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) 2000-2003 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
  33 static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
  34 {
  35         nodepda_t *tmp_nodepda;
  36
  37         tmp_nodepda = NODEPDA(nasid_to_cnodeid(nasid));
  38         return &tmp_nodepda->bte_if[interface];
  39
  40 }
  41
  42 /************************************************************************
  43  * Block Transfer Engine copy related functions.
  44  *
  45  ***********************************************************************/
  46
  47 /*
  48  * bte_copy(src, dest, len, mode, notification)
  49  *
  50  * Use the block transfer engine to move kernel memory from src to dest
  51  * using the assigned mode.
  52  *
  53  * Paramaters:
  54  *   src - physical address of the transfer source.
  55  *   dest - physical address of the transfer destination.
  56  *   len - number of bytes to transfer from source to dest.
  57  *   mode - hardware defined.  See reference information
  58  *          for IBCT0/1 in the SHUB Programmers Reference
  59  *   notification - kernel virtual address of the notification cache
  60  *                  line.  If NULL, the default is used and
  61  *                  the bte_copy is synchronous.
  62  *
  63  * NOTE:  This function requires src, dest, and len to
  64  * be cacheline aligned.
  65  */
  66 bte_result_t bte_copy(u64 src, u64 dest, u64 len, u64 mode, void *notification)
  67 {
  68         u64 transfer_size;
  69         u64 transfer_stat;
  70         struct bteinfo_s *bte;
  71         bte_result_t bte_status;
  72         unsigned long irq_flags;
  73         unsigned long itc_end = 0;
  74         struct bteinfo_s *btes_to_try[MAX_INTERFACES_TO_TRY];
  75         int bte_if_index;
  76         int bte_pri, bte_sec;
  77
  78         BTE_PRINTK(("bte_copy(0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%p)\n",
  79                     src, dest, len, mode, notification));
  80
  81         if (len == 0) {
  82                 return BTE_SUCCESS;
  83         }
  84
  85         BUG_ON((len & L1_CACHE_MASK) ||
  86                  (src & L1_CACHE_MASK) || (dest & L1_CACHE_MASK));
  87         BUG_ON(!(len < ((BTE_LEN_MASK + 1) << L1_CACHE_SHIFT)));
  88
  89         /* CPU 0 (per node) tries bte0 first, CPU 1 try bte1 first */
  90         if (cpuid_to_subnode(smp_processor_id()) == 0) {
  91                 bte_pri = 0;
  92                 bte_sec = 1;
  93         } else {
  94                 bte_pri = 1;
  95                 bte_sec = 0;
  96         }
  97
  98         if (mode & BTE_USE_DEST) {
  99                 /* try remote then local */
 100                 btes_to_try[0] = bte_if_on_node(NASID_GET(dest), bte_pri);
 101                 btes_to_try[1] = bte_if_on_node(NASID_GET(dest), bte_sec);
 102                 if (mode & BTE_USE_ANY) {
 103                         btes_to_try[2] = bte_if_on_node(get_nasid(), bte_pri);
 104                         btes_to_try[3] = bte_if_on_node(get_nasid(), bte_sec);
 105                 } else {
 106                         btes_to_try[2] = NULL;
 107                         btes_to_try[3] = NULL;
 108                 }
 109         } else {
 110                 /* try local then remote */
 111                 btes_to_try[0] = bte_if_on_node(get_nasid(), bte_pri);
 112                 btes_to_try[1] = bte_if_on_node(get_nasid(), bte_sec);
 113                 if (mode & BTE_USE_ANY) {
 114                         btes_to_try[2] = bte_if_on_node(NASID_GET(dest), bte_pri);
 115                         btes_to_try[3] = bte_if_on_node(NASID_GET(dest), bte_sec);
 116                 } else {
 117                         btes_to_try[2] = NULL;
 118                         btes_to_try[3] = NULL;
 119                 }
 120         }
 121
 122 retry_bteop:
 123         do {
 124                 local_irq_save(irq_flags);
 125
 126                 bte_if_index = 0;
 127
 128                 /* Attempt to lock one of the BTE interfaces. */
 129                 while (bte_if_index < MAX_INTERFACES_TO_TRY) {
 130                         bte = btes_to_try[bte_if_index++];
 131
 132                         if (bte == NULL) {
 133                                 continue;
 134                         }
 135
 136                         if (spin_trylock(&bte->spinlock)) {
 137                                 if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
 138                                     (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
 139                                         /* Got the lock but BTE still busy */
 140                                         spin_unlock(&bte->spinlock);
 141                                 } else {
 142                                         /* we got the lock and it's not busy */
 143                                         break;
 144                                 }
 145                         }
 146                         bte = NULL;
 147                 }
 148
 149                 if (bte != NULL) {
 150                         break;
 151                 }
 152
 153                 local_irq_restore(irq_flags);
 154
 155                 if (!(mode & BTE_WACQUIRE)) {
 156                         return BTEFAIL_NOTAVAIL;
 157                 }
 158         } while (1);
 159
 160         if (notification == NULL) {
 161                 /* User does not want to be notified. */
 162                 bte->most_rcnt_na = &bte->notify;
 163         } else {
 164                 bte->most_rcnt_na = notification;
 165         }
 166
 167         /* Calculate the number of cache lines to transfer. */
 168         transfer_size = ((len >> L1_CACHE_SHIFT) & BTE_LEN_MASK);
 169
 170         /* Initialize the notification to a known value. */
 171         *bte->most_rcnt_na = BTE_WORD_BUSY;
 172
 173         /* Set the status reg busy bit and transfer length */
 174         BTE_PRINTKV(("IBLS = 0x%lx\n", IBLS_BUSY | transfer_size));
 175         BTE_LNSTAT_STORE(bte, IBLS_BUSY | transfer_size);
 176
 177         /* Set the source and destination registers */
 178         BTE_PRINTKV(("IBSA = 0x%lx)\n", (TO_PHYS(src))));
 179         BTE_SRC_STORE(bte, TO_PHYS(src));
 180         BTE_PRINTKV(("IBDA = 0x%lx)\n", (TO_PHYS(dest))));
 181         BTE_DEST_STORE(bte, TO_PHYS(dest));
 182
 183         /* Set the notification register */
 184         BTE_PRINTKV(("IBNA = 0x%lx)\n",
 185                      TO_PHYS(ia64_tpa((unsigned long)bte->most_rcnt_na))));
 186         BTE_NOTIF_STORE(bte,
 187                         TO_PHYS(ia64_tpa((unsigned long)bte->most_rcnt_na)));
 188
 189         /* Initiate the transfer */
 190         BTE_PRINTK(("IBCT = 0x%lx)\n", BTE_VALID_MODE(mode)));
 191         BTE_CTRL_STORE(bte, BTE_VALID_MODE(mode));
 192
 193         itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
 194
 195         spin_unlock_irqrestore(&bte->spinlock, irq_flags);
 196
 197         if (notification != NULL) {
 198                 return BTE_SUCCESS;
 199         }
 200
 201         while ((transfer_stat = *bte->most_rcnt_na) == BTE_WORD_BUSY) {
 202                 if (ia64_get_itc() > itc_end) {
 203                         BTE_PRINTK(("BTE timeout nasid 0x%x bte%d IBLS = 0x%lx na 0x%lx\n",
 204                                 NASID_GET(bte->bte_base_addr), bte->bte_num,
 205                                 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na) );
 206                         bte->bte_error_count++;
 207                         bte->bh_error = IBLS_ERROR;
 208                         bte_error_handler((unsigned long)NODEPDA(bte->bte_cnode));
 209                         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 210                         goto retry_bteop;
 211                 }
 212         }
 213
 214         BTE_PRINTKV((" Delay Done.  IBLS = 0x%lx, most_rcnt_na = 0x%lx\n",
 215                      BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 216
 217         if (transfer_stat & IBLS_ERROR) {
 218                 bte_status = transfer_stat & ~IBLS_ERROR;
 219         } else {
 220                 bte_status = BTE_SUCCESS;
 221         }
 222         *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
 223
 224         BTE_PRINTK(("Returning status is 0x%lx and most_rcnt_na is 0x%lx\n",
 225                     BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
 226
 227         return bte_status;
 228 }
 229
 230 EXPORT_SYMBOL(bte_copy);
 231
 232 /*
 233  * bte_unaligned_copy(src, dest, len, mode)
 234  *
 235  * use the block transfer engine to move kernel
 236  * memory from src to dest using the assigned mode.
 237  *
 238  * Paramaters:
 239  *   src - physical address of the transfer source.
 240  *   dest - physical address of the transfer destination.
 241  *   len - number of bytes to transfer from source to dest.
 242  *   mode - hardware defined.  See reference information
 243  *          for IBCT0/1 in the SGI documentation.
 244  *
 245  * NOTE: If the source, dest, and len are all cache line aligned,
 246  * then it would be _FAR_ preferrable to use bte_copy instead.
 247  */
 248 bte_result_t bte_unaligned_copy(u64 src, u64 dest, u64 len, u64 mode)
 249 {
 250         int destFirstCacheOffset;
 251         u64 headBteSource;
 252         u64 headBteLen;
 253         u64 headBcopySrcOffset;
 254         u64 headBcopyDest;
 255         u64 headBcopyLen;
 256         u64 footBteSource;
 257         u64 footBteLen;
 258         u64 footBcopyDest;
 259         u64 footBcopyLen;
 260         bte_result_t rv;
 261         char *bteBlock, *bteBlock_unaligned;
 262
 263         if (len == 0) {
 264                 return BTE_SUCCESS;
 265         }
 266
 267         /* temporary buffer used during unaligned transfers */
 268         bteBlock_unaligned = kmalloc(len + 3 * L1_CACHE_BYTES,
 269                                      GFP_KERNEL | GFP_DMA);
 270         if (bteBlock_unaligned == NULL) {
 271                 return BTEFAIL_NOTAVAIL;
 272         }
 273         bteBlock = (char *)L1_CACHE_ALIGN((u64) bteBlock_unaligned);
 274
 275         headBcopySrcOffset = src & L1_CACHE_MASK;
 276         destFirstCacheOffset = dest & L1_CACHE_MASK;
 277
 278         /*
 279          * At this point, the transfer is broken into
 280          * (up to) three sections.  The first section is
 281          * from the start address to the first physical
 282          * cache line, the second is from the first physical
 283          * cache line to the last complete cache line,
 284          * and the third is from the last cache line to the
 285          * end of the buffer.  The first and third sections
 286          * are handled by bte copying into a temporary buffer
 287          * and then bcopy'ing the necessary section into the
 288          * final location.  The middle section is handled with
 289          * a standard bte copy.
 290          *
 291          * One nasty exception to the above rule is when the
 292          * source and destination are not symetrically
 293          * mis-aligned.  If the source offset from the first
 294          * cache line is different from the destination offset,
 295          * we make the first section be the entire transfer
 296          * and the bcopy the entire block into place.
 297          */
 298         if (headBcopySrcOffset == destFirstCacheOffset) {
 299
 300                 /*
 301                  * Both the source and destination are the same
 302                  * distance from a cache line boundary so we can
 303                  * use the bte to transfer the bulk of the
 304                  * data.
 305                  */
 306                 headBteSource = src & ~L1_CACHE_MASK;
 307                 headBcopyDest = dest;
 308                 if (headBcopySrcOffset) {
 309                         headBcopyLen =
 310                             (len >
 311                              (L1_CACHE_BYTES -
 312                               headBcopySrcOffset) ? L1_CACHE_BYTES
 313                              - headBcopySrcOffset : len);
 314                         headBteLen = L1_CACHE_BYTES;
 315                 } else {
 316                         headBcopyLen = 0;
 317                         headBteLen = 0;
 318                 }
 319
 320                 if (len > headBcopyLen) {
 321                         footBcopyLen = (len - headBcopyLen) & L1_CACHE_MASK;
 322                         footBteLen = L1_CACHE_BYTES;
 323
 324                         footBteSource = src + len - footBcopyLen;
 325                         footBcopyDest = dest + len - footBcopyLen;
 326
 327                         if (footBcopyDest == (headBcopyDest + headBcopyLen)) {
 328                                 /*
 329                                  * We have two contigous bcopy
 330                                  * blocks.  Merge them.
 331                                  */
 332                                 headBcopyLen += footBcopyLen;
 333                                 headBteLen += footBteLen;
 334                         } else if (footBcopyLen > 0) {
 335                                 rv = bte_copy(footBteSource,
 336                                               ia64_tpa((unsigned long)bteBlock),
 337                                               footBteLen, mode, NULL);
 338                                 if (rv != BTE_SUCCESS) {
 339                                         kfree(bteBlock_unaligned);
 340                                         return rv;
 341                                 }
 342
 343                                 memcpy(__va(footBcopyDest),
 344                                        (char *)bteBlock, footBcopyLen);
 345                         }
 346                 } else {
 347                         footBcopyLen = 0;
 348                         footBteLen = 0;
 349                 }
 350
 351                 if (len > (headBcopyLen + footBcopyLen)) {
 352                         /* now transfer the middle. */
 353                         rv = bte_copy((src + headBcopyLen),
 354                                       (dest +
 355                                        headBcopyLen),
 356                                       (len - headBcopyLen -
 357                                        footBcopyLen), mode, NULL);
 358                         if (rv != BTE_SUCCESS) {
 359                                 kfree(bteBlock_unaligned);
 360                                 return rv;
 361                         }
 362
 363                 }
 364         } else {
 365
 366                 /*
 367                  * The transfer is not symetric, we will
 368                  * allocate a buffer large enough for all the
 369                  * data, bte_copy into that buffer and then
 370                  * bcopy to the destination.
 371                  */
 372
 373                 /* Add the leader from source */
 374                 headBteLen = len + (src & L1_CACHE_MASK);
 375                 /* Add the trailing bytes from footer. */
 376                 headBteLen += L1_CACHE_BYTES - (headBteLen & L1_CACHE_MASK);
 377                 headBteSource = src & ~L1_CACHE_MASK;
 378                 headBcopySrcOffset = src & L1_CACHE_MASK;
 379                 headBcopyDest = dest;
 380                 headBcopyLen = len;
 381         }
 382
 383         if (headBcopyLen > 0) {
 384                 rv = bte_copy(headBteSource,
 385                               ia64_tpa((unsigned long)bteBlock), headBteLen,
 386                               mode, NULL);
 387                 if (rv != BTE_SUCCESS) {
 388                         kfree(bteBlock_unaligned);
 389                         return rv;
 390                 }
 391
 392                 memcpy(__va(headBcopyDest), ((char *)bteBlock +
 393                                              headBcopySrcOffset), headBcopyLen);
 394         }
 395         kfree(bteBlock_unaligned);
 396         return BTE_SUCCESS;
 397 }
 398
 399 EXPORT_SYMBOL(bte_unaligned_copy);
 400
 401 /************************************************************************
 402  * Block Transfer Engine initialization functions.
 403  *
 404  ***********************************************************************/
 405
 406 /*
 407  * bte_init_node(nodepda, cnode)
 408  *
 409  * Initialize the nodepda structure with BTE base addresses and
 410  * spinlocks.
 411  */
 412 void bte_init_node(nodepda_t * mynodepda, cnodeid_t cnode)
 413 {
 414         int i;
 415
 416         /*
 417          * Indicate that all the block transfer engines on this node
 418          * are available.
 419          */
 420
 421         /*
 422          * Allocate one bte_recover_t structure per node.  It holds
 423          * the recovery lock for node.  All the bte interface structures
 424          * will point at this one bte_recover structure to get the lock.
 425          */
 426         spin_lock_init(&mynodepda->bte_recovery_lock);
 427         init_timer(&mynodepda->bte_recovery_timer);
 428         mynodepda->bte_recovery_timer.function = bte_error_handler;
 429         mynodepda->bte_recovery_timer.data = (unsigned long)mynodepda;
 430
 431         for (i = 0; i < BTES_PER_NODE; i++) {
 432                 /* Which link status register should we use? */
 433                 unsigned long link_status = (i == 0 ? IIO_IBLS0 : IIO_IBLS1);
 434                 mynodepda->bte_if[i].bte_base_addr = (u64 *)
 435                     REMOTE_HUB_ADDR(cnodeid_to_nasid(cnode), link_status);
 436
 437                 /*
 438                  * Initialize the notification and spinlock
 439                  * so the first transfer can occur.
 440                  */
 441                 mynodepda->bte_if[i].most_rcnt_na =
 442                     &(mynodepda->bte_if[i].notify);
 443                 mynodepda->bte_if[i].notify = BTE_WORD_AVAILABLE;
 444                 spin_lock_init(&mynodepda->bte_if[i].spinlock);
 445
 446                 mynodepda->bte_if[i].bte_cnode = cnode;
 447                 mynodepda->bte_if[i].bte_error_count = 0;
 448                 mynodepda->bte_if[i].bte_num = i;
 449                 mynodepda->bte_if[i].cleanup_active = 0;
 450                 mynodepda->bte_if[i].bh_error = 0;
 451         }
 452
 453 }