err.no Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_init_chip.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/pci.h>
  35 #include <linux/netdevice.h>
  36 #include <linux/vmalloc.h>
  37
  38 #include "ipath_kernel.h"
  39 #include "ips_common.h"
  40
  41 /*
  42  * min buffers we want to have per port, after driver
  43  */
  44 #define IPATH_MIN_USER_PORT_BUFCNT 8
  45
  46 /*
  47  * Number of ports we are configured to use (to allow for more pio
  48  * buffers per port, etc.)  Zero means use chip value.
  49  */
  50 static ushort ipath_cfgports;
  51
  52 module_param_named(cfgports, ipath_cfgports, ushort, S_IRUGO);
  53 MODULE_PARM_DESC(cfgports, "Set max number of ports to use");
  54
  55 /*
  56  * Number of buffers reserved for driver (layered drivers and SMA
  57  * send).  Reserved at end of buffer list.   Initialized based on
  58  * number of PIO buffers if not set via module interface.
  59  * The problem with this is that it's global, but we'll use different
  60  * numbers for different chip types.  So the default value is not
  61  * very useful.  I've redefined it for the 1.3 release so that it's
  62  * zero unless set by the user to something else, in which case we
  63  * try to respect it.
  64  */
  65 static ushort ipath_kpiobufs;
  66
  67 static int ipath_set_kpiobufs(const char *val, struct kernel_param *kp);
  68
  69 module_param_call(kpiobufs, ipath_set_kpiobufs, param_get_ushort,
  70                   &ipath_kpiobufs, S_IWUSR | S_IRUGO);
  71 MODULE_PARM_DESC(kpiobufs, "Set number of PIO buffers for driver");
  72
  73 /**
  74  * create_port0_egr - allocate the eager TID buffers
  75  * @dd: the infinipath device
  76  *
  77  * This code is now quite different for user and kernel, because
  78  * the kernel uses skb's, for the accelerated network performance.
  79  * This is the kernel (port0) version.
  80  *
  81  * Allocate the eager TID buffers and program them into infinipath.
  82  * We use the network layer alloc_skb() allocator to allocate the
  83  * memory, and either use the buffers as is for things like SMA
  84  * packets, or pass the buffers up to the ipath layered driver and
  85  * thence the network layer, replacing them as we do so (see
  86  * ipath_rcv_layer()).
  87  */
  88 static int create_port0_egr(struct ipath_devdata *dd)
  89 {
  90         unsigned e, egrcnt;
  91         struct sk_buff **skbs;
  92         int ret;
  93
  94         egrcnt = dd->ipath_rcvegrcnt;
  95
  96         skbs = vmalloc(sizeof(*dd->ipath_port0_skbs) * egrcnt);
  97         if (skbs == NULL) {
  98                 ipath_dev_err(dd, "allocation error for eager TID "
  99                               "skb array\n");
 100                 ret = -ENOMEM;
 101                 goto bail;
 102         }
 103         for (e = 0; e < egrcnt; e++) {
 104                 /*
 105                  * This is a bit tricky in that we allocate extra
 106                  * space for 2 bytes of the 14 byte ethernet header.
 107                  * These two bytes are passed in the ipath header so
 108                  * the rest of the data is word aligned.  We allocate
 109                  * 4 bytes so that the data buffer stays word aligned.
 110                  * See ipath_kreceive() for more details.
 111                  */
 112                 skbs[e] = ipath_alloc_skb(dd, GFP_KERNEL);
 113                 if (!skbs[e]) {
 114                         ipath_dev_err(dd, "SKB allocation error for "
 115                                       "eager TID %u\n", e);
 116                         while (e != 0)
 117                                 dev_kfree_skb(skbs[--e]);
 118                         vfree(skbs);
 119                         ret = -ENOMEM;
 120                         goto bail;
 121                 }
 122         }
 123         /*
 124          * After loop above, so we can test non-NULL to see if ready
 125          * to use at receive, etc.
 126          */
 127         dd->ipath_port0_skbs = skbs;
 128
 129         for (e = 0; e < egrcnt; e++) {
 130                 unsigned long phys =
 131                         virt_to_phys(dd->ipath_port0_skbs[e]->data);
 132                 dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
 133                                     ((char __iomem *) dd->ipath_kregbase +
 134                                      dd->ipath_rcvegrbase), 0, phys);
 135         }
 136
 137         ret = 0;
 138
 139 bail:
 140         return ret;
 141 }
 142
 143 static int bringup_link(struct ipath_devdata *dd)
 144 {
 145         u64 val, ibc;
 146         int ret = 0;
 147
 148         /* hold IBC in reset */
 149         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
 150         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 151                          dd->ipath_control);
 152
 153         /*
 154          * Note that prior to try 14 or 15 of IB, the credit scaling
 155          * wasn't working, because it was swapped for writes with the
 156          * 1 bit default linkstate field
 157          */
 158
 159         /* ignore pbc and align word */
 160         val = dd->ipath_piosize2k - 2 * sizeof(u32);
 161         /*
 162          * for ICRC, which we only send in diag test pkt mode, and we
 163          * don't need to worry about that for mtu
 164          */
 165         val += 1;
 166         /*
 167          * Set the IBC maxpktlength to the size of our pio buffers the
 168          * maxpktlength is in words.  This is *not* the IB data MTU.
 169          */
 170         ibc = (val / sizeof(u32)) << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
 171         /* in KB */
 172         ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
 173         /*
 174          * How often flowctrl sent.  More or less in usecs; balance against
 175          * watermark value, so that in theory senders always get a flow
 176          * control update in time to not let the IB link go idle.
 177          */
 178         ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
 179         /* max error tolerance */
 180         ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
 181         /* use "real" buffer space for */
 182         ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
 183         /* IB credit flow control. */
 184         ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
 185         /* initially come up waiting for TS1, without sending anything. */
 186         dd->ipath_ibcctrl = ibc;
 187         /*
 188          * Want to start out with both LINKCMD and LINKINITCMD in NOP
 189          * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
 190          * to stay a NOP
 191          */
 192         ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
 193                 INFINIPATH_IBCC_LINKINITCMD_SHIFT;
 194         ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
 195                    (unsigned long long) ibc);
 196         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
 197
 198         // be sure chip saw it
 199         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 200
 201         ret = dd->ipath_f_bringup_serdes(dd);
 202
 203         if (ret)
 204                 dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
 205                          "not usable\n");
 206         else {
 207                 /* enable IBC */
 208                 dd->ipath_control |= INFINIPATH_C_LINKENABLE;
 209                 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 210                                  dd->ipath_control);
 211         }
 212
 213         return ret;
 214 }
 215
 216 static int init_chip_first(struct ipath_devdata *dd,
 217                            struct ipath_portdata **pdp)
 218 {
 219         struct ipath_portdata *pd = NULL;
 220         int ret = 0;
 221         u64 val;
 222
 223         /*
 224          * skip cfgports stuff because we are not allocating memory,
 225          * and we don't want problems if the portcnt changed due to
 226          * cfgports.  We do still check and report a difference, if
 227          * not same (should be impossible).
 228          */
 229         dd->ipath_portcnt =
 230                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
 231         if (!ipath_cfgports)
 232                 dd->ipath_cfgports = dd->ipath_portcnt;
 233         else if (ipath_cfgports <= dd->ipath_portcnt) {
 234                 dd->ipath_cfgports = ipath_cfgports;
 235                 ipath_dbg("Configured to use %u ports out of %u in chip\n",
 236                           dd->ipath_cfgports, dd->ipath_portcnt);
 237         } else {
 238                 dd->ipath_cfgports = dd->ipath_portcnt;
 239                 ipath_dbg("Tried to configured to use %u ports; chip "
 240                           "only supports %u\n", ipath_cfgports,
 241                           dd->ipath_portcnt);
 242         }
 243         dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_cfgports,
 244                                GFP_KERNEL);
 245
 246         if (!dd->ipath_pd) {
 247                 ipath_dev_err(dd, "Unable to allocate portdata array, "
 248                               "failing\n");
 249                 ret = -ENOMEM;
 250                 goto done;
 251         }
 252
 253         dd->ipath_lastegrheads = kzalloc(sizeof(*dd->ipath_lastegrheads)
 254                                          * dd->ipath_cfgports,
 255                                          GFP_KERNEL);
 256         dd->ipath_lastrcvhdrqtails =
 257                 kzalloc(sizeof(*dd->ipath_lastrcvhdrqtails)
 258                         * dd->ipath_cfgports, GFP_KERNEL);
 259
 260         if (!dd->ipath_lastegrheads || !dd->ipath_lastrcvhdrqtails) {
 261                 ipath_dev_err(dd, "Unable to allocate head arrays, "
 262                               "failing\n");
 263                 ret = -ENOMEM;
 264                 goto done;
 265         }
 266
 267         dd->ipath_pd[0] = kzalloc(sizeof(*pd), GFP_KERNEL);
 268
 269         if (!dd->ipath_pd[0]) {
 270                 ipath_dev_err(dd, "Unable to allocate portdata for port "
 271                               "0, failing\n");
 272                 ret = -ENOMEM;
 273                 goto done;
 274         }
 275         pd = dd->ipath_pd[0];
 276         pd->port_dd = dd;
 277         pd->port_port = 0;
 278         pd->port_cnt = 1;
 279         /* The port 0 pkey table is used by the layer interface. */
 280         pd->port_pkeys[0] = IPS_DEFAULT_P_KEY;
 281         dd->ipath_rcvtidcnt =
 282                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 283         dd->ipath_rcvtidbase =
 284                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 285         dd->ipath_rcvegrcnt =
 286                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 287         dd->ipath_rcvegrbase =
 288                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 289         dd->ipath_palign =
 290                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
 291         dd->ipath_piobufbase =
 292                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
 293         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
 294         dd->ipath_piosize2k = val & ~0U;
 295         dd->ipath_piosize4k = val >> 32;
 296         dd->ipath_ibmtu = 4096; /* default to largest legal MTU */
 297         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
 298         dd->ipath_piobcnt2k = val & ~0U;
 299         dd->ipath_piobcnt4k = val >> 32;
 300         dd->ipath_pio2kbase =
 301                 (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 302                                  (dd->ipath_piobufbase & 0xffffffff));
 303         if (dd->ipath_piobcnt4k) {
 304                 dd->ipath_pio4kbase = (u32 __iomem *)
 305                         (((char __iomem *) dd->ipath_kregbase) +
 306                          (dd->ipath_piobufbase >> 32));
 307                 /*
 308                  * 4K buffers take 2 pages; we use roundup just to be
 309                  * paranoid; we calculate it once here, rather than on
 310                  * ever buf allocate
 311                  */
 312                 dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
 313                                           dd->ipath_palign);
 314                 ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
 315                           "(%x aligned)\n",
 316                           dd->ipath_piobcnt2k, dd->ipath_piosize2k,
 317                           dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
 318                           dd->ipath_piosize4k, dd->ipath_pio4kbase,
 319                           dd->ipath_4kalign);
 320         }
 321         else ipath_dbg("%u 2k piobufs @ %p\n",
 322                        dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
 323
 324         spin_lock_init(&dd->ipath_tid_lock);
 325
 326 done:
 327         *pdp = pd;
 328         return ret;
 329 }
 330
 331 /**
 332  * init_chip_reset - re-initialize after a reset, or enable
 333  * @dd: the infinipath device
 334  * @pdp: output for port data
 335  *
 336  * sanity check at least some of the values after reset, and
 337  * ensure no receive or transmit (explictly, in case reset
 338  * failed
 339  */
 340 static int init_chip_reset(struct ipath_devdata *dd,
 341                            struct ipath_portdata **pdp)
 342 {
 343         struct ipath_portdata *pd;
 344         u32 rtmp;
 345
 346         *pdp = pd = dd->ipath_pd[0];
 347         /* ensure chip does no sends or receives while we re-initialize */
 348         dd->ipath_control = dd->ipath_sendctrl = dd->ipath_rcvctrl = 0U;
 349         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, 0);
 350         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0);
 351         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0);
 352
 353         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
 354         if (dd->ipath_portcnt != rtmp)
 355                 dev_info(&dd->pcidev->dev, "portcnt was %u before "
 356                          "reset, now %u, using original\n",
 357                          dd->ipath_portcnt, rtmp);
 358         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 359         if (rtmp != dd->ipath_rcvtidcnt)
 360                 dev_info(&dd->pcidev->dev, "tidcnt was %u before "
 361                          "reset, now %u, using original\n",
 362                          dd->ipath_rcvtidcnt, rtmp);
 363         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 364         if (rtmp != dd->ipath_rcvtidbase)
 365                 dev_info(&dd->pcidev->dev, "tidbase was %u before "
 366                          "reset, now %u, using original\n",
 367                          dd->ipath_rcvtidbase, rtmp);
 368         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 369         if (rtmp != dd->ipath_rcvegrcnt)
 370                 dev_info(&dd->pcidev->dev, "egrcnt was %u before "
 371                          "reset, now %u, using original\n",
 372                          dd->ipath_rcvegrcnt, rtmp);
 373         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 374         if (rtmp != dd->ipath_rcvegrbase)
 375                 dev_info(&dd->pcidev->dev, "egrbase was %u before "
 376                          "reset, now %u, using original\n",
 377                          dd->ipath_rcvegrbase, rtmp);
 378
 379         return 0;
 380 }
 381
 382 static int init_pioavailregs(struct ipath_devdata *dd)
 383 {
 384         int ret;
 385
 386         dd->ipath_pioavailregs_dma = dma_alloc_coherent(
 387                 &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
 388                 GFP_KERNEL);
 389         if (!dd->ipath_pioavailregs_dma) {
 390                 ipath_dev_err(dd, "failed to allocate PIOavail reg area "
 391                               "in memory\n");
 392                 ret = -ENOMEM;
 393                 goto done;
 394         }
 395
 396         /*
 397          * we really want L2 cache aligned, but for current CPUs of
 398          * interest, they are the same.
 399          */
 400         dd->ipath_statusp = (u64 *)
 401                 ((char *)dd->ipath_pioavailregs_dma +
 402                  ((2 * L1_CACHE_BYTES +
 403                    dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
 404         /* copy the current value now that it's really allocated */
 405         *dd->ipath_statusp = dd->_ipath_status;
 406         /*
 407          * setup buffer to hold freeze msg, accessible to apps,
 408          * following statusp
 409          */
 410         dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
 411         /* and its length */
 412         dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
 413
 414         if (dd->ipath_unit * 64 > (IPATH_PORT0_RCVHDRTAIL_SIZE - 64)) {
 415                 ipath_dev_err(dd, "unit %u too large for port 0 "
 416                               "rcvhdrtail buffer size\n", dd->ipath_unit);
 417                 ret = -ENODEV;
 418         }
 419         else
 420                 ret = 0;
 421
 422         /* so we can get current tail in ipath_kreceive(), per chip */
 423         dd->ipath_hdrqtailptr = &ipath_port0_rcvhdrtail[
 424                 dd->ipath_unit * (64 / sizeof(*ipath_port0_rcvhdrtail))];
 425 done:
 426         return ret;
 427 }
 428
 429 /**
 430  * init_shadow_tids - allocate the shadow TID array
 431  * @dd: the infinipath device
 432  *
 433  * allocate the shadow TID array, so we can ipath_munlock previous
 434  * entries.  It may make more sense to move the pageshadow to the
 435  * port data structure, so we only allocate memory for ports actually
 436  * in use, since we at 8k per port, now.
 437  */
 438 static void init_shadow_tids(struct ipath_devdata *dd)
 439 {
 440         dd->ipath_pageshadow = (struct page **)
 441                 vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 442                         sizeof(struct page *));
 443         if (!dd->ipath_pageshadow)
 444                 ipath_dev_err(dd, "failed to allocate shadow page * "
 445                               "array, no expected sends!\n");
 446         else
 447                 memset(dd->ipath_pageshadow, 0,
 448                        dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 449                        sizeof(struct page *));
 450 }
 451
 452 static void enable_chip(struct ipath_devdata *dd,
 453                         struct ipath_portdata *pd, int reinit)
 454 {
 455         u32 val;
 456         int i;
 457
 458         if (!reinit) {
 459                 init_waitqueue_head(&ipath_sma_state_wait);
 460         }
 461         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 462                          dd->ipath_rcvctrl);
 463
 464         /* Enable PIO send, and update of PIOavail regs to memory. */
 465         dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
 466                 INFINIPATH_S_PIOBUFAVAILUPD;
 467         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 468                          dd->ipath_sendctrl);
 469
 470         /*
 471          * enable port 0 receive, and receive interrupt.  other ports
 472          * done as user opens and inits them.
 473          */
 474         dd->ipath_rcvctrl = INFINIPATH_R_TAILUPD |
 475                 (1ULL << INFINIPATH_R_PORTENABLE_SHIFT) |
 476                 (1ULL << INFINIPATH_R_INTRAVAIL_SHIFT);
 477         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 478                          dd->ipath_rcvctrl);
 479
 480         /*
 481          * now ready for use.  this should be cleared whenever we
 482          * detect a reset, or initiate one.
 483          */
 484         dd->ipath_flags |= IPATH_INITTED;
 485
 486         /*
 487          * init our shadow copies of head from tail values, and write
 488          * head values to match.
 489          */
 490         val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
 491         (void)ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
 492         dd->ipath_port0head = ipath_read_ureg32(dd, ur_rcvhdrtail, 0);
 493
 494         /* Initialize so we interrupt on next packet received */
 495         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
 496                                dd->ipath_rhdrhead_intr_off |
 497                                dd->ipath_port0head, 0);
 498
 499         /*
 500          * by now pioavail updates to memory should have occurred, so
 501          * copy them into our working/shadow registers; this is in
 502          * case something went wrong with abort, but mostly to get the
 503          * initial values of the generation bit correct.
 504          */
 505         for (i = 0; i < dd->ipath_pioavregs; i++) {
 506                 __le64 val;
 507
 508                 /*
 509                  * Chip Errata bug 6641; even and odd qwords>3 are swapped.
 510                  */
 511                 if (i > 3) {
 512                         if (i & 1)
 513                                 val = dd->ipath_pioavailregs_dma[i - 1];
 514                         else
 515                                 val = dd->ipath_pioavailregs_dma[i + 1];
 516                 }
 517                 else
 518                         val = dd->ipath_pioavailregs_dma[i];
 519                 dd->ipath_pioavailshadow[i] = le64_to_cpu(val);
 520         }
 521         /* can get counters, stats, etc. */
 522         dd->ipath_flags |= IPATH_PRESENT;
 523 }
 524
 525 static int init_housekeeping(struct ipath_devdata *dd,
 526                              struct ipath_portdata **pdp, int reinit)
 527 {
 528         char boardn[32];
 529         int ret = 0;
 530
 531         /*
 532          * have to clear shadow copies of registers at init that are
 533          * not otherwise set here, or all kinds of bizarre things
 534          * happen with driver on chip reset
 535          */
 536         dd->ipath_rcvhdrsize = 0;
 537
 538         /*
 539          * Don't clear ipath_flags as 8bit mode was set before
 540          * entering this func. However, we do set the linkstate to
 541          * unknown, so we can watch for a transition.
 542          * PRESENT is set because we want register reads to work,
 543          * and the kernel infrastructure saw it in config space;
 544          * We clear it if we have failures.
 545          */
 546         dd->ipath_flags |= IPATH_LINKUNK | IPATH_PRESENT;
 547         dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
 548                              IPATH_LINKDOWN | IPATH_LINKINIT);
 549
 550         ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
 551         dd->ipath_revision =
 552                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
 553
 554         /*
 555          * set up fundamental info we need to use the chip; we assume
 556          * if the revision reg and these regs are OK, we don't need to
 557          * special case the rest
 558          */
 559         dd->ipath_sregbase =
 560                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
 561         dd->ipath_cregbase =
 562                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
 563         dd->ipath_uregbase =
 564                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
 565         ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
 566                    "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
 567                    dd->ipath_uregbase, dd->ipath_cregbase);
 568         if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
 569             || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
 570             || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
 571             || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
 572                 ipath_dev_err(dd, "Register read failures from chip, "
 573                               "giving up initialization\n");
 574                 dd->ipath_flags &= ~IPATH_PRESENT;
 575                 ret = -ENODEV;
 576                 goto done;
 577         }
 578
 579         /* clear the initial reset flag, in case first driver load */
 580         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 581                          INFINIPATH_E_RESET);
 582
 583         if (reinit)
 584                 ret = init_chip_reset(dd, pdp);
 585         else
 586                 ret = init_chip_first(dd, pdp);
 587
 588         if (ret)
 589                 goto done;
 590
 591         ipath_cdbg(VERBOSE, "Revision %llx (PCI %x), %u ports, %u tids, "
 592                    "%u egrtids\n", (unsigned long long) dd->ipath_revision,
 593                    dd->ipath_pcirev, dd->ipath_portcnt, dd->ipath_rcvtidcnt,
 594                    dd->ipath_rcvegrcnt);
 595
 596         if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
 597              INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
 598                 ipath_dev_err(dd, "Driver only handles version %d, "
 599                               "chip swversion is %d (%llx), failng\n",
 600                               IPATH_CHIP_SWVERSION,
 601                               (int)(dd->ipath_revision >>
 602                                     INFINIPATH_R_SOFTWARE_SHIFT) &
 603                               INFINIPATH_R_SOFTWARE_MASK,
 604                               (unsigned long long) dd->ipath_revision);
 605                 ret = -ENOSYS;
 606                 goto done;
 607         }
 608         dd->ipath_majrev = (u8) ((dd->ipath_revision >>
 609                                   INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
 610                                  INFINIPATH_R_CHIPREVMAJOR_MASK);
 611         dd->ipath_minrev = (u8) ((dd->ipath_revision >>
 612                                   INFINIPATH_R_CHIPREVMINOR_SHIFT) &
 613                                  INFINIPATH_R_CHIPREVMINOR_MASK);
 614         dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
 615                                     INFINIPATH_R_BOARDID_SHIFT) &
 616                                    INFINIPATH_R_BOARDID_MASK);
 617
 618         ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
 619
 620         snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
 621                  "Driver %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
 622                  "SW Compat %u\n",
 623                  IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
 624                  (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
 625                  INFINIPATH_R_ARCH_MASK,
 626                  dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
 627                  (unsigned)(dd->ipath_revision >>
 628                             INFINIPATH_R_SOFTWARE_SHIFT) &
 629                  INFINIPATH_R_SOFTWARE_MASK);
 630
 631         ipath_dbg("%s", dd->ipath_boardversion);
 632
 633 done:
 634         return ret;
 635 }
 636
 637
 638 /**
 639  * ipath_init_chip - do the actual initialization sequence on the chip
 640  * @dd: the infinipath device
 641  * @reinit: reinitializing, so don't allocate new memory
 642  *
 643  * Do the actual initialization sequence on the chip.  This is done
 644  * both from the init routine called from the PCI infrastructure, and
 645  * when we reset the chip, or detect that it was reset internally,
 646  * or it's administratively re-enabled.
 647  *
 648  * Memory allocation here and in called routines is only done in
 649  * the first case (reinit == 0).  We have to be careful, because even
 650  * without memory allocation, we need to re-write all the chip registers
 651  * TIDs, etc. after the reset or enable has completed.
 652  */
 653 int ipath_init_chip(struct ipath_devdata *dd, int reinit)
 654 {
 655         int ret = 0, i;
 656         u32 val32, kpiobufs;
 657         u64 val, atmp;
 658         struct ipath_portdata *pd = NULL; /* keep gcc4 happy */
 659
 660         ret = init_housekeeping(dd, &pd, reinit);
 661         if (ret)
 662                 goto done;
 663
 664         /*
 665          * we ignore most issues after reporting them, but have to specially
 666          * handle hardware-disabled chips.
 667          */
 668         if (ret == 2) {
 669                 /* unique error, known to ipath_init_one */
 670                 ret = -EPERM;
 671                 goto done;
 672         }
 673
 674         /*
 675          * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
 676          * but then it no longer nicely fits power of two, and since
 677          * we now use routines that backend onto __get_free_pages, the
 678          * rest would be wasted.
 679          */
 680         dd->ipath_rcvhdrcnt = dd->ipath_rcvegrcnt;
 681         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
 682                          dd->ipath_rcvhdrcnt);
 683
 684         /*
 685          * Set up the shadow copies of the piobufavail registers,
 686          * which we compare against the chip registers for now, and
 687          * the in memory DMA'ed copies of the registers.  This has to
 688          * be done early, before we calculate lastport, etc.
 689          */
 690         val = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
 691         /*
 692          * calc number of pioavail registers, and save it; we have 2
 693          * bits per buffer.
 694          */
 695         dd->ipath_pioavregs = ALIGN(val, sizeof(u64) * BITS_PER_BYTE / 2)
 696                 / (sizeof(u64) * BITS_PER_BYTE / 2);
 697         if (ipath_kpiobufs == 0) {
 698                 /* not set by user, or set explictly to default  */
 699                 if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) > 128)
 700                         kpiobufs = 32;
 701                 else
 702                         kpiobufs = 16;
 703         }
 704         else
 705                 kpiobufs = ipath_kpiobufs;
 706
 707         if (kpiobufs >
 708             (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
 709              (dd->ipath_cfgports * IPATH_MIN_USER_PORT_BUFCNT))) {
 710                 i = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
 711                         (dd->ipath_cfgports * IPATH_MIN_USER_PORT_BUFCNT);
 712                 if (i < 0)
 713                         i = 0;
 714                 dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs for "
 715                          "kernel leaves too few for %d user ports "
 716                          "(%d each); using %u\n", kpiobufs,
 717                          dd->ipath_cfgports - 1,
 718                          IPATH_MIN_USER_PORT_BUFCNT, i);
 719                 /*
 720                  * shouldn't change ipath_kpiobufs, because could be
 721                  * different for different devices...
 722                  */
 723                 kpiobufs = i;
 724         }
 725         dd->ipath_lastport_piobuf =
 726                 dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - kpiobufs;
 727         dd->ipath_pbufsport = dd->ipath_cfgports > 1
 728                 ? dd->ipath_lastport_piobuf / (dd->ipath_cfgports - 1)
 729                 : 0;
 730         val32 = dd->ipath_lastport_piobuf -
 731                 (dd->ipath_pbufsport * (dd->ipath_cfgports - 1));
 732         if (val32 > 0) {
 733                 ipath_dbg("allocating %u pbufs/port leaves %u unused, "
 734                           "add to kernel\n", dd->ipath_pbufsport, val32);
 735                 dd->ipath_lastport_piobuf -= val32;
 736                 ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
 737                           dd->ipath_pbufsport, val32);
 738         }
 739         dd->ipath_lastpioindex = dd->ipath_lastport_piobuf;
 740         ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
 741                    "each for %u user ports\n", kpiobufs,
 742                    dd->ipath_piobcnt2k + dd->ipath_piobcnt4k,
 743                    dd->ipath_pbufsport, dd->ipath_cfgports - 1);
 744
 745         dd->ipath_f_early_init(dd);
 746
 747         /* early_init sets rcvhdrentsize and rcvhdrsize, so this must be
 748          * done after early_init */
 749         dd->ipath_hdrqlast =
 750                 dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
 751         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
 752                          dd->ipath_rcvhdrentsize);
 753         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
 754                          dd->ipath_rcvhdrsize);
 755
 756         if (!reinit) {
 757                 ret = init_pioavailregs(dd);
 758                 init_shadow_tids(dd);
 759                 if (ret)
 760                         goto done;
 761         }
 762
 763         (void)ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
 764                                dd->ipath_pioavailregs_phys);
 765         /*
 766          * this is to detect s/w errors, which the h/w works around by
 767          * ignoring the low 6 bits of address, if it wasn't aligned.
 768          */
 769         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
 770         if (val != dd->ipath_pioavailregs_phys) {
 771                 ipath_dev_err(dd, "Catastrophic software error, "
 772                               "SendPIOAvailAddr written as %lx, "
 773                               "read back as %llx\n",
 774                               (unsigned long) dd->ipath_pioavailregs_phys,
 775                               (unsigned long long) val);
 776                 ret = -EINVAL;
 777                 goto done;
 778         }
 779
 780         val = ipath_port0_rcvhdrtail_dma + dd->ipath_unit * 64;
 781
 782         /* verify that the alignment requirement was met */
 783         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
 784                               0, val);
 785         atmp = ipath_read_kreg64_port(
 786                 dd, dd->ipath_kregs->kr_rcvhdrtailaddr, 0);
 787         if (val != atmp) {
 788                 ipath_dev_err(dd, "Catastrophic software error, "
 789                               "RcvHdrTailAddr0 written as %llx, "
 790                               "read back as %llx from %x\n",
 791                               (unsigned long long) val,
 792                               (unsigned long long) atmp,
 793                               dd->ipath_kregs->kr_rcvhdrtailaddr);
 794                 ret = -EINVAL;
 795                 goto done;
 796         }
 797
 798         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
 799
 800         /*
 801          * make sure we are not in freeze, and PIO send enabled, so
 802          * writes to pbc happen
 803          */
 804         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
 805         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 806                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 807         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
 808         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 809                          INFINIPATH_S_PIOENABLE);
 810
 811         /*
 812          * before error clears, since we expect serdes pll errors during
 813          * this, the first time after reset
 814          */
 815         if (bringup_link(dd)) {
 816                 dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
 817                 ret = -ENETDOWN;
 818                 goto done;
 819         }
 820
 821         /*
 822          * clear any "expected" hwerrs from reset and/or initialization
 823          * clear any that aren't enabled (at least this once), and then
 824          * set the enable mask
 825          */
 826         dd->ipath_f_init_hwerrors(dd);
 827         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 828                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 829         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 830                          dd->ipath_hwerrmask);
 831
 832         dd->ipath_maskederrs = dd->ipath_ignorederrs;
 833         /* clear all */
 834         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
 835         /* enable errors that are masked, at least this first time. */
 836         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 837                          ~dd->ipath_maskederrs);
 838         /* clear any interrups up to this point (ints still not enabled) */
 839         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
 840
 841         ipath_stats.sps_lid[dd->ipath_unit] = dd->ipath_lid;
 842
 843         /*
 844          * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
 845          * re-init, the simplest way to handle this is to free
 846          * existing, and re-allocate.
 847          */
 848         if (reinit)
 849                 ipath_free_pddata(dd, 0, 0);
 850         dd->ipath_f_tidtemplate(dd);
 851         ret = ipath_create_rcvhdrq(dd, pd);
 852         if (!ret)
 853                 ret = create_port0_egr(dd);
 854         if (ret)
 855                 ipath_dev_err(dd, "failed to allocate port 0 (kernel) "
 856                               "rcvhdrq and/or egr bufs\n");
 857         else
 858                 enable_chip(dd, pd, reinit);
 859
 860         /*
 861          * cause retrigger of pending interrupts ignored during init,
 862          * even if we had errors
 863          */
 864         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 865
 866         if(!dd->ipath_stats_timer_active) {
 867                 /*
 868                  * first init, or after an admin disable/enable
 869                  * set up stats retrieval timer, even if we had errors
 870                  * in last portion of setup
 871                  */
 872                 init_timer(&dd->ipath_stats_timer);
 873                 dd->ipath_stats_timer.function = ipath_get_faststats;
 874                 dd->ipath_stats_timer.data = (unsigned long) dd;
 875                 /* every 5 seconds; */
 876                 dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
 877                 /* takes ~16 seconds to overflow at full IB 4x bandwdith */
 878                 add_timer(&dd->ipath_stats_timer);
 879                 dd->ipath_stats_timer_active = 1;
 880         }
 881
 882 done:
 883         if (!ret) {
 884                 *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
 885                 if (!dd->ipath_f_intrsetup(dd)) {
 886                         /* now we can enable all interrupts from the chip */
 887                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 888                                          -1LL);
 889                         /* force re-interrupt of any pending interrupts. */
 890                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
 891                                          0ULL);
 892                         /* chip is usable; mark it as initialized */
 893                         *dd->ipath_statusp |= IPATH_STATUS_INITTED;
 894                 } else
 895                         ipath_dev_err(dd, "No interrupts enabled, couldn't "
 896                                       "setup interrupt address\n");
 897
 898                 if (dd->ipath_cfgports > ipath_stats.sps_nports)
 899                         /*
 900                          * sps_nports is a global, so, we set it to
 901                          * the highest number of ports of any of the
 902                          * chips we find; we never decrement it, at
 903                          * least for now.  Since this might have changed
 904                          * over disable/enable or prior to reset, always
 905                          * do the check and potentially adjust.
 906                          */
 907                         ipath_stats.sps_nports = dd->ipath_cfgports;
 908         } else
 909                 ipath_dbg("Failed (%d) to initialize chip\n", ret);
 910
 911         /* if ret is non-zero, we probably should do some cleanup
 912            here... */
 913         return ret;
 914 }
 915
 916 static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
 917 {
 918         struct ipath_devdata *dd;
 919         unsigned long flags;
 920         unsigned short val;
 921         int ret;
 922
 923         ret = ipath_parse_ushort(str, &val);
 924
 925         spin_lock_irqsave(&ipath_devs_lock, flags);
 926
 927         if (ret < 0)
 928                 goto bail;
 929
 930         if (val == 0) {
 931                 ret = -EINVAL;
 932                 goto bail;
 933         }
 934
 935         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 936                 if (dd->ipath_kregbase)
 937                         continue;
 938                 if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
 939                            (dd->ipath_cfgports *
 940                             IPATH_MIN_USER_PORT_BUFCNT)))
 941                 {
 942                         ipath_dev_err(
 943                                 dd,
 944                                 "Allocating %d PIO bufs for kernel leaves "
 945                                 "too few for %d user ports (%d each)\n",
 946                                 val, dd->ipath_cfgports - 1,
 947                                 IPATH_MIN_USER_PORT_BUFCNT);
 948                         ret = -EINVAL;
 949                         goto bail;
 950                 }
 951                 dd->ipath_lastport_piobuf =
 952                         dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
 953         }
 954
 955         ret = 0;
 956 bail:
 957         spin_unlock_irqrestore(&ipath_devs_lock, flags);
 958
 959         return ret;
 960 }