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) 2000-2005 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   8
   9 #include <linux/config.h>
  10 #include <linux/module.h>
  11 #include <asm/sn/nodepda.h>
  12 #include <asm/sn/addrs.h>
  13 #include <asm/sn/arch.h>
  14 #include <asm/sn/sn_cpuid.h>
  15 #include <asm/sn/pda.h>
  16 #include <asm/sn/shubio.h>
  17 #include <asm/nodedata.h>
  18 #include <asm/delay.h>
  19
  20 #include <linux/bootmem.h>
  21 #include <linux/string.h>
  22 #include <linux/sched.h>
  23
  24 #include <asm/sn/bte.h>
  25
  26 #ifndef L1_CACHE_MASK
  27 #define L1_CACHE_MASK (L1_CACHE_BYTES - 1)
  28 #endif
  29
  30 /* two interfaces on two btes */
  31 #define MAX_INTERFACES_TO_TRY           4
  32 #define MAX_NODES_TO_TRY                2
  33
  34 static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
  35 {
  36         nodepda_t *tmp_nodepda;
  37
  38         if (nasid_to_cnodeid(nasid) == -1)
  39                 return (struct bteinfo_s *)NULL;;
  40
  41         tmp_nodepda = NODEPDA(nasid_to_cnodeid(nasid));
  42         return &tmp_nodepda->bte_if[interface];
  43
  44 }
  45
  46 static inline void bte_start_transfer(struct bteinfo_s *bte, u64 len, u64 mode)
  47 {
  48         if (is_shub2()) {
  49                 BTE_CTRL_STORE(bte, (IBLS_BUSY | ((len) | (mode) << 24)));
  50         } else {
  51                 BTE_LNSTAT_STORE(bte, len);
  52                 BTE_CTRL_STORE(bte, mode);
  53         }
  54 }
  55
  56 /************************************************************************
  57  * Block Transfer Engine copy related functions.
  58  *
  59  ***********************************************************************/
  60
  61 /*
  62  * bte_copy(src, dest, len, mode, notification)
  63  *
  64  * Use the block transfer engine to move kernel memory from src to dest
  65  * using the assigned mode.
  66  *
  67  * Paramaters:
  68  *   src - physical address of the transfer source.
  69  *   dest - physical address of the transfer destination.
  70  *   len - number of bytes to transfer from source to dest.
  71  *   mode - hardware defined.  See reference information
  72  *          for IBCT0/1 in the SHUB Programmers Reference
  73  *   notification - kernel virtual address of the notification cache
  74  *                  line.  If NULL, the default is used and
  75  *                  the bte_copy is synchronous.
  76  *
  77  * NOTE:  This function requires src, dest, and len to
  78  * be cacheline aligned.
  79  */
  80 bte_result_t bte_copy(u64 src, u64 dest, u64 len, u64 mode, void *notification)
  81 {
  82         u64 transfer_size;
  83         u64 transfer_stat;
  84         u64 notif_phys_addr;
  85         struct bteinfo_s *bte;
  86         bte_result_t bte_status;
  87         unsigned long irq_flags;
  88         unsigned long itc_end = 0;
  89         int nasid_to_try[MAX_NODES_TO_TRY];
  90         int my_nasid = get_nasid();
  91         int bte_if_index, nasid_index;
  92         int bte_first, btes_per_node = BTES_PER_NODE;
  93
  94         BTE_PRINTK(("bte_copy(0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%p)\n",
  95                     src, dest, len, mode, notification));
  96
  97         if (len == 0) {
  98                 return BTE_SUCCESS;
  99         }
 100
 101         BUG_ON((len & L1_CACHE_MASK) ||
 102                  (src & L1_CACHE_MASK) || (dest & L1_CACHE_MASK));
 103         BUG_ON(!(len < ((BTE_LEN_MASK + 1) << L1_CACHE_SHIFT)));
 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                                 continue;
 141                         }
 142
 143                         if (spin_trylock(&bte->spinlock)) {
 144                                 if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
 145                                     (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
 146                                         /* Got the lock but BTE still busy */
 147                                         spin_unlock(&bte->spinlock);
 148                                 } else {
 149                                         /* we got the lock and it's not busy */
 150                                         break;
 151                                 }
 152                         }
 153
 154                         bte_if_index = (bte_if_index + 1) % btes_per_node; /* Next interface */
 155                         if (bte_if_index == bte_first) {
 156                                 /*
 157                                  * We've tried all interfaces on this node
 158                                  */
 159                                 nasid_index++;
 160                         }
 161
 162                         bte = NULL;
 163                 }
 164
 165                 if (bte != NULL) {
 166                         break;
 167                 }
 168
 169                 local_irq_restore(irq_flags);
 170
 171                 if (!(mode & BTE_WACQUIRE)) {
 172                         return BTEFAIL_NOTAVAIL;
 173                 }
 174         } while (1);
 175
 176         if (notification == NULL) {
 177                 /* User does not want to be notified. */
 178                 bte->most_rcnt_na = &bte->notify;
 179         } else {
 180                 bte->most_rcnt_na = notification;
 181         }
 182
 183         /* Calculate the number of cache lines to transfer. */
 184         transfer_size = ((len >> L1_CACHE_SHIFT) & BTE_LEN_MASK);
 185
 186         /* Initialize the notification to a known value. */
 187         *bte->most_rcnt_na = BTE_WORD_BUSY;
 188         notif_phys_addr = TO_PHYS(ia64_tpa((unsigned long)bte->most_rcnt_na));
 189
 190         if (is_shub2()) {
 191                 src = SH2_TIO_PHYS_TO_DMA(src);
 192                 dest = SH2_TIO_PHYS_TO_DMA(dest);
 193                 notif_phys_addr = SH2_TIO_PHYS_TO_DMA(notif_phys_addr);
 194         }
 195         /* Set the source and destination registers */
 196         BTE_PRINTKV(("IBSA = 0x%lx)\n", (TO_PHYS(src))));
 197         BTE_SRC_STORE(bte, TO_PHYS(src));
 198         BTE_PRINTKV(("IBDA = 0x%lx)\n", (TO_PHYS(dest))));
 199         BTE_DEST_STORE(bte, TO_PHYS(dest));
 200
 201         /* Set the notification register */
 202         BTE_PRINTKV(("IBNA = 0x%lx)\n", notif_phys_addr));
 203         BTE_NOTIF_STORE(bte, notif_phys_addr);
 204
 205         /* Initiate the transfer */
 206         BTE_PRINTK(("IBCT = 0x%lx)\n", BTE_VALID_MODE(mode)));
 207         bte_start_transfer(bte, transfer_size, BTE_VALID_MODE(mode));
 208
 209         itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
 210
 211         spin_unlock_irqrestore(&bte->spinlock, irq_flags);
 212
 213         if (notification != NULL) {
 214                 return BTE_SUCCESS;
 215         }
 216
 217         while ((transfer_stat = *bte->most_rcnt_na) == BTE_WORD_BUSY) {
 218                 cpu_relax();
 219                 if (ia64_get_itc() > itc_end) {
 220                         BTE_PRINTK(("BTE timeout nasid 0x%x bte%d IBLS = 0x%lx na 0x%lx\n",
 221                                 NASID_GET(bte->bte_base_addr), bte->bte_num,
 222                                 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na) );
 223                         bte->bte_error_count++;
 224                         bte->bh_error = IBLS_ERROR;
 225                         bte_error_handler((unsigned long)NODEPDA(bte->bte_cnode));
 226                         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 227                         goto retry_bteop;
 228                 }
 229         }
 230
 231         BTE_PRINTKV((" Delay Done.  IBLS = 0x%lx, most_rcnt_na = 0x%lx\n",
 232                      BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 233
 234         if (transfer_stat & IBLS_ERROR) {
 235                 bte_status = transfer_stat & ~IBLS_ERROR;
 236         } else {
 237                 bte_status = BTE_SUCCESS;
 238         }
 239         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 240
 241         BTE_PRINTK(("Returning status is 0x%lx and most_rcnt_na is 0x%lx\n",
 242                     BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 243
 244         return bte_status;
 245 }
 246
 247 EXPORT_SYMBOL(bte_copy);
 248
 249 /*
 250  * bte_unaligned_copy(src, dest, len, mode)
 251  *
 252  * use the block transfer engine to move kernel
 253  * memory from src to dest using the assigned mode.
 254  *
 255  * Paramaters:
 256  *   src - physical address of the transfer source.
 257  *   dest - physical address of the transfer destination.
 258  *   len - number of bytes to transfer from source to dest.
 259  *   mode - hardware defined.  See reference information
 260  *          for IBCT0/1 in the SGI documentation.
 261  *
 262  * NOTE: If the source, dest, and len are all cache line aligned,
 263  * then it would be _FAR_ preferrable to use bte_copy instead.
 264  */
 265 bte_result_t bte_unaligned_copy(u64 src, u64 dest, u64 len, u64 mode)
 266 {
 267         int destFirstCacheOffset;
 268         u64 headBteSource;
 269         u64 headBteLen;
 270         u64 headBcopySrcOffset;
 271         u64 headBcopyDest;
 272         u64 headBcopyLen;
 273         u64 footBteSource;
 274         u64 footBteLen;
 275         u64 footBcopyDest;
 276         u64 footBcopyLen;
 277         bte_result_t rv;
 278         char *bteBlock, *bteBlock_unaligned;
 279
 280         if (len == 0) {
 281                 return BTE_SUCCESS;
 282         }
 283
 284         /* temporary buffer used during unaligned transfers */
 285         bteBlock_unaligned = kmalloc(len + 3 * L1_CACHE_BYTES,
 286                                      GFP_KERNEL | GFP_DMA);
 287         if (bteBlock_unaligned == NULL) {
 288                 return BTEFAIL_NOTAVAIL;
 289         }
 290         bteBlock = (char *)L1_CACHE_ALIGN((u64) bteBlock_unaligned);
 291
 292         headBcopySrcOffset = src & L1_CACHE_MASK;
 293         destFirstCacheOffset = dest & L1_CACHE_MASK;
 294
 295         /*
 296          * At this point, the transfer is broken into
 297          * (up to) three sections.  The first section is
 298          * from the start address to the first physical
 299          * cache line, the second is from the first physical
 300          * cache line to the last complete cache line,
 301          * and the third is from the last cache line to the
 302          * end of the buffer.  The first and third sections
 303          * are handled by bte copying into a temporary buffer
 304          * and then bcopy'ing the necessary section into the
 305          * final location.  The middle section is handled with
 306          * a standard bte copy.
 307          *
 308          * One nasty exception to the above rule is when the
 309          * source and destination are not symetrically
 310          * mis-aligned.  If the source offset from the first
 311          * cache line is different from the destination offset,
 312          * we make the first section be the entire transfer
 313          * and the bcopy the entire block into place.
 314          */
 315         if (headBcopySrcOffset == destFirstCacheOffset) {
 316
 317                 /*
 318                  * Both the source and destination are the same
 319                  * distance from a cache line boundary so we can
 320                  * use the bte to transfer the bulk of the
 321                  * data.
 322                  */
 323                 headBteSource = src & ~L1_CACHE_MASK;
 324                 headBcopyDest = dest;
 325                 if (headBcopySrcOffset) {
 326                         headBcopyLen =
 327                             (len >
 328                              (L1_CACHE_BYTES -
 329                               headBcopySrcOffset) ? L1_CACHE_BYTES
 330                              - headBcopySrcOffset : len);
 331                         headBteLen = L1_CACHE_BYTES;
 332                 } else {
 333                         headBcopyLen = 0;
 334                         headBteLen = 0;
 335                 }
 336
 337                 if (len > headBcopyLen) {
 338                         footBcopyLen = (len - headBcopyLen) & L1_CACHE_MASK;
 339                         footBteLen = L1_CACHE_BYTES;
 340
 341                         footBteSource = src + len - footBcopyLen;
 342                         footBcopyDest = dest + len - footBcopyLen;
 343
 344                         if (footBcopyDest == (headBcopyDest + headBcopyLen)) {
 345                                 /*
 346                                  * We have two contigous bcopy
 347                                  * blocks.  Merge them.
 348                                  */
 349                                 headBcopyLen += footBcopyLen;
 350                                 headBteLen += footBteLen;
 351                         } else if (footBcopyLen > 0) {
 352                                 rv = bte_copy(footBteSource,
 353                                               ia64_tpa((unsigned long)bteBlock),
 354                                               footBteLen, mode, NULL);
 355                                 if (rv != BTE_SUCCESS) {
 356                                         kfree(bteBlock_unaligned);
 357                                         return rv;
 358                                 }
 359
 360                                 memcpy(__va(footBcopyDest),
 361                                        (char *)bteBlock, footBcopyLen);
 362                         }
 363                 } else {
 364                         footBcopyLen = 0;
 365                         footBteLen = 0;
 366                 }
 367
 368                 if (len > (headBcopyLen + footBcopyLen)) {
 369                         /* now transfer the middle. */
 370                         rv = bte_copy((src + headBcopyLen),
 371                                       (dest +
 372                                        headBcopyLen),
 373                                       (len - headBcopyLen -
 374                                        footBcopyLen), mode, NULL);
 375                         if (rv != BTE_SUCCESS) {
 376                                 kfree(bteBlock_unaligned);
 377                                 return rv;
 378                         }
 379
 380                 }
 381         } else {
 382
 383                 /*
 384                  * The transfer is not symetric, we will
 385                  * allocate a buffer large enough for all the
 386                  * data, bte_copy into that buffer and then
 387                  * bcopy to the destination.
 388                  */
 389
 390                 /* Add the leader from source */
 391                 headBteLen = len + (src & L1_CACHE_MASK);
 392                 /* Add the trailing bytes from footer. */
 393                 headBteLen += L1_CACHE_BYTES - (headBteLen & L1_CACHE_MASK);
 394                 headBteSource = src & ~L1_CACHE_MASK;
 395                 headBcopySrcOffset = src & L1_CACHE_MASK;
 396                 headBcopyDest = dest;
 397                 headBcopyLen = len;
 398         }
 399
 400         if (headBcopyLen > 0) {
 401                 rv = bte_copy(headBteSource,
 402                               ia64_tpa((unsigned long)bteBlock), headBteLen,
 403                               mode, NULL);
 404                 if (rv != BTE_SUCCESS) {
 405                         kfree(bteBlock_unaligned);
 406                         return rv;
 407                 }
 408
 409                 memcpy(__va(headBcopyDest), ((char *)bteBlock +
 410                                              headBcopySrcOffset), headBcopyLen);
 411         }
 412         kfree(bteBlock_unaligned);
 413         return BTE_SUCCESS;
 414 }
 415
 416 EXPORT_SYMBOL(bte_unaligned_copy);
 417
 418 /************************************************************************
 419  * Block Transfer Engine initialization functions.
 420  *
 421  ***********************************************************************/
 422
 423 /*
 424  * bte_init_node(nodepda, cnode)
 425  *
 426  * Initialize the nodepda structure with BTE base addresses and
 427  * spinlocks.
 428  */
 429 void bte_init_node(nodepda_t * mynodepda, cnodeid_t cnode)
 430 {
 431         int i;
 432
 433         /*
 434          * Indicate that all the block transfer engines on this node
 435          * are available.
 436          */
 437
 438         /*
 439          * Allocate one bte_recover_t structure per node.  It holds
 440          * the recovery lock for node.  All the bte interface structures
 441          * will point at this one bte_recover structure to get the lock.
 442          */
 443         spin_lock_init(&mynodepda->bte_recovery_lock);
 444         init_timer(&mynodepda->bte_recovery_timer);
 445         mynodepda->bte_recovery_timer.function = bte_error_handler;
 446         mynodepda->bte_recovery_timer.data = (unsigned long)mynodepda;
 447
 448         for (i = 0; i < BTES_PER_NODE; i++) {
 449                 u64 *base_addr;
 450
 451                 /* Which link status register should we use? */
 452                 base_addr = (u64 *)
 453                     REMOTE_HUB_ADDR(cnodeid_to_nasid(cnode), BTE_BASE_ADDR(i));
 454                 mynodepda->bte_if[i].bte_base_addr = base_addr;
 455                 mynodepda->bte_if[i].bte_source_addr = BTE_SOURCE_ADDR(base_addr);
 456                 mynodepda->bte_if[i].bte_destination_addr = BTE_DEST_ADDR(base_addr);
 457                 mynodepda->bte_if[i].bte_control_addr = BTE_CTRL_ADDR(base_addr);
 458                 mynodepda->bte_if[i].bte_notify_addr = BTE_NOTIF_ADDR(base_addr);
 459
 460                 /*
 461                  * Initialize the notification and spinlock
 462                  * so the first transfer can occur.
 463                  */
 464                 mynodepda->bte_if[i].most_rcnt_na =
 465                     &(mynodepda->bte_if[i].notify);
 466                 mynodepda->bte_if[i].notify = BTE_WORD_AVAILABLE;
 467                 spin_lock_init(&mynodepda->bte_if[i].spinlock);
 468
 469                 mynodepda->bte_if[i].bte_cnode = cnode;
 470                 mynodepda->bte_if[i].bte_error_count = 0;
 471                 mynodepda->bte_if[i].bte_num = i;
 472                 mynodepda->bte_if[i].cleanup_active = 0;
 473                 mynodepda->bte_if[i].bh_error = 0;
 474         }
 475
 476 }