arch/ia64/sn/kernel/bte.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) 2000-2007 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   8
   9 #include <linux/module.h>
  10 #include <asm/sn/nodepda.h>
  11 #include <asm/sn/addrs.h>
  12 #include <asm/sn/arch.h>
  13 #include <asm/sn/sn_cpuid.h>
  14 #include <asm/sn/pda.h>
  15 #include <asm/sn/shubio.h>
  16 #include <asm/nodedata.h>
  17 #include <asm/delay.h>
  18
  19 #include <linux/bootmem.h>
  20 #include <linux/string.h>
  21 #include <linux/sched.h>
  22
  23 #include <asm/sn/bte.h>
  24
  25 #ifndef L1_CACHE_MASK
  26 #define L1_CACHE_MASK (L1_CACHE_BYTES - 1)
  27 #endif
  28
  29 /* two interfaces on two btes */
  30 #define MAX_INTERFACES_TO_TRY           4
  31 #define MAX_NODES_TO_TRY                2
  32
  33 static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
  34 {
  35         nodepda_t *tmp_nodepda;
  36
  37         if (nasid_to_cnodeid(nasid) == -1)
  38                 return (struct bteinfo_s *)NULL;
  39
  40         tmp_nodepda = NODEPDA(nasid_to_cnodeid(nasid));
  41         return &tmp_nodepda->bte_if[interface];
  42
  43 }
  44
  45 static inline void bte_start_transfer(struct bteinfo_s *bte, u64 len, u64 mode)
  46 {
  47         if (is_shub2()) {
  48                 BTE_CTRL_STORE(bte, (IBLS_BUSY | ((len) | (mode) << 24)));
  49         } else {
  50                 BTE_LNSTAT_STORE(bte, len);
  51                 BTE_CTRL_STORE(bte, mode);
  52         }
  53 }
  54
  55 /************************************************************************
  56  * Block Transfer Engine copy related functions.
  57  *
  58  ***********************************************************************/
  59
  60 /*
  61  * bte_copy(src, dest, len, mode, notification)
  62  *
  63  * Use the block transfer engine to move kernel memory from src to dest
  64  * using the assigned mode.
  65  *
  66  * Parameters:
  67  *   src - physical address of the transfer source.
  68  *   dest - physical address of the transfer destination.
  69  *   len - number of bytes to transfer from source to dest.
  70  *   mode - hardware defined.  See reference information
  71  *          for IBCT0/1 in the SHUB Programmers Reference
  72  *   notification - kernel virtual address of the notification cache
  73  *                  line.  If NULL, the default is used and
  74  *                  the bte_copy is synchronous.
  75  *
  76  * NOTE:  This function requires src, dest, and len to
  77  * be cacheline aligned.
  78  */
  79 bte_result_t bte_copy(u64 src, u64 dest, u64 len, u64 mode, void *notification)
  80 {
  81         u64 transfer_size;
  82         u64 transfer_stat;
  83         u64 notif_phys_addr;
  84         struct bteinfo_s *bte;
  85         bte_result_t bte_status;
  86         unsigned long irq_flags;
  87         unsigned long itc_end = 0;
  88         int nasid_to_try[MAX_NODES_TO_TRY];
  89         int my_nasid = cpuid_to_nasid(raw_smp_processor_id());
  90         int bte_if_index, nasid_index;
  91         int bte_first, btes_per_node = BTES_PER_NODE;
  92
  93         BTE_PRINTK(("bte_copy(0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%p)\n",
  94                     src, dest, len, mode, notification));
  95
  96         if (len == 0) {
  97                 return BTE_SUCCESS;
  98         }
  99
 100         BUG_ON(len & L1_CACHE_MASK);
 101         BUG_ON(src & L1_CACHE_MASK);
 102         BUG_ON(dest & L1_CACHE_MASK);
 103         BUG_ON(len > BTE_MAX_XFER);
 104
 105         /*
 106          * Start with interface corresponding to cpu number
 107          */
 108         bte_first = raw_smp_processor_id() % btes_per_node;
 109
 110         if (mode & BTE_USE_DEST) {
 111                 /* try remote then local */
 112                 nasid_to_try[0] = NASID_GET(dest);
 113                 if (mode & BTE_USE_ANY) {
 114                         nasid_to_try[1] = my_nasid;
 115                 } else {
 116                         nasid_to_try[1] = (int)NULL;
 117                 }
 118         } else {
 119                 /* try local then remote */
 120                 nasid_to_try[0] = my_nasid;
 121                 if (mode & BTE_USE_ANY) {
 122                         nasid_to_try[1] = NASID_GET(dest);
 123                 } else {
 124                         nasid_to_try[1] = (int)NULL;
 125                 }
 126         }
 127
 128 retry_bteop:
 129         do {
 130                 local_irq_save(irq_flags);
 131
 132                 bte_if_index = bte_first;
 133                 nasid_index = 0;
 134
 135                 /* Attempt to lock one of the BTE interfaces. */
 136                 while (nasid_index < MAX_NODES_TO_TRY) {
 137                         bte = bte_if_on_node(nasid_to_try[nasid_index],bte_if_index);
 138
 139                         if (bte == NULL) {
 140                                 nasid_index++;
 141                                 continue;
 142                         }
 143
 144                         if (spin_trylock(&bte->spinlock)) {
 145                                 if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
 146                                     (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
 147                                         /* Got the lock but BTE still busy */
 148                                         spin_unlock(&bte->spinlock);
 149                                 } else {
 150                                         /* we got the lock and it's not busy */
 151                                         break;
 152                                 }
 153                         }
 154
 155                         bte_if_index = (bte_if_index + 1) % btes_per_node; /* Next interface */
 156                         if (bte_if_index == bte_first) {
 157                                 /*
 158                                  * We've tried all interfaces on this node
 159                                  */
 160                                 nasid_index++;
 161                         }
 162
 163                         bte = NULL;
 164                 }
 165
 166                 if (bte != NULL) {
 167                         break;
 168                 }
 169
 170                 local_irq_restore(irq_flags);
 171
 172                 if (!(mode & BTE_WACQUIRE)) {
 173                         return BTEFAIL_NOTAVAIL;
 174                 }
 175         } while (1);
 176
 177         if (notification == NULL) {
 178                 /* User does not want to be notified. */
 179                 bte->most_rcnt_na = &bte->notify;
 180         } else {
 181                 bte->most_rcnt_na = notification;
 182         }
 183
 184         /* Calculate the number of cache lines to transfer. */
 185         transfer_size = ((len >> L1_CACHE_SHIFT) & BTE_LEN_MASK);
 186
 187         /* Initialize the notification to a known value. */
 188         *bte->most_rcnt_na = BTE_WORD_BUSY;
 189         notif_phys_addr = (u64)bte->most_rcnt_na;
 190
 191         /* Set the source and destination registers */
 192         BTE_PRINTKV(("IBSA = 0x%lx)\n", src));
 193         BTE_SRC_STORE(bte, src);
 194         BTE_PRINTKV(("IBDA = 0x%lx)\n", dest));
 195         BTE_DEST_STORE(bte, dest);
 196
 197         /* Set the notification register */
 198         BTE_PRINTKV(("IBNA = 0x%lx)\n", notif_phys_addr));
 199         BTE_NOTIF_STORE(bte, notif_phys_addr);
 200
 201         /* Initiate the transfer */
 202         BTE_PRINTK(("IBCT = 0x%lx)\n", BTE_VALID_MODE(mode)));
 203         bte_start_transfer(bte, transfer_size, BTE_VALID_MODE(mode));
 204
 205         itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
 206
 207         spin_unlock_irqrestore(&bte->spinlock, irq_flags);
 208
 209         if (notification != NULL) {
 210                 return BTE_SUCCESS;
 211         }
 212
 213         while ((transfer_stat = *bte->most_rcnt_na) == BTE_WORD_BUSY) {
 214                 cpu_relax();
 215                 if (ia64_get_itc() > itc_end) {
 216                         BTE_PRINTK(("BTE timeout nasid 0x%x bte%d IBLS = 0x%lx na 0x%lx\n",
 217                                 NASID_GET(bte->bte_base_addr), bte->bte_num,
 218                                 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na) );
 219                         bte->bte_error_count++;
 220                         bte->bh_error = IBLS_ERROR;
 221                         bte_error_handler((unsigned long)NODEPDA(bte->bte_cnode));
 222                         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 223                         goto retry_bteop;
 224                 }
 225         }
 226
 227         BTE_PRINTKV((" Delay Done.  IBLS = 0x%lx, most_rcnt_na = 0x%lx\n",
 228                      BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 229
 230         if (transfer_stat & IBLS_ERROR) {
 231                 bte_status = BTE_GET_ERROR_STATUS(transfer_stat);
 232         } else {
 233                 bte_status = BTE_SUCCESS;
 234         }
 235         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 236
 237         BTE_PRINTK(("Returning status is 0x%lx and most_rcnt_na is 0x%lx\n",
 238                     BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 239
 240         return bte_status;
 241 }
 242
 243 EXPORT_SYMBOL(bte_copy);
 244
 245 /*
 246  * bte_unaligned_copy(src, dest, len, mode)
 247  *
 248  * use the block transfer engine to move kernel
 249  * memory from src to dest using the assigned mode.
 250  *
 251  * Parameters:
 252  *   src - physical address of the transfer source.
 253  *   dest - physical address of the transfer destination.
 254  *   len - number of bytes to transfer from source to dest.
 255  *   mode - hardware defined.  See reference information
 256  *          for IBCT0/1 in the SGI documentation.
 257  *
 258  * NOTE: If the source, dest, and len are all cache line aligned,
 259  * then it would be _FAR_ preferable to use bte_copy instead.
 260  */
 261 bte_result_t bte_unaligned_copy(u64 src, u64 dest, u64 len, u64 mode)
 262 {
 263         int destFirstCacheOffset;
 264         u64 headBteSource;
 265         u64 headBteLen;
 266         u64 headBcopySrcOffset;
 267         u64 headBcopyDest;
 268         u64 headBcopyLen;
 269         u64 footBteSource;
 270         u64 footBteLen;
 271         u64 footBcopyDest;
 272         u64 footBcopyLen;
 273         bte_result_t rv;
 274         char *bteBlock, *bteBlock_unaligned;
 275
 276         if (len == 0) {
 277                 return BTE_SUCCESS;
 278         }
 279
 280         /* temporary buffer used during unaligned transfers */
 281         bteBlock_unaligned = kmalloc(len + 3 * L1_CACHE_BYTES, GFP_KERNEL);
 282         if (bteBlock_unaligned == NULL) {
 283                 return BTEFAIL_NOTAVAIL;
 284         }
 285         bteBlock = (char *)L1_CACHE_ALIGN((u64) bteBlock_unaligned);
 286
 287         headBcopySrcOffset = src & L1_CACHE_MASK;
 288         destFirstCacheOffset = dest & L1_CACHE_MASK;
 289
 290         /*
 291          * At this point, the transfer is broken into
 292          * (up to) three sections.  The first section is
 293          * from the start address to the first physical
 294          * cache line, the second is from the first physical
 295          * cache line to the last complete cache line,
 296          * and the third is from the last cache line to the
 297          * end of the buffer.  The first and third sections
 298          * are handled by bte copying into a temporary buffer
 299          * and then bcopy'ing the necessary section into the
 300          * final location.  The middle section is handled with
 301          * a standard bte copy.
 302          *
 303          * One nasty exception to the above rule is when the
 304          * source and destination are not symmetrically
 305          * mis-aligned.  If the source offset from the first
 306          * cache line is different from the destination offset,
 307          * we make the first section be the entire transfer
 308          * and the bcopy the entire block into place.
 309          */
 310         if (headBcopySrcOffset == destFirstCacheOffset) {
 311
 312                 /*
 313                  * Both the source and destination are the same
 314                  * distance from a cache line boundary so we can
 315                  * use the bte to transfer the bulk of the
 316                  * data.
 317                  */
 318                 headBteSource = src & ~L1_CACHE_MASK;
 319                 headBcopyDest = dest;
 320                 if (headBcopySrcOffset) {
 321                         headBcopyLen =
 322                             (len >
 323                              (L1_CACHE_BYTES -
 324                               headBcopySrcOffset) ? L1_CACHE_BYTES
 325                              - headBcopySrcOffset : len);
 326                         headBteLen = L1_CACHE_BYTES;
 327                 } else {
 328                         headBcopyLen = 0;
 329                         headBteLen = 0;
 330                 }
 331
 332                 if (len > headBcopyLen) {
 333                         footBcopyLen = (len - headBcopyLen) & L1_CACHE_MASK;
 334                         footBteLen = L1_CACHE_BYTES;
 335
 336                         footBteSource = src + len - footBcopyLen;
 337                         footBcopyDest = dest + len - footBcopyLen;
 338
 339                         if (footBcopyDest == (headBcopyDest + headBcopyLen)) {
 340                                 /*
 341                                  * We have two contiguous bcopy
 342                                  * blocks.  Merge them.
 343                                  */
 344                                 headBcopyLen += footBcopyLen;
 345                                 headBteLen += footBteLen;
 346                         } else if (footBcopyLen > 0) {
 347                                 rv = bte_copy(footBteSource,
 348                                               ia64_tpa((unsigned long)bteBlock),
 349                                               footBteLen, mode, NULL);
 350                                 if (rv != BTE_SUCCESS) {
 351                                         kfree(bteBlock_unaligned);
 352                                         return rv;
 353                                 }
 354
 355                                 memcpy(__va(footBcopyDest),
 356                                        (char *)bteBlock, footBcopyLen);
 357                         }
 358                 } else {
 359                         footBcopyLen = 0;
 360                         footBteLen = 0;
 361                 }
 362
 363                 if (len > (headBcopyLen + footBcopyLen)) {
 364                         /* now transfer the middle. */
 365                         rv = bte_copy((src + headBcopyLen),
 366                                       (dest +
 367                                        headBcopyLen),
 368                                       (len - headBcopyLen -
 369                                        footBcopyLen), mode, NULL);
 370                         if (rv != BTE_SUCCESS) {
 371                                 kfree(bteBlock_unaligned);
 372                                 return rv;
 373                         }
 374
 375                 }
 376         } else {
 377
 378                 /*
 379                  * The transfer is not symmetric, we will
 380                  * allocate a buffer large enough for all the
 381                  * data, bte_copy into that buffer and then
 382                  * bcopy to the destination.
 383                  */
 384
 385                 headBcopySrcOffset = src & L1_CACHE_MASK;
 386                 headBcopyDest = dest;
 387                 headBcopyLen = len;
 388
 389                 headBteSource = src - headBcopySrcOffset;
 390                 /* Add the leading and trailing bytes from source */
 391                 headBteLen = L1_CACHE_ALIGN(len + headBcopySrcOffset);
 392         }
 393
 394         if (headBcopyLen > 0) {
 395                 rv = bte_copy(headBteSource,
 396                               ia64_tpa((unsigned long)bteBlock), headBteLen,
 397                               mode, NULL);
 398                 if (rv != BTE_SUCCESS) {
 399                         kfree(bteBlock_unaligned);
 400                         return rv;
 401                 }
 402
 403                 memcpy(__va(headBcopyDest), ((char *)bteBlock +
 404                                              headBcopySrcOffset), headBcopyLen);
 405         }
 406         kfree(bteBlock_unaligned);
 407         return BTE_SUCCESS;
 408 }
 409
 410 EXPORT_SYMBOL(bte_unaligned_copy);
 411
 412 /************************************************************************
 413  * Block Transfer Engine initialization functions.
 414  *
 415  ***********************************************************************/
 416
 417 /*
 418  * bte_init_node(nodepda, cnode)
 419  *
 420  * Initialize the nodepda structure with BTE base addresses and
 421  * spinlocks.
 422  */
 423 void bte_init_node(nodepda_t * mynodepda, cnodeid_t cnode)
 424 {
 425         int i;
 426
 427         /*
 428          * Indicate that all the block transfer engines on this node
 429          * are available.
 430          */
 431
 432         /*
 433          * Allocate one bte_recover_t structure per node.  It holds
 434          * the recovery lock for node.  All the bte interface structures
 435          * will point at this one bte_recover structure to get the lock.
 436          */
 437         spin_lock_init(&mynodepda->bte_recovery_lock);
 438         init_timer(&mynodepda->bte_recovery_timer);
 439         mynodepda->bte_recovery_timer.function = bte_error_handler;
 440         mynodepda->bte_recovery_timer.data = (unsigned long)mynodepda;
 441
 442         for (i = 0; i < BTES_PER_NODE; i++) {
 443                 u64 *base_addr;
 444
 445                 /* Which link status register should we use? */
 446                 base_addr = (u64 *)
 447                     REMOTE_HUB_ADDR(cnodeid_to_nasid(cnode), BTE_BASE_ADDR(i));
 448                 mynodepda->bte_if[i].bte_base_addr = base_addr;
 449                 mynodepda->bte_if[i].bte_source_addr = BTE_SOURCE_ADDR(base_addr);
 450                 mynodepda->bte_if[i].bte_destination_addr = BTE_DEST_ADDR(base_addr);
 451                 mynodepda->bte_if[i].bte_control_addr = BTE_CTRL_ADDR(base_addr);
 452                 mynodepda->bte_if[i].bte_notify_addr = BTE_NOTIF_ADDR(base_addr);
 453
 454                 /*
 455                  * Initialize the notification and spinlock
 456                  * so the first transfer can occur.
 457                  */
 458                 mynodepda->bte_if[i].most_rcnt_na =
 459                     &(mynodepda->bte_if[i].notify);
 460                 mynodepda->bte_if[i].notify = BTE_WORD_AVAILABLE;
 461                 spin_lock_init(&mynodepda->bte_if[i].spinlock);
 462
 463                 mynodepda->bte_if[i].bte_cnode = cnode;
 464                 mynodepda->bte_if[i].bte_error_count = 0;
 465                 mynodepda->bte_if[i].bte_num = i;
 466                 mynodepda->bte_if[i].cleanup_active = 0;
 467                 mynodepda->bte_if[i].bh_error = 0;
 468         }
 469
 470 }