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_layer.h"
  43 #include "ipath_verbs.h"
  44 #include "ipath_common.h"
  45
  46 static void ipath_update_pio_bufs(struct ipath_devdata *);
  47
  48 const char *ipath_get_unit_name(int unit)
  49 {
  50         static char iname[16];
  51         snprintf(iname, sizeof iname, "infinipath%u", unit);
  52         return iname;
  53 }
  54
  55 #define DRIVER_LOAD_MSG "QLogic " IPATH_DRV_NAME " loaded: "
  56 #define PFX IPATH_DRV_NAME ": "
  57
  58 /*
  59  * The size has to be longer than this string, so we can append
  60  * board/chip information to it in the init code.
  61  */
  62 const char ipath_core_version[] = IPATH_IDSTR "\n";
  63
  64 static struct idr unit_table;
  65 DEFINE_SPINLOCK(ipath_devs_lock);
  66 LIST_HEAD(ipath_dev_list);
  67
  68 wait_queue_head_t ipath_sma_state_wait;
  69
  70 unsigned ipath_debug = __IPATH_INFO;
  71
  72 module_param_named(debug, ipath_debug, uint, S_IWUSR | S_IRUGO);
  73 MODULE_PARM_DESC(debug, "mask for debug prints");
  74 EXPORT_SYMBOL_GPL(ipath_debug);
  75
  76 MODULE_LICENSE("GPL");
  77 MODULE_AUTHOR("QLogic <support@pathscale.com>");
  78 MODULE_DESCRIPTION("QLogic InfiniPath driver");
  79
  80 const char *ipath_ibcstatus_str[] = {
  81         "Disabled",
  82         "LinkUp",
  83         "PollActive",
  84         "PollQuiet",
  85         "SleepDelay",
  86         "SleepQuiet",
  87         "LState6",              /* unused */
  88         "LState7",              /* unused */
  89         "CfgDebounce",
  90         "CfgRcvfCfg",
  91         "CfgWaitRmt",
  92         "CfgIdle",
  93         "RecovRetrain",
  94         "LState0xD",            /* unused */
  95         "RecovWaitRmt",
  96         "RecovIdle",
  97 };
  98
  99 /*
 100  * These variables are initialized in the chip-specific files
 101  * but are defined here.
 102  */
 103 u16 ipath_gpio_sda_num, ipath_gpio_scl_num;
 104 u64 ipath_gpio_sda, ipath_gpio_scl;
 105 u64 infinipath_i_bitsextant;
 106 ipath_err_t infinipath_e_bitsextant, infinipath_hwe_bitsextant;
 107 u32 infinipath_i_rcvavail_mask, infinipath_i_rcvurg_mask;
 108
 109 static void __devexit ipath_remove_one(struct pci_dev *);
 110 static int __devinit ipath_init_one(struct pci_dev *,
 111                                     const struct pci_device_id *);
 112
 113 /* Only needed for registration, nothing else needs this info */
 114 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
 115 #define PCI_DEVICE_ID_INFINIPATH_HT 0xd
 116 #define PCI_DEVICE_ID_INFINIPATH_PE800 0x10
 117
 118 static const struct pci_device_id ipath_pci_tbl[] = {
 119         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_HT) },
 120         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_PE800) },
 121         { 0, }
 122 };
 123
 124 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
 125
 126 static struct pci_driver ipath_driver = {
 127         .name = IPATH_DRV_NAME,
 128         .probe = ipath_init_one,
 129         .remove = __devexit_p(ipath_remove_one),
 130         .id_table = ipath_pci_tbl,
 131 };
 132
 133
 134 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
 135                              u32 *bar0, u32 *bar1)
 136 {
 137         int ret;
 138
 139         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
 140         if (ret)
 141                 ipath_dev_err(dd, "failed to read bar0 before enable: "
 142                               "error %d\n", -ret);
 143
 144         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
 145         if (ret)
 146                 ipath_dev_err(dd, "failed to read bar1 before enable: "
 147                               "error %d\n", -ret);
 148
 149         ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
 150 }
 151
 152 static void ipath_free_devdata(struct pci_dev *pdev,
 153                                struct ipath_devdata *dd)
 154 {
 155         unsigned long flags;
 156
 157         pci_set_drvdata(pdev, NULL);
 158
 159         if (dd->ipath_unit != -1) {
 160                 spin_lock_irqsave(&ipath_devs_lock, flags);
 161                 idr_remove(&unit_table, dd->ipath_unit);
 162                 list_del(&dd->ipath_list);
 163                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
 164         }
 165         vfree(dd);
 166 }
 167
 168 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
 169 {
 170         unsigned long flags;
 171         struct ipath_devdata *dd;
 172         int ret;
 173
 174         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
 175                 dd = ERR_PTR(-ENOMEM);
 176                 goto bail;
 177         }
 178
 179         dd = vmalloc(sizeof(*dd));
 180         if (!dd) {
 181                 dd = ERR_PTR(-ENOMEM);
 182                 goto bail;
 183         }
 184         memset(dd, 0, sizeof(*dd));
 185         dd->ipath_unit = -1;
 186
 187         spin_lock_irqsave(&ipath_devs_lock, flags);
 188
 189         ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
 190         if (ret < 0) {
 191                 printk(KERN_ERR IPATH_DRV_NAME
 192                        ": Could not allocate unit ID: error %d\n", -ret);
 193                 ipath_free_devdata(pdev, dd);
 194                 dd = ERR_PTR(ret);
 195                 goto bail_unlock;
 196         }
 197
 198         dd->pcidev = pdev;
 199         pci_set_drvdata(pdev, dd);
 200
 201         list_add(&dd->ipath_list, &ipath_dev_list);
 202
 203 bail_unlock:
 204         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 205
 206 bail:
 207         return dd;
 208 }
 209
 210 static inline struct ipath_devdata *__ipath_lookup(int unit)
 211 {
 212         return idr_find(&unit_table, unit);
 213 }
 214
 215 struct ipath_devdata *ipath_lookup(int unit)
 216 {
 217         struct ipath_devdata *dd;
 218         unsigned long flags;
 219
 220         spin_lock_irqsave(&ipath_devs_lock, flags);
 221         dd = __ipath_lookup(unit);
 222         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 223
 224         return dd;
 225 }
 226
 227 int ipath_count_units(int *npresentp, int *nupp, u32 *maxportsp)
 228 {
 229         int nunits, npresent, nup;
 230         struct ipath_devdata *dd;
 231         unsigned long flags;
 232         u32 maxports;
 233
 234         nunits = npresent = nup = maxports = 0;
 235
 236         spin_lock_irqsave(&ipath_devs_lock, flags);
 237
 238         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 239                 nunits++;
 240                 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
 241                         npresent++;
 242                 if (dd->ipath_lid &&
 243                     !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
 244                                          | IPATH_LINKUNK)))
 245                         nup++;
 246                 if (dd->ipath_cfgports > maxports)
 247                         maxports = dd->ipath_cfgports;
 248         }
 249
 250         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 251
 252         if (npresentp)
 253                 *npresentp = npresent;
 254         if (nupp)
 255                 *nupp = nup;
 256         if (maxportsp)
 257                 *maxportsp = maxports;
 258
 259         return nunits;
 260 }
 261
 262 /*
 263  * These next two routines are placeholders in case we don't have per-arch
 264  * code for controlling write combining.  If explicit control of write
 265  * combining is not available, performance will probably be awful.
 266  */
 267
 268 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
 269 {
 270         return -EOPNOTSUPP;
 271 }
 272
 273 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
 274 {
 275 }
 276
 277 static int __devinit ipath_init_one(struct pci_dev *pdev,
 278                                     const struct pci_device_id *ent)
 279 {
 280         int ret, len, j;
 281         struct ipath_devdata *dd;
 282         unsigned long long addr;
 283         u32 bar0 = 0, bar1 = 0;
 284         u8 rev;
 285
 286         dd = ipath_alloc_devdata(pdev);
 287         if (IS_ERR(dd)) {
 288                 ret = PTR_ERR(dd);
 289                 printk(KERN_ERR IPATH_DRV_NAME
 290                        ": Could not allocate devdata: error %d\n", -ret);
 291                 goto bail;
 292         }
 293
 294         ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
 295
 296         read_bars(dd, pdev, &bar0, &bar1);
 297
 298         ret = pci_enable_device(pdev);
 299         if (ret) {
 300                 /* This can happen iff:
 301                  *
 302                  * We did a chip reset, and then failed to reprogram the
 303                  * BAR, or the chip reset due to an internal error.  We then
 304                  * unloaded the driver and reloaded it.
 305                  *
 306                  * Both reset cases set the BAR back to initial state.  For
 307                  * the latter case, the AER sticky error bit at offset 0x718
 308                  * should be set, but the Linux kernel doesn't yet know
 309                  * about that, it appears.  If the original BAR was retained
 310                  * in the kernel data structures, this may be OK.
 311                  */
 312                 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
 313                               dd->ipath_unit, -ret);
 314                 goto bail_devdata;
 315         }
 316         addr = pci_resource_start(pdev, 0);
 317         len = pci_resource_len(pdev, 0);
 318         ipath_cdbg(VERBOSE, "regbase (0) %llx len %d irq %x, vend %x/%x "
 319                    "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
 320                    ent->device, ent->driver_data);
 321
 322         read_bars(dd, pdev, &bar0, &bar1);
 323
 324         if (!bar1 && !(bar0 & ~0xf)) {
 325                 if (addr) {
 326                         dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
 327                                  "rewriting as %llx\n", addr);
 328                         ret = pci_write_config_dword(
 329                                 pdev, PCI_BASE_ADDRESS_0, addr);
 330                         if (ret) {
 331                                 ipath_dev_err(dd, "rewrite of BAR0 "
 332                                               "failed: err %d\n", -ret);
 333                                 goto bail_disable;
 334                         }
 335                         ret = pci_write_config_dword(
 336                                 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
 337                         if (ret) {
 338                                 ipath_dev_err(dd, "rewrite of BAR1 "
 339                                               "failed: err %d\n", -ret);
 340                                 goto bail_disable;
 341                         }
 342                 } else {
 343                         ipath_dev_err(dd, "BAR is 0 (probable RESET), "
 344                                       "not usable until reboot\n");
 345                         ret = -ENODEV;
 346                         goto bail_disable;
 347                 }
 348         }
 349
 350         ret = pci_request_regions(pdev, IPATH_DRV_NAME);
 351         if (ret) {
 352                 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
 353                          "err %d\n", dd->ipath_unit, -ret);
 354                 goto bail_disable;
 355         }
 356
 357         ret = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
 358         if (ret) {
 359                 /*
 360                  * if the 64 bit setup fails, try 32 bit.  Some systems
 361                  * do not setup 64 bit maps on systems with 2GB or less
 362                  * memory installed.
 363                  */
 364                 ret = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
 365                 if (ret) {
 366                         dev_info(&pdev->dev,
 367                                 "Unable to set DMA mask for unit %u: %d\n",
 368                                 dd->ipath_unit, ret);
 369                         goto bail_regions;
 370                 }
 371                 else {
 372                         ipath_dbg("No 64bit DMA mask, used 32 bit mask\n");
 373                         ret = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
 374                         if (ret)
 375                                 dev_info(&pdev->dev,
 376                                         "Unable to set DMA consistent mask "
 377                                         "for unit %u: %d\n",
 378                                         dd->ipath_unit, ret);
 379
 380                 }
 381         }
 382         else {
 383                 ret = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
 384                 if (ret)
 385                         dev_info(&pdev->dev,
 386                                 "Unable to set DMA consistent mask "
 387                                 "for unit %u: %d\n",
 388                                 dd->ipath_unit, ret);
 389         }
 390
 391         pci_set_master(pdev);
 392
 393         /*
 394          * Save BARs to rewrite after device reset.  Save all 64 bits of
 395          * BAR, just in case.
 396          */
 397         dd->ipath_pcibar0 = addr;
 398         dd->ipath_pcibar1 = addr >> 32;
 399         dd->ipath_deviceid = ent->device;       /* save for later use */
 400         dd->ipath_vendorid = ent->vendor;
 401
 402         /* setup the chip-specific functions, as early as possible. */
 403         switch (ent->device) {
 404         case PCI_DEVICE_ID_INFINIPATH_HT:
 405                 ipath_init_ht400_funcs(dd);
 406                 break;
 407         case PCI_DEVICE_ID_INFINIPATH_PE800:
 408                 ipath_init_pe800_funcs(dd);
 409                 break;
 410         default:
 411                 ipath_dev_err(dd, "Found unknown QLogic deviceid 0x%x, "
 412                               "failing\n", ent->device);
 413                 return -ENODEV;
 414         }
 415
 416         for (j = 0; j < 6; j++) {
 417                 if (!pdev->resource[j].start)
 418                         continue;
 419                 ipath_cdbg(VERBOSE, "BAR %d start %llx, end %llx, len %llx\n",
 420                            j, (unsigned long long)pdev->resource[j].start,
 421                            (unsigned long long)pdev->resource[j].end,
 422                            (unsigned long long)pci_resource_len(pdev, j));
 423         }
 424
 425         if (!addr) {
 426                 ipath_dev_err(dd, "No valid address in BAR 0!\n");
 427                 ret = -ENODEV;
 428                 goto bail_regions;
 429         }
 430
 431         dd->ipath_deviceid = ent->device;       /* save for later use */
 432         dd->ipath_vendorid = ent->vendor;
 433
 434         ret = pci_read_config_byte(pdev, PCI_REVISION_ID, &rev);
 435         if (ret) {
 436                 ipath_dev_err(dd, "Failed to read PCI revision ID unit "
 437                               "%u: err %d\n", dd->ipath_unit, -ret);
 438                 goto bail_regions;      /* shouldn't ever happen */
 439         }
 440         dd->ipath_pcirev = rev;
 441
 442 #if defined(__powerpc__)
 443         /* There isn't a generic way to specify writethrough mappings */
 444         dd->ipath_kregbase = __ioremap(addr, len,
 445                 (_PAGE_NO_CACHE|_PAGE_WRITETHRU));
 446 #else
 447         dd->ipath_kregbase = ioremap_nocache(addr, len);
 448 #endif
 449
 450         if (!dd->ipath_kregbase) {
 451                 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
 452                           addr);
 453                 ret = -ENOMEM;
 454                 goto bail_iounmap;
 455         }
 456         dd->ipath_kregend = (u64 __iomem *)
 457                 ((void __iomem *)dd->ipath_kregbase + len);
 458         dd->ipath_physaddr = addr;      /* used for io_remap, etc. */
 459         /* for user mmap */
 460         ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p\n",
 461                    addr, dd->ipath_kregbase);
 462
 463         /*
 464          * clear ipath_flags here instead of in ipath_init_chip as it is set
 465          * by ipath_setup_htconfig.
 466          */
 467         dd->ipath_flags = 0;
 468         dd->ipath_lli_counter = 0;
 469         dd->ipath_lli_errors = 0;
 470
 471         if (dd->ipath_f_bus(dd, pdev))
 472                 ipath_dev_err(dd, "Failed to setup config space; "
 473                               "continuing anyway\n");
 474
 475         /*
 476          * set up our interrupt handler; IRQF_SHARED probably not needed,
 477          * since MSI interrupts shouldn't be shared but won't  hurt for now.
 478          * check 0 irq after we return from chip-specific bus setup, since
 479          * that can affect this due to setup
 480          */
 481         if (!pdev->irq)
 482                 ipath_dev_err(dd, "irq is 0, BIOS error?  Interrupts won't "
 483                               "work\n");
 484         else {
 485                 ret = request_irq(pdev->irq, ipath_intr, IRQF_SHARED,
 486                                   IPATH_DRV_NAME, dd);
 487                 if (ret) {
 488                         ipath_dev_err(dd, "Couldn't setup irq handler, "
 489                                       "irq=%u: %d\n", pdev->irq, ret);
 490                         goto bail_iounmap;
 491                 }
 492         }
 493
 494         ret = ipath_init_chip(dd, 0);   /* do the chip-specific init */
 495         if (ret)
 496                 goto bail_iounmap;
 497
 498         ret = ipath_enable_wc(dd);
 499
 500         if (ret) {
 501                 ipath_dev_err(dd, "Write combining not enabled "
 502                               "(err %d): performance may be poor\n",
 503                               -ret);
 504                 ret = 0;
 505         }
 506
 507         ipath_device_create_group(&pdev->dev, dd);
 508         ipathfs_add_device(dd);
 509         ipath_user_add(dd);
 510         ipath_diag_add(dd);
 511         ipath_layer_add(dd);
 512         ipath_register_ib_device(dd);
 513
 514         goto bail;
 515
 516 bail_iounmap:
 517         iounmap((volatile void __iomem *) dd->ipath_kregbase);
 518
 519 bail_regions:
 520         pci_release_regions(pdev);
 521
 522 bail_disable:
 523         pci_disable_device(pdev);
 524
 525 bail_devdata:
 526         ipath_free_devdata(pdev, dd);
 527
 528 bail:
 529         return ret;
 530 }
 531
 532 static void __devexit ipath_remove_one(struct pci_dev *pdev)
 533 {
 534         struct ipath_devdata *dd;
 535
 536         ipath_cdbg(VERBOSE, "removing, pdev=%p\n", pdev);
 537         if (!pdev)
 538                 return;
 539
 540         dd = pci_get_drvdata(pdev);
 541         ipath_unregister_ib_device(dd->verbs_dev);
 542         ipath_layer_remove(dd);
 543         ipath_diag_remove(dd);
 544         ipath_user_remove(dd);
 545         ipathfs_remove_device(dd);
 546         ipath_device_remove_group(&pdev->dev, dd);
 547         ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
 548                    "unit %u\n", dd, (u32) dd->ipath_unit);
 549         if (dd->ipath_kregbase) {
 550                 ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n",
 551                            dd->ipath_kregbase);
 552                 iounmap((volatile void __iomem *) dd->ipath_kregbase);
 553                 dd->ipath_kregbase = NULL;
 554         }
 555         pci_release_regions(pdev);
 556         ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
 557         pci_disable_device(pdev);
 558
 559         ipath_free_devdata(pdev, dd);
 560 }
 561
 562 /* general driver use */
 563 DEFINE_MUTEX(ipath_mutex);
 564
 565 static DEFINE_SPINLOCK(ipath_pioavail_lock);
 566
 567 /**
 568  * ipath_disarm_piobufs - cancel a range of PIO buffers
 569  * @dd: the infinipath device
 570  * @first: the first PIO buffer to cancel
 571  * @cnt: the number of PIO buffers to cancel
 572  *
 573  * cancel a range of PIO buffers, used when they might be armed, but
 574  * not triggered.  Used at init to ensure buffer state, and also user
 575  * process close, in case it died while writing to a PIO buffer
 576  * Also after errors.
 577  */
 578 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
 579                           unsigned cnt)
 580 {
 581         unsigned i, last = first + cnt;
 582         u64 sendctrl, sendorig;
 583
 584         ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
 585         sendorig = dd->ipath_sendctrl | INFINIPATH_S_DISARM;
 586         for (i = first; i < last; i++) {
 587                 sendctrl = sendorig |
 588                         (i << INFINIPATH_S_DISARMPIOBUF_SHIFT);
 589                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 590                                  sendctrl);
 591         }
 592
 593         /*
 594          * Write it again with current value, in case ipath_sendctrl changed
 595          * while we were looping; no critical bits that would require
 596          * locking.
 597          *
 598          * Write a 0, and then the original value, reading scratch in
 599          * between.  This seems to avoid a chip timing race that causes
 600          * pioavail updates to memory to stop.
 601          */
 602         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 603                          0);
 604         sendorig = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 605         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 606                          dd->ipath_sendctrl);
 607 }
 608
 609 /**
 610  * ipath_wait_linkstate - wait for an IB link state change to occur
 611  * @dd: the infinipath device
 612  * @state: the state to wait for
 613  * @msecs: the number of milliseconds to wait
 614  *
 615  * wait up to msecs milliseconds for IB link state change to occur for
 616  * now, take the easy polling route.  Currently used only by
 617  * ipath_layer_set_linkstate.  Returns 0 if state reached, otherwise
 618  * -ETIMEDOUT state can have multiple states set, for any of several
 619  * transitions.
 620  */
 621 int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state, int msecs)
 622 {
 623         dd->ipath_sma_state_wanted = state;
 624         wait_event_interruptible_timeout(ipath_sma_state_wait,
 625                                          (dd->ipath_flags & state),
 626                                          msecs_to_jiffies(msecs));
 627         dd->ipath_sma_state_wanted = 0;
 628
 629         if (!(dd->ipath_flags & state)) {
 630                 u64 val;
 631                 ipath_cdbg(SMA, "Didn't reach linkstate %s within %u ms\n",
 632                            /* test INIT ahead of DOWN, both can be set */
 633                            (state & IPATH_LINKINIT) ? "INIT" :
 634                            ((state & IPATH_LINKDOWN) ? "DOWN" :
 635                             ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
 636                            msecs);
 637                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 638                 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
 639                            (unsigned long long) ipath_read_kreg64(
 640                                    dd, dd->ipath_kregs->kr_ibcctrl),
 641                            (unsigned long long) val,
 642                            ipath_ibcstatus_str[val & 0xf]);
 643         }
 644         return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
 645 }
 646
 647 void ipath_decode_err(char *buf, size_t blen, ipath_err_t err)
 648 {
 649         *buf = '\0';
 650         if (err & INFINIPATH_E_RHDRLEN)
 651                 strlcat(buf, "rhdrlen ", blen);
 652         if (err & INFINIPATH_E_RBADTID)
 653                 strlcat(buf, "rbadtid ", blen);
 654         if (err & INFINIPATH_E_RBADVERSION)
 655                 strlcat(buf, "rbadversion ", blen);
 656         if (err & INFINIPATH_E_RHDR)
 657                 strlcat(buf, "rhdr ", blen);
 658         if (err & INFINIPATH_E_RLONGPKTLEN)
 659                 strlcat(buf, "rlongpktlen ", blen);
 660         if (err & INFINIPATH_E_RSHORTPKTLEN)
 661                 strlcat(buf, "rshortpktlen ", blen);
 662         if (err & INFINIPATH_E_RMAXPKTLEN)
 663                 strlcat(buf, "rmaxpktlen ", blen);
 664         if (err & INFINIPATH_E_RMINPKTLEN)
 665                 strlcat(buf, "rminpktlen ", blen);
 666         if (err & INFINIPATH_E_RFORMATERR)
 667                 strlcat(buf, "rformaterr ", blen);
 668         if (err & INFINIPATH_E_RUNSUPVL)
 669                 strlcat(buf, "runsupvl ", blen);
 670         if (err & INFINIPATH_E_RUNEXPCHAR)
 671                 strlcat(buf, "runexpchar ", blen);
 672         if (err & INFINIPATH_E_RIBFLOW)
 673                 strlcat(buf, "ribflow ", blen);
 674         if (err & INFINIPATH_E_REBP)
 675                 strlcat(buf, "EBP ", blen);
 676         if (err & INFINIPATH_E_SUNDERRUN)
 677                 strlcat(buf, "sunderrun ", blen);
 678         if (err & INFINIPATH_E_SPIOARMLAUNCH)
 679                 strlcat(buf, "spioarmlaunch ", blen);
 680         if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
 681                 strlcat(buf, "sunexperrpktnum ", blen);
 682         if (err & INFINIPATH_E_SDROPPEDDATAPKT)
 683                 strlcat(buf, "sdroppeddatapkt ", blen);
 684         if (err & INFINIPATH_E_SDROPPEDSMPPKT)
 685                 strlcat(buf, "sdroppedsmppkt ", blen);
 686         if (err & INFINIPATH_E_SMAXPKTLEN)
 687                 strlcat(buf, "smaxpktlen ", blen);
 688         if (err & INFINIPATH_E_SMINPKTLEN)
 689                 strlcat(buf, "sminpktlen ", blen);
 690         if (err & INFINIPATH_E_SUNSUPVL)
 691                 strlcat(buf, "sunsupVL ", blen);
 692         if (err & INFINIPATH_E_SPKTLEN)
 693                 strlcat(buf, "spktlen ", blen);
 694         if (err & INFINIPATH_E_INVALIDADDR)
 695                 strlcat(buf, "invalidaddr ", blen);
 696         if (err & INFINIPATH_E_RICRC)
 697                 strlcat(buf, "CRC ", blen);
 698         if (err & INFINIPATH_E_RVCRC)
 699                 strlcat(buf, "VCRC ", blen);
 700         if (err & INFINIPATH_E_RRCVEGRFULL)
 701                 strlcat(buf, "rcvegrfull ", blen);
 702         if (err & INFINIPATH_E_RRCVHDRFULL)
 703                 strlcat(buf, "rcvhdrfull ", blen);
 704         if (err & INFINIPATH_E_IBSTATUSCHANGED)
 705                 strlcat(buf, "ibcstatuschg ", blen);
 706         if (err & INFINIPATH_E_RIBLOSTLINK)
 707                 strlcat(buf, "riblostlink ", blen);
 708         if (err & INFINIPATH_E_HARDWARE)
 709                 strlcat(buf, "hardware ", blen);
 710         if (err & INFINIPATH_E_RESET)
 711                 strlcat(buf, "reset ", blen);
 712 }
 713
 714 /**
 715  * get_rhf_errstring - decode RHF errors
 716  * @err: the err number
 717  * @msg: the output buffer
 718  * @len: the length of the output buffer
 719  *
 720  * only used one place now, may want more later
 721  */
 722 static void get_rhf_errstring(u32 err, char *msg, size_t len)
 723 {
 724         /* if no errors, and so don't need to check what's first */
 725         *msg = '\0';
 726
 727         if (err & INFINIPATH_RHF_H_ICRCERR)
 728                 strlcat(msg, "icrcerr ", len);
 729         if (err & INFINIPATH_RHF_H_VCRCERR)
 730                 strlcat(msg, "vcrcerr ", len);
 731         if (err & INFINIPATH_RHF_H_PARITYERR)
 732                 strlcat(msg, "parityerr ", len);
 733         if (err & INFINIPATH_RHF_H_LENERR)
 734                 strlcat(msg, "lenerr ", len);
 735         if (err & INFINIPATH_RHF_H_MTUERR)
 736                 strlcat(msg, "mtuerr ", len);
 737         if (err & INFINIPATH_RHF_H_IHDRERR)
 738                 /* infinipath hdr checksum error */
 739                 strlcat(msg, "ipathhdrerr ", len);
 740         if (err & INFINIPATH_RHF_H_TIDERR)
 741                 strlcat(msg, "tiderr ", len);
 742         if (err & INFINIPATH_RHF_H_MKERR)
 743                 /* bad port, offset, etc. */
 744                 strlcat(msg, "invalid ipathhdr ", len);
 745         if (err & INFINIPATH_RHF_H_IBERR)
 746                 strlcat(msg, "iberr ", len);
 747         if (err & INFINIPATH_RHF_L_SWA)
 748                 strlcat(msg, "swA ", len);
 749         if (err & INFINIPATH_RHF_L_SWB)
 750                 strlcat(msg, "swB ", len);
 751 }
 752
 753 /**
 754  * ipath_get_egrbuf - get an eager buffer
 755  * @dd: the infinipath device
 756  * @bufnum: the eager buffer to get
 757  * @err: unused
 758  *
 759  * must only be called if ipath_pd[port] is known to be allocated
 760  */
 761 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum,
 762                                      int err)
 763 {
 764         return dd->ipath_port0_skbs ?
 765                 (void *)dd->ipath_port0_skbs[bufnum]->data : NULL;
 766 }
 767
 768 /**
 769  * ipath_alloc_skb - allocate an skb and buffer with possible constraints
 770  * @dd: the infinipath device
 771  * @gfp_mask: the sk_buff SFP mask
 772  */
 773 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
 774                                 gfp_t gfp_mask)
 775 {
 776         struct sk_buff *skb;
 777         u32 len;
 778
 779         /*
 780          * Only fully supported way to handle this is to allocate lots
 781          * extra, align as needed, and then do skb_reserve().  That wastes
 782          * a lot of memory...  I'll have to hack this into infinipath_copy
 783          * also.
 784          */
 785
 786         /*
 787          * We need 4 extra bytes for unaligned transfer copying
 788          */
 789         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 790                 /* we need a 4KB multiple alignment, and there is no way
 791                  * to do it except to allocate extra and then skb_reserve
 792                  * enough to bring it up to the right alignment.
 793                  */
 794                 len = dd->ipath_ibmaxlen + 4 + (1 << 11) - 1;
 795         }
 796         else
 797                 len = dd->ipath_ibmaxlen + 4;
 798         skb = __dev_alloc_skb(len, gfp_mask);
 799         if (!skb) {
 800                 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
 801                               len);
 802                 goto bail;
 803         }
 804         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 805                 u32 una = ((1 << 11) - 1) & (unsigned long)(skb->data + 4);
 806                 if (una)
 807                         skb_reserve(skb, 4 + (1 << 11) - una);
 808                 else
 809                         skb_reserve(skb, 4);
 810         } else
 811                 skb_reserve(skb, 4);
 812
 813 bail:
 814         return skb;
 815 }
 816
 817 /**
 818  * ipath_rcv_layer - receive a packet for the layered (ethernet) driver
 819  * @dd: the infinipath device
 820  * @etail: the sk_buff number
 821  * @tlen: the total packet length
 822  * @hdr: the ethernet header
 823  *
 824  * Separate routine for better overall optimization
 825  */
 826 static void ipath_rcv_layer(struct ipath_devdata *dd, u32 etail,
 827                             u32 tlen, struct ether_header *hdr)
 828 {
 829         u32 elen;
 830         u8 pad, *bthbytes;
 831         struct sk_buff *skb, *nskb;
 832
 833         if (dd->ipath_port0_skbs &&
 834                         hdr->sub_opcode == IPATH_ITH4X_OPCODE_ENCAP) {
 835                 /*
 836                  * Allocate a new sk_buff to replace the one we give
 837                  * to the network stack.
 838                  */
 839                 nskb = ipath_alloc_skb(dd, GFP_ATOMIC);
 840                 if (!nskb) {
 841                         /* count OK packets that we drop */
 842                         ipath_stats.sps_krdrops++;
 843                         return;
 844                 }
 845
 846                 bthbytes = (u8 *) hdr->bth;
 847                 pad = (bthbytes[1] >> 4) & 3;
 848                 /* +CRC32 */
 849                 elen = tlen - (sizeof(*hdr) + pad + sizeof(u32));
 850
 851                 skb = dd->ipath_port0_skbs[etail];
 852                 dd->ipath_port0_skbs[etail] = nskb;
 853                 skb_put(skb, elen);
 854
 855                 dd->ipath_f_put_tid(dd, etail + (u64 __iomem *)
 856                                     ((char __iomem *) dd->ipath_kregbase
 857                                      + dd->ipath_rcvegrbase), 0,
 858                                     virt_to_phys(nskb->data));
 859
 860                 __ipath_layer_rcv(dd, hdr, skb);
 861
 862                 /* another ether packet received */
 863                 ipath_stats.sps_ether_rpkts++;
 864         }
 865         else if (hdr->sub_opcode == IPATH_ITH4X_OPCODE_LID_ARP)
 866                 __ipath_layer_rcv_lid(dd, hdr);
 867 }
 868
 869 static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
 870                              u32 eflags,
 871                              u32 l,
 872                              u32 etail,
 873                              u64 *rc)
 874 {
 875         char emsg[128];
 876         struct ipath_message_header *hdr;
 877
 878         get_rhf_errstring(eflags, emsg, sizeof emsg);
 879         hdr = (struct ipath_message_header *)&rc[1];
 880         ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
 881                    "tlen=%x opcode=%x egridx=%x: %s\n",
 882                    eflags, l,
 883                    ipath_hdrget_rcv_type((__le32 *) rc),
 884                    ipath_hdrget_length_in_bytes((__le32 *) rc),
 885                    be32_to_cpu(hdr->bth[0]) >> 24,
 886                    etail, emsg);
 887
 888         /* Count local link integrity errors. */
 889         if (eflags & (INFINIPATH_RHF_H_ICRCERR | INFINIPATH_RHF_H_VCRCERR)) {
 890                 u8 n = (dd->ipath_ibcctrl >>
 891                         INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
 892                         INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
 893
 894                 if (++dd->ipath_lli_counter > n) {
 895                         dd->ipath_lli_counter = 0;
 896                         dd->ipath_lli_errors++;
 897                 }
 898         }
 899 }
 900
 901 /*
 902  * ipath_kreceive - receive a packet
 903  * @dd: the infinipath device
 904  *
 905  * called from interrupt handler for errors or receive interrupt
 906  */
 907 void ipath_kreceive(struct ipath_devdata *dd)
 908 {
 909         u64 *rc;
 910         void *ebuf;
 911         const u32 rsize = dd->ipath_rcvhdrentsize;      /* words */
 912         const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
 913         u32 etail = -1, l, hdrqtail;
 914         struct ipath_message_header *hdr;
 915         u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
 916         static u64 totcalls;    /* stats, may eventually remove */
 917
 918         if (!dd->ipath_hdrqtailptr) {
 919                 ipath_dev_err(dd,
 920                               "hdrqtailptr not set, can't do receives\n");
 921                 goto bail;
 922         }
 923
 924         /* There is already a thread processing this queue. */
 925         if (test_and_set_bit(0, &dd->ipath_rcv_pending))
 926                 goto bail;
 927
 928         l = dd->ipath_port0head;
 929         hdrqtail = (u32) le64_to_cpu(*dd->ipath_hdrqtailptr);
 930         if (l == hdrqtail)
 931                 goto done;
 932
 933 reloop:
 934         for (i = 0; l != hdrqtail; i++) {
 935                 u32 qp;
 936                 u8 *bthbytes;
 937
 938                 rc = (u64 *) (dd->ipath_pd[0]->port_rcvhdrq + (l << 2));
 939                 hdr = (struct ipath_message_header *)&rc[1];
 940                 /*
 941                  * could make a network order version of IPATH_KD_QP, and
 942                  * do the obvious shift before masking to speed this up.
 943                  */
 944                 qp = ntohl(hdr->bth[1]) & 0xffffff;
 945                 bthbytes = (u8 *) hdr->bth;
 946
 947                 eflags = ipath_hdrget_err_flags((__le32 *) rc);
 948                 etype = ipath_hdrget_rcv_type((__le32 *) rc);
 949                 /* total length */
 950                 tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
 951                 ebuf = NULL;
 952                 if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
 953                         /*
 954                          * it turns out that the chips uses an eager buffer
 955                          * for all non-expected packets, whether it "needs"
 956                          * one or not.  So always get the index, but don't
 957                          * set ebuf (so we try to copy data) unless the
 958                          * length requires it.
 959                          */
 960                         etail = ipath_hdrget_index((__le32 *) rc);
 961                         if (tlen > sizeof(*hdr) ||
 962                             etype == RCVHQ_RCV_TYPE_NON_KD)
 963                                 ebuf = ipath_get_egrbuf(dd, etail, 0);
 964                 }
 965
 966                 /*
 967                  * both tiderr and ipathhdrerr are set for all plain IB
 968                  * packets; only ipathhdrerr should be set.
 969                  */
 970
 971                 if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
 972                     RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
 973                             hdr->iph.ver_port_tid_offset) !=
 974                     IPS_PROTO_VERSION) {
 975                         ipath_cdbg(PKT, "Bad InfiniPath protocol version "
 976                                    "%x\n", etype);
 977                 }
 978
 979                 if (unlikely(eflags))
 980                         ipath_rcv_hdrerr(dd, eflags, l, etail, rc);
 981                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
 982                                 ipath_ib_rcv(dd->verbs_dev, rc + 1, ebuf,
 983                                              tlen);
 984                                 if (dd->ipath_lli_counter)
 985                                         dd->ipath_lli_counter--;
 986
 987                 } else if (etype == RCVHQ_RCV_TYPE_EAGER) {
 988                         if (qp == IPATH_KD_QP &&
 989                             bthbytes[0] == ipath_layer_rcv_opcode &&
 990                             ebuf)
 991                                 ipath_rcv_layer(dd, etail, tlen,
 992                                                 (struct ether_header *)hdr);
 993                         else
 994                                 ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
 995                                            "qp=%x), len %x; ignored\n",
 996                                            etype, bthbytes[0], qp, tlen);
 997                 }
 998                 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
 999                         ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
1000                                   be32_to_cpu(hdr->bth[0]) & 0xff);
1001                 else {
1002                         /*
1003                          * error packet, type of error  unknown.
1004                          * Probably type 3, but we don't know, so don't
1005                          * even try to print the opcode, etc.
1006                          */
1007                         ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
1008                                   "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
1009                                   "hdr %llx %llx %llx %llx %llx\n",
1010                                   etail, tlen, (unsigned long) rc, l,
1011                                   (unsigned long long) rc[0],
1012                                   (unsigned long long) rc[1],
1013                                   (unsigned long long) rc[2],
1014                                   (unsigned long long) rc[3],
1015                                   (unsigned long long) rc[4],
1016                                   (unsigned long long) rc[5]);
1017                 }
1018                 l += rsize;
1019                 if (l >= maxcnt)
1020                         l = 0;
1021                 if (etype != RCVHQ_RCV_TYPE_EXPECTED)
1022                     updegr = 1;
1023                 /*
1024                  * update head regs on last packet, and every 16 packets.
1025                  * Reduce bus traffic, while still trying to prevent
1026                  * rcvhdrq overflows, for when the queue is nearly full
1027                  */
1028                 if (l == hdrqtail || (i && !(i&0xf))) {
1029                         u64 lval;
1030                         if (l == hdrqtail) /* PE-800 interrupt only on last */
1031                                 lval = dd->ipath_rhdrhead_intr_off | l;
1032                         else
1033                                 lval = l;
1034                         (void)ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
1035                         if (updegr) {
1036                                 (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
1037                                                        etail, 0);
1038                                 updegr = 0;
1039                         }
1040                 }
1041         }
1042
1043         if (!dd->ipath_rhdrhead_intr_off && !reloop) {
1044                 /* HT-400 workaround; we can have a race clearing chip
1045                  * interrupt with another interrupt about to be delivered,
1046                  * and can clear it before it is delivered on the GPIO
1047                  * workaround.  By doing the extra check here for the
1048                  * in-memory tail register updating while we were doing
1049                  * earlier packets, we "almost" guarantee we have covered
1050                  * that case.
1051                  */
1052                 u32 hqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
1053                 if (hqtail != hdrqtail) {
1054                         hdrqtail = hqtail;
1055                         reloop = 1; /* loop 1 extra time at most */
1056                         goto reloop;
1057                 }
1058         }
1059
1060         pkttot += i;
1061
1062         dd->ipath_port0head = l;
1063
1064         if (pkttot > ipath_stats.sps_maxpkts_call)
1065                 ipath_stats.sps_maxpkts_call = pkttot;
1066         ipath_stats.sps_port0pkts += pkttot;
1067         ipath_stats.sps_avgpkts_call =
1068                 ipath_stats.sps_port0pkts / ++totcalls;
1069
1070 done:
1071         clear_bit(0, &dd->ipath_rcv_pending);
1072         smp_mb__after_clear_bit();
1073
1074 bail:;
1075 }
1076
1077 /**
1078  * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1079  * @dd: the infinipath device
1080  *
1081  * called whenever our local copy indicates we have run out of send buffers
1082  * NOTE: This can be called from interrupt context by some code
1083  * and from non-interrupt context by ipath_getpiobuf().
1084  */
1085
1086 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1087 {
1088         unsigned long flags;
1089         int i;
1090         const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1091
1092         /* If the generation (check) bits have changed, then we update the
1093          * busy bit for the corresponding PIO buffer.  This algorithm will
1094          * modify positions to the value they already have in some cases
1095          * (i.e., no change), but it's faster than changing only the bits
1096          * that have changed.
1097          *
1098          * We would like to do this atomicly, to avoid spinlocks in the
1099          * critical send path, but that's not really possible, given the
1100          * type of changes, and that this routine could be called on
1101          * multiple cpu's simultaneously, so we lock in this routine only,
1102          * to avoid conflicting updates; all we change is the shadow, and
1103          * it's a single 64 bit memory location, so by definition the update
1104          * is atomic in terms of what other cpu's can see in testing the
1105          * bits.  The spin_lock overhead isn't too bad, since it only
1106          * happens when all buffers are in use, so only cpu overhead, not
1107          * latency or bandwidth is affected.
1108          */
1109 #define _IPATH_ALL_CHECKBITS 0x5555555555555555ULL
1110         if (!dd->ipath_pioavailregs_dma) {
1111                 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1112                 return;
1113         }
1114         if (ipath_debug & __IPATH_VERBDBG) {
1115                 /* only if packet debug and verbose */
1116                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1117                 unsigned long *shadow = dd->ipath_pioavailshadow;
1118
1119                 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1120                            "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1121                            "s3=%lx\n",
1122                            (unsigned long long) le64_to_cpu(dma[0]),
1123                            shadow[0],
1124                            (unsigned long long) le64_to_cpu(dma[1]),
1125                            shadow[1],
1126                            (unsigned long long) le64_to_cpu(dma[2]),
1127                            shadow[2],
1128                            (unsigned long long) le64_to_cpu(dma[3]),
1129                            shadow[3]);
1130                 if (piobregs > 4)
1131                         ipath_cdbg(
1132                                 PKT, "2nd group, dma4=%llx shad4=%lx, "
1133                                 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1134                                 "d7=%llx s7=%lx\n",
1135                                 (unsigned long long) le64_to_cpu(dma[4]),
1136                                 shadow[4],
1137                                 (unsigned long long) le64_to_cpu(dma[5]),
1138                                 shadow[5],
1139                                 (unsigned long long) le64_to_cpu(dma[6]),
1140                                 shadow[6],
1141                                 (unsigned long long) le64_to_cpu(dma[7]),
1142                                 shadow[7]);
1143         }
1144         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1145         for (i = 0; i < piobregs; i++) {
1146                 u64 pchbusy, pchg, piov, pnew;
1147                 /*
1148                  * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1149                  */
1150                 if (i > 3) {
1151                         if (i & 1)
1152                                 piov = le64_to_cpu(
1153                                         dd->ipath_pioavailregs_dma[i - 1]);
1154                         else
1155                                 piov = le64_to_cpu(
1156                                         dd->ipath_pioavailregs_dma[i + 1]);
1157                 } else
1158                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1159                 pchg = _IPATH_ALL_CHECKBITS &
1160                         ~(dd->ipath_pioavailshadow[i] ^ piov);
1161                 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1162                 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1163                         pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1164                         pnew |= piov & pchbusy;
1165                         dd->ipath_pioavailshadow[i] = pnew;
1166                 }
1167         }
1168         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1169 }
1170
1171 /**
1172  * ipath_setrcvhdrsize - set the receive header size
1173  * @dd: the infinipath device
1174  * @rhdrsize: the receive header size
1175  *
1176  * called from user init code, and also layered driver init
1177  */
1178 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1179 {
1180         int ret = 0;
1181
1182         if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1183                 if (dd->ipath_rcvhdrsize != rhdrsize) {
1184                         dev_info(&dd->pcidev->dev,
1185                                  "Error: can't set protocol header "
1186                                  "size %u, already %u\n",
1187                                  rhdrsize, dd->ipath_rcvhdrsize);
1188                         ret = -EAGAIN;
1189                 } else
1190                         ipath_cdbg(VERBOSE, "Reuse same protocol header "
1191                                    "size %u\n", dd->ipath_rcvhdrsize);
1192         } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1193                                (sizeof(u64) / sizeof(u32)))) {
1194                 ipath_dbg("Error: can't set protocol header size %u "
1195                           "(> max %u)\n", rhdrsize,
1196                           dd->ipath_rcvhdrentsize -
1197                           (u32) (sizeof(u64) / sizeof(u32)));
1198                 ret = -EOVERFLOW;
1199         } else {
1200                 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1201                 dd->ipath_rcvhdrsize = rhdrsize;
1202                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1203                                  dd->ipath_rcvhdrsize);
1204                 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1205                            dd->ipath_rcvhdrsize);
1206         }
1207         return ret;
1208 }
1209
1210 /**
1211  * ipath_getpiobuf - find an available pio buffer
1212  * @dd: the infinipath device
1213  * @pbufnum: the buffer number is placed here
1214  *
1215  * do appropriate marking as busy, etc.
1216  * returns buffer number if one found (>=0), negative number is error.
1217  * Used by ipath_sma_send_pkt and ipath_layer_send
1218  */
1219 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
1220 {
1221         int i, j, starti, updated = 0;
1222         unsigned piobcnt, iter;
1223         unsigned long flags;
1224         unsigned long *shadow = dd->ipath_pioavailshadow;
1225         u32 __iomem *buf;
1226
1227         piobcnt = (unsigned)(dd->ipath_piobcnt2k
1228                              + dd->ipath_piobcnt4k);
1229         starti = dd->ipath_lastport_piobuf;
1230         iter = piobcnt - starti;
1231         if (dd->ipath_upd_pio_shadow) {
1232                 /*
1233                  * Minor optimization.  If we had no buffers on last call,
1234                  * start out by doing the update; continue and do scan even
1235                  * if no buffers were updated, to be paranoid
1236                  */
1237                 ipath_update_pio_bufs(dd);
1238                 /* we scanned here, don't do it at end of scan */
1239                 updated = 1;
1240                 i = starti;
1241         } else
1242                 i = dd->ipath_lastpioindex;
1243
1244 rescan:
1245         /*
1246          * while test_and_set_bit() is atomic, we do that and then the
1247          * change_bit(), and the pair is not.  See if this is the cause
1248          * of the remaining armlaunch errors.
1249          */
1250         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1251         for (j = 0; j < iter; j++, i++) {
1252                 if (i >= piobcnt)
1253                         i = starti;
1254                 /*
1255                  * To avoid bus lock overhead, we first find a candidate
1256                  * buffer, then do the test and set, and continue if that
1257                  * fails.
1258                  */
1259                 if (test_bit((2 * i) + 1, shadow) ||
1260                     test_and_set_bit((2 * i) + 1, shadow))
1261                         continue;
1262                 /* flip generation bit */
1263                 change_bit(2 * i, shadow);
1264                 break;
1265         }
1266         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1267
1268         if (j == iter) {
1269                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1270
1271                 /*
1272                  * first time through; shadow exhausted, but may be real
1273                  * buffers available, so go see; if any updated, rescan
1274                  * (once)
1275                  */
1276                 if (!updated) {
1277                         ipath_update_pio_bufs(dd);
1278                         updated = 1;
1279                         i = starti;
1280                         goto rescan;
1281                 }
1282                 dd->ipath_upd_pio_shadow = 1;
1283                 /*
1284                  * not atomic, but if we lose one once in a while, that's OK
1285                  */
1286                 ipath_stats.sps_nopiobufs++;
1287                 if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1288                         ipath_dbg(
1289                                 "%u pio sends with no bufavail; dmacopy: "
1290                                 "%llx %llx %llx %llx; shadow:  "
1291                                 "%lx %lx %lx %lx\n",
1292                                 dd->ipath_consec_nopiobuf,
1293                                 (unsigned long long) le64_to_cpu(dma[0]),
1294                                 (unsigned long long) le64_to_cpu(dma[1]),
1295                                 (unsigned long long) le64_to_cpu(dma[2]),
1296                                 (unsigned long long) le64_to_cpu(dma[3]),
1297                                 shadow[0], shadow[1], shadow[2],
1298                                 shadow[3]);
1299                         /*
1300                          * 4 buffers per byte, 4 registers above, cover rest
1301                          * below
1302                          */
1303                         if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1304                             (sizeof(shadow[0]) * 4 * 4))
1305                                 ipath_dbg("2nd group: dmacopy: %llx %llx "
1306                                           "%llx %llx; shadow: %lx %lx "
1307                                           "%lx %lx\n",
1308                                           (unsigned long long)
1309                                           le64_to_cpu(dma[4]),
1310                                           (unsigned long long)
1311                                           le64_to_cpu(dma[5]),
1312                                           (unsigned long long)
1313                                           le64_to_cpu(dma[6]),
1314                                           (unsigned long long)
1315                                           le64_to_cpu(dma[7]),
1316                                           shadow[4], shadow[5],
1317                                           shadow[6], shadow[7]);
1318                 }
1319                 buf = NULL;
1320                 goto bail;
1321         }
1322
1323         if (updated)
1324                 /*
1325                  * ran out of bufs, now some (at least this one we just
1326                  * got) are now available, so tell the layered driver.
1327                  */
1328                 __ipath_layer_intr(dd, IPATH_LAYER_INT_SEND_CONTINUE);
1329
1330         /*
1331          * set next starting place.  Since it's just an optimization,
1332          * it doesn't matter who wins on this, so no locking
1333          */
1334         dd->ipath_lastpioindex = i + 1;
1335         if (dd->ipath_upd_pio_shadow)
1336                 dd->ipath_upd_pio_shadow = 0;
1337         if (dd->ipath_consec_nopiobuf)
1338                 dd->ipath_consec_nopiobuf = 0;
1339         if (i < dd->ipath_piobcnt2k)
1340                 buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1341                                        i * dd->ipath_palign);
1342         else
1343                 buf = (u32 __iomem *)
1344                         (dd->ipath_pio4kbase +
1345                          (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1346         ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1347                    i, (i < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1348         if (pbufnum)
1349                 *pbufnum = i;
1350
1351 bail:
1352         return buf;
1353 }
1354
1355 /**
1356  * ipath_create_rcvhdrq - create a receive header queue
1357  * @dd: the infinipath device
1358  * @pd: the port data
1359  *
1360  * this must be contiguous memory (from an i/o perspective), and must be
1361  * DMA'able (which means for some systems, it will go through an IOMMU,
1362  * or be forced into a low address range).
1363  */
1364 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1365                          struct ipath_portdata *pd)
1366 {
1367         int ret = 0;
1368
1369         if (!pd->port_rcvhdrq) {
1370                 dma_addr_t phys_hdrqtail;
1371                 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1372                 int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1373                                 sizeof(u32), PAGE_SIZE);
1374
1375                 pd->port_rcvhdrq = dma_alloc_coherent(
1376                         &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1377                         gfp_flags);
1378
1379                 if (!pd->port_rcvhdrq) {
1380                         ipath_dev_err(dd, "attempt to allocate %d bytes "
1381                                       "for port %u rcvhdrq failed\n",
1382                                       amt, pd->port_port);
1383                         ret = -ENOMEM;
1384                         goto bail;
1385                 }
1386                 pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
1387                         &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
1388                 if (!pd->port_rcvhdrtail_kvaddr) {
1389                         ipath_dev_err(dd, "attempt to allocate 1 page "
1390                                       "for port %u rcvhdrqtailaddr failed\n",
1391                                       pd->port_port);
1392                         ret = -ENOMEM;
1393                         goto bail;
1394                 }
1395                 pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
1396
1397                 pd->port_rcvhdrq_size = amt;
1398
1399                 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1400                            "for port %u rcvhdr Q\n",
1401                            amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1402                            (unsigned long) pd->port_rcvhdrq_phys,
1403                            (unsigned long) pd->port_rcvhdrq_size,
1404                            pd->port_port);
1405
1406                 ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
1407                            pd->port_port,
1408                            (unsigned long long) phys_hdrqtail);
1409         }
1410         else
1411                 ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
1412                            "hdrtailaddr@%p %llx physical\n",
1413                            pd->port_port, pd->port_rcvhdrq,
1414                            pd->port_rcvhdrq_phys, pd->port_rcvhdrtail_kvaddr,
1415                            (unsigned long long)pd->port_rcvhdrqtailaddr_phys);
1416
1417         /* clear for security and sanity on each use */
1418         memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
1419         memset((void *)pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1420
1421         /*
1422          * tell chip each time we init it, even if we are re-using previous
1423          * memory (we zero the register at process close)
1424          */
1425         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1426                               pd->port_port, pd->port_rcvhdrqtailaddr_phys);
1427         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1428                               pd->port_port, pd->port_rcvhdrq_phys);
1429
1430         ret = 0;
1431 bail:
1432         return ret;
1433 }
1434
1435 int ipath_waitfor_complete(struct ipath_devdata *dd, ipath_kreg reg_id,
1436                            u64 bits_to_wait_for, u64 * valp)
1437 {
1438         unsigned long timeout;
1439         u64 lastval, val;
1440         int ret;
1441
1442         lastval = ipath_read_kreg64(dd, reg_id);
1443         /* wait a ridiculously long time */
1444         timeout = jiffies + msecs_to_jiffies(5);
1445         do {
1446                 val = ipath_read_kreg64(dd, reg_id);
1447                 /* set so they have something, even on failures. */
1448                 *valp = val;
1449                 if ((val & bits_to_wait_for) == bits_to_wait_for) {
1450                         ret = 0;
1451                         break;
1452                 }
1453                 if (val != lastval)
1454                         ipath_cdbg(VERBOSE, "Changed from %llx to %llx, "
1455                                    "waiting for %llx bits\n",
1456                                    (unsigned long long) lastval,
1457                                    (unsigned long long) val,
1458                                    (unsigned long long) bits_to_wait_for);
1459                 cond_resched();
1460                 if (time_after(jiffies, timeout)) {
1461                         ipath_dbg("Didn't get bits %llx in register 0x%x, "
1462                                   "got %llx\n",
1463                                   (unsigned long long) bits_to_wait_for,
1464                                   reg_id, (unsigned long long) *valp);
1465                         ret = -ENODEV;
1466                         break;
1467                 }
1468         } while (1);
1469
1470         return ret;
1471 }
1472
1473 /**
1474  * ipath_waitfor_mdio_cmdready - wait for last command to complete
1475  * @dd: the infinipath device
1476  *
1477  * Like ipath_waitfor_complete(), but we wait for the CMDVALID bit to go
1478  * away indicating the last command has completed.  It doesn't return data
1479  */
1480 int ipath_waitfor_mdio_cmdready(struct ipath_devdata *dd)
1481 {
1482         unsigned long timeout;
1483         u64 val;
1484         int ret;
1485
1486         /* wait a ridiculously long time */
1487         timeout = jiffies + msecs_to_jiffies(5);
1488         do {
1489                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_mdio);
1490                 if (!(val & IPATH_MDIO_CMDVALID)) {
1491                         ret = 0;
1492                         break;
1493                 }
1494                 cond_resched();
1495                 if (time_after(jiffies, timeout)) {
1496                         ipath_dbg("CMDVALID stuck in mdio reg? (%llx)\n",
1497                                   (unsigned long long) val);
1498                         ret = -ENODEV;
1499                         break;
1500                 }
1501         } while (1);
1502
1503         return ret;
1504 }
1505
1506 void ipath_set_ib_lstate(struct ipath_devdata *dd, int which)
1507 {
1508         static const char *what[4] = {
1509                 [0] = "DOWN",
1510                 [INFINIPATH_IBCC_LINKCMD_INIT] = "INIT",
1511                 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1512                 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1513         };
1514         int linkcmd = (which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
1515                         INFINIPATH_IBCC_LINKCMD_MASK;
1516
1517         ipath_cdbg(SMA, "Trying to move unit %u to %s, current ltstate "
1518                    "is %s\n", dd->ipath_unit,
1519                    what[linkcmd],
1520                    ipath_ibcstatus_str[
1521                            (ipath_read_kreg64
1522                             (dd, dd->ipath_kregs->kr_ibcstatus) >>
1523                             INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
1524                            INFINIPATH_IBCS_LINKTRAININGSTATE_MASK]);
1525         /* flush all queued sends when going to DOWN or INIT, to be sure that
1526          * they don't block SMA and other MAD packets */
1527         if (!linkcmd || linkcmd == INFINIPATH_IBCC_LINKCMD_INIT) {
1528                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1529                                  INFINIPATH_S_ABORT);
1530                 ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
1531                                     (unsigned)(dd->ipath_piobcnt2k +
1532                                     dd->ipath_piobcnt4k) -
1533                                     dd->ipath_lastport_piobuf);
1534         }
1535
1536         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1537                          dd->ipath_ibcctrl | which);
1538 }
1539
1540 /**
1541  * ipath_read_kreg64_port - read a device's per-port 64-bit kernel register
1542  * @dd: the infinipath device
1543  * @regno: the register number to read
1544  * @port: the port containing the register
1545  *
1546  * Registers that vary with the chip implementation constants (port)
1547  * use this routine.
1548  */
1549 u64 ipath_read_kreg64_port(const struct ipath_devdata *dd, ipath_kreg regno,
1550                            unsigned port)
1551 {
1552         u16 where;
1553
1554         if (port < dd->ipath_portcnt &&
1555             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1556              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1557                 where = regno + port;
1558         else
1559                 where = -1;
1560
1561         return ipath_read_kreg64(dd, where);
1562 }
1563
1564 /**
1565  * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
1566  * @dd: the infinipath device
1567  * @regno: the register number to write
1568  * @port: the port containing the register
1569  * @value: the value to write
1570  *
1571  * Registers that vary with the chip implementation constants (port)
1572  * use this routine.
1573  */
1574 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
1575                           unsigned port, u64 value)
1576 {
1577         u16 where;
1578
1579         if (port < dd->ipath_portcnt &&
1580             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1581              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1582                 where = regno + port;
1583         else
1584                 where = -1;
1585
1586         ipath_write_kreg(dd, where, value);
1587 }
1588
1589 /**
1590  * ipath_shutdown_device - shut down a device
1591  * @dd: the infinipath device
1592  *
1593  * This is called to make the device quiet when we are about to
1594  * unload the driver, and also when the device is administratively
1595  * disabled.   It does not free any data structures.
1596  * Everything it does has to be setup again by ipath_init_chip(dd,1)
1597  */
1598 void ipath_shutdown_device(struct ipath_devdata *dd)
1599 {
1600         u64 val;
1601
1602         ipath_dbg("Shutting down the device\n");
1603
1604         dd->ipath_flags |= IPATH_LINKUNK;
1605         dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
1606                              IPATH_LINKINIT | IPATH_LINKARMED |
1607                              IPATH_LINKACTIVE);
1608         *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
1609                                 IPATH_STATUS_IB_READY);
1610
1611         /* mask interrupts, but not errors */
1612         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
1613
1614         dd->ipath_rcvctrl = 0;
1615         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1616                          dd->ipath_rcvctrl);
1617
1618         /*
1619          * gracefully stop all sends allowing any in progress to trickle out
1620          * first.
1621          */
1622         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0ULL);
1623         /* flush it */
1624         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1625         /*
1626          * enough for anything that's going to trickle out to have actually
1627          * done so.
1628          */
1629         udelay(5);
1630
1631         /*
1632          * abort any armed or launched PIO buffers that didn't go. (self
1633          * clearing).  Will cause any packet currently being transmitted to
1634          * go out with an EBP, and may also cause a short packet error on
1635          * the receiver.
1636          */
1637         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1638                          INFINIPATH_S_ABORT);
1639
1640         ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1641                             INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1642
1643         /*
1644          * we are shutting down, so tell the layered driver.  We don't do
1645          * this on just a link state change, much like ethernet, a cable
1646          * unplug, etc. doesn't change driver state
1647          */
1648         ipath_layer_intr(dd, IPATH_LAYER_INT_IF_DOWN);
1649
1650         /* disable IBC */
1651         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
1652         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
1653                          dd->ipath_control | INFINIPATH_C_FREEZEMODE);
1654
1655         /*
1656          * clear SerdesEnable and turn the leds off; do this here because
1657          * we are unloading, so don't count on interrupts to move along
1658          * Turn the LEDs off explictly for the same reason.
1659          */
1660         dd->ipath_f_quiet_serdes(dd);
1661         dd->ipath_f_setextled(dd, 0, 0);
1662
1663         if (dd->ipath_stats_timer_active) {
1664                 del_timer_sync(&dd->ipath_stats_timer);
1665                 dd->ipath_stats_timer_active = 0;
1666         }
1667
1668         /*
1669          * clear all interrupts and errors, so that the next time the driver
1670          * is loaded or device is enabled, we know that whatever is set
1671          * happened while we were unloaded
1672          */
1673         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
1674                          ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
1675         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
1676         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
1677 }
1678
1679 /**
1680  * ipath_free_pddata - free a port's allocated data
1681  * @dd: the infinipath device
1682  * @pd: the portdata structure
1683  *
1684  * free up any allocated data for a port
1685  * This should not touch anything that would affect a simultaneous
1686  * re-allocation of port data, because it is called after ipath_mutex
1687  * is released (and can be called from reinit as well).
1688  * It should never change any chip state, or global driver state.
1689  * (The only exception to global state is freeing the port0 port0_skbs.)
1690  */
1691 void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
1692 {
1693         if (!pd)
1694                 return;
1695
1696         if (pd->port_rcvhdrq) {
1697                 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
1698                            "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
1699                            (unsigned long) pd->port_rcvhdrq_size);
1700                 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
1701                                   pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1702                 pd->port_rcvhdrq = NULL;
1703                 if (pd->port_rcvhdrtail_kvaddr) {
1704                         dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1705                                          (void *)pd->port_rcvhdrtail_kvaddr,
1706                                          pd->port_rcvhdrqtailaddr_phys);
1707                         pd->port_rcvhdrtail_kvaddr = NULL;
1708                 }
1709         }
1710         if (pd->port_port && pd->port_rcvegrbuf) {
1711                 unsigned e;
1712
1713                 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
1714                         void *base = pd->port_rcvegrbuf[e];
1715                         size_t size = pd->port_rcvegrbuf_size;
1716
1717                         ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
1718                                    "chunk %u/%u\n", base,
1719                                    (unsigned long) size,
1720                                    e, pd->port_rcvegrbuf_chunks);
1721                         dma_free_coherent(&dd->pcidev->dev, size,
1722                                 base, pd->port_rcvegrbuf_phys[e]);
1723                 }
1724                 vfree(pd->port_rcvegrbuf);
1725                 pd->port_rcvegrbuf = NULL;
1726                 vfree(pd->port_rcvegrbuf_phys);
1727                 pd->port_rcvegrbuf_phys = NULL;
1728                 pd->port_rcvegrbuf_chunks = 0;
1729         } else if (pd->port_port == 0 && dd->ipath_port0_skbs) {
1730                 unsigned e;
1731                 struct sk_buff **skbs = dd->ipath_port0_skbs;
1732
1733                 dd->ipath_port0_skbs = NULL;
1734                 ipath_cdbg(VERBOSE, "free closed port %d ipath_port0_skbs "
1735                            "@ %p\n", pd->port_port, skbs);
1736                 for (e = 0; e < dd->ipath_rcvegrcnt; e++)
1737                         if (skbs[e])
1738                                 dev_kfree_skb(skbs[e]);
1739                 vfree(skbs);
1740         }
1741         kfree(pd->port_tid_pg_list);
1742         kfree(pd);
1743 }
1744
1745 static int __init infinipath_init(void)
1746 {
1747         int ret;
1748
1749         ipath_dbg(KERN_INFO DRIVER_LOAD_MSG "%s", ipath_core_version);
1750
1751         /*
1752          * These must be called before the driver is registered with
1753          * the PCI subsystem.
1754          */
1755         idr_init(&unit_table);
1756         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
1757                 ret = -ENOMEM;
1758                 goto bail;
1759         }
1760
1761         ret = pci_register_driver(&ipath_driver);
1762         if (ret < 0) {
1763                 printk(KERN_ERR IPATH_DRV_NAME
1764                        ": Unable to register driver: error %d\n", -ret);
1765                 goto bail_unit;
1766         }
1767
1768         ret = ipath_driver_create_group(&ipath_driver.driver);
1769         if (ret < 0) {
1770                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create driver "
1771                        "sysfs entries: error %d\n", -ret);
1772                 goto bail_pci;
1773         }
1774
1775         ret = ipath_init_ipathfs();
1776         if (ret < 0) {
1777                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
1778                        "ipathfs: error %d\n", -ret);
1779                 goto bail_group;
1780         }
1781
1782         goto bail;
1783
1784 bail_group:
1785         ipath_driver_remove_group(&ipath_driver.driver);
1786
1787 bail_pci:
1788         pci_unregister_driver(&ipath_driver);
1789
1790 bail_unit:
1791         idr_destroy(&unit_table);
1792
1793 bail:
1794         return ret;
1795 }
1796
1797 static void cleanup_device(struct ipath_devdata *dd)
1798 {
1799         int port;
1800
1801         ipath_shutdown_device(dd);
1802
1803         if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
1804                 /* can't do anything more with chip; needs re-init */
1805                 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
1806                 if (dd->ipath_kregbase) {
1807                         /*
1808                          * if we haven't already cleaned up before these are
1809                          * to ensure any register reads/writes "fail" until
1810                          * re-init
1811                          */
1812                         dd->ipath_kregbase = NULL;
1813                         dd->ipath_uregbase = 0;
1814                         dd->ipath_sregbase = 0;
1815                         dd->ipath_cregbase = 0;
1816                         dd->ipath_kregsize = 0;
1817                 }
1818                 ipath_disable_wc(dd);
1819         }
1820
1821         if (dd->ipath_pioavailregs_dma) {
1822                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1823                                   (void *) dd->ipath_pioavailregs_dma,
1824                                   dd->ipath_pioavailregs_phys);
1825                 dd->ipath_pioavailregs_dma = NULL;
1826         }
1827         if (dd->ipath_dummy_hdrq) {
1828                 dma_free_coherent(&dd->pcidev->dev,
1829                         dd->ipath_pd[0]->port_rcvhdrq_size,
1830                         dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
1831                 dd->ipath_dummy_hdrq = NULL;
1832         }
1833
1834         if (dd->ipath_pageshadow) {
1835                 struct page **tmpp = dd->ipath_pageshadow;
1836                 int i, cnt = 0;
1837
1838                 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
1839                            "locked\n");
1840                 for (port = 0; port < dd->ipath_cfgports; port++) {
1841                         int port_tidbase = port * dd->ipath_rcvtidcnt;
1842                         int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1843                         for (i = port_tidbase; i < maxtid; i++) {
1844                                 if (!tmpp[i])
1845                                         continue;
1846                                 ipath_release_user_pages(&tmpp[i], 1);
1847                                 tmpp[i] = NULL;
1848                                 cnt++;
1849                         }
1850                 }
1851                 if (cnt) {
1852                         ipath_stats.sps_pageunlocks += cnt;
1853                         ipath_cdbg(VERBOSE, "There were still %u expTID "
1854                                    "entries locked\n", cnt);
1855                 }
1856                 if (ipath_stats.sps_pagelocks ||
1857                     ipath_stats.sps_pageunlocks)
1858                         ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
1859                                    "unlocked via ipath_m{un}lock\n",
1860                                    (unsigned long long)
1861                                    ipath_stats.sps_pagelocks,
1862                                    (unsigned long long)
1863                                    ipath_stats.sps_pageunlocks);
1864
1865                 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
1866                            dd->ipath_pageshadow);
1867                 vfree(dd->ipath_pageshadow);
1868                 dd->ipath_pageshadow = NULL;
1869         }
1870
1871         /*
1872          * free any resources still in use (usually just kernel ports)
1873          * at unload; we do for portcnt, not cfgports, because cfgports
1874          * could have changed while we were loaded.
1875          */
1876         for (port = 0; port < dd->ipath_portcnt; port++) {
1877                 struct ipath_portdata *pd = dd->ipath_pd[port];
1878                 dd->ipath_pd[port] = NULL;
1879                 ipath_free_pddata(dd, pd);
1880         }
1881         kfree(dd->ipath_pd);
1882         /*
1883          * debuggability, in case some cleanup path tries to use it
1884          * after this
1885          */
1886         dd->ipath_pd = NULL;
1887 }
1888
1889 static void __exit infinipath_cleanup(void)
1890 {
1891         struct ipath_devdata *dd, *tmp;
1892         unsigned long flags;
1893
1894         ipath_exit_ipathfs();
1895
1896         ipath_driver_remove_group(&ipath_driver.driver);
1897
1898         spin_lock_irqsave(&ipath_devs_lock, flags);
1899
1900         /*
1901          * turn off rcv, send, and interrupts for all ports, all drivers
1902          * should also hard reset the chip here?
1903          * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
1904          * for all versions of the driver, if they were allocated
1905          */
1906         list_for_each_entry_safe(dd, tmp, &ipath_dev_list, ipath_list) {
1907                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
1908
1909                 if (dd->ipath_kregbase)
1910                         cleanup_device(dd);
1911
1912                 if (dd->pcidev) {
1913                         if (dd->pcidev->irq) {
1914                                 ipath_cdbg(VERBOSE,
1915                                            "unit %u free_irq of irq %x\n",
1916                                            dd->ipath_unit, dd->pcidev->irq);
1917                                 free_irq(dd->pcidev->irq, dd);
1918                         } else
1919                                 ipath_dbg("irq is 0, not doing free_irq "
1920                                           "for unit %u\n", dd->ipath_unit);
1921
1922                         /*
1923                          * we check for NULL here, because it's outside
1924                          * the kregbase check, and we need to call it
1925                          * after the free_irq.  Thus it's possible that
1926                          * the function pointers were never initialized.
1927                          */
1928                         if (dd->ipath_f_cleanup)
1929                                 /* clean up chip-specific stuff */
1930                                 dd->ipath_f_cleanup(dd);
1931
1932                         dd->pcidev = NULL;
1933                 }
1934                 spin_lock_irqsave(&ipath_devs_lock, flags);
1935         }
1936
1937         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1938
1939         ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
1940         pci_unregister_driver(&ipath_driver);
1941
1942         idr_destroy(&unit_table);
1943 }
1944
1945 /**
1946  * ipath_reset_device - reset the chip if possible
1947  * @unit: the device to reset
1948  *
1949  * Whether or not reset is successful, we attempt to re-initialize the chip
1950  * (that is, much like a driver unload/reload).  We clear the INITTED flag
1951  * so that the various entry points will fail until we reinitialize.  For
1952  * now, we only allow this if no user ports are open that use chip resources
1953  */
1954 int ipath_reset_device(int unit)
1955 {
1956         int ret, i;
1957         struct ipath_devdata *dd = ipath_lookup(unit);
1958
1959         if (!dd) {
1960                 ret = -ENODEV;
1961                 goto bail;
1962         }
1963
1964         dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
1965
1966         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
1967                 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
1968                          "not initialized or not present\n", unit);
1969                 ret = -ENXIO;
1970                 goto bail;
1971         }
1972
1973         if (dd->ipath_pd)
1974                 for (i = 1; i < dd->ipath_cfgports; i++) {
1975                         if (dd->ipath_pd[i] && dd->ipath_pd[i]->port_cnt) {
1976                                 ipath_dbg("unit %u port %d is in use "
1977                                           "(PID %u cmd %s), can't reset\n",
1978                                           unit, i,
1979                                           dd->ipath_pd[i]->port_pid,
1980                                           dd->ipath_pd[i]->port_comm);
1981                                 ret = -EBUSY;
1982                                 goto bail;
1983                         }
1984                 }
1985
1986         dd->ipath_flags &= ~IPATH_INITTED;
1987         ret = dd->ipath_f_reset(dd);
1988         if (ret != 1)
1989                 ipath_dbg("reset was not successful\n");
1990         ipath_dbg("Trying to reinitialize unit %u after reset attempt\n",
1991                   unit);
1992         ret = ipath_init_chip(dd, 1);
1993         if (ret)
1994                 ipath_dev_err(dd, "Reinitialize unit %u after "
1995                               "reset failed with %d\n", unit, ret);
1996         else
1997                 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
1998                          "resetting\n", unit);
1999
2000 bail:
2001         return ret;
2002 }
2003
2004 module_init(infinipath_init);
2005 module_exit(infinipath_cleanup);