SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_driver.c

   1 /*
   2  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/spinlock.h>
  34 #include <linux/idr.h>
  35 #include <linux/pci.h>
  36 #include <linux/delay.h>
  37 #include <linux/netdevice.h>
  38 #include <linux/vmalloc.h>
  39
  40 #include "ipath_kernel.h"
  41 #include "ips_common.h"
  42 #include "ipath_layer.h"
  43
  44 static void ipath_update_pio_bufs(struct ipath_devdata *);
  45
  46 const char *ipath_get_unit_name(int unit)
  47 {
  48         static char iname[16];
  49         snprintf(iname, sizeof iname, "infinipath%u", unit);
  50         return iname;
  51 }
  52
  53 EXPORT_SYMBOL_GPL(ipath_get_unit_name);
  54
  55 #define DRIVER_LOAD_MSG "PathScale " 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("PathScale <support@pathscale.com>");
  78 MODULE_DESCRIPTION("Pathscale 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,
 120                     PCI_DEVICE_ID_INFINIPATH_HT)},
 121         {PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE,
 122                     PCI_DEVICE_ID_INFINIPATH_PE800)},
 123 };
 124
 125 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
 126
 127 static struct pci_driver ipath_driver = {
 128         .name = IPATH_DRV_NAME,
 129         .probe = ipath_init_one,
 130         .remove = __devexit_p(ipath_remove_one),
 131         .id_table = ipath_pci_tbl,
 132 };
 133
 134 /*
 135  * This is where port 0's rcvhdrtail register is written back; we also
 136  * want nothing else sharing the cache line, so make it a cache line
 137  * in size.  Used for all units.
 138  */
 139 volatile __le64 *ipath_port0_rcvhdrtail;
 140 dma_addr_t ipath_port0_rcvhdrtail_dma;
 141 static int port0_rcvhdrtail_refs;
 142
 143 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
 144                              u32 *bar0, u32 *bar1)
 145 {
 146         int ret;
 147
 148         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
 149         if (ret)
 150                 ipath_dev_err(dd, "failed to read bar0 before enable: "
 151                               "error %d\n", -ret);
 152
 153         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
 154         if (ret)
 155                 ipath_dev_err(dd, "failed to read bar1 before enable: "
 156                               "error %d\n", -ret);
 157
 158         ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
 159 }
 160
 161 static void ipath_free_devdata(struct pci_dev *pdev,
 162                                struct ipath_devdata *dd)
 163 {
 164         unsigned long flags;
 165
 166         pci_set_drvdata(pdev, NULL);
 167
 168         if (dd->ipath_unit != -1) {
 169                 spin_lock_irqsave(&ipath_devs_lock, flags);
 170                 idr_remove(&unit_table, dd->ipath_unit);
 171                 list_del(&dd->ipath_list);
 172                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
 173         }
 174         dma_free_coherent(&pdev->dev, sizeof(*dd), dd, dd->ipath_dma_addr);
 175 }
 176
 177 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
 178 {
 179         unsigned long flags;
 180         struct ipath_devdata *dd;
 181         dma_addr_t dma_addr;
 182         int ret;
 183
 184         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
 185                 dd = ERR_PTR(-ENOMEM);
 186                 goto bail;
 187         }
 188
 189         dd = dma_alloc_coherent(&pdev->dev, sizeof(*dd), &dma_addr,
 190                                 GFP_KERNEL);
 191
 192         if (!dd) {
 193                 dd = ERR_PTR(-ENOMEM);
 194                 goto bail;
 195         }
 196
 197         dd->ipath_dma_addr = dma_addr;
 198         dd->ipath_unit = -1;
 199
 200         spin_lock_irqsave(&ipath_devs_lock, flags);
 201
 202         ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
 203         if (ret < 0) {
 204                 printk(KERN_ERR IPATH_DRV_NAME
 205                        ": Could not allocate unit ID: error %d\n", -ret);
 206                 ipath_free_devdata(pdev, dd);
 207                 dd = ERR_PTR(ret);
 208                 goto bail_unlock;
 209         }
 210
 211         dd->pcidev = pdev;
 212         pci_set_drvdata(pdev, dd);
 213
 214         list_add(&dd->ipath_list, &ipath_dev_list);
 215
 216 bail_unlock:
 217         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 218
 219 bail:
 220         return dd;
 221 }
 222
 223 static inline struct ipath_devdata *__ipath_lookup(int unit)
 224 {
 225         return idr_find(&unit_table, unit);
 226 }
 227
 228 struct ipath_devdata *ipath_lookup(int unit)
 229 {
 230         struct ipath_devdata *dd;
 231         unsigned long flags;
 232
 233         spin_lock_irqsave(&ipath_devs_lock, flags);
 234         dd = __ipath_lookup(unit);
 235         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 236
 237         return dd;
 238 }
 239
 240 int ipath_count_units(int *npresentp, int *nupp, u32 *maxportsp)
 241 {
 242         int nunits, npresent, nup;
 243         struct ipath_devdata *dd;
 244         unsigned long flags;
 245         u32 maxports;
 246
 247         nunits = npresent = nup = maxports = 0;
 248
 249         spin_lock_irqsave(&ipath_devs_lock, flags);
 250
 251         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 252                 nunits++;
 253                 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
 254                         npresent++;
 255                 if (dd->ipath_lid &&
 256                     !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
 257                                          | IPATH_LINKUNK)))
 258                         nup++;
 259                 if (dd->ipath_cfgports > maxports)
 260                         maxports = dd->ipath_cfgports;
 261         }
 262
 263         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 264
 265         if (npresentp)
 266                 *npresentp = npresent;
 267         if (nupp)
 268                 *nupp = nup;
 269         if (maxportsp)
 270                 *maxportsp = maxports;
 271
 272         return nunits;
 273 }
 274
 275 static int init_port0_rcvhdrtail(struct pci_dev *pdev)
 276 {
 277         int ret;
 278
 279         mutex_lock(&ipath_mutex);
 280
 281         if (!ipath_port0_rcvhdrtail) {
 282                 ipath_port0_rcvhdrtail =
 283                         dma_alloc_coherent(&pdev->dev,
 284                                            IPATH_PORT0_RCVHDRTAIL_SIZE,
 285                                            &ipath_port0_rcvhdrtail_dma,
 286                                            GFP_KERNEL);
 287
 288                 if (!ipath_port0_rcvhdrtail) {
 289                         ret = -ENOMEM;
 290                         goto bail;
 291                 }
 292         }
 293         port0_rcvhdrtail_refs++;
 294         ret = 0;
 295
 296 bail:
 297         mutex_unlock(&ipath_mutex);
 298
 299         return ret;
 300 }
 301
 302 static void cleanup_port0_rcvhdrtail(struct pci_dev *pdev)
 303 {
 304         mutex_lock(&ipath_mutex);
 305
 306         if (!--port0_rcvhdrtail_refs) {
 307                 dma_free_coherent(&pdev->dev, IPATH_PORT0_RCVHDRTAIL_SIZE,
 308                                   (void *) ipath_port0_rcvhdrtail,
 309                                   ipath_port0_rcvhdrtail_dma);
 310                 ipath_port0_rcvhdrtail = NULL;
 311         }
 312
 313         mutex_unlock(&ipath_mutex);
 314 }
 315
 316 /*
 317  * These next two routines are placeholders in case we don't have per-arch
 318  * code for controlling write combining.  If explicit control of write
 319  * combining is not available, performance will probably be awful.
 320  */
 321
 322 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
 323 {
 324         return -EOPNOTSUPP;
 325 }
 326
 327 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
 328 {
 329 }
 330
 331 static int __devinit ipath_init_one(struct pci_dev *pdev,
 332                                     const struct pci_device_id *ent)
 333 {
 334         int ret, len, j;
 335         struct ipath_devdata *dd;
 336         unsigned long long addr;
 337         u32 bar0 = 0, bar1 = 0;
 338         u8 rev;
 339
 340         ret = init_port0_rcvhdrtail(pdev);
 341         if (ret < 0) {
 342                 printk(KERN_ERR IPATH_DRV_NAME
 343                        ": Could not allocate port0_rcvhdrtail: error %d\n",
 344                        -ret);
 345                 goto bail;
 346         }
 347
 348         dd = ipath_alloc_devdata(pdev);
 349         if (IS_ERR(dd)) {
 350                 ret = PTR_ERR(dd);
 351                 printk(KERN_ERR IPATH_DRV_NAME
 352                        ": Could not allocate devdata: error %d\n", -ret);
 353                 goto bail_rcvhdrtail;
 354         }
 355
 356         ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
 357
 358         read_bars(dd, pdev, &bar0, &bar1);
 359
 360         ret = pci_enable_device(pdev);
 361         if (ret) {
 362                 /* This can happen iff:
 363                  *
 364                  * We did a chip reset, and then failed to reprogram the
 365                  * BAR, or the chip reset due to an internal error.  We then
 366                  * unloaded the driver and reloaded it.
 367                  *
 368                  * Both reset cases set the BAR back to initial state.  For
 369                  * the latter case, the AER sticky error bit at offset 0x718
 370                  * should be set, but the Linux kernel doesn't yet know
 371                  * about that, it appears.  If the original BAR was retained
 372                  * in the kernel data structures, this may be OK.
 373                  */
 374                 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
 375                               dd->ipath_unit, -ret);
 376                 goto bail_devdata;
 377         }
 378         addr = pci_resource_start(pdev, 0);
 379         len = pci_resource_len(pdev, 0);
 380         ipath_cdbg(VERBOSE, "regbase (0) %llx len %d irq %x, vend %x/%x "
 381                    "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
 382                    ent->device, ent->driver_data);
 383
 384         read_bars(dd, pdev, &bar0, &bar1);
 385
 386         if (!bar1 && !(bar0 & ~0xf)) {
 387                 if (addr) {
 388                         dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
 389                                  "rewriting as %llx\n", addr);
 390                         ret = pci_write_config_dword(
 391                                 pdev, PCI_BASE_ADDRESS_0, addr);
 392                         if (ret) {
 393                                 ipath_dev_err(dd, "rewrite of BAR0 "
 394                                               "failed: err %d\n", -ret);
 395                                 goto bail_disable;
 396                         }
 397                         ret = pci_write_config_dword(
 398                                 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
 399                         if (ret) {
 400                                 ipath_dev_err(dd, "rewrite of BAR1 "
 401                                               "failed: err %d\n", -ret);
 402                                 goto bail_disable;
 403                         }
 404                 } else {
 405                         ipath_dev_err(dd, "BAR is 0 (probable RESET), "
 406                                       "not usable until reboot\n");
 407                         ret = -ENODEV;
 408                         goto bail_disable;
 409                 }
 410         }
 411
 412         ret = pci_request_regions(pdev, IPATH_DRV_NAME);
 413         if (ret) {
 414                 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
 415                          "err %d\n", dd->ipath_unit, -ret);
 416                 goto bail_disable;
 417         }
 418
 419         ret = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
 420         if (ret) {
 421                 dev_info(&pdev->dev, "pci_set_dma_mask unit %u "
 422                          "fails: %d\n", dd->ipath_unit, ret);
 423                 goto bail_regions;
 424         }
 425
 426         pci_set_master(pdev);
 427
 428         /*
 429          * Save BARs to rewrite after device reset.  Save all 64 bits of
 430          * BAR, just in case.
 431          */
 432         dd->ipath_pcibar0 = addr;
 433         dd->ipath_pcibar1 = addr >> 32;
 434         dd->ipath_deviceid = ent->device;       /* save for later use */
 435         dd->ipath_vendorid = ent->vendor;
 436
 437         /* setup the chip-specific functions, as early as possible. */
 438         switch (ent->device) {
 439         case PCI_DEVICE_ID_INFINIPATH_HT:
 440                 ipath_init_ht400_funcs(dd);
 441                 break;
 442         case PCI_DEVICE_ID_INFINIPATH_PE800:
 443                 ipath_init_pe800_funcs(dd);
 444                 break;
 445         default:
 446                 ipath_dev_err(dd, "Found unknown PathScale deviceid 0x%x, "
 447                               "failing\n", ent->device);
 448                 return -ENODEV;
 449         }
 450
 451         for (j = 0; j < 6; j++) {
 452                 if (!pdev->resource[j].start)
 453                         continue;
 454                 ipath_cdbg(VERBOSE, "BAR %d start %lx, end %lx, len %lx\n",
 455                            j, pdev->resource[j].start,
 456                            pdev->resource[j].end,
 457                            pci_resource_len(pdev, j));
 458         }
 459
 460         if (!addr) {
 461                 ipath_dev_err(dd, "No valid address in BAR 0!\n");
 462                 ret = -ENODEV;
 463                 goto bail_regions;
 464         }
 465
 466         dd->ipath_deviceid = ent->device;       /* save for later use */
 467         dd->ipath_vendorid = ent->vendor;
 468
 469         ret = pci_read_config_byte(pdev, PCI_REVISION_ID, &rev);
 470         if (ret) {
 471                 ipath_dev_err(dd, "Failed to read PCI revision ID unit "
 472                               "%u: err %d\n", dd->ipath_unit, -ret);
 473                 goto bail_regions;      /* shouldn't ever happen */
 474         }
 475         dd->ipath_pcirev = rev;
 476
 477         dd->ipath_kregbase = ioremap_nocache(addr, len);
 478
 479         if (!dd->ipath_kregbase) {
 480                 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
 481                           addr);
 482                 ret = -ENOMEM;
 483                 goto bail_iounmap;
 484         }
 485         dd->ipath_kregend = (u64 __iomem *)
 486                 ((void __iomem *)dd->ipath_kregbase + len);
 487         dd->ipath_physaddr = addr;      /* used for io_remap, etc. */
 488         /* for user mmap */
 489         dd->ipath_kregvirt = (u64 __iomem *) phys_to_virt(addr);
 490         ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p "
 491                    "kregvirt %p\n", addr, dd->ipath_kregbase,
 492                    dd->ipath_kregvirt);
 493
 494         /*
 495          * clear ipath_flags here instead of in ipath_init_chip as it is set
 496          * by ipath_setup_htconfig.
 497          */
 498         dd->ipath_flags = 0;
 499
 500         if (dd->ipath_f_bus(dd, pdev))
 501                 ipath_dev_err(dd, "Failed to setup config space; "
 502                               "continuing anyway\n");
 503
 504         /*
 505          * set up our interrupt handler; SA_SHIRQ probably not needed,
 506          * since MSI interrupts shouldn't be shared but won't  hurt for now.
 507          * check 0 irq after we return from chip-specific bus setup, since
 508          * that can affect this due to setup
 509          */
 510         if (!pdev->irq)
 511                 ipath_dev_err(dd, "irq is 0, BIOS error?  Interrupts won't "
 512                               "work\n");
 513         else {
 514                 ret = request_irq(pdev->irq, ipath_intr, SA_SHIRQ,
 515                                   IPATH_DRV_NAME, dd);
 516                 if (ret) {
 517                         ipath_dev_err(dd, "Couldn't setup irq handler, "
 518                                       "irq=%u: %d\n", pdev->irq, ret);
 519                         goto bail_iounmap;
 520                 }
 521         }
 522
 523         ret = ipath_init_chip(dd, 0);   /* do the chip-specific init */
 524         if (ret)
 525                 goto bail_iounmap;
 526
 527         ret = ipath_enable_wc(dd);
 528
 529         if (ret) {
 530                 ipath_dev_err(dd, "Write combining not enabled "
 531                               "(err %d): performance may be poor\n",
 532                               -ret);
 533                 ret = 0;
 534         }
 535
 536         ipath_device_create_group(&pdev->dev, dd);
 537         ipathfs_add_device(dd);
 538         ipath_user_add(dd);
 539         ipath_layer_add(dd);
 540
 541         goto bail;
 542
 543 bail_iounmap:
 544         iounmap((volatile void __iomem *) dd->ipath_kregbase);
 545
 546 bail_regions:
 547         pci_release_regions(pdev);
 548
 549 bail_disable:
 550         pci_disable_device(pdev);
 551
 552 bail_devdata:
 553         ipath_free_devdata(pdev, dd);
 554
 555 bail_rcvhdrtail:
 556         cleanup_port0_rcvhdrtail(pdev);
 557
 558 bail:
 559         return ret;
 560 }
 561
 562 static void __devexit ipath_remove_one(struct pci_dev *pdev)
 563 {
 564         struct ipath_devdata *dd;
 565
 566         ipath_cdbg(VERBOSE, "removing, pdev=%p\n", pdev);
 567         if (!pdev)
 568                 return;
 569
 570         dd = pci_get_drvdata(pdev);
 571         ipath_layer_del(dd);
 572         ipath_user_del(dd);
 573         ipathfs_remove_device(dd);
 574         ipath_device_remove_group(&pdev->dev, dd);
 575         ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
 576                    "unit %u\n", dd, (u32) dd->ipath_unit);
 577         if (dd->ipath_kregbase) {
 578                 ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n",
 579                            dd->ipath_kregbase);
 580                 iounmap((volatile void __iomem *) dd->ipath_kregbase);
 581                 dd->ipath_kregbase = NULL;
 582         }
 583         pci_release_regions(pdev);
 584         ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
 585         pci_disable_device(pdev);
 586
 587         ipath_free_devdata(pdev, dd);
 588         cleanup_port0_rcvhdrtail(pdev);
 589 }
 590
 591 /* general driver use */
 592 DEFINE_MUTEX(ipath_mutex);
 593
 594 static DEFINE_SPINLOCK(ipath_pioavail_lock);
 595
 596 /**
 597  * ipath_disarm_piobufs - cancel a range of PIO buffers
 598  * @dd: the infinipath device
 599  * @first: the first PIO buffer to cancel
 600  * @cnt: the number of PIO buffers to cancel
 601  *
 602  * cancel a range of PIO buffers, used when they might be armed, but
 603  * not triggered.  Used at init to ensure buffer state, and also user
 604  * process close, in case it died while writing to a PIO buffer
 605  * Also after errors.
 606  */
 607 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
 608                           unsigned cnt)
 609 {
 610         unsigned i, last = first + cnt;
 611         u64 sendctrl, sendorig;
 612
 613         ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
 614         sendorig = dd->ipath_sendctrl | INFINIPATH_S_DISARM;
 615         for (i = first; i < last; i++) {
 616                 sendctrl = sendorig |
 617                         (i << INFINIPATH_S_DISARMPIOBUF_SHIFT);
 618                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 619                                  sendctrl);
 620         }
 621
 622         /*
 623          * Write it again with current value, in case ipath_sendctrl changed
 624          * while we were looping; no critical bits that would require
 625          * locking.
 626          *
 627          * Write a 0, and then the original value, reading scratch in
 628          * between.  This seems to avoid a chip timing race that causes
 629          * pioavail updates to memory to stop.
 630          */
 631         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 632                          0);
 633         sendorig = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 634         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 635                          dd->ipath_sendctrl);
 636 }
 637
 638 /**
 639  * ipath_wait_linkstate - wait for an IB link state change to occur
 640  * @dd: the infinipath device
 641  * @state: the state to wait for
 642  * @msecs: the number of milliseconds to wait
 643  *
 644  * wait up to msecs milliseconds for IB link state change to occur for
 645  * now, take the easy polling route.  Currently used only by
 646  * ipath_layer_set_linkstate.  Returns 0 if state reached, otherwise
 647  * -ETIMEDOUT state can have multiple states set, for any of several
 648  * transitions.
 649  */
 650 int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state, int msecs)
 651 {
 652         dd->ipath_sma_state_wanted = state;
 653         wait_event_interruptible_timeout(ipath_sma_state_wait,
 654                                          (dd->ipath_flags & state),
 655                                          msecs_to_jiffies(msecs));
 656         dd->ipath_sma_state_wanted = 0;
 657
 658         if (!(dd->ipath_flags & state)) {
 659                 u64 val;
 660                 ipath_cdbg(SMA, "Didn't reach linkstate %s within %u ms\n",
 661                            /* test INIT ahead of DOWN, both can be set */
 662                            (state & IPATH_LINKINIT) ? "INIT" :
 663                            ((state & IPATH_LINKDOWN) ? "DOWN" :
 664                             ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
 665                            msecs);
 666                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
 667                 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
 668                            (unsigned long long) ipath_read_kreg64(
 669                                    dd, dd->ipath_kregs->kr_ibcctrl),
 670                            (unsigned long long) val,
 671                            ipath_ibcstatus_str[val & 0xf]);
 672         }
 673         return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
 674 }
 675
 676 void ipath_decode_err(char *buf, size_t blen, ipath_err_t err)
 677 {
 678         *buf = '\0';
 679         if (err & INFINIPATH_E_RHDRLEN)
 680                 strlcat(buf, "rhdrlen ", blen);
 681         if (err & INFINIPATH_E_RBADTID)
 682                 strlcat(buf, "rbadtid ", blen);
 683         if (err & INFINIPATH_E_RBADVERSION)
 684                 strlcat(buf, "rbadversion ", blen);
 685         if (err & INFINIPATH_E_RHDR)
 686                 strlcat(buf, "rhdr ", blen);
 687         if (err & INFINIPATH_E_RLONGPKTLEN)
 688                 strlcat(buf, "rlongpktlen ", blen);
 689         if (err & INFINIPATH_E_RSHORTPKTLEN)
 690                 strlcat(buf, "rshortpktlen ", blen);
 691         if (err & INFINIPATH_E_RMAXPKTLEN)
 692                 strlcat(buf, "rmaxpktlen ", blen);
 693         if (err & INFINIPATH_E_RMINPKTLEN)
 694                 strlcat(buf, "rminpktlen ", blen);
 695         if (err & INFINIPATH_E_RFORMATERR)
 696                 strlcat(buf, "rformaterr ", blen);
 697         if (err & INFINIPATH_E_RUNSUPVL)
 698                 strlcat(buf, "runsupvl ", blen);
 699         if (err & INFINIPATH_E_RUNEXPCHAR)
 700                 strlcat(buf, "runexpchar ", blen);
 701         if (err & INFINIPATH_E_RIBFLOW)
 702                 strlcat(buf, "ribflow ", blen);
 703         if (err & INFINIPATH_E_REBP)
 704                 strlcat(buf, "EBP ", blen);
 705         if (err & INFINIPATH_E_SUNDERRUN)
 706                 strlcat(buf, "sunderrun ", blen);
 707         if (err & INFINIPATH_E_SPIOARMLAUNCH)
 708                 strlcat(buf, "spioarmlaunch ", blen);
 709         if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
 710                 strlcat(buf, "sunexperrpktnum ", blen);
 711         if (err & INFINIPATH_E_SDROPPEDDATAPKT)
 712                 strlcat(buf, "sdroppeddatapkt ", blen);
 713         if (err & INFINIPATH_E_SDROPPEDSMPPKT)
 714                 strlcat(buf, "sdroppedsmppkt ", blen);
 715         if (err & INFINIPATH_E_SMAXPKTLEN)
 716                 strlcat(buf, "smaxpktlen ", blen);
 717         if (err & INFINIPATH_E_SMINPKTLEN)
 718                 strlcat(buf, "sminpktlen ", blen);
 719         if (err & INFINIPATH_E_SUNSUPVL)
 720                 strlcat(buf, "sunsupVL ", blen);
 721         if (err & INFINIPATH_E_SPKTLEN)
 722                 strlcat(buf, "spktlen ", blen);
 723         if (err & INFINIPATH_E_INVALIDADDR)
 724                 strlcat(buf, "invalidaddr ", blen);
 725         if (err & INFINIPATH_E_RICRC)
 726                 strlcat(buf, "CRC ", blen);
 727         if (err & INFINIPATH_E_RVCRC)
 728                 strlcat(buf, "VCRC ", blen);
 729         if (err & INFINIPATH_E_RRCVEGRFULL)
 730                 strlcat(buf, "rcvegrfull ", blen);
 731         if (err & INFINIPATH_E_RRCVHDRFULL)
 732                 strlcat(buf, "rcvhdrfull ", blen);
 733         if (err & INFINIPATH_E_IBSTATUSCHANGED)
 734                 strlcat(buf, "ibcstatuschg ", blen);
 735         if (err & INFINIPATH_E_RIBLOSTLINK)
 736                 strlcat(buf, "riblostlink ", blen);
 737         if (err & INFINIPATH_E_HARDWARE)
 738                 strlcat(buf, "hardware ", blen);
 739         if (err & INFINIPATH_E_RESET)
 740                 strlcat(buf, "reset ", blen);
 741 }
 742
 743 /**
 744  * get_rhf_errstring - decode RHF errors
 745  * @err: the err number
 746  * @msg: the output buffer
 747  * @len: the length of the output buffer
 748  *
 749  * only used one place now, may want more later
 750  */
 751 static void get_rhf_errstring(u32 err, char *msg, size_t len)
 752 {
 753         /* if no errors, and so don't need to check what's first */
 754         *msg = '\0';
 755
 756         if (err & INFINIPATH_RHF_H_ICRCERR)
 757                 strlcat(msg, "icrcerr ", len);
 758         if (err & INFINIPATH_RHF_H_VCRCERR)
 759                 strlcat(msg, "vcrcerr ", len);
 760         if (err & INFINIPATH_RHF_H_PARITYERR)
 761                 strlcat(msg, "parityerr ", len);
 762         if (err & INFINIPATH_RHF_H_LENERR)
 763                 strlcat(msg, "lenerr ", len);
 764         if (err & INFINIPATH_RHF_H_MTUERR)
 765                 strlcat(msg, "mtuerr ", len);
 766         if (err & INFINIPATH_RHF_H_IHDRERR)
 767                 /* infinipath hdr checksum error */
 768                 strlcat(msg, "ipathhdrerr ", len);
 769         if (err & INFINIPATH_RHF_H_TIDERR)
 770                 strlcat(msg, "tiderr ", len);
 771         if (err & INFINIPATH_RHF_H_MKERR)
 772                 /* bad port, offset, etc. */
 773                 strlcat(msg, "invalid ipathhdr ", len);
 774         if (err & INFINIPATH_RHF_H_IBERR)
 775                 strlcat(msg, "iberr ", len);
 776         if (err & INFINIPATH_RHF_L_SWA)
 777                 strlcat(msg, "swA ", len);
 778         if (err & INFINIPATH_RHF_L_SWB)
 779                 strlcat(msg, "swB ", len);
 780 }
 781
 782 /**
 783  * ipath_get_egrbuf - get an eager buffer
 784  * @dd: the infinipath device
 785  * @bufnum: the eager buffer to get
 786  * @err: unused
 787  *
 788  * must only be called if ipath_pd[port] is known to be allocated
 789  */
 790 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum,
 791                                      int err)
 792 {
 793         return dd->ipath_port0_skbs ?
 794                 (void *)dd->ipath_port0_skbs[bufnum]->data : NULL;
 795 }
 796
 797 /**
 798  * ipath_alloc_skb - allocate an skb and buffer with possible constraints
 799  * @dd: the infinipath device
 800  * @gfp_mask: the sk_buff SFP mask
 801  */
 802 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
 803                                 gfp_t gfp_mask)
 804 {
 805         struct sk_buff *skb;
 806         u32 len;
 807
 808         /*
 809          * Only fully supported way to handle this is to allocate lots
 810          * extra, align as needed, and then do skb_reserve().  That wastes
 811          * a lot of memory...  I'll have to hack this into infinipath_copy
 812          * also.
 813          */
 814
 815         /*
 816          * We need 4 extra bytes for unaligned transfer copying
 817          */
 818         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 819                 /* we need a 4KB multiple alignment, and there is no way
 820                  * to do it except to allocate extra and then skb_reserve
 821                  * enough to bring it up to the right alignment.
 822                  */
 823                 len = dd->ipath_ibmaxlen + 4 + (1 << 11) - 1;
 824         }
 825         else
 826                 len = dd->ipath_ibmaxlen + 4;
 827         skb = __dev_alloc_skb(len, gfp_mask);
 828         if (!skb) {
 829                 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
 830                               len);
 831                 goto bail;
 832         }
 833         if (dd->ipath_flags & IPATH_4BYTE_TID) {
 834                 u32 una = ((1 << 11) - 1) & (unsigned long)(skb->data + 4);
 835                 if (una)
 836                         skb_reserve(skb, 4 + (1 << 11) - una);
 837                 else
 838                         skb_reserve(skb, 4);
 839         } else
 840                 skb_reserve(skb, 4);
 841
 842 bail:
 843         return skb;
 844 }
 845
 846 /**
 847  * ipath_rcv_layer - receive a packet for the layered (ethernet) driver
 848  * @dd: the infinipath device
 849  * @etail: the sk_buff number
 850  * @tlen: the total packet length
 851  * @hdr: the ethernet header
 852  *
 853  * Separate routine for better overall optimization
 854  */
 855 static void ipath_rcv_layer(struct ipath_devdata *dd, u32 etail,
 856                             u32 tlen, struct ether_header *hdr)
 857 {
 858         u32 elen;
 859         u8 pad, *bthbytes;
 860         struct sk_buff *skb, *nskb;
 861
 862         if (dd->ipath_port0_skbs && hdr->sub_opcode == OPCODE_ENCAP) {
 863                 /*
 864                  * Allocate a new sk_buff to replace the one we give
 865                  * to the network stack.
 866                  */
 867                 nskb = ipath_alloc_skb(dd, GFP_ATOMIC);
 868                 if (!nskb) {
 869                         /* count OK packets that we drop */
 870                         ipath_stats.sps_krdrops++;
 871                         return;
 872                 }
 873
 874                 bthbytes = (u8 *) hdr->bth;
 875                 pad = (bthbytes[1] >> 4) & 3;
 876                 /* +CRC32 */
 877                 elen = tlen - (sizeof(*hdr) + pad + sizeof(u32));
 878
 879                 skb = dd->ipath_port0_skbs[etail];
 880                 dd->ipath_port0_skbs[etail] = nskb;
 881                 skb_put(skb, elen);
 882
 883                 dd->ipath_f_put_tid(dd, etail + (u64 __iomem *)
 884                                     ((char __iomem *) dd->ipath_kregbase
 885                                      + dd->ipath_rcvegrbase), 0,
 886                                     virt_to_phys(nskb->data));
 887
 888                 __ipath_layer_rcv(dd, hdr, skb);
 889
 890                 /* another ether packet received */
 891                 ipath_stats.sps_ether_rpkts++;
 892         }
 893         else if (hdr->sub_opcode == OPCODE_LID_ARP)
 894                 __ipath_layer_rcv_lid(dd, hdr);
 895 }
 896
 897 /*
 898  * ipath_kreceive - receive a packet
 899  * @dd: the infinipath device
 900  *
 901  * called from interrupt handler for errors or receive interrupt
 902  */
 903 void ipath_kreceive(struct ipath_devdata *dd)
 904 {
 905         u64 *rc;
 906         void *ebuf;
 907         const u32 rsize = dd->ipath_rcvhdrentsize;      /* words */
 908         const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
 909         u32 etail = -1, l, hdrqtail;
 910         struct ips_message_header *hdr;
 911         u32 eflags, i, etype, tlen, pkttot = 0;
 912         static u64 totcalls;    /* stats, may eventually remove */
 913         char emsg[128];
 914
 915         if (!dd->ipath_hdrqtailptr) {
 916                 ipath_dev_err(dd,
 917                               "hdrqtailptr not set, can't do receives\n");
 918                 goto bail;
 919         }
 920
 921         /* There is already a thread processing this queue. */
 922         if (test_and_set_bit(0, &dd->ipath_rcv_pending))
 923                 goto bail;
 924
 925         if (dd->ipath_port0head ==
 926             (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
 927                 goto done;
 928
 929 gotmore:
 930         /*
 931          * read only once at start.  If in flood situation, this helps
 932          * performance slightly.  If more arrive while we are processing,
 933          * we'll come back here and do them
 934          */
 935         hdrqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
 936
 937         for (i = 0, l = dd->ipath_port0head; l != hdrqtail; i++) {
 938                 u32 qp;
 939                 u8 *bthbytes;
 940
 941                 rc = (u64 *) (dd->ipath_pd[0]->port_rcvhdrq + (l << 2));
 942                 hdr = (struct ips_message_header *)&rc[1];
 943                 /*
 944                  * could make a network order version of IPATH_KD_QP, and
 945                  * do the obvious shift before masking to speed this up.
 946                  */
 947                 qp = ntohl(hdr->bth[1]) & 0xffffff;
 948                 bthbytes = (u8 *) hdr->bth;
 949
 950                 eflags = ips_get_hdr_err_flags((__le32 *) rc);
 951                 etype = ips_get_rcv_type((__le32 *) rc);
 952                 /* total length */
 953                 tlen = ips_get_length_in_bytes((__le32 *) rc);
 954                 ebuf = NULL;
 955                 if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
 956                         /*
 957                          * it turns out that the chips uses an eager buffer
 958                          * for all non-expected packets, whether it "needs"
 959                          * one or not.  So always get the index, but don't
 960                          * set ebuf (so we try to copy data) unless the
 961                          * length requires it.
 962                          */
 963                         etail = ips_get_index((__le32 *) rc);
 964                         if (tlen > sizeof(*hdr) ||
 965                             etype == RCVHQ_RCV_TYPE_NON_KD)
 966                                 ebuf = ipath_get_egrbuf(dd, etail, 0);
 967                 }
 968
 969                 /*
 970                  * both tiderr and ipathhdrerr are set for all plain IB
 971                  * packets; only ipathhdrerr should be set.
 972                  */
 973
 974                 if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
 975                     RCVHQ_RCV_TYPE_ERROR && ips_get_ipath_ver(
 976                             hdr->iph.ver_port_tid_offset) !=
 977                     IPS_PROTO_VERSION) {
 978                         ipath_cdbg(PKT, "Bad InfiniPath protocol version "
 979                                    "%x\n", etype);
 980                 }
 981
 982                 if (eflags & ~(INFINIPATH_RHF_H_TIDERR |
 983                                INFINIPATH_RHF_H_IHDRERR)) {
 984                         get_rhf_errstring(eflags, emsg, sizeof emsg);
 985                         ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
 986                                    "tlen=%x opcode=%x egridx=%x: %s\n",
 987                                    eflags, l, etype, tlen, bthbytes[0],
 988                                    ips_get_index((__le32 *) rc), emsg);
 989                 } else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
 990                                 int ret = __ipath_verbs_rcv(dd, rc + 1,
 991                                                             ebuf, tlen);
 992                                 if (ret == -ENODEV)
 993                                         ipath_cdbg(VERBOSE,
 994                                                    "received IB packet, "
 995                                                    "not SMA (QP=%x)\n", qp);
 996                 } else if (etype == RCVHQ_RCV_TYPE_EAGER) {
 997                         if (qp == IPATH_KD_QP &&
 998                             bthbytes[0] == ipath_layer_rcv_opcode &&
 999                             ebuf)
1000                                 ipath_rcv_layer(dd, etail, tlen,
1001                                                 (struct ether_header *)hdr);
1002                         else
1003                                 ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1004                                            "qp=%x), len %x; ignored\n",
1005                                            etype, bthbytes[0], qp, tlen);
1006                 }
1007                 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
1008                         ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
1009                                   be32_to_cpu(hdr->bth[0]) & 0xff);
1010                 else if (eflags & (INFINIPATH_RHF_H_TIDERR |
1011                                    INFINIPATH_RHF_H_IHDRERR)) {
1012                         /*
1013                          * This is a type 3 packet, only the LRH is in the
1014                          * rcvhdrq, the rest of the header is in the eager
1015                          * buffer.
1016                          */
1017                         u8 opcode;
1018                         if (ebuf) {
1019                                 bthbytes = (u8 *) ebuf;
1020                                 opcode = *bthbytes;
1021                         }
1022                         else
1023                                 opcode = 0;
1024                         get_rhf_errstring(eflags, emsg, sizeof emsg);
1025                         ipath_dbg("Err %x (%s), opcode %x, egrbuf %x, "
1026                                   "len %x\n", eflags, emsg, opcode, etail,
1027                                   tlen);
1028                 } else {
1029                         /*
1030                          * error packet, type of error  unknown.
1031                          * Probably type 3, but we don't know, so don't
1032                          * even try to print the opcode, etc.
1033                          */
1034                         ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
1035                                   "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
1036                                   "hdr %llx %llx %llx %llx %llx\n",
1037                                   etail, tlen, (unsigned long) rc, l,
1038                                   (unsigned long long) rc[0],
1039                                   (unsigned long long) rc[1],
1040                                   (unsigned long long) rc[2],
1041                                   (unsigned long long) rc[3],
1042                                   (unsigned long long) rc[4],
1043                                   (unsigned long long) rc[5]);
1044                 }
1045                 l += rsize;
1046                 if (l >= maxcnt)
1047                         l = 0;
1048                 /*
1049                  * update for each packet, to help prevent overflows if we
1050                  * have lots of packets.
1051                  */
1052                 (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1053                                        dd->ipath_rhdrhead_intr_off | l, 0);
1054                 if (etype != RCVHQ_RCV_TYPE_EXPECTED)
1055                         (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
1056                                                etail, 0);
1057         }
1058
1059         pkttot += i;
1060
1061         dd->ipath_port0head = l;
1062
1063         if (hdrqtail != (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
1064                 /* more arrived while we handled first batch */
1065                 goto gotmore;
1066
1067         if (pkttot > ipath_stats.sps_maxpkts_call)
1068                 ipath_stats.sps_maxpkts_call = pkttot;
1069         ipath_stats.sps_port0pkts += pkttot;
1070         ipath_stats.sps_avgpkts_call =
1071                 ipath_stats.sps_port0pkts / ++totcalls;
1072
1073 done:
1074         clear_bit(0, &dd->ipath_rcv_pending);
1075         smp_mb__after_clear_bit();
1076
1077 bail:;
1078 }
1079
1080 /**
1081  * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1082  * @dd: the infinipath device
1083  *
1084  * called whenever our local copy indicates we have run out of send buffers
1085  * NOTE: This can be called from interrupt context by some code
1086  * and from non-interrupt context by ipath_getpiobuf().
1087  */
1088
1089 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1090 {
1091         unsigned long flags;
1092         int i;
1093         const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1094
1095         /* If the generation (check) bits have changed, then we update the
1096          * busy bit for the corresponding PIO buffer.  This algorithm will
1097          * modify positions to the value they already have in some cases
1098          * (i.e., no change), but it's faster than changing only the bits
1099          * that have changed.
1100          *
1101          * We would like to do this atomicly, to avoid spinlocks in the
1102          * critical send path, but that's not really possible, given the
1103          * type of changes, and that this routine could be called on
1104          * multiple cpu's simultaneously, so we lock in this routine only,
1105          * to avoid conflicting updates; all we change is the shadow, and
1106          * it's a single 64 bit memory location, so by definition the update
1107          * is atomic in terms of what other cpu's can see in testing the
1108          * bits.  The spin_lock overhead isn't too bad, since it only
1109          * happens when all buffers are in use, so only cpu overhead, not
1110          * latency or bandwidth is affected.
1111          */
1112 #define _IPATH_ALL_CHECKBITS 0x5555555555555555ULL
1113         if (!dd->ipath_pioavailregs_dma) {
1114                 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1115                 return;
1116         }
1117         if (ipath_debug & __IPATH_VERBDBG) {
1118                 /* only if packet debug and verbose */
1119                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1120                 unsigned long *shadow = dd->ipath_pioavailshadow;
1121
1122                 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1123                            "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1124                            "s3=%lx\n",
1125                            (unsigned long long) le64_to_cpu(dma[0]),
1126                            shadow[0],
1127                            (unsigned long long) le64_to_cpu(dma[1]),
1128                            shadow[1],
1129                            (unsigned long long) le64_to_cpu(dma[2]),
1130                            shadow[2],
1131                            (unsigned long long) le64_to_cpu(dma[3]),
1132                            shadow[3]);
1133                 if (piobregs > 4)
1134                         ipath_cdbg(
1135                                 PKT, "2nd group, dma4=%llx shad4=%lx, "
1136                                 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1137                                 "d7=%llx s7=%lx\n",
1138                                 (unsigned long long) le64_to_cpu(dma[4]),
1139                                 shadow[4],
1140                                 (unsigned long long) le64_to_cpu(dma[5]),
1141                                 shadow[5],
1142                                 (unsigned long long) le64_to_cpu(dma[6]),
1143                                 shadow[6],
1144                                 (unsigned long long) le64_to_cpu(dma[7]),
1145                                 shadow[7]);
1146         }
1147         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1148         for (i = 0; i < piobregs; i++) {
1149                 u64 pchbusy, pchg, piov, pnew;
1150                 /*
1151                  * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1152                  */
1153                 if (i > 3) {
1154                         if (i & 1)
1155                                 piov = le64_to_cpu(
1156                                         dd->ipath_pioavailregs_dma[i - 1]);
1157                         else
1158                                 piov = le64_to_cpu(
1159                                         dd->ipath_pioavailregs_dma[i + 1]);
1160                 } else
1161                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1162                 pchg = _IPATH_ALL_CHECKBITS &
1163                         ~(dd->ipath_pioavailshadow[i] ^ piov);
1164                 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1165                 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1166                         pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1167                         pnew |= piov & pchbusy;
1168                         dd->ipath_pioavailshadow[i] = pnew;
1169                 }
1170         }
1171         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1172 }
1173
1174 /**
1175  * ipath_setrcvhdrsize - set the receive header size
1176  * @dd: the infinipath device
1177  * @rhdrsize: the receive header size
1178  *
1179  * called from user init code, and also layered driver init
1180  */
1181 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1182 {
1183         int ret = 0;
1184
1185         if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1186                 if (dd->ipath_rcvhdrsize != rhdrsize) {
1187                         dev_info(&dd->pcidev->dev,
1188                                  "Error: can't set protocol header "
1189                                  "size %u, already %u\n",
1190                                  rhdrsize, dd->ipath_rcvhdrsize);
1191                         ret = -EAGAIN;
1192                 } else
1193                         ipath_cdbg(VERBOSE, "Reuse same protocol header "
1194                                    "size %u\n", dd->ipath_rcvhdrsize);
1195         } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1196                                (sizeof(u64) / sizeof(u32)))) {
1197                 ipath_dbg("Error: can't set protocol header size %u "
1198                           "(> max %u)\n", rhdrsize,
1199                           dd->ipath_rcvhdrentsize -
1200                           (u32) (sizeof(u64) / sizeof(u32)));
1201                 ret = -EOVERFLOW;
1202         } else {
1203                 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1204                 dd->ipath_rcvhdrsize = rhdrsize;
1205                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1206                                  dd->ipath_rcvhdrsize);
1207                 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1208                            dd->ipath_rcvhdrsize);
1209         }
1210         return ret;
1211 }
1212
1213 /**
1214  * ipath_getpiobuf - find an available pio buffer
1215  * @dd: the infinipath device
1216  * @pbufnum: the buffer number is placed here
1217  *
1218  * do appropriate marking as busy, etc.
1219  * returns buffer number if one found (>=0), negative number is error.
1220  * Used by ipath_sma_send_pkt and ipath_layer_send
1221  */
1222 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
1223 {
1224         int i, j, starti, updated = 0;
1225         unsigned piobcnt, iter;
1226         unsigned long flags;
1227         unsigned long *shadow = dd->ipath_pioavailshadow;
1228         u32 __iomem *buf;
1229
1230         piobcnt = (unsigned)(dd->ipath_piobcnt2k
1231                              + dd->ipath_piobcnt4k);
1232         starti = dd->ipath_lastport_piobuf;
1233         iter = piobcnt - starti;
1234         if (dd->ipath_upd_pio_shadow) {
1235                 /*
1236                  * Minor optimization.  If we had no buffers on last call,
1237                  * start out by doing the update; continue and do scan even
1238                  * if no buffers were updated, to be paranoid
1239                  */
1240                 ipath_update_pio_bufs(dd);
1241                 /* we scanned here, don't do it at end of scan */
1242                 updated = 1;
1243                 i = starti;
1244         } else
1245                 i = dd->ipath_lastpioindex;
1246
1247 rescan:
1248         /*
1249          * while test_and_set_bit() is atomic, we do that and then the
1250          * change_bit(), and the pair is not.  See if this is the cause
1251          * of the remaining armlaunch errors.
1252          */
1253         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1254         for (j = 0; j < iter; j++, i++) {
1255                 if (i >= piobcnt)
1256                         i = starti;
1257                 /*
1258                  * To avoid bus lock overhead, we first find a candidate
1259                  * buffer, then do the test and set, and continue if that
1260                  * fails.
1261                  */
1262                 if (test_bit((2 * i) + 1, shadow) ||
1263                     test_and_set_bit((2 * i) + 1, shadow))
1264                         continue;
1265                 /* flip generation bit */
1266                 change_bit(2 * i, shadow);
1267                 break;
1268         }
1269         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1270
1271         if (j == iter) {
1272                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1273
1274                 /*
1275                  * first time through; shadow exhausted, but may be real
1276                  * buffers available, so go see; if any updated, rescan
1277                  * (once)
1278                  */
1279                 if (!updated) {
1280                         ipath_update_pio_bufs(dd);
1281                         updated = 1;
1282                         i = starti;
1283                         goto rescan;
1284                 }
1285                 dd->ipath_upd_pio_shadow = 1;
1286                 /*
1287                  * not atomic, but if we lose one once in a while, that's OK
1288                  */
1289                 ipath_stats.sps_nopiobufs++;
1290                 if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1291                         ipath_dbg(
1292                                 "%u pio sends with no bufavail; dmacopy: "
1293                                 "%llx %llx %llx %llx; shadow:  "
1294                                 "%lx %lx %lx %lx\n",
1295                                 dd->ipath_consec_nopiobuf,
1296                                 (unsigned long long) le64_to_cpu(dma[0]),
1297                                 (unsigned long long) le64_to_cpu(dma[1]),
1298                                 (unsigned long long) le64_to_cpu(dma[2]),
1299                                 (unsigned long long) le64_to_cpu(dma[3]),
1300                                 shadow[0], shadow[1], shadow[2],
1301                                 shadow[3]);
1302                         /*
1303                          * 4 buffers per byte, 4 registers above, cover rest
1304                          * below
1305                          */
1306                         if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1307                             (sizeof(shadow[0]) * 4 * 4))
1308                                 ipath_dbg("2nd group: dmacopy: %llx %llx "
1309                                           "%llx %llx; shadow: %lx %lx "
1310                                           "%lx %lx\n",
1311                                           (unsigned long long)
1312                                           le64_to_cpu(dma[4]),
1313                                           (unsigned long long)
1314                                           le64_to_cpu(dma[5]),
1315                                           (unsigned long long)
1316                                           le64_to_cpu(dma[6]),
1317                                           (unsigned long long)
1318                                           le64_to_cpu(dma[7]),
1319                                           shadow[4], shadow[5],
1320                                           shadow[6], shadow[7]);
1321                 }
1322                 buf = NULL;
1323                 goto bail;
1324         }
1325
1326         if (updated)
1327                 /*
1328                  * ran out of bufs, now some (at least this one we just
1329                  * got) are now available, so tell the layered driver.
1330                  */
1331                 __ipath_layer_intr(dd, IPATH_LAYER_INT_SEND_CONTINUE);
1332
1333         /*
1334          * set next starting place.  Since it's just an optimization,
1335          * it doesn't matter who wins on this, so no locking
1336          */
1337         dd->ipath_lastpioindex = i + 1;
1338         if (dd->ipath_upd_pio_shadow)
1339                 dd->ipath_upd_pio_shadow = 0;
1340         if (dd->ipath_consec_nopiobuf)
1341                 dd->ipath_consec_nopiobuf = 0;
1342         if (i < dd->ipath_piobcnt2k)
1343                 buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1344                                        i * dd->ipath_palign);
1345         else
1346                 buf = (u32 __iomem *)
1347                         (dd->ipath_pio4kbase +
1348                          (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1349         ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1350                    i, (i < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1351         if (pbufnum)
1352                 *pbufnum = i;
1353
1354 bail:
1355         return buf;
1356 }
1357
1358 /**
1359  * ipath_create_rcvhdrq - create a receive header queue
1360  * @dd: the infinipath device
1361  * @pd: the port data
1362  *
1363  * this *must* be physically contiguous memory, and for now,
1364  * that limits it to what kmalloc can do.
1365  */
1366 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1367                          struct ipath_portdata *pd)
1368 {
1369         int ret = 0, amt;
1370
1371         amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1372                     sizeof(u32), PAGE_SIZE);
1373         if (!pd->port_rcvhdrq) {
1374                 /*
1375                  * not using REPEAT isn't viable; at 128KB, we can easily
1376                  * fail this.  The problem with REPEAT is we can block here
1377                  * "forever".  There isn't an inbetween, unfortunately.  We
1378                  * could reduce the risk by never freeing the rcvhdrq except
1379                  * at unload, but even then, the first time a port is used,
1380                  * we could delay for some time...
1381                  */
1382                 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1383
1384                 pd->port_rcvhdrq = dma_alloc_coherent(
1385                         &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1386                         gfp_flags);
1387
1388                 if (!pd->port_rcvhdrq) {
1389                         ipath_dev_err(dd, "attempt to allocate %d bytes "
1390                                       "for port %u rcvhdrq failed\n",
1391                                       amt, pd->port_port);
1392                         ret = -ENOMEM;
1393                         goto bail;
1394                 }
1395
1396                 pd->port_rcvhdrq_size = amt;
1397
1398                 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1399                            "for port %u rcvhdr Q\n",
1400                            amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1401                            (unsigned long) pd->port_rcvhdrq_phys,
1402                            (unsigned long) pd->port_rcvhdrq_size,
1403                            pd->port_port);
1404         } else {
1405                 /*
1406                  * clear for security, sanity, and/or debugging, each
1407                  * time we reuse
1408                  */
1409                 memset(pd->port_rcvhdrq, 0, amt);
1410         }
1411
1412         /*
1413          * tell chip each time we init it, even if we are re-using previous
1414          * memory (we zero it at process close)
1415          */
1416         ipath_cdbg(VERBOSE, "writing port %d rcvhdraddr as %lx\n",
1417                    pd->port_port, (unsigned long) pd->port_rcvhdrq_phys);
1418         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1419                               pd->port_port, pd->port_rcvhdrq_phys);
1420
1421         ret = 0;
1422 bail:
1423         return ret;
1424 }
1425
1426 int ipath_waitfor_complete(struct ipath_devdata *dd, ipath_kreg reg_id,
1427                            u64 bits_to_wait_for, u64 * valp)
1428 {
1429         unsigned long timeout;
1430         u64 lastval, val;
1431         int ret;
1432
1433         lastval = ipath_read_kreg64(dd, reg_id);
1434         /* wait a ridiculously long time */
1435         timeout = jiffies + msecs_to_jiffies(5);
1436         do {
1437                 val = ipath_read_kreg64(dd, reg_id);
1438                 /* set so they have something, even on failures. */
1439                 *valp = val;
1440                 if ((val & bits_to_wait_for) == bits_to_wait_for) {
1441                         ret = 0;
1442                         break;
1443                 }
1444                 if (val != lastval)
1445                         ipath_cdbg(VERBOSE, "Changed from %llx to %llx, "
1446                                    "waiting for %llx bits\n",
1447                                    (unsigned long long) lastval,
1448                                    (unsigned long long) val,
1449                                    (unsigned long long) bits_to_wait_for);
1450                 cond_resched();
1451                 if (time_after(jiffies, timeout)) {
1452                         ipath_dbg("Didn't get bits %llx in register 0x%x, "
1453                                   "got %llx\n",
1454                                   (unsigned long long) bits_to_wait_for,
1455                                   reg_id, (unsigned long long) *valp);
1456                         ret = -ENODEV;
1457                         break;
1458                 }
1459         } while (1);
1460
1461         return ret;
1462 }
1463
1464 /**
1465  * ipath_waitfor_mdio_cmdready - wait for last command to complete
1466  * @dd: the infinipath device
1467  *
1468  * Like ipath_waitfor_complete(), but we wait for the CMDVALID bit to go
1469  * away indicating the last command has completed.  It doesn't return data
1470  */
1471 int ipath_waitfor_mdio_cmdready(struct ipath_devdata *dd)
1472 {
1473         unsigned long timeout;
1474         u64 val;
1475         int ret;
1476
1477         /* wait a ridiculously long time */
1478         timeout = jiffies + msecs_to_jiffies(5);
1479         do {
1480                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_mdio);
1481                 if (!(val & IPATH_MDIO_CMDVALID)) {
1482                         ret = 0;
1483                         break;
1484                 }
1485                 cond_resched();
1486                 if (time_after(jiffies, timeout)) {
1487                         ipath_dbg("CMDVALID stuck in mdio reg? (%llx)\n",
1488                                   (unsigned long long) val);
1489                         ret = -ENODEV;
1490                         break;
1491                 }
1492         } while (1);
1493
1494         return ret;
1495 }
1496
1497 void ipath_set_ib_lstate(struct ipath_devdata *dd, int which)
1498 {
1499         static const char *what[4] = {
1500                 [0] = "DOWN",
1501                 [INFINIPATH_IBCC_LINKCMD_INIT] = "INIT",
1502                 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1503                 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1504         };
1505         ipath_cdbg(SMA, "Trying to move unit %u to %s, current ltstate "
1506                    "is %s\n", dd->ipath_unit,
1507                    what[(which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
1508                         INFINIPATH_IBCC_LINKCMD_MASK],
1509                    ipath_ibcstatus_str[
1510                            (ipath_read_kreg64
1511                             (dd, dd->ipath_kregs->kr_ibcstatus) >>
1512                             INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
1513                            INFINIPATH_IBCS_LINKTRAININGSTATE_MASK]);
1514
1515         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1516                          dd->ipath_ibcctrl | which);
1517 }
1518
1519 /**
1520  * ipath_read_kreg64_port - read a device's per-port 64-bit kernel register
1521  * @dd: the infinipath device
1522  * @regno: the register number to read
1523  * @port: the port containing the register
1524  *
1525  * Registers that vary with the chip implementation constants (port)
1526  * use this routine.
1527  */
1528 u64 ipath_read_kreg64_port(const struct ipath_devdata *dd, ipath_kreg regno,
1529                            unsigned port)
1530 {
1531         u16 where;
1532
1533         if (port < dd->ipath_portcnt &&
1534             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1535              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1536                 where = regno + port;
1537         else
1538                 where = -1;
1539
1540         return ipath_read_kreg64(dd, where);
1541 }
1542
1543 /**
1544  * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
1545  * @dd: the infinipath device
1546  * @regno: the register number to write
1547  * @port: the port containing the register
1548  * @value: the value to write
1549  *
1550  * Registers that vary with the chip implementation constants (port)
1551  * use this routine.
1552  */
1553 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
1554                           unsigned port, u64 value)
1555 {
1556         u16 where;
1557
1558         if (port < dd->ipath_portcnt &&
1559             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1560              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1561                 where = regno + port;
1562         else
1563                 where = -1;
1564
1565         ipath_write_kreg(dd, where, value);
1566 }
1567
1568 /**
1569  * ipath_shutdown_device - shut down a device
1570  * @dd: the infinipath device
1571  *
1572  * This is called to make the device quiet when we are about to
1573  * unload the driver, and also when the device is administratively
1574  * disabled.   It does not free any data structures.
1575  * Everything it does has to be setup again by ipath_init_chip(dd,1)
1576  */
1577 void ipath_shutdown_device(struct ipath_devdata *dd)
1578 {
1579         u64 val;
1580
1581         ipath_dbg("Shutting down the device\n");
1582
1583         dd->ipath_flags |= IPATH_LINKUNK;
1584         dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
1585                              IPATH_LINKINIT | IPATH_LINKARMED |
1586                              IPATH_LINKACTIVE);
1587         *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
1588                                 IPATH_STATUS_IB_READY);
1589
1590         /* mask interrupts, but not errors */
1591         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
1592
1593         dd->ipath_rcvctrl = 0;
1594         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1595                          dd->ipath_rcvctrl);
1596
1597         /*
1598          * gracefully stop all sends allowing any in progress to trickle out
1599          * first.
1600          */
1601         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0ULL);
1602         /* flush it */
1603         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1604         /*
1605          * enough for anything that's going to trickle out to have actually
1606          * done so.
1607          */
1608         udelay(5);
1609
1610         /*
1611          * abort any armed or launched PIO buffers that didn't go. (self
1612          * clearing).  Will cause any packet currently being transmitted to
1613          * go out with an EBP, and may also cause a short packet error on
1614          * the receiver.
1615          */
1616         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1617                          INFINIPATH_S_ABORT);
1618
1619         ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1620                             INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1621
1622         /*
1623          * we are shutting down, so tell the layered driver.  We don't do
1624          * this on just a link state change, much like ethernet, a cable
1625          * unplug, etc. doesn't change driver state
1626          */
1627         ipath_layer_intr(dd, IPATH_LAYER_INT_IF_DOWN);
1628
1629         /* disable IBC */
1630         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
1631         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
1632                          dd->ipath_control);
1633
1634         /*
1635          * clear SerdesEnable and turn the leds off; do this here because
1636          * we are unloading, so don't count on interrupts to move along
1637          * Turn the LEDs off explictly for the same reason.
1638          */
1639         dd->ipath_f_quiet_serdes(dd);
1640         dd->ipath_f_setextled(dd, 0, 0);
1641
1642         if (dd->ipath_stats_timer_active) {
1643                 del_timer_sync(&dd->ipath_stats_timer);
1644                 dd->ipath_stats_timer_active = 0;
1645         }
1646
1647         /*
1648          * clear all interrupts and errors, so that the next time the driver
1649          * is loaded or device is enabled, we know that whatever is set
1650          * happened while we were unloaded
1651          */
1652         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
1653                          ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
1654         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
1655         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
1656 }
1657
1658 /**
1659  * ipath_free_pddata - free a port's allocated data
1660  * @dd: the infinipath device
1661  * @port: the port
1662  * @freehdrq: free the port data structure if true
1663  *
1664  * when closing, free up any allocated data for a port, if the
1665  * reference count goes to zero
1666  * Note: this also optionally frees the portdata itself!
1667  * Any changes here have to be matched up with the reinit case
1668  * of ipath_init_chip(), which calls this routine on reinit after reset.
1669  */
1670 void ipath_free_pddata(struct ipath_devdata *dd, u32 port, int freehdrq)
1671 {
1672         struct ipath_portdata *pd = dd->ipath_pd[port];
1673
1674         if (!pd)
1675                 return;
1676         if (freehdrq)
1677                 /*
1678                  * only clear and free portdata if we are going to also
1679                  * release the hdrq, otherwise we leak the hdrq on each
1680                  * open/close cycle
1681                  */
1682                 dd->ipath_pd[port] = NULL;
1683         if (freehdrq && pd->port_rcvhdrq) {
1684                 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
1685                            "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
1686                            (unsigned long) pd->port_rcvhdrq_size);
1687                 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
1688                                   pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1689                 pd->port_rcvhdrq = NULL;
1690         }
1691         if (port && pd->port_rcvegrbuf) {
1692                 /* always free this */
1693                 if (pd->port_rcvegrbuf) {
1694                         unsigned e;
1695
1696                         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
1697                                 void *base = pd->port_rcvegrbuf[e];
1698                                 size_t size = pd->port_rcvegrbuf_size;
1699
1700                                 ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
1701                                            "chunk %u/%u\n", base,
1702                                            (unsigned long) size,
1703                                            e, pd->port_rcvegrbuf_chunks);
1704                                 dma_free_coherent(
1705                                         &dd->pcidev->dev, size, base,
1706                                         pd->port_rcvegrbuf_phys[e]);
1707                         }
1708                         vfree(pd->port_rcvegrbuf);
1709                         pd->port_rcvegrbuf = NULL;
1710                         vfree(pd->port_rcvegrbuf_phys);
1711                         pd->port_rcvegrbuf_phys = NULL;
1712                 }
1713                 pd->port_rcvegrbuf_chunks = 0;
1714         } else if (port == 0 && dd->ipath_port0_skbs) {
1715                 unsigned e;
1716                 struct sk_buff **skbs = dd->ipath_port0_skbs;
1717
1718                 dd->ipath_port0_skbs = NULL;
1719                 ipath_cdbg(VERBOSE, "free closed port %d ipath_port0_skbs "
1720                            "@ %p\n", pd->port_port, skbs);
1721                 for (e = 0; e < dd->ipath_rcvegrcnt; e++)
1722                         if (skbs[e])
1723                                 dev_kfree_skb(skbs[e]);
1724                 vfree(skbs);
1725         }
1726         if (freehdrq) {
1727                 kfree(pd->port_tid_pg_list);
1728                 kfree(pd);
1729         }
1730 }
1731
1732 int __init infinipath_init(void)
1733 {
1734         int ret;
1735
1736         ipath_dbg(KERN_INFO DRIVER_LOAD_MSG "%s", ipath_core_version);
1737
1738         /*
1739          * These must be called before the driver is registered with
1740          * the PCI subsystem.
1741          */
1742         idr_init(&unit_table);
1743         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
1744                 ret = -ENOMEM;
1745                 goto bail;
1746         }
1747
1748         ret = pci_register_driver(&ipath_driver);
1749         if (ret < 0) {
1750                 printk(KERN_ERR IPATH_DRV_NAME
1751                        ": Unable to register driver: error %d\n", -ret);
1752                 goto bail_unit;
1753         }
1754
1755         ret = ipath_driver_create_group(&ipath_driver.driver);
1756         if (ret < 0) {
1757                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create driver "
1758                        "sysfs entries: error %d\n", -ret);
1759                 goto bail_pci;
1760         }
1761
1762         ret = ipath_init_ipathfs();
1763         if (ret < 0) {
1764                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
1765                        "ipathfs: error %d\n", -ret);
1766                 goto bail_group;
1767         }
1768
1769         goto bail;
1770
1771 bail_group:
1772         ipath_driver_remove_group(&ipath_driver.driver);
1773
1774 bail_pci:
1775         pci_unregister_driver(&ipath_driver);
1776
1777 bail_unit:
1778         idr_destroy(&unit_table);
1779
1780 bail:
1781         return ret;
1782 }
1783
1784 static void cleanup_device(struct ipath_devdata *dd)
1785 {
1786         int port;
1787
1788         ipath_shutdown_device(dd);
1789
1790         if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
1791                 /* can't do anything more with chip; needs re-init */
1792                 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
1793                 if (dd->ipath_kregbase) {
1794                         /*
1795                          * if we haven't already cleaned up before these are
1796                          * to ensure any register reads/writes "fail" until
1797                          * re-init
1798                          */
1799                         dd->ipath_kregbase = NULL;
1800                         dd->ipath_kregvirt = NULL;
1801                         dd->ipath_uregbase = 0;
1802                         dd->ipath_sregbase = 0;
1803                         dd->ipath_cregbase = 0;
1804                         dd->ipath_kregsize = 0;
1805                 }
1806                 ipath_disable_wc(dd);
1807         }
1808
1809         if (dd->ipath_pioavailregs_dma) {
1810                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1811                                   (void *) dd->ipath_pioavailregs_dma,
1812                                   dd->ipath_pioavailregs_phys);
1813                 dd->ipath_pioavailregs_dma = NULL;
1814         }
1815
1816         if (dd->ipath_pageshadow) {
1817                 struct page **tmpp = dd->ipath_pageshadow;
1818                 int i, cnt = 0;
1819
1820                 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
1821                            "locked\n");
1822                 for (port = 0; port < dd->ipath_cfgports; port++) {
1823                         int port_tidbase = port * dd->ipath_rcvtidcnt;
1824                         int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1825                         for (i = port_tidbase; i < maxtid; i++) {
1826                                 if (!tmpp[i])
1827                                         continue;
1828                                 ipath_release_user_pages(&tmpp[i], 1);
1829                                 tmpp[i] = NULL;
1830                                 cnt++;
1831                         }
1832                 }
1833                 if (cnt) {
1834                         ipath_stats.sps_pageunlocks += cnt;
1835                         ipath_cdbg(VERBOSE, "There were still %u expTID "
1836                                    "entries locked\n", cnt);
1837                 }
1838                 if (ipath_stats.sps_pagelocks ||
1839                     ipath_stats.sps_pageunlocks)
1840                         ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
1841                                    "unlocked via ipath_m{un}lock\n",
1842                                    (unsigned long long)
1843                                    ipath_stats.sps_pagelocks,
1844                                    (unsigned long long)
1845                                    ipath_stats.sps_pageunlocks);
1846
1847                 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
1848                            dd->ipath_pageshadow);
1849                 vfree(dd->ipath_pageshadow);
1850                 dd->ipath_pageshadow = NULL;
1851         }
1852
1853         /*
1854          * free any resources still in use (usually just kernel ports)
1855          * at unload
1856          */
1857         for (port = 0; port < dd->ipath_cfgports; port++)
1858                 ipath_free_pddata(dd, port, 1);
1859         kfree(dd->ipath_pd);
1860         /*
1861          * debuggability, in case some cleanup path tries to use it
1862          * after this
1863          */
1864         dd->ipath_pd = NULL;
1865 }
1866
1867 static void __exit infinipath_cleanup(void)
1868 {
1869         struct ipath_devdata *dd, *tmp;
1870         unsigned long flags;
1871
1872         ipath_exit_ipathfs();
1873
1874         ipath_driver_remove_group(&ipath_driver.driver);
1875
1876         spin_lock_irqsave(&ipath_devs_lock, flags);
1877
1878         /*
1879          * turn off rcv, send, and interrupts for all ports, all drivers
1880          * should also hard reset the chip here?
1881          * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
1882          * for all versions of the driver, if they were allocated
1883          */
1884         list_for_each_entry_safe(dd, tmp, &ipath_dev_list, ipath_list) {
1885                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
1886
1887                 if (dd->ipath_kregbase)
1888                         cleanup_device(dd);
1889
1890                 if (dd->pcidev) {
1891                         if (dd->pcidev->irq) {
1892                                 ipath_cdbg(VERBOSE,
1893                                            "unit %u free_irq of irq %x\n",
1894                                            dd->ipath_unit, dd->pcidev->irq);
1895                                 free_irq(dd->pcidev->irq, dd);
1896                         } else
1897                                 ipath_dbg("irq is 0, not doing free_irq "
1898                                           "for unit %u\n", dd->ipath_unit);
1899                         dd->pcidev = NULL;
1900                 }
1901
1902                 /*
1903                  * we check for NULL here, because it's outside the kregbase
1904                  * check, and we need to call it after the free_irq.  Thus
1905                  * it's possible that the function pointers were never
1906                  * initialized.
1907                  */
1908                 if (dd->ipath_f_cleanup)
1909                         /* clean up chip-specific stuff */
1910                         dd->ipath_f_cleanup(dd);
1911
1912                 spin_lock_irqsave(&ipath_devs_lock, flags);
1913         }
1914
1915         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1916
1917         ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
1918         pci_unregister_driver(&ipath_driver);
1919
1920         idr_destroy(&unit_table);
1921 }
1922
1923 /**
1924  * ipath_reset_device - reset the chip if possible
1925  * @unit: the device to reset
1926  *
1927  * Whether or not reset is successful, we attempt to re-initialize the chip
1928  * (that is, much like a driver unload/reload).  We clear the INITTED flag
1929  * so that the various entry points will fail until we reinitialize.  For
1930  * now, we only allow this if no user ports are open that use chip resources
1931  */
1932 int ipath_reset_device(int unit)
1933 {
1934         int ret, i;
1935         struct ipath_devdata *dd = ipath_lookup(unit);
1936
1937         if (!dd) {
1938                 ret = -ENODEV;
1939                 goto bail;
1940         }
1941
1942         dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
1943
1944         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
1945                 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
1946                          "not initialized or not present\n", unit);
1947                 ret = -ENXIO;
1948                 goto bail;
1949         }
1950
1951         if (dd->ipath_pd)
1952                 for (i = 1; i < dd->ipath_portcnt; i++) {
1953                         if (dd->ipath_pd[i] && dd->ipath_pd[i]->port_cnt) {
1954                                 ipath_dbg("unit %u port %d is in use "
1955                                           "(PID %u cmd %s), can't reset\n",
1956                                           unit, i,
1957                                           dd->ipath_pd[i]->port_pid,
1958                                           dd->ipath_pd[i]->port_comm);
1959                                 ret = -EBUSY;
1960                                 goto bail;
1961                         }
1962                 }
1963
1964         dd->ipath_flags &= ~IPATH_INITTED;
1965         ret = dd->ipath_f_reset(dd);
1966         if (ret != 1)
1967                 ipath_dbg("reset was not successful\n");
1968         ipath_dbg("Trying to reinitialize unit %u after reset attempt\n",
1969                   unit);
1970         ret = ipath_init_chip(dd, 1);
1971         if (ret)
1972                 ipath_dev_err(dd, "Reinitialize unit %u after "
1973                               "reset failed with %d\n", unit, ret);
1974         else
1975                 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
1976                          "resetting\n", unit);
1977
1978 bail:
1979         return ret;
1980 }
1981
1982 module_init(infinipath_init);
1983 module_exit(infinipath_cleanup);