drivers/infiniband/hw/ipath/ipath_driver.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/spinlock.h>
  35 #include <linux/idr.h>
  36 #include <linux/pci.h>
  37 #include <linux/delay.h>
  38 #include <linux/netdevice.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "ipath_kernel.h"
  42 #include "ipath_verbs.h"
  43 #include "ipath_common.h"
  44
  45 static void ipath_update_pio_bufs(struct ipath_devdata *);
  46
  47 const char *ipath_get_unit_name(int unit)
  48 {
  49         static char iname[16];
  50         snprintf(iname, sizeof iname, "infinipath%u", unit);
  51         return iname;
  52 }
  53
  54 #define DRIVER_LOAD_MSG "QLogic " IPATH_DRV_NAME " loaded: "
  55 #define PFX IPATH_DRV_NAME ": "
  56
  57 /*
  58  * The size has to be longer than this string, so we can append
  59  * board/chip information to it in the init code.
  60  */
  61 const char ib_ipath_version[] = IPATH_IDSTR "\n";
  62
  63 static struct idr unit_table;
  64 DEFINE_SPINLOCK(ipath_devs_lock);
  65 LIST_HEAD(ipath_dev_list);
  66
  67 wait_queue_head_t ipath_state_wait;
  68
  69 unsigned ipath_debug = __IPATH_INFO;
  70
  71 module_param_named(debug, ipath_debug, uint, S_IWUSR | S_IRUGO);
  72 MODULE_PARM_DESC(debug, "mask for debug prints");
  73 EXPORT_SYMBOL_GPL(ipath_debug);
  74
  75 MODULE_LICENSE("GPL");
  76 MODULE_AUTHOR("QLogic <support@pathscale.com>");
  77 MODULE_DESCRIPTION("QLogic InfiniPath driver");
  78
  79 const char *ipath_ibcstatus_str[] = {
  80         "Disabled",
  81         "LinkUp",
  82         "PollActive",
  83         "PollQuiet",
  84         "SleepDelay",
  85         "SleepQuiet",
  86         "LState6",              /* unused */
  87         "LState7",              /* unused */
  88         "CfgDebounce",
  89         "CfgRcvfCfg",
  90         "CfgWaitRmt",
  91         "CfgIdle",
  92         "RecovRetrain",
  93         "LState0xD",            /* unused */
  94         "RecovWaitRmt",
  95         "RecovIdle",
  96 };
  97
  98 static void __devexit ipath_remove_one(struct pci_dev *);
  99 static int __devinit ipath_init_one(struct pci_dev *,
 100                                     const struct pci_device_id *);
 101
 102 /* Only needed for registration, nothing else needs this info */
 103 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
 104 #define PCI_DEVICE_ID_INFINIPATH_HT 0xd
 105 #define PCI_DEVICE_ID_INFINIPATH_PE800 0x10
 106
 107 static const struct pci_device_id ipath_pci_tbl[] = {
 108         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_HT) },
 109         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_PE800) },
 110         { 0, }
 111 };
 112
 113 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
 114
 115 static struct pci_driver ipath_driver = {
 116         .name = IPATH_DRV_NAME,
 117         .probe = ipath_init_one,
 118         .remove = __devexit_p(ipath_remove_one),
 119         .id_table = ipath_pci_tbl,
 120 };
 121
 122
 123 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
 124                              u32 *bar0, u32 *bar1)
 125 {
 126         int ret;
 127
 128         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
 129         if (ret)
 130                 ipath_dev_err(dd, "failed to read bar0 before enable: "
 131                               "error %d\n", -ret);
 132
 133         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
 134         if (ret)
 135                 ipath_dev_err(dd, "failed to read bar1 before enable: "
 136                               "error %d\n", -ret);
 137
 138         ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
 139 }
 140
 141 static void ipath_free_devdata(struct pci_dev *pdev,
 142                                struct ipath_devdata *dd)
 143 {
 144         unsigned long flags;
 145
 146         pci_set_drvdata(pdev, NULL);
 147
 148         if (dd->ipath_unit != -1) {
 149                 spin_lock_irqsave(&ipath_devs_lock, flags);
 150                 idr_remove(&unit_table, dd->ipath_unit);
 151                 list_del(&dd->ipath_list);
 152                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
 153         }
 154         vfree(dd);
 155 }
 156
 157 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
 158 {
 159         unsigned long flags;
 160         struct ipath_devdata *dd;
 161         int ret;
 162
 163         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
 164                 dd = ERR_PTR(-ENOMEM);
 165                 goto bail;
 166         }
 167
 168         dd = vmalloc(sizeof(*dd));
 169         if (!dd) {
 170                 dd = ERR_PTR(-ENOMEM);
 171                 goto bail;
 172         }
 173         memset(dd, 0, sizeof(*dd));
 174         dd->ipath_unit = -1;
 175
 176         spin_lock_irqsave(&ipath_devs_lock, flags);
 177
 178         ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
 179         if (ret < 0) {
 180                 printk(KERN_ERR IPATH_DRV_NAME
 181                        ": Could not allocate unit ID: error %d\n", -ret);
 182                 ipath_free_devdata(pdev, dd);
 183                 dd = ERR_PTR(ret);
 184                 goto bail_unlock;
 185         }
 186
 187         dd->pcidev = pdev;
 188         pci_set_drvdata(pdev, dd);
 189
 190         list_add(&dd->ipath_list, &ipath_dev_list);
 191
 192 bail_unlock:
 193         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 194
 195 bail:
 196         return dd;
 197 }
 198
 199 static inline struct ipath_devdata *__ipath_lookup(int unit)
 200 {
 201         return idr_find(&unit_table, unit);
 202 }
 203
 204 struct ipath_devdata *ipath_lookup(int unit)
 205 {
 206         struct ipath_devdata *dd;
 207         unsigned long flags;
 208
 209         spin_lock_irqsave(&ipath_devs_lock, flags);
 210         dd = __ipath_lookup(unit);
 211         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 212
 213         return dd;
 214 }
 215
 216 int ipath_count_units(int *npresentp, int *nupp, u32 *maxportsp)
 217 {
 218         int nunits, npresent, nup;
 219         struct ipath_devdata *dd;
 220         unsigned long flags;
 221         u32 maxports;
 222
 223         nunits = npresent = nup = maxports = 0;
 224
 225         spin_lock_irqsave(&ipath_devs_lock, flags);
 226
 227         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 228                 nunits++;
 229                 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
 230                         npresent++;
 231                 if (dd->ipath_lid &&
 232                     !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
 233                                          | IPATH_LINKUNK)))
 234                         nup++;
 235                 if (dd->ipath_cfgports > maxports)
 236                         maxports = dd->ipath_cfgports;
 237         }
 238
 239         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 240
 241         if (npresentp)
 242                 *npresentp = npresent;
 243         if (nupp)
 244                 *nupp = nup;
 245         if (maxportsp)
 246                 *maxportsp = maxports;
 247
 248         return nunits;
 249 }
 250
 251 /*
 252  * These next two routines are placeholders in case we don't have per-arch
 253  * code for controlling write combining.  If explicit control of write
 254  * combining is not available, performance will probably be awful.
 255  */
 256
 257 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
 258 {
 259         return -EOPNOTSUPP;
 260 }
 261
 262 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
 263 {
 264 }
 265
 266 static int __devinit ipath_init_one(struct pci_dev *pdev,
 267                                     const struct pci_device_id *ent)
 268 {
 269         int ret, len, j;
 270         struct ipath_devdata *dd;
 271         unsigned long long addr;
 272         u32 bar0 = 0, bar1 = 0;
 273         u8 rev;
 274
 275         dd = ipath_alloc_devdata(pdev);
 276         if (IS_ERR(dd)) {
 277                 ret = PTR_ERR(dd);
 278                 printk(KERN_ERR IPATH_DRV_NAME
 279                        ": Could not allocate devdata: error %d\n", -ret);
 280                 goto bail;
 281         }
 282
 283         ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
 284
 285         read_bars(dd, pdev, &bar0, &bar1);
 286
 287         ret = pci_enable_device(pdev);
 288         if (ret) {
 289                 /* This can happen iff:
 290                  *
 291                  * We did a chip reset, and then failed to reprogram the
 292                  * BAR, or the chip reset due to an internal error.  We then
 293                  * unloaded the driver and reloaded it.
 294                  *
 295                  * Both reset cases set the BAR back to initial state.  For
 296                  * the latter case, the AER sticky error bit at offset 0x718
 297                  * should be set, but the Linux kernel doesn't yet know
 298                  * about that, it appears.  If the original BAR was retained
 299                  * in the kernel data structures, this may be OK.
 300                  */
 301                 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
 302                               dd->ipath_unit, -ret);
 303                 goto bail_devdata;
 304         }
 305         addr = pci_resource_start(pdev, 0);
 306         len = pci_resource_len(pdev, 0);
 307         ipath_cdbg(VERBOSE, "regbase (0) %llx len %d pdev->irq %d, vend %x/%x "
 308                    "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
 309                    ent->device, ent->driver_data);
 310
 311         read_bars(dd, pdev, &bar0, &bar1);
 312
 313         if (!bar1 && !(bar0 & ~0xf)) {
 314                 if (addr) {
 315                         dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
 316                                  "rewriting as %llx\n", addr);
 317                         ret = pci_write_config_dword(
 318                                 pdev, PCI_BASE_ADDRESS_0, addr);
 319                         if (ret) {
 320                                 ipath_dev_err(dd, "rewrite of BAR0 "
 321                                               "failed: err %d\n", -ret);
 322                                 goto bail_disable;
 323                         }
 324                         ret = pci_write_config_dword(
 325                                 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
 326                         if (ret) {
 327                                 ipath_dev_err(dd, "rewrite of BAR1 "
 328                                               "failed: err %d\n", -ret);
 329                                 goto bail_disable;
 330                         }
 331                 } else {
 332                         ipath_dev_err(dd, "BAR is 0 (probable RESET), "
 333                                       "not usable until reboot\n");
 334                         ret = -ENODEV;
 335                         goto bail_disable;
 336                 }
 337         }
 338
 339         ret = pci_request_regions(pdev, IPATH_DRV_NAME);
 340         if (ret) {
 341                 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
 342                          "err %d\n", dd->ipath_unit, -ret);
 343                 goto bail_disable;
 344         }
 345
 346         ret = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
 347         if (ret) {
 348                 /*
 349                  * if the 64 bit setup fails, try 32 bit.  Some systems
 350                  * do not setup 64 bit maps on systems with 2GB or less
 351                  * memory installed.
 352                  */
 353                 ret = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
 354                 if (ret) {
 355                         dev_info(&pdev->dev,
 356                                 "Unable to set DMA mask for unit %u: %d\n",
 357                                 dd->ipath_unit, ret);
 358                         goto bail_regions;
 359                 }
 360                 else {
 361                         ipath_dbg("No 64bit DMA mask, used 32 bit mask\n");
 362                         ret = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
 363                         if (ret)
 364                                 dev_info(&pdev->dev,
 365                                         "Unable to set DMA consistent mask "
 366                                         "for unit %u: %d\n",
 367                                         dd->ipath_unit, ret);
 368
 369                 }
 370         }
 371         else {
 372                 ret = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
 373                 if (ret)
 374                         dev_info(&pdev->dev,
 375                                 "Unable to set DMA consistent mask "
 376                                 "for unit %u: %d\n",
 377                                 dd->ipath_unit, ret);
 378         }
 379
 380         pci_set_master(pdev);
 381
 382         /*
 383          * Save BARs to rewrite after device reset.  Save all 64 bits of
 384          * BAR, just in case.
 385          */
 386         dd->ipath_pcibar0 = addr;
 387         dd->ipath_pcibar1 = addr >> 32;
 388         dd->ipath_deviceid = ent->device;       /* save for later use */
 389         dd->ipath_vendorid = ent->vendor;
 390
 391         /* setup the chip-specific functions, as early as possible. */
 392         switch (ent->device) {
 393 #ifdef CONFIG_HT_IRQ
 394         case PCI_DEVICE_ID_INFINIPATH_HT:
 395                 ipath_init_iba6110_funcs(dd);
 396                 break;
 397 #endif
 398 #ifdef CONFIG_PCI_MSI
 399         case PCI_DEVICE_ID_INFINIPATH_PE800:
 400                 ipath_init_iba6120_funcs(dd);
 401                 break;
 402 #endif
 403         default:
 404                 ipath_dev_err(dd, "Found unknown QLogic deviceid 0x%x, "
 405                               "failing\n", ent->device);
 406                 return -ENODEV;
 407         }
 408
 409         for (j = 0; j < 6; j++) {
 410                 if (!pdev->resource[j].start)
 411                         continue;
 412                 ipath_cdbg(VERBOSE, "BAR %d start %llx, end %llx, len %llx\n",
 413                            j, (unsigned long long)pdev->resource[j].start,
 414                            (unsigned long long)pdev->resource[j].end,
 415                            (unsigned long long)pci_resource_len(pdev, j));
 416         }
 417
 418         if (!addr) {
 419                 ipath_dev_err(dd, "No valid address in BAR 0!\n");
 420                 ret = -ENODEV;
 421                 goto bail_regions;
 422         }
 423
 424         dd->ipath_deviceid = ent->device;       /* save for later use */
 425         dd->ipath_vendorid = ent->vendor;
 426
 427         ret = pci_read_config_byte(pdev, PCI_REVISION_ID, &rev);
 428         if (ret) {
 429                 ipath_dev_err(dd, "Failed to read PCI revision ID unit "
 430                               "%u: err %d\n", dd->ipath_unit, -ret);
 431                 goto bail_regions;      /* shouldn't ever happen */
 432         }
 433         dd->ipath_pcirev = rev;
 434
 435 #if defined(__powerpc__)
 436         /* There isn't a generic way to specify writethrough mappings */
 437         dd->ipath_kregbase = __ioremap(addr, len,
 438                 (_PAGE_NO_CACHE|_PAGE_WRITETHRU));
 439 #else
 440         dd->ipath_kregbase = ioremap_nocache(addr, len);
 441 #endif
 442
 443         if (!dd->ipath_kregbase) {
 444                 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
 445                           addr);
 446                 ret = -ENOMEM;
 447                 goto bail_iounmap;
 448         }
 449         dd->ipath_kregend = (u64 __iomem *)
 450                 ((void __iomem *)dd->ipath_kregbase + len);
 451         dd->ipath_physaddr = addr;      /* used for io_remap, etc. */
 452         /* for user mmap */
 453         ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p\n",
 454                    addr, dd->ipath_kregbase);
 455
 456         /*
 457          * clear ipath_flags here instead of in ipath_init_chip as it is set
 458          * by ipath_setup_htconfig.
 459          */
 460         dd->ipath_flags = 0;
 461         dd->ipath_lli_counter = 0;
 462         dd->ipath_lli_errors = 0;
 463
 464         if (dd->ipath_f_bus(dd, pdev))
 465                 ipath_dev_err(dd, "Failed to setup config space; "
 466                               "continuing anyway\n");
 467
 468         /*
 469          * set up our interrupt handler; IRQF_SHARED probably not needed,
 470          * since MSI interrupts shouldn't be shared but won't  hurt for now.
 471          * check 0 irq after we return from chip-specific bus setup, since
 472          * that can affect this due to setup
 473          */
 474         if (!dd->ipath_irq)
 475                 ipath_dev_err(dd, "irq is 0, BIOS error?  Interrupts won't "
 476                               "work\n");
 477         else {
 478                 ret = request_irq(dd->ipath_irq, ipath_intr, IRQF_SHARED,
 479                                   IPATH_DRV_NAME, dd);
 480                 if (ret) {
 481                         ipath_dev_err(dd, "Couldn't setup irq handler, "
 482                                       "irq=%d: %d\n", dd->ipath_irq, ret);
 483                         goto bail_iounmap;
 484                 }
 485         }
 486
 487         ret = ipath_init_chip(dd, 0);   /* do the chip-specific init */
 488         if (ret)
 489                 goto bail_irqsetup;
 490
 491         ret = ipath_enable_wc(dd);
 492
 493         if (ret) {
 494                 ipath_dev_err(dd, "Write combining not enabled "
 495                               "(err %d): performance may be poor\n",
 496                               -ret);
 497                 ret = 0;
 498         }
 499
 500         ipath_device_create_group(&pdev->dev, dd);
 501         ipathfs_add_device(dd);
 502         ipath_user_add(dd);
 503         ipath_diag_add(dd);
 504         ipath_register_ib_device(dd);
 505
 506         goto bail;
 507
 508 bail_irqsetup:
 509         if (pdev->irq) free_irq(pdev->irq, dd);
 510
 511 bail_iounmap:
 512         iounmap((volatile void __iomem *) dd->ipath_kregbase);
 513
 514 bail_regions:
 515         pci_release_regions(pdev);
 516
 517 bail_disable:
 518         pci_disable_device(pdev);
 519
 520 bail_devdata:
 521         ipath_free_devdata(pdev, dd);
 522
 523 bail:
 524         return ret;
 525 }
 526
 527 static void __devexit cleanup_device(struct ipath_devdata *dd)
 528 {
 529         int port;
 530
 531         ipath_shutdown_device(dd);
 532
 533         if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
 534                 /* can't do anything more with chip; needs re-init */
 535                 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
 536                 if (dd->ipath_kregbase) {
 537                         /*
 538                          * if we haven't already cleaned up before these are
 539                          * to ensure any register reads/writes "fail" until
 540                          * re-init
 541                          */
 542                         dd->ipath_kregbase = NULL;
 543                         dd->ipath_uregbase = 0;
 544                         dd->ipath_sregbase = 0;
 545                         dd->ipath_cregbase = 0;
 546                         dd->ipath_kregsize = 0;
 547                 }
 548                 ipath_disable_wc(dd);
 549         }
 550
 551         if (dd->ipath_pioavailregs_dma) {
 552                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
 553                                   (void *) dd->ipath_pioavailregs_dma,
 554                                   dd->ipath_pioavailregs_phys);
 555                 dd->ipath_pioavailregs_dma = NULL;
 556         }
 557         if (dd->ipath_dummy_hdrq) {
 558                 dma_free_coherent(&dd->pcidev->dev,
 559                         dd->ipath_pd[0]->port_rcvhdrq_size,
 560                         dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
 561                 dd->ipath_dummy_hdrq = NULL;
 562         }
 563
 564         if (dd->ipath_pageshadow) {
 565                 struct page **tmpp = dd->ipath_pageshadow;
 566                 dma_addr_t *tmpd = dd->ipath_physshadow;
 567                 int i, cnt = 0;
 568
 569                 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
 570                            "locked\n");
 571                 for (port = 0; port < dd->ipath_cfgports; port++) {
 572                         int port_tidbase = port * dd->ipath_rcvtidcnt;
 573                         int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
 574                         for (i = port_tidbase; i < maxtid; i++) {
 575                                 if (!tmpp[i])
 576                                         continue;
 577                                 pci_unmap_page(dd->pcidev, tmpd[i],
 578                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 579                                 ipath_release_user_pages(&tmpp[i], 1);
 580                                 tmpp[i] = NULL;
 581                                 cnt++;
 582                         }
 583                 }
 584                 if (cnt) {
 585                         ipath_stats.sps_pageunlocks += cnt;
 586                         ipath_cdbg(VERBOSE, "There were still %u expTID "
 587                                    "entries locked\n", cnt);
 588                 }
 589                 if (ipath_stats.sps_pagelocks ||
 590                     ipath_stats.sps_pageunlocks)
 591                         ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
 592                                    "unlocked via ipath_m{un}lock\n",
 593                                    (unsigned long long)
 594                                    ipath_stats.sps_pagelocks,
 595                                    (unsigned long long)
 596                                    ipath_stats.sps_pageunlocks);
 597
 598                 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
 599                            dd->ipath_pageshadow);
 600                 vfree(dd->ipath_pageshadow);
 601                 dd->ipath_pageshadow = NULL;
 602         }
 603
 604         /*
 605          * free any resources still in use (usually just kernel ports)
 606          * at unload; we do for portcnt, not cfgports, because cfgports
 607          * could have changed while we were loaded.
 608          */
 609         for (port = 0; port < dd->ipath_portcnt; port++) {
 610                 struct ipath_portdata *pd = dd->ipath_pd[port];
 611                 dd->ipath_pd[port] = NULL;
 612                 ipath_free_pddata(dd, pd);
 613         }
 614         kfree(dd->ipath_pd);
 615         /*
 616          * debuggability, in case some cleanup path tries to use it
 617          * after this
 618          */
 619         dd->ipath_pd = NULL;
 620 }
 621
 622 static void __devexit ipath_remove_one(struct pci_dev *pdev)
 623 {
 624         struct ipath_devdata *dd = pci_get_drvdata(pdev);
 625
 626         ipath_cdbg(VERBOSE, "removing, pdev=%p, dd=%p\n", pdev, dd);
 627
 628         if (dd->verbs_dev)
 629                 ipath_unregister_ib_device(dd->verbs_dev);
 630
 631         ipath_diag_remove(dd);
 632         ipath_user_remove(dd);
 633         ipathfs_remove_device(dd);
 634         ipath_device_remove_group(&pdev->dev, dd);
 635
 636         ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
 637                    "unit %u\n", dd, (u32) dd->ipath_unit);
 638
 639         cleanup_device(dd);
 640
 641         /*
 642          * turn off rcv, send, and interrupts for all ports, all drivers
 643          * should also hard reset the chip here?
 644          * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
 645          * for all versions of the driver, if they were allocated
 646          */
 647         if (dd->ipath_irq) {
 648                 ipath_cdbg(VERBOSE, "unit %u free irq %d\n",
 649                            dd->ipath_unit, dd->ipath_irq);
 650                 dd->ipath_f_free_irq(dd);
 651         } else
 652                 ipath_dbg("irq is 0, not doing free_irq "
 653                           "for unit %u\n", dd->ipath_unit);
 654         /*
 655          * we check for NULL here, because it's outside
 656          * the kregbase check, and we need to call it
 657          * after the free_irq.  Thus it's possible that
 658          * the function pointers were never initialized.
 659          */
 660         if (dd->ipath_f_cleanup)
 661                 /* clean up chip-specific stuff */
 662                 dd->ipath_f_cleanup(dd);
 663
 664         ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n", dd->ipath_kregbase);
 665         iounmap((volatile void __iomem *) dd->ipath_kregbase);
 666         pci_release_regions(pdev);
 667         ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
 668         pci_disable_device(pdev);
 669
 670         ipath_free_devdata(pdev, dd);
 671 }
 672
 673 /* general driver use */
 674 DEFINE_MUTEX(ipath_mutex);
 675
 676 static DEFINE_SPINLOCK(ipath_pioavail_lock);
 677
 678 /**
 679  * ipath_disarm_piobufs - cancel a range of PIO buffers
 680  * @dd: the infinipath device
 681  * @first: the first PIO buffer to cancel
 682  * @cnt: the number of PIO buffers to cancel
 683  *
 684  * cancel a range of PIO buffers, used when they might be armed, but
 685  * not triggered.  Used at init to ensure buffer state, and also user
 686  * process close, in case it died while writing to a PIO buffer
 687  * Also after errors.
 688  */
 689 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
 690                           unsigned cnt)
 691 {
 692         unsigned i, last = first + cnt;
 693         u64 sendctrl, sendorig;
 694
 695         ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
 696         sendorig = dd->ipath_sendctrl | INFINIPATH_S_DISARM;
 697         for (i = first; i < last; i++) {
 698                 sendctrl = sendorig |
 699                         (i << INFINIPATH_S_DISARMPIOBUF_SHIFT);
 700                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 701                                  sendctrl);
 702         }
 703
 704         /*
 705          * Write it again with current value, in case ipath_sendctrl changed
 706          * while we were looping; no critical bits that would require
 707          * locking.
 708          *
 709          * Write a 0, and then the original value, reading scratch in
 710          * between.  This seems to avoid a chip timing race that causes
 711          * pioavail updates to memory to stop.
 712          */
 713         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 714                          0);
 715         sendorig = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 716         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 717                          dd->ipath_sendctrl);
 718 }
 719
 720 /**
 721  * ipath_wait_linkstate - wait for an IB link state change to occur
 722  * @dd: the infinipath device
 723  * @state: the state to wait for
 724  * @msecs: the number of milliseconds to wait
 725  *
 726  * wait up to msecs milliseconds for IB link state change to occur for
 727  * now, take the easy polling route.  Currently used only by
 728  * ipath_set_linkstate.  Returns 0 if state reached, otherwise
 729  * -ETIMEDOUT state can have multiple states set, for any of several
 730  * transitions.
 731  */
 732 static int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state,
 733                                 int msecs)
 734 {
 735         dd->ipath_state_wanted = state;
 736         wait_event_interruptible_timeout(ipath_state_wait,
 737                                          (dd->ipath_flags & state),
 738                                          msecs_to_jiffies(msecs));
 739         dd->ipath_state_wanted = 0;
 740
 741         if (!(dd->ipath_flags & state)) {
 742                 u64 val;
 743                 ipath_cdbg(VERBOSE, "Didn't reach linkstate %s within %u"
 744                            " ms\n",
 745                            /* test INIT ahead of DOWN, both can be set */
 746                            (state & IPATH_LINKINIT) ? "INIT" :
 747                            ((state & IPATH_LINKDOWN) ? "DOWN" :
 748                             ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
 749                            msecs);
 750                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 751                 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
 752                            (unsigned long long) ipath_read_kreg64(
 753                                    dd, dd->ipath_kregs->kr_ibcctrl),
 754                            (unsigned long long) val,
 755                            ipath_ibcstatus_str[val & 0xf]);
 756         }
 757         return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
 758 }
 759
 760 /*
 761  * Decode the error status into strings, deciding whether to always
 762  * print * it or not depending on "normal packet errors" vs everything
 763  * else.   Return 1 if "real" errors, otherwise 0 if only packet
 764  * errors, so caller can decide what to print with the string.
 765  */
 766 int ipath_decode_err(char *buf, size_t blen, ipath_err_t err)
 767 {
 768         int iserr = 1;
 769         *buf = '\0';
 770         if (err & INFINIPATH_E_PKTERRS) {
 771                 if (!(err & ~INFINIPATH_E_PKTERRS))
 772                         iserr = 0; // if only packet errors.
 773                 if (ipath_debug & __IPATH_ERRPKTDBG) {
 774                         if (err & INFINIPATH_E_REBP)
 775                                 strlcat(buf, "EBP ", blen);
 776                         if (err & INFINIPATH_E_RVCRC)
 777                                 strlcat(buf, "VCRC ", blen);
 778                         if (err & INFINIPATH_E_RICRC) {
 779                                 strlcat(buf, "CRC ", blen);
 780                                 // clear for check below, so only once
 781                                 err &= INFINIPATH_E_RICRC;
 782                         }
 783                         if (err & INFINIPATH_E_RSHORTPKTLEN)
 784                                 strlcat(buf, "rshortpktlen ", blen);
 785                         if (err & INFINIPATH_E_SDROPPEDDATAPKT)
 786                                 strlcat(buf, "sdroppeddatapkt ", blen);
 787                         if (err & INFINIPATH_E_SPKTLEN)
 788                                 strlcat(buf, "spktlen ", blen);
 789                 }
 790                 if ((err & INFINIPATH_E_RICRC) &&
 791                         !(err&(INFINIPATH_E_RVCRC|INFINIPATH_E_REBP)))
 792                         strlcat(buf, "CRC ", blen);
 793                 if (!iserr)
 794                         goto done;
 795         }
 796         if (err & INFINIPATH_E_RHDRLEN)
 797                 strlcat(buf, "rhdrlen ", blen);
 798         if (err & INFINIPATH_E_RBADTID)
 799                 strlcat(buf, "rbadtid ", blen);
 800         if (err & INFINIPATH_E_RBADVERSION)
 801                 strlcat(buf, "rbadversion ", blen);
 802         if (err & INFINIPATH_E_RHDR)
 803                 strlcat(buf, "rhdr ", blen);
 804         if (err & INFINIPATH_E_RLONGPKTLEN)
 805                 strlcat(buf, "rlongpktlen ", blen);
 806         if (err & INFINIPATH_E_RMAXPKTLEN)
 807                 strlcat(buf, "rmaxpktlen ", blen);
 808         if (err & INFINIPATH_E_RMINPKTLEN)
 809                 strlcat(buf, "rminpktlen ", blen);
 810         if (err & INFINIPATH_E_SMINPKTLEN)
 811                 strlcat(buf, "sminpktlen ", blen);
 812         if (err & INFINIPATH_E_RFORMATERR)
 813                 strlcat(buf, "rformaterr ", blen);
 814         if (err & INFINIPATH_E_RUNSUPVL)
 815                 strlcat(buf, "runsupvl ", blen);
 816         if (err & INFINIPATH_E_RUNEXPCHAR)
 817                 strlcat(buf, "runexpchar ", blen);
 818         if (err & INFINIPATH_E_RIBFLOW)
 819                 strlcat(buf, "ribflow ", blen);
 820         if (err & INFINIPATH_E_SUNDERRUN)
 821                 strlcat(buf, "sunderrun ", blen);
 822         if (err & INFINIPATH_E_SPIOARMLAUNCH)
 823                 strlcat(buf, "spioarmlaunch ", blen);
 824         if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
 825                 strlcat(buf, "sunexperrpktnum ", blen);
 826         if (err & INFINIPATH_E_SDROPPEDSMPPKT)
 827                 strlcat(buf, "sdroppedsmppkt ", blen);
 828         if (err & INFINIPATH_E_SMAXPKTLEN)
 829                 strlcat(buf, "smaxpktlen ", blen);
 830         if (err & INFINIPATH_E_SUNSUPVL)
 831                 strlcat(buf, "sunsupVL ", blen);
 832         if (err & INFINIPATH_E_INVALIDADDR)
 833                 strlcat(buf, "invalidaddr ", blen);
 834         if (err & INFINIPATH_E_RRCVEGRFULL)
 835                 strlcat(buf, "rcvegrfull ", blen);
 836         if (err & INFINIPATH_E_RRCVHDRFULL)
 837                 strlcat(buf, "rcvhdrfull ", blen);
 838         if (err & INFINIPATH_E_IBSTATUSCHANGED)
 839                 strlcat(buf, "ibcstatuschg ", blen);
 840         if (err & INFINIPATH_E_RIBLOSTLINK)
 841                 strlcat(buf, "riblostlink ", blen);
 842         if (err & INFINIPATH_E_HARDWARE)
 843                 strlcat(buf, "hardware ", blen);
 844         if (err & INFINIPATH_E_RESET)
 845                 strlcat(buf, "reset ", blen);
 846 done:
 847         return iserr;
 848 }
 849
 850 /**
 851  * get_rhf_errstring - decode RHF errors
 852  * @err: the err number
 853  * @msg: the output buffer
 854  * @len: the length of the output buffer
 855  *
 856  * only used one place now, may want more later
 857  */
 858 static void get_rhf_errstring(u32 err, char *msg, size_t len)
 859 {
 860         /* if no errors, and so don't need to check what's first */
 861         *msg = '\0';
 862
 863         if (err & INFINIPATH_RHF_H_ICRCERR)
 864                 strlcat(msg, "icrcerr ", len);
 865         if (err & INFINIPATH_RHF_H_VCRCERR)
 866                 strlcat(msg, "vcrcerr ", len);
 867         if (err & INFINIPATH_RHF_H_PARITYERR)
 868                 strlcat(msg, "parityerr ", len);
 869         if (err & INFINIPATH_RHF_H_LENERR)
 870                 strlcat(msg, "lenerr ", len);
 871         if (err & INFINIPATH_RHF_H_MTUERR)
 872                 strlcat(msg, "mtuerr ", len);
 873         if (err & INFINIPATH_RHF_H_IHDRERR)
 874                 /* infinipath hdr checksum error */
 875                 strlcat(msg, "ipathhdrerr ", len);
 876         if (err & INFINIPATH_RHF_H_TIDERR)
 877                 strlcat(msg, "tiderr ", len);
 878         if (err & INFINIPATH_RHF_H_MKERR)
 879                 /* bad port, offset, etc. */
 880                 strlcat(msg, "invalid ipathhdr ", len);
 881         if (err & INFINIPATH_RHF_H_IBERR)
 882                 strlcat(msg, "iberr ", len);
 883         if (err & INFINIPATH_RHF_L_SWA)
 884                 strlcat(msg, "swA ", len);
 885         if (err & INFINIPATH_RHF_L_SWB)
 886                 strlcat(msg, "swB ", len);
 887 }
 888
 889 /**
 890  * ipath_get_egrbuf - get an eager buffer
 891  * @dd: the infinipath device
 892  * @bufnum: the eager buffer to get
 893  * @err: unused
 894  *
 895  * must only be called if ipath_pd[port] is known to be allocated
 896  */
 897 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum,
 898                                      int err)
 899 {
 900         return dd->ipath_port0_skbinfo ?
 901                 (void *) dd->ipath_port0_skbinfo[bufnum].skb->data : NULL;
 902 }
 903
 904 /**
 905  * ipath_alloc_skb - allocate an skb and buffer with possible constraints
 906  * @dd: the infinipath device
 907  * @gfp_mask: the sk_buff SFP mask
 908  */
 909 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
 910                                 gfp_t gfp_mask)
 911 {
 912         struct sk_buff *skb;
 913         u32 len;
 914
 915         /*
 916          * Only fully supported way to handle this is to allocate lots
 917          * extra, align as needed, and then do skb_reserve().  That wastes
 918          * a lot of memory...  I'll have to hack this into infinipath_copy
 919          * also.
 920          */
 921
 922         /*
 923          * We need 2 extra bytes for ipath_ether data sent in the
 924          * key header.  In order to keep everything dword aligned,
 925          * we'll reserve 4 bytes.
 926          */
 927         len = dd->ipath_ibmaxlen + 4;
 928
 929         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 930                 /* We need a 2KB multiple alignment, and there is no way
 931                  * to do it except to allocate extra and then skb_reserve
 932                  * enough to bring it up to the right alignment.
 933                  */
 934                 len += 2047;
 935         }
 936
 937         skb = __dev_alloc_skb(len, gfp_mask);
 938         if (!skb) {
 939                 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
 940                               len);
 941                 goto bail;
 942         }
 943
 944         skb_reserve(skb, 4);
 945
 946         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 947                 u32 una = (unsigned long)skb->data & 2047;
 948                 if (una)
 949                         skb_reserve(skb, 2048 - una);
 950         }
 951
 952 bail:
 953         return skb;
 954 }
 955
 956 static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
 957                              u32 eflags,
 958                              u32 l,
 959                              u32 etail,
 960                              u64 *rc)
 961 {
 962         char emsg[128];
 963         struct ipath_message_header *hdr;
 964
 965         get_rhf_errstring(eflags, emsg, sizeof emsg);
 966         hdr = (struct ipath_message_header *)&rc[1];
 967         ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
 968                    "tlen=%x opcode=%x egridx=%x: %s\n",
 969                    eflags, l,
 970                    ipath_hdrget_rcv_type((__le32 *) rc),
 971                    ipath_hdrget_length_in_bytes((__le32 *) rc),
 972                    be32_to_cpu(hdr->bth[0]) >> 24,
 973                    etail, emsg);
 974
 975         /* Count local link integrity errors. */
 976         if (eflags & (INFINIPATH_RHF_H_ICRCERR | INFINIPATH_RHF_H_VCRCERR)) {
 977                 u8 n = (dd->ipath_ibcctrl >>
 978                         INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
 979                         INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
 980
 981                 if (++dd->ipath_lli_counter > n) {
 982                         dd->ipath_lli_counter = 0;
 983                         dd->ipath_lli_errors++;
 984                 }
 985         }
 986 }
 987
 988 /*
 989  * ipath_kreceive - receive a packet
 990  * @dd: the infinipath device
 991  *
 992  * called from interrupt handler for errors or receive interrupt
 993  */
 994 void ipath_kreceive(struct ipath_devdata *dd)
 995 {
 996         u64 *rc;
 997         void *ebuf;
 998         const u32 rsize = dd->ipath_rcvhdrentsize;      /* words */
 999         const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
1000         u32 etail = -1, l, hdrqtail;
1001         struct ipath_message_header *hdr;
1002         u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
1003         static u64 totcalls;    /* stats, may eventually remove */
1004
1005         if (!dd->ipath_hdrqtailptr) {
1006                 ipath_dev_err(dd,
1007                               "hdrqtailptr not set, can't do receives\n");
1008                 goto bail;
1009         }
1010
1011         /* There is already a thread processing this queue. */
1012         if (test_and_set_bit(0, &dd->ipath_rcv_pending))
1013                 goto bail;
1014
1015         l = dd->ipath_port0head;
1016         hdrqtail = (u32) le64_to_cpu(*dd->ipath_hdrqtailptr);
1017         if (l == hdrqtail)
1018                 goto done;
1019
1020 reloop:
1021         for (i = 0; l != hdrqtail; i++) {
1022                 u32 qp;
1023                 u8 *bthbytes;
1024
1025                 rc = (u64 *) (dd->ipath_pd[0]->port_rcvhdrq + (l << 2));
1026                 hdr = (struct ipath_message_header *)&rc[1];
1027                 /*
1028                  * could make a network order version of IPATH_KD_QP, and
1029                  * do the obvious shift before masking to speed this up.
1030                  */
1031                 qp = ntohl(hdr->bth[1]) & 0xffffff;
1032                 bthbytes = (u8 *) hdr->bth;
1033
1034                 eflags = ipath_hdrget_err_flags((__le32 *) rc);
1035                 etype = ipath_hdrget_rcv_type((__le32 *) rc);
1036                 /* total length */
1037                 tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
1038                 ebuf = NULL;
1039                 if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
1040                         /*
1041                          * it turns out that the chips uses an eager buffer
1042                          * for all non-expected packets, whether it "needs"
1043                          * one or not.  So always get the index, but don't
1044                          * set ebuf (so we try to copy data) unless the
1045                          * length requires it.
1046                          */
1047                         etail = ipath_hdrget_index((__le32 *) rc);
1048                         if (tlen > sizeof(*hdr) ||
1049                             etype == RCVHQ_RCV_TYPE_NON_KD)
1050                                 ebuf = ipath_get_egrbuf(dd, etail, 0);
1051                 }
1052
1053                 /*
1054                  * both tiderr and ipathhdrerr are set for all plain IB
1055                  * packets; only ipathhdrerr should be set.
1056                  */
1057
1058                 if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
1059                     RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
1060                             hdr->iph.ver_port_tid_offset) !=
1061                     IPS_PROTO_VERSION) {
1062                         ipath_cdbg(PKT, "Bad InfiniPath protocol version "
1063                                    "%x\n", etype);
1064                 }
1065
1066                 if (unlikely(eflags))
1067                         ipath_rcv_hdrerr(dd, eflags, l, etail, rc);
1068                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
1069                         ipath_ib_rcv(dd->verbs_dev, rc + 1, ebuf, tlen);
1070                         if (dd->ipath_lli_counter)
1071                                 dd->ipath_lli_counter--;
1072                         ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1073                                    "qp=%x), len %x; ignored\n",
1074                                    etype, bthbytes[0], qp, tlen);
1075                 }
1076                 else if (etype == RCVHQ_RCV_TYPE_EAGER)
1077                         ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1078                                    "qp=%x), len %x; ignored\n",
1079                                    etype, bthbytes[0], qp, tlen);
1080                 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
1081                         ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
1082                                   be32_to_cpu(hdr->bth[0]) & 0xff);
1083                 else {
1084                         /*
1085                          * error packet, type of error  unknown.
1086                          * Probably type 3, but we don't know, so don't
1087                          * even try to print the opcode, etc.
1088                          */
1089                         ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
1090                                   "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
1091                                   "hdr %llx %llx %llx %llx %llx\n",
1092                                   etail, tlen, (unsigned long) rc, l,
1093                                   (unsigned long long) rc[0],
1094                                   (unsigned long long) rc[1],
1095                                   (unsigned long long) rc[2],
1096                                   (unsigned long long) rc[3],
1097                                   (unsigned long long) rc[4],
1098                                   (unsigned long long) rc[5]);
1099                 }
1100                 l += rsize;
1101                 if (l >= maxcnt)
1102                         l = 0;
1103                 if (etype != RCVHQ_RCV_TYPE_EXPECTED)
1104                     updegr = 1;
1105                 /*
1106                  * update head regs on last packet, and every 16 packets.
1107                  * Reduce bus traffic, while still trying to prevent
1108                  * rcvhdrq overflows, for when the queue is nearly full
1109                  */
1110                 if (l == hdrqtail || (i && !(i&0xf))) {
1111                         u64 lval;
1112                         if (l == hdrqtail)
1113                                 /* request IBA6120 interrupt only on last */
1114                                 lval = dd->ipath_rhdrhead_intr_off | l;
1115                         else
1116                                 lval = l;
1117                         (void)ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
1118                         if (updegr) {
1119                                 (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
1120                                                        etail, 0);
1121                                 updegr = 0;
1122                         }
1123                 }
1124         }
1125
1126         if (!dd->ipath_rhdrhead_intr_off && !reloop) {
1127                 /* IBA6110 workaround; we can have a race clearing chip
1128                  * interrupt with another interrupt about to be delivered,
1129                  * and can clear it before it is delivered on the GPIO
1130                  * workaround.  By doing the extra check here for the
1131                  * in-memory tail register updating while we were doing
1132                  * earlier packets, we "almost" guarantee we have covered
1133                  * that case.
1134                  */
1135                 u32 hqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
1136                 if (hqtail != hdrqtail) {
1137                         hdrqtail = hqtail;
1138                         reloop = 1; /* loop 1 extra time at most */
1139                         goto reloop;
1140                 }
1141         }
1142
1143         pkttot += i;
1144
1145         dd->ipath_port0head = l;
1146
1147         if (pkttot > ipath_stats.sps_maxpkts_call)
1148                 ipath_stats.sps_maxpkts_call = pkttot;
1149         ipath_stats.sps_port0pkts += pkttot;
1150         ipath_stats.sps_avgpkts_call =
1151                 ipath_stats.sps_port0pkts / ++totcalls;
1152
1153 done:
1154         clear_bit(0, &dd->ipath_rcv_pending);
1155         smp_mb__after_clear_bit();
1156
1157 bail:;
1158 }
1159
1160 /**
1161  * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1162  * @dd: the infinipath device
1163  *
1164  * called whenever our local copy indicates we have run out of send buffers
1165  * NOTE: This can be called from interrupt context by some code
1166  * and from non-interrupt context by ipath_getpiobuf().
1167  */
1168
1169 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1170 {
1171         unsigned long flags;
1172         int i;
1173         const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1174
1175         /* If the generation (check) bits have changed, then we update the
1176          * busy bit for the corresponding PIO buffer.  This algorithm will
1177          * modify positions to the value they already have in some cases
1178          * (i.e., no change), but it's faster than changing only the bits
1179          * that have changed.
1180          *
1181          * We would like to do this atomicly, to avoid spinlocks in the
1182          * critical send path, but that's not really possible, given the
1183          * type of changes, and that this routine could be called on
1184          * multiple cpu's simultaneously, so we lock in this routine only,
1185          * to avoid conflicting updates; all we change is the shadow, and
1186          * it's a single 64 bit memory location, so by definition the update
1187          * is atomic in terms of what other cpu's can see in testing the
1188          * bits.  The spin_lock overhead isn't too bad, since it only
1189          * happens when all buffers are in use, so only cpu overhead, not
1190          * latency or bandwidth is affected.
1191          */
1192 #define _IPATH_ALL_CHECKBITS 0x5555555555555555ULL
1193         if (!dd->ipath_pioavailregs_dma) {
1194                 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1195                 return;
1196         }
1197         if (ipath_debug & __IPATH_VERBDBG) {
1198                 /* only if packet debug and verbose */
1199                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1200                 unsigned long *shadow = dd->ipath_pioavailshadow;
1201
1202                 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1203                            "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1204                            "s3=%lx\n",
1205                            (unsigned long long) le64_to_cpu(dma[0]),
1206                            shadow[0],
1207                            (unsigned long long) le64_to_cpu(dma[1]),
1208                            shadow[1],
1209                            (unsigned long long) le64_to_cpu(dma[2]),
1210                            shadow[2],
1211                            (unsigned long long) le64_to_cpu(dma[3]),
1212                            shadow[3]);
1213                 if (piobregs > 4)
1214                         ipath_cdbg(
1215                                 PKT, "2nd group, dma4=%llx shad4=%lx, "
1216                                 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1217                                 "d7=%llx s7=%lx\n",
1218                                 (unsigned long long) le64_to_cpu(dma[4]),
1219                                 shadow[4],
1220                                 (unsigned long long) le64_to_cpu(dma[5]),
1221                                 shadow[5],
1222                                 (unsigned long long) le64_to_cpu(dma[6]),
1223                                 shadow[6],
1224                                 (unsigned long long) le64_to_cpu(dma[7]),
1225                                 shadow[7]);
1226         }
1227         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1228         for (i = 0; i < piobregs; i++) {
1229                 u64 pchbusy, pchg, piov, pnew;
1230                 /*
1231                  * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1232                  */
1233                 if (i > 3) {
1234                         if (i & 1)
1235                                 piov = le64_to_cpu(
1236                                         dd->ipath_pioavailregs_dma[i - 1]);
1237                         else
1238                                 piov = le64_to_cpu(
1239                                         dd->ipath_pioavailregs_dma[i + 1]);
1240                 } else
1241                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1242                 pchg = _IPATH_ALL_CHECKBITS &
1243                         ~(dd->ipath_pioavailshadow[i] ^ piov);
1244                 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1245                 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1246                         pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1247                         pnew |= piov & pchbusy;
1248                         dd->ipath_pioavailshadow[i] = pnew;
1249                 }
1250         }
1251         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1252 }
1253
1254 /**
1255  * ipath_setrcvhdrsize - set the receive header size
1256  * @dd: the infinipath device
1257  * @rhdrsize: the receive header size
1258  *
1259  * called from user init code, and also layered driver init
1260  */
1261 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1262 {
1263         int ret = 0;
1264
1265         if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1266                 if (dd->ipath_rcvhdrsize != rhdrsize) {
1267                         dev_info(&dd->pcidev->dev,
1268                                  "Error: can't set protocol header "
1269                                  "size %u, already %u\n",
1270                                  rhdrsize, dd->ipath_rcvhdrsize);
1271                         ret = -EAGAIN;
1272                 } else
1273                         ipath_cdbg(VERBOSE, "Reuse same protocol header "
1274                                    "size %u\n", dd->ipath_rcvhdrsize);
1275         } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1276                                (sizeof(u64) / sizeof(u32)))) {
1277                 ipath_dbg("Error: can't set protocol header size %u "
1278                           "(> max %u)\n", rhdrsize,
1279                           dd->ipath_rcvhdrentsize -
1280                           (u32) (sizeof(u64) / sizeof(u32)));
1281                 ret = -EOVERFLOW;
1282         } else {
1283                 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1284                 dd->ipath_rcvhdrsize = rhdrsize;
1285                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1286                                  dd->ipath_rcvhdrsize);
1287                 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1288                            dd->ipath_rcvhdrsize);
1289         }
1290         return ret;
1291 }
1292
1293 /**
1294  * ipath_getpiobuf - find an available pio buffer
1295  * @dd: the infinipath device
1296  * @pbufnum: the buffer number is placed here
1297  *
1298  * do appropriate marking as busy, etc.
1299  * returns buffer number if one found (>=0), negative number is error.
1300  * Used by ipath_layer_send
1301  */
1302 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
1303 {
1304         int i, j, starti, updated = 0;
1305         unsigned piobcnt, iter;
1306         unsigned long flags;
1307         unsigned long *shadow = dd->ipath_pioavailshadow;
1308         u32 __iomem *buf;
1309
1310         piobcnt = (unsigned)(dd->ipath_piobcnt2k
1311                              + dd->ipath_piobcnt4k);
1312         starti = dd->ipath_lastport_piobuf;
1313         iter = piobcnt - starti;
1314         if (dd->ipath_upd_pio_shadow) {
1315                 /*
1316                  * Minor optimization.  If we had no buffers on last call,
1317                  * start out by doing the update; continue and do scan even
1318                  * if no buffers were updated, to be paranoid
1319                  */
1320                 ipath_update_pio_bufs(dd);
1321                 /* we scanned here, don't do it at end of scan */
1322                 updated = 1;
1323                 i = starti;
1324         } else
1325                 i = dd->ipath_lastpioindex;
1326
1327 rescan:
1328         /*
1329          * while test_and_set_bit() is atomic, we do that and then the
1330          * change_bit(), and the pair is not.  See if this is the cause
1331          * of the remaining armlaunch errors.
1332          */
1333         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1334         for (j = 0; j < iter; j++, i++) {
1335                 if (i >= piobcnt)
1336                         i = starti;
1337                 /*
1338                  * To avoid bus lock overhead, we first find a candidate
1339                  * buffer, then do the test and set, and continue if that
1340                  * fails.
1341                  */
1342                 if (test_bit((2 * i) + 1, shadow) ||
1343                     test_and_set_bit((2 * i) + 1, shadow))
1344                         continue;
1345                 /* flip generation bit */
1346                 change_bit(2 * i, shadow);
1347                 break;
1348         }
1349         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1350
1351         if (j == iter) {
1352                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1353
1354                 /*
1355                  * first time through; shadow exhausted, but may be real
1356                  * buffers available, so go see; if any updated, rescan
1357                  * (once)
1358                  */
1359                 if (!updated) {
1360                         ipath_update_pio_bufs(dd);
1361                         updated = 1;
1362                         i = starti;
1363                         goto rescan;
1364                 }
1365                 dd->ipath_upd_pio_shadow = 1;
1366                 /*
1367                  * not atomic, but if we lose one once in a while, that's OK
1368                  */
1369                 ipath_stats.sps_nopiobufs++;
1370                 if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1371                         ipath_dbg(
1372                                 "%u pio sends with no bufavail; dmacopy: "
1373                                 "%llx %llx %llx %llx; shadow:  "
1374                                 "%lx %lx %lx %lx\n",
1375                                 dd->ipath_consec_nopiobuf,
1376                                 (unsigned long long) le64_to_cpu(dma[0]),
1377                                 (unsigned long long) le64_to_cpu(dma[1]),
1378                                 (unsigned long long) le64_to_cpu(dma[2]),
1379                                 (unsigned long long) le64_to_cpu(dma[3]),
1380                                 shadow[0], shadow[1], shadow[2],
1381                                 shadow[3]);
1382                         /*
1383                          * 4 buffers per byte, 4 registers above, cover rest
1384                          * below
1385                          */
1386                         if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1387                             (sizeof(shadow[0]) * 4 * 4))
1388                                 ipath_dbg("2nd group: dmacopy: %llx %llx "
1389                                           "%llx %llx; shadow: %lx %lx "
1390                                           "%lx %lx\n",
1391                                           (unsigned long long)
1392                                           le64_to_cpu(dma[4]),
1393                                           (unsigned long long)
1394                                           le64_to_cpu(dma[5]),
1395                                           (unsigned long long)
1396                                           le64_to_cpu(dma[6]),
1397                                           (unsigned long long)
1398                                           le64_to_cpu(dma[7]),
1399                                           shadow[4], shadow[5],
1400                                           shadow[6], shadow[7]);
1401                 }
1402                 buf = NULL;
1403                 goto bail;
1404         }
1405
1406         /*
1407          * set next starting place.  Since it's just an optimization,
1408          * it doesn't matter who wins on this, so no locking
1409          */
1410         dd->ipath_lastpioindex = i + 1;
1411         if (dd->ipath_upd_pio_shadow)
1412                 dd->ipath_upd_pio_shadow = 0;
1413         if (dd->ipath_consec_nopiobuf)
1414                 dd->ipath_consec_nopiobuf = 0;
1415         if (i < dd->ipath_piobcnt2k)
1416                 buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1417                                        i * dd->ipath_palign);
1418         else
1419                 buf = (u32 __iomem *)
1420                         (dd->ipath_pio4kbase +
1421                          (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1422         ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1423                    i, (i < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1424         if (pbufnum)
1425                 *pbufnum = i;
1426
1427 bail:
1428         return buf;
1429 }
1430
1431 /**
1432  * ipath_create_rcvhdrq - create a receive header queue
1433  * @dd: the infinipath device
1434  * @pd: the port data
1435  *
1436  * this must be contiguous memory (from an i/o perspective), and must be
1437  * DMA'able (which means for some systems, it will go through an IOMMU,
1438  * or be forced into a low address range).
1439  */
1440 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1441                          struct ipath_portdata *pd)
1442 {
1443         int ret = 0;
1444
1445         if (!pd->port_rcvhdrq) {
1446                 dma_addr_t phys_hdrqtail;
1447                 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1448                 int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1449                                 sizeof(u32), PAGE_SIZE);
1450
1451                 pd->port_rcvhdrq = dma_alloc_coherent(
1452                         &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1453                         gfp_flags);
1454
1455                 if (!pd->port_rcvhdrq) {
1456                         ipath_dev_err(dd, "attempt to allocate %d bytes "
1457                                       "for port %u rcvhdrq failed\n",
1458                                       amt, pd->port_port);
1459                         ret = -ENOMEM;
1460                         goto bail;
1461                 }
1462                 pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
1463                         &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
1464                 if (!pd->port_rcvhdrtail_kvaddr) {
1465                         ipath_dev_err(dd, "attempt to allocate 1 page "
1466                                       "for port %u rcvhdrqtailaddr failed\n",
1467                                       pd->port_port);
1468                         ret = -ENOMEM;
1469                         dma_free_coherent(&dd->pcidev->dev, amt,
1470                                           pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1471                         pd->port_rcvhdrq = NULL;
1472                         goto bail;
1473                 }
1474                 pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
1475
1476                 pd->port_rcvhdrq_size = amt;
1477
1478                 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1479                            "for port %u rcvhdr Q\n",
1480                            amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1481                            (unsigned long) pd->port_rcvhdrq_phys,
1482                            (unsigned long) pd->port_rcvhdrq_size,
1483                            pd->port_port);
1484
1485                 ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
1486                            pd->port_port,
1487                            (unsigned long long) phys_hdrqtail);
1488         }
1489         else
1490                 ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
1491                            "hdrtailaddr@%p %llx physical\n",
1492                            pd->port_port, pd->port_rcvhdrq,
1493                            (unsigned long long) pd->port_rcvhdrq_phys,
1494                            pd->port_rcvhdrtail_kvaddr, (unsigned long long)
1495                            pd->port_rcvhdrqtailaddr_phys);
1496
1497         /* clear for security and sanity on each use */
1498         memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
1499         memset(pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1500
1501         /*
1502          * tell chip each time we init it, even if we are re-using previous
1503          * memory (we zero the register at process close)
1504          */
1505         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1506                               pd->port_port, pd->port_rcvhdrqtailaddr_phys);
1507         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1508                               pd->port_port, pd->port_rcvhdrq_phys);
1509
1510         ret = 0;
1511 bail:
1512         return ret;
1513 }
1514
1515 int ipath_waitfor_complete(struct ipath_devdata *dd, ipath_kreg reg_id,
1516                            u64 bits_to_wait_for, u64 * valp)
1517 {
1518         unsigned long timeout;
1519         u64 lastval, val;
1520         int ret;
1521
1522         lastval = ipath_read_kreg64(dd, reg_id);
1523         /* wait a ridiculously long time */
1524         timeout = jiffies + msecs_to_jiffies(5);
1525         do {
1526                 val = ipath_read_kreg64(dd, reg_id);
1527                 /* set so they have something, even on failures. */
1528                 *valp = val;
1529                 if ((val & bits_to_wait_for) == bits_to_wait_for) {
1530                         ret = 0;
1531                         break;
1532                 }
1533                 if (val != lastval)
1534                         ipath_cdbg(VERBOSE, "Changed from %llx to %llx, "
1535                                    "waiting for %llx bits\n",
1536                                    (unsigned long long) lastval,
1537                                    (unsigned long long) val,
1538                                    (unsigned long long) bits_to_wait_for);
1539                 cond_resched();
1540                 if (time_after(jiffies, timeout)) {
1541                         ipath_dbg("Didn't get bits %llx in register 0x%x, "
1542                                   "got %llx\n",
1543                                   (unsigned long long) bits_to_wait_for,
1544                                   reg_id, (unsigned long long) *valp);
1545                         ret = -ENODEV;
1546                         break;
1547                 }
1548         } while (1);
1549
1550         return ret;
1551 }
1552
1553 /**
1554  * ipath_waitfor_mdio_cmdready - wait for last command to complete
1555  * @dd: the infinipath device
1556  *
1557  * Like ipath_waitfor_complete(), but we wait for the CMDVALID bit to go
1558  * away indicating the last command has completed.  It doesn't return data
1559  */
1560 int ipath_waitfor_mdio_cmdready(struct ipath_devdata *dd)
1561 {
1562         unsigned long timeout;
1563         u64 val;
1564         int ret;
1565
1566         /* wait a ridiculously long time */
1567         timeout = jiffies + msecs_to_jiffies(5);
1568         do {
1569                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_mdio);
1570                 if (!(val & IPATH_MDIO_CMDVALID)) {
1571                         ret = 0;
1572                         break;
1573                 }
1574                 cond_resched();
1575                 if (time_after(jiffies, timeout)) {
1576                         ipath_dbg("CMDVALID stuck in mdio reg? (%llx)\n",
1577                                   (unsigned long long) val);
1578                         ret = -ENODEV;
1579                         break;
1580                 }
1581         } while (1);
1582
1583         return ret;
1584 }
1585
1586 static void ipath_set_ib_lstate(struct ipath_devdata *dd, int which)
1587 {
1588         static const char *what[4] = {
1589                 [0] = "DOWN",
1590                 [INFINIPATH_IBCC_LINKCMD_INIT] = "INIT",
1591                 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1592                 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1593         };
1594         int linkcmd = (which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
1595                         INFINIPATH_IBCC_LINKCMD_MASK;
1596
1597         ipath_cdbg(VERBOSE, "Trying to move unit %u to %s, current ltstate "
1598                    "is %s\n", dd->ipath_unit,
1599                    what[linkcmd],
1600                    ipath_ibcstatus_str[
1601                            (ipath_read_kreg64
1602                             (dd, dd->ipath_kregs->kr_ibcstatus) >>
1603                             INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
1604                            INFINIPATH_IBCS_LINKTRAININGSTATE_MASK]);
1605         /* flush all queued sends when going to DOWN or INIT, to be sure that
1606          * they don't block MAD packets */
1607         if (!linkcmd || linkcmd == INFINIPATH_IBCC_LINKCMD_INIT) {
1608                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1609                                  INFINIPATH_S_ABORT);
1610                 ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
1611                                     (unsigned)(dd->ipath_piobcnt2k +
1612                                     dd->ipath_piobcnt4k) -
1613                                     dd->ipath_lastport_piobuf);
1614         }
1615
1616         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1617                          dd->ipath_ibcctrl | which);
1618 }
1619
1620 int ipath_set_linkstate(struct ipath_devdata *dd, u8 newstate)
1621 {
1622         u32 lstate;
1623         int ret;
1624
1625         switch (newstate) {
1626         case IPATH_IB_LINKDOWN:
1627                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_POLL <<
1628                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1629                 /* don't wait */
1630                 ret = 0;
1631                 goto bail;
1632
1633         case IPATH_IB_LINKDOWN_SLEEP:
1634                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_SLEEP <<
1635                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1636                 /* don't wait */
1637                 ret = 0;
1638                 goto bail;
1639
1640         case IPATH_IB_LINKDOWN_DISABLE:
1641                 ipath_set_ib_lstate(dd,
1642                                     INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1643                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1644                 /* don't wait */
1645                 ret = 0;
1646                 goto bail;
1647
1648         case IPATH_IB_LINKINIT:
1649                 if (dd->ipath_flags & IPATH_LINKINIT) {
1650                         ret = 0;
1651                         goto bail;
1652                 }
1653                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_INIT <<
1654                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1655                 lstate = IPATH_LINKINIT;
1656                 break;
1657
1658         case IPATH_IB_LINKARM:
1659                 if (dd->ipath_flags & IPATH_LINKARMED) {
1660                         ret = 0;
1661                         goto bail;
1662                 }
1663                 if (!(dd->ipath_flags &
1664                       (IPATH_LINKINIT | IPATH_LINKACTIVE))) {
1665                         ret = -EINVAL;
1666                         goto bail;
1667                 }
1668                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ARMED <<
1669                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1670                 /*
1671                  * Since the port can transition to ACTIVE by receiving
1672                  * a non VL 15 packet, wait for either state.
1673                  */
1674                 lstate = IPATH_LINKARMED | IPATH_LINKACTIVE;
1675                 break;
1676
1677         case IPATH_IB_LINKACTIVE:
1678                 if (dd->ipath_flags & IPATH_LINKACTIVE) {
1679                         ret = 0;
1680                         goto bail;
1681                 }
1682                 if (!(dd->ipath_flags & IPATH_LINKARMED)) {
1683                         ret = -EINVAL;
1684                         goto bail;
1685                 }
1686                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ACTIVE <<
1687                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1688                 lstate = IPATH_LINKACTIVE;
1689                 break;
1690
1691         case IPATH_IB_LINK_LOOPBACK:
1692                 dev_info(&dd->pcidev->dev, "Enabling IB local loopback\n");
1693                 dd->ipath_ibcctrl |= INFINIPATH_IBCC_LOOPBACK;
1694                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1695                                  dd->ipath_ibcctrl);
1696                 ret = 0;
1697                 goto bail; // no state change to wait for
1698
1699         case IPATH_IB_LINK_EXTERNAL:
1700                 dev_info(&dd->pcidev->dev, "Disabling IB local loopback (normal)\n");
1701                 dd->ipath_ibcctrl &= ~INFINIPATH_IBCC_LOOPBACK;
1702                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1703                                  dd->ipath_ibcctrl);
1704                 ret = 0;
1705                 goto bail; // no state change to wait for
1706
1707         default:
1708                 ipath_dbg("Invalid linkstate 0x%x requested\n", newstate);
1709                 ret = -EINVAL;
1710                 goto bail;
1711         }
1712         ret = ipath_wait_linkstate(dd, lstate, 2000);
1713
1714 bail:
1715         return ret;
1716 }
1717
1718 /**
1719  * ipath_set_mtu - set the MTU
1720  * @dd: the infinipath device
1721  * @arg: the new MTU
1722  *
1723  * we can handle "any" incoming size, the issue here is whether we
1724  * need to restrict our outgoing size.   For now, we don't do any
1725  * sanity checking on this, and we don't deal with what happens to
1726  * programs that are already running when the size changes.
1727  * NOTE: changing the MTU will usually cause the IBC to go back to
1728  * link initialize (IPATH_IBSTATE_INIT) state...
1729  */
1730 int ipath_set_mtu(struct ipath_devdata *dd, u16 arg)
1731 {
1732         u32 piosize;
1733         int changed = 0;
1734         int ret;
1735
1736         /*
1737          * mtu is IB data payload max.  It's the largest power of 2 less
1738          * than piosize (or even larger, since it only really controls the
1739          * largest we can receive; we can send the max of the mtu and
1740          * piosize).  We check that it's one of the valid IB sizes.
1741          */
1742         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
1743             arg != 4096) {
1744                 ipath_dbg("Trying to set invalid mtu %u, failing\n", arg);
1745                 ret = -EINVAL;
1746                 goto bail;
1747         }
1748         if (dd->ipath_ibmtu == arg) {
1749                 ret = 0;        /* same as current */
1750                 goto bail;
1751         }
1752
1753         piosize = dd->ipath_ibmaxlen;
1754         dd->ipath_ibmtu = arg;
1755
1756         if (arg >= (piosize - IPATH_PIO_MAXIBHDR)) {
1757                 /* Only if it's not the initial value (or reset to it) */
1758                 if (piosize != dd->ipath_init_ibmaxlen) {
1759                         dd->ipath_ibmaxlen = piosize;
1760                         changed = 1;
1761                 }
1762         } else if ((arg + IPATH_PIO_MAXIBHDR) != dd->ipath_ibmaxlen) {
1763                 piosize = arg + IPATH_PIO_MAXIBHDR;
1764                 ipath_cdbg(VERBOSE, "ibmaxlen was 0x%x, setting to 0x%x "
1765                            "(mtu 0x%x)\n", dd->ipath_ibmaxlen, piosize,
1766                            arg);
1767                 dd->ipath_ibmaxlen = piosize;
1768                 changed = 1;
1769         }
1770
1771         if (changed) {
1772                 /*
1773                  * set the IBC maxpktlength to the size of our pio
1774                  * buffers in words
1775                  */
1776                 u64 ibc = dd->ipath_ibcctrl;
1777                 ibc &= ~(INFINIPATH_IBCC_MAXPKTLEN_MASK <<
1778                          INFINIPATH_IBCC_MAXPKTLEN_SHIFT);
1779
1780                 piosize = piosize - 2 * sizeof(u32);    /* ignore pbc */
1781                 dd->ipath_ibmaxlen = piosize;
1782                 piosize /= sizeof(u32); /* in words */
1783                 /*
1784                  * for ICRC, which we only send in diag test pkt mode, and
1785                  * we don't need to worry about that for mtu
1786                  */
1787                 piosize += 1;
1788
1789                 ibc |= piosize << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
1790                 dd->ipath_ibcctrl = ibc;
1791                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1792                                  dd->ipath_ibcctrl);
1793                 dd->ipath_f_tidtemplate(dd);
1794         }
1795
1796         ret = 0;
1797
1798 bail:
1799         return ret;
1800 }
1801
1802 int ipath_set_lid(struct ipath_devdata *dd, u32 arg, u8 lmc)
1803 {
1804         dd->ipath_lid = arg;
1805         dd->ipath_lmc = lmc;
1806
1807         return 0;
1808 }
1809
1810
1811 /**
1812  * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
1813  * @dd: the infinipath device
1814  * @regno: the register number to write
1815  * @port: the port containing the register
1816  * @value: the value to write
1817  *
1818  * Registers that vary with the chip implementation constants (port)
1819  * use this routine.
1820  */
1821 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
1822                           unsigned port, u64 value)
1823 {
1824         u16 where;
1825
1826         if (port < dd->ipath_portcnt &&
1827             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1828              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1829                 where = regno + port;
1830         else
1831                 where = -1;
1832
1833         ipath_write_kreg(dd, where, value);
1834 }
1835
1836 /**
1837  * ipath_shutdown_device - shut down a device
1838  * @dd: the infinipath device
1839  *
1840  * This is called to make the device quiet when we are about to
1841  * unload the driver, and also when the device is administratively
1842  * disabled.   It does not free any data structures.
1843  * Everything it does has to be setup again by ipath_init_chip(dd,1)
1844  */
1845 void ipath_shutdown_device(struct ipath_devdata *dd)
1846 {
1847         ipath_dbg("Shutting down the device\n");
1848
1849         dd->ipath_flags |= IPATH_LINKUNK;
1850         dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
1851                              IPATH_LINKINIT | IPATH_LINKARMED |
1852                              IPATH_LINKACTIVE);
1853         *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
1854                                 IPATH_STATUS_IB_READY);
1855
1856         /* mask interrupts, but not errors */
1857         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
1858
1859         dd->ipath_rcvctrl = 0;
1860         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1861                          dd->ipath_rcvctrl);
1862
1863         /*
1864          * gracefully stop all sends allowing any in progress to trickle out
1865          * first.
1866          */
1867         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0ULL);
1868         /* flush it */
1869         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1870         /*
1871          * enough for anything that's going to trickle out to have actually
1872          * done so.
1873          */
1874         udelay(5);
1875
1876         /*
1877          * abort any armed or launched PIO buffers that didn't go. (self
1878          * clearing).  Will cause any packet currently being transmitted to
1879          * go out with an EBP, and may also cause a short packet error on
1880          * the receiver.
1881          */
1882         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1883                          INFINIPATH_S_ABORT);
1884
1885         ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1886                             INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1887
1888         /* disable IBC */
1889         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
1890         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
1891                          dd->ipath_control | INFINIPATH_C_FREEZEMODE);
1892
1893         /*
1894          * clear SerdesEnable and turn the leds off; do this here because
1895          * we are unloading, so don't count on interrupts to move along
1896          * Turn the LEDs off explictly for the same reason.
1897          */
1898         dd->ipath_f_quiet_serdes(dd);
1899         dd->ipath_f_setextled(dd, 0, 0);
1900
1901         if (dd->ipath_stats_timer_active) {
1902                 del_timer_sync(&dd->ipath_stats_timer);
1903                 dd->ipath_stats_timer_active = 0;
1904         }
1905
1906         /*
1907          * clear all interrupts and errors, so that the next time the driver
1908          * is loaded or device is enabled, we know that whatever is set
1909          * happened while we were unloaded
1910          */
1911         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
1912                          ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
1913         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
1914         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
1915 }
1916
1917 /**
1918  * ipath_free_pddata - free a port's allocated data
1919  * @dd: the infinipath device
1920  * @pd: the portdata structure
1921  *
1922  * free up any allocated data for a port
1923  * This should not touch anything that would affect a simultaneous
1924  * re-allocation of port data, because it is called after ipath_mutex
1925  * is released (and can be called from reinit as well).
1926  * It should never change any chip state, or global driver state.
1927  * (The only exception to global state is freeing the port0 port0_skbs.)
1928  */
1929 void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
1930 {
1931         if (!pd)
1932                 return;
1933
1934         if (pd->port_rcvhdrq) {
1935                 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
1936                            "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
1937                            (unsigned long) pd->port_rcvhdrq_size);
1938                 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
1939                                   pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1940                 pd->port_rcvhdrq = NULL;
1941                 if (pd->port_rcvhdrtail_kvaddr) {
1942                         dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1943                                          pd->port_rcvhdrtail_kvaddr,
1944                                          pd->port_rcvhdrqtailaddr_phys);
1945                         pd->port_rcvhdrtail_kvaddr = NULL;
1946                 }
1947         }
1948         if (pd->port_port && pd->port_rcvegrbuf) {
1949                 unsigned e;
1950
1951                 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
1952                         void *base = pd->port_rcvegrbuf[e];
1953                         size_t size = pd->port_rcvegrbuf_size;
1954
1955                         ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
1956                                    "chunk %u/%u\n", base,
1957                                    (unsigned long) size,
1958                                    e, pd->port_rcvegrbuf_chunks);
1959                         dma_free_coherent(&dd->pcidev->dev, size,
1960                                 base, pd->port_rcvegrbuf_phys[e]);
1961                 }
1962                 kfree(pd->port_rcvegrbuf);
1963                 pd->port_rcvegrbuf = NULL;
1964                 kfree(pd->port_rcvegrbuf_phys);
1965                 pd->port_rcvegrbuf_phys = NULL;
1966                 pd->port_rcvegrbuf_chunks = 0;
1967         } else if (pd->port_port == 0 && dd->ipath_port0_skbinfo) {
1968                 unsigned e;
1969                 struct ipath_skbinfo *skbinfo = dd->ipath_port0_skbinfo;
1970
1971                 dd->ipath_port0_skbinfo = NULL;
1972                 ipath_cdbg(VERBOSE, "free closed port %d "
1973                            "ipath_port0_skbinfo @ %p\n", pd->port_port,
1974                            skbinfo);
1975                 for (e = 0; e < dd->ipath_rcvegrcnt; e++)
1976                 if (skbinfo[e].skb) {
1977                         pci_unmap_single(dd->pcidev, skbinfo[e].phys,
1978                                          dd->ipath_ibmaxlen,
1979                                          PCI_DMA_FROMDEVICE);
1980                         dev_kfree_skb(skbinfo[e].skb);
1981                 }
1982                 vfree(skbinfo);
1983         }
1984         kfree(pd->port_tid_pg_list);
1985         vfree(pd->subport_uregbase);
1986         vfree(pd->subport_rcvegrbuf);
1987         vfree(pd->subport_rcvhdr_base);
1988         kfree(pd);
1989 }
1990
1991 static int __init infinipath_init(void)
1992 {
1993         int ret;
1994
1995         if (ipath_debug & __IPATH_DBG)
1996                 printk(KERN_INFO DRIVER_LOAD_MSG "%s", ib_ipath_version);
1997
1998         /*
1999          * These must be called before the driver is registered with
2000          * the PCI subsystem.
2001          */
2002         idr_init(&unit_table);
2003         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
2004                 ret = -ENOMEM;
2005                 goto bail;
2006         }
2007
2008         ret = pci_register_driver(&ipath_driver);
2009         if (ret < 0) {
2010                 printk(KERN_ERR IPATH_DRV_NAME
2011                        ": Unable to register driver: error %d\n", -ret);
2012                 goto bail_unit;
2013         }
2014
2015         ret = ipath_driver_create_group(&ipath_driver.driver);
2016         if (ret < 0) {
2017                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create driver "
2018                        "sysfs entries: error %d\n", -ret);
2019                 goto bail_pci;
2020         }
2021
2022         ret = ipath_init_ipathfs();
2023         if (ret < 0) {
2024                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
2025                        "ipathfs: error %d\n", -ret);
2026                 goto bail_group;
2027         }
2028
2029         goto bail;
2030
2031 bail_group:
2032         ipath_driver_remove_group(&ipath_driver.driver);
2033
2034 bail_pci:
2035         pci_unregister_driver(&ipath_driver);
2036
2037 bail_unit:
2038         idr_destroy(&unit_table);
2039
2040 bail:
2041         return ret;
2042 }
2043
2044 static void __exit infinipath_cleanup(void)
2045 {
2046         ipath_exit_ipathfs();
2047
2048         ipath_driver_remove_group(&ipath_driver.driver);
2049
2050         ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
2051         pci_unregister_driver(&ipath_driver);
2052
2053         idr_destroy(&unit_table);
2054 }
2055
2056 /**
2057  * ipath_reset_device - reset the chip if possible
2058  * @unit: the device to reset
2059  *
2060  * Whether or not reset is successful, we attempt to re-initialize the chip
2061  * (that is, much like a driver unload/reload).  We clear the INITTED flag
2062  * so that the various entry points will fail until we reinitialize.  For
2063  * now, we only allow this if no user ports are open that use chip resources
2064  */
2065 int ipath_reset_device(int unit)
2066 {
2067         int ret, i;
2068         struct ipath_devdata *dd = ipath_lookup(unit);
2069
2070         if (!dd) {
2071                 ret = -ENODEV;
2072                 goto bail;
2073         }
2074
2075         dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
2076
2077         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
2078                 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
2079                          "not initialized or not present\n", unit);
2080                 ret = -ENXIO;
2081                 goto bail;
2082         }
2083
2084         if (dd->ipath_pd)
2085                 for (i = 1; i < dd->ipath_cfgports; i++) {
2086                         if (dd->ipath_pd[i] && dd->ipath_pd[i]->port_cnt) {
2087                                 ipath_dbg("unit %u port %d is in use "
2088                                           "(PID %u cmd %s), can't reset\n",
2089                                           unit, i,
2090                                           dd->ipath_pd[i]->port_pid,
2091                                           dd->ipath_pd[i]->port_comm);
2092                                 ret = -EBUSY;
2093                                 goto bail;
2094                         }
2095                 }
2096
2097         dd->ipath_flags &= ~IPATH_INITTED;
2098         ret = dd->ipath_f_reset(dd);
2099         if (ret != 1)
2100                 ipath_dbg("reset was not successful\n");
2101         ipath_dbg("Trying to reinitialize unit %u after reset attempt\n",
2102                   unit);
2103         ret = ipath_init_chip(dd, 1);
2104         if (ret)
2105                 ipath_dev_err(dd, "Reinitialize unit %u after "
2106                               "reset failed with %d\n", unit, ret);
2107         else
2108                 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
2109                          "resetting\n", unit);
2110
2111 bail:
2112         return ret;
2113 }
2114
2115 int ipath_set_rx_pol_inv(struct ipath_devdata *dd, u8 new_pol_inv)
2116 {
2117         u64 val;
2118         if ( new_pol_inv > INFINIPATH_XGXS_RX_POL_MASK ) {
2119                 return -1;
2120         }
2121         if ( dd->ipath_rx_pol_inv != new_pol_inv ) {
2122                 dd->ipath_rx_pol_inv = new_pol_inv;
2123                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig);
2124                 val &= ~(INFINIPATH_XGXS_RX_POL_MASK <<
2125                          INFINIPATH_XGXS_RX_POL_SHIFT);
2126                 val |= ((u64)dd->ipath_rx_pol_inv) <<
2127                         INFINIPATH_XGXS_RX_POL_SHIFT;
2128                 ipath_write_kreg(dd, dd->ipath_kregs->kr_xgxsconfig, val);
2129         }
2130         return 0;
2131 }
2132 module_init(infinipath_init);
2133 module_exit(infinipath_cleanup);