
1232         part_nasid_pa = (u64) XPC_RP_PART_NASIDS(part->remote_rp_pa);
1233
1234         bte_res = xp_bte_copy(part_nasid_pa, (u64) nasid_mask,
1235                         xp_nasid_mask_bytes, (BTE_NOTIFY | BTE_WACQUIRE), NULL);
1236
1237         return xpc_map_bte_errors(bte_res);
1238 }
1239