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 "ipath_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 (verbs and layered drivers.)
  57  * 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 verbs
  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 ipath_skbinfo *skbinfo;
  92         int ret;
  93
  94         egrcnt = dd->ipath_rcvegrcnt;
  95
  96         skbinfo = vmalloc(sizeof(*dd->ipath_port0_skbinfo) * egrcnt);
  97         if (skbinfo == 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                 skbinfo[e].skb = ipath_alloc_skb(dd, GFP_KERNEL);
 113                 if (!skbinfo[e].skb) {
 114                         ipath_dev_err(dd, "SKB allocation error for "
 115                                       "eager TID %u\n", e);
 116                         while (e != 0)
 117                                 dev_kfree_skb(skbinfo[--e].skb);
 118                         vfree(skbinfo);
 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_skbinfo = skbinfo;
 128
 129         for (e = 0; e < egrcnt; e++) {
 130                 dd->ipath_port0_skbinfo[e].phys =
 131                   ipath_map_single(dd->pcidev,
 132                                    dd->ipath_port0_skbinfo[e].skb->data,
 133                                    dd->ipath_ibmaxlen, PCI_DMA_FROMDEVICE);
 134                 dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
 135                                     ((char __iomem *) dd->ipath_kregbase +
 136                                      dd->ipath_rcvegrbase), 0,
 137                                     dd->ipath_port0_skbinfo[e].phys);
 138         }
 139
 140         ret = 0;
 141
 142 bail:
 143         return ret;
 144 }
 145
 146 static int bringup_link(struct ipath_devdata *dd)
 147 {
 148         u64 val, ibc;
 149         int ret = 0;
 150
 151         /* hold IBC in reset */
 152         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
 153         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 154                          dd->ipath_control);
 155
 156         /*
 157          * Note that prior to try 14 or 15 of IB, the credit scaling
 158          * wasn't working, because it was swapped for writes with the
 159          * 1 bit default linkstate field
 160          */
 161
 162         /* ignore pbc and align word */
 163         val = dd->ipath_piosize2k - 2 * sizeof(u32);
 164         /*
 165          * for ICRC, which we only send in diag test pkt mode, and we
 166          * don't need to worry about that for mtu
 167          */
 168         val += 1;
 169         /*
 170          * Set the IBC maxpktlength to the size of our pio buffers the
 171          * maxpktlength is in words.  This is *not* the IB data MTU.
 172          */
 173         ibc = (val / sizeof(u32)) << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
 174         /* in KB */
 175         ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
 176         /*
 177          * How often flowctrl sent.  More or less in usecs; balance against
 178          * watermark value, so that in theory senders always get a flow
 179          * control update in time to not let the IB link go idle.
 180          */
 181         ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
 182         /* max error tolerance */
 183         ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
 184         /* use "real" buffer space for */
 185         ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
 186         /* IB credit flow control. */
 187         ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
 188         /* initially come up waiting for TS1, without sending anything. */
 189         dd->ipath_ibcctrl = ibc;
 190         /*
 191          * Want to start out with both LINKCMD and LINKINITCMD in NOP
 192          * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
 193          * to stay a NOP
 194          */
 195         ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
 196                 INFINIPATH_IBCC_LINKINITCMD_SHIFT;
 197         ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
 198                    (unsigned long long) ibc);
 199         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
 200
 201         // be sure chip saw it
 202         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 203
 204         ret = dd->ipath_f_bringup_serdes(dd);
 205
 206         if (ret)
 207                 dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
 208                          "not usable\n");
 209         else {
 210                 /* enable IBC */
 211                 dd->ipath_control |= INFINIPATH_C_LINKENABLE;
 212                 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 213                                  dd->ipath_control);
 214         }
 215
 216         return ret;
 217 }
 218
 219 static struct ipath_portdata *create_portdata0(struct ipath_devdata *dd)
 220 {
 221         struct ipath_portdata *pd = NULL;
 222
 223         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 224         if (pd) {
 225                 pd->port_dd = dd;
 226                 pd->port_cnt = 1;
 227                 /* The port 0 pkey table is used by the layer interface. */
 228                 pd->port_pkeys[0] = IPATH_DEFAULT_P_KEY;
 229         }
 230         return pd;
 231 }
 232
 233 static int init_chip_first(struct ipath_devdata *dd,
 234                            struct ipath_portdata **pdp)
 235 {
 236         struct ipath_portdata *pd = NULL;
 237         int ret = 0;
 238         u64 val;
 239
 240         /*
 241          * skip cfgports stuff because we are not allocating memory,
 242          * and we don't want problems if the portcnt changed due to
 243          * cfgports.  We do still check and report a difference, if
 244          * not same (should be impossible).
 245          */
 246         dd->ipath_portcnt =
 247                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
 248         if (!ipath_cfgports)
 249                 dd->ipath_cfgports = dd->ipath_portcnt;
 250         else if (ipath_cfgports <= dd->ipath_portcnt) {
 251                 dd->ipath_cfgports = ipath_cfgports;
 252                 ipath_dbg("Configured to use %u ports out of %u in chip\n",
 253                           dd->ipath_cfgports, dd->ipath_portcnt);
 254         } else {
 255                 dd->ipath_cfgports = dd->ipath_portcnt;
 256                 ipath_dbg("Tried to configured to use %u ports; chip "
 257                           "only supports %u\n", ipath_cfgports,
 258                           dd->ipath_portcnt);
 259         }
 260         /*
 261          * Allocate full portcnt array, rather than just cfgports, because
 262          * cleanup iterates across all possible ports.
 263          */
 264         dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_portcnt,
 265                                GFP_KERNEL);
 266
 267         if (!dd->ipath_pd) {
 268                 ipath_dev_err(dd, "Unable to allocate portdata array, "
 269                               "failing\n");
 270                 ret = -ENOMEM;
 271                 goto done;
 272         }
 273
 274         dd->ipath_lastegrheads = kzalloc(sizeof(*dd->ipath_lastegrheads)
 275                                          * dd->ipath_cfgports,
 276                                          GFP_KERNEL);
 277         dd->ipath_lastrcvhdrqtails =
 278                 kzalloc(sizeof(*dd->ipath_lastrcvhdrqtails)
 279                         * dd->ipath_cfgports, GFP_KERNEL);
 280
 281         if (!dd->ipath_lastegrheads || !dd->ipath_lastrcvhdrqtails) {
 282                 ipath_dev_err(dd, "Unable to allocate head arrays, "
 283                               "failing\n");
 284                 ret = -ENOMEM;
 285                 goto done;
 286         }
 287
 288         pd = create_portdata0(dd);
 289
 290         if (!pd) {
 291                 ipath_dev_err(dd, "Unable to allocate portdata for port "
 292                               "0, failing\n");
 293                 ret = -ENOMEM;
 294                 goto done;
 295         }
 296         dd->ipath_pd[0] = pd;
 297
 298         dd->ipath_rcvtidcnt =
 299                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 300         dd->ipath_rcvtidbase =
 301                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 302         dd->ipath_rcvegrcnt =
 303                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 304         dd->ipath_rcvegrbase =
 305                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 306         dd->ipath_palign =
 307                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
 308         dd->ipath_piobufbase =
 309                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
 310         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
 311         dd->ipath_piosize2k = val & ~0U;
 312         dd->ipath_piosize4k = val >> 32;
 313         /*
 314          * Note: the chips support a maximum MTU of 4096, but the driver
 315          * hasn't implemented this feature yet, so set the initial value
 316          * to 2048.
 317          */
 318         dd->ipath_ibmtu = 2048;
 319         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
 320         dd->ipath_piobcnt2k = val & ~0U;
 321         dd->ipath_piobcnt4k = val >> 32;
 322         dd->ipath_pio2kbase =
 323                 (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 324                                  (dd->ipath_piobufbase & 0xffffffff));
 325         if (dd->ipath_piobcnt4k) {
 326                 dd->ipath_pio4kbase = (u32 __iomem *)
 327                         (((char __iomem *) dd->ipath_kregbase) +
 328                          (dd->ipath_piobufbase >> 32));
 329                 /*
 330                  * 4K buffers take 2 pages; we use roundup just to be
 331                  * paranoid; we calculate it once here, rather than on
 332                  * ever buf allocate
 333                  */
 334                 dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
 335                                           dd->ipath_palign);
 336                 ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
 337                           "(%x aligned)\n",
 338                           dd->ipath_piobcnt2k, dd->ipath_piosize2k,
 339                           dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
 340                           dd->ipath_piosize4k, dd->ipath_pio4kbase,
 341                           dd->ipath_4kalign);
 342         }
 343         else ipath_dbg("%u 2k piobufs @ %p\n",
 344                        dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
 345
 346         spin_lock_init(&dd->ipath_tid_lock);
 347
 348         spin_lock_init(&dd->ipath_gpio_lock);
 349         spin_lock_init(&dd->ipath_eep_st_lock);
 350         sema_init(&dd->ipath_eep_sem, 1);
 351
 352 done:
 353         *pdp = pd;
 354         return ret;
 355 }
 356
 357 /**
 358  * init_chip_reset - re-initialize after a reset, or enable
 359  * @dd: the infinipath device
 360  * @pdp: output for port data
 361  *
 362  * sanity check at least some of the values after reset, and
 363  * ensure no receive or transmit (explictly, in case reset
 364  * failed
 365  */
 366 static int init_chip_reset(struct ipath_devdata *dd,
 367                            struct ipath_portdata **pdp)
 368 {
 369         u32 rtmp;
 370
 371         *pdp = dd->ipath_pd[0];
 372         /* ensure chip does no sends or receives while we re-initialize */
 373         dd->ipath_control = dd->ipath_sendctrl = dd->ipath_rcvctrl = 0U;
 374         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, 0);
 375         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0);
 376         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0);
 377
 378         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
 379         if (dd->ipath_portcnt != rtmp)
 380                 dev_info(&dd->pcidev->dev, "portcnt was %u before "
 381                          "reset, now %u, using original\n",
 382                          dd->ipath_portcnt, rtmp);
 383         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 384         if (rtmp != dd->ipath_rcvtidcnt)
 385                 dev_info(&dd->pcidev->dev, "tidcnt was %u before "
 386                          "reset, now %u, using original\n",
 387                          dd->ipath_rcvtidcnt, rtmp);
 388         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 389         if (rtmp != dd->ipath_rcvtidbase)
 390                 dev_info(&dd->pcidev->dev, "tidbase was %u before "
 391                          "reset, now %u, using original\n",
 392                          dd->ipath_rcvtidbase, rtmp);
 393         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 394         if (rtmp != dd->ipath_rcvegrcnt)
 395                 dev_info(&dd->pcidev->dev, "egrcnt was %u before "
 396                          "reset, now %u, using original\n",
 397                          dd->ipath_rcvegrcnt, rtmp);
 398         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 399         if (rtmp != dd->ipath_rcvegrbase)
 400                 dev_info(&dd->pcidev->dev, "egrbase was %u before "
 401                          "reset, now %u, using original\n",
 402                          dd->ipath_rcvegrbase, rtmp);
 403
 404         return 0;
 405 }
 406
 407 static int init_pioavailregs(struct ipath_devdata *dd)
 408 {
 409         int ret;
 410
 411         dd->ipath_pioavailregs_dma = dma_alloc_coherent(
 412                 &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
 413                 GFP_KERNEL);
 414         if (!dd->ipath_pioavailregs_dma) {
 415                 ipath_dev_err(dd, "failed to allocate PIOavail reg area "
 416                               "in memory\n");
 417                 ret = -ENOMEM;
 418                 goto done;
 419         }
 420
 421         /*
 422          * we really want L2 cache aligned, but for current CPUs of
 423          * interest, they are the same.
 424          */
 425         dd->ipath_statusp = (u64 *)
 426                 ((char *)dd->ipath_pioavailregs_dma +
 427                  ((2 * L1_CACHE_BYTES +
 428                    dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
 429         /* copy the current value now that it's really allocated */
 430         *dd->ipath_statusp = dd->_ipath_status;
 431         /*
 432          * setup buffer to hold freeze msg, accessible to apps,
 433          * following statusp
 434          */
 435         dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
 436         /* and its length */
 437         dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
 438
 439         ret = 0;
 440
 441 done:
 442         return ret;
 443 }
 444
 445 /**
 446  * init_shadow_tids - allocate the shadow TID array
 447  * @dd: the infinipath device
 448  *
 449  * allocate the shadow TID array, so we can ipath_munlock previous
 450  * entries.  It may make more sense to move the pageshadow to the
 451  * port data structure, so we only allocate memory for ports actually
 452  * in use, since we at 8k per port, now.
 453  */
 454 static void init_shadow_tids(struct ipath_devdata *dd)
 455 {
 456         struct page **pages;
 457         dma_addr_t *addrs;
 458
 459         pages = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 460                         sizeof(struct page *));
 461         if (!pages) {
 462                 ipath_dev_err(dd, "failed to allocate shadow page * "
 463                               "array, no expected sends!\n");
 464                 dd->ipath_pageshadow = NULL;
 465                 return;
 466         }
 467
 468         addrs = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 469                         sizeof(dma_addr_t));
 470         if (!addrs) {
 471                 ipath_dev_err(dd, "failed to allocate shadow dma handle "
 472                               "array, no expected sends!\n");
 473                 vfree(dd->ipath_pageshadow);
 474                 dd->ipath_pageshadow = NULL;
 475                 return;
 476         }
 477
 478         memset(pages, 0, dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 479                sizeof(struct page *));
 480
 481         dd->ipath_pageshadow = pages;
 482         dd->ipath_physshadow = addrs;
 483 }
 484
 485 static void enable_chip(struct ipath_devdata *dd,
 486                         struct ipath_portdata *pd, int reinit)
 487 {
 488         u32 val;
 489         int i;
 490
 491         if (!reinit)
 492                 init_waitqueue_head(&ipath_state_wait);
 493
 494         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 495                          dd->ipath_rcvctrl);
 496
 497         /* Enable PIO send, and update of PIOavail regs to memory. */
 498         dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
 499                 INFINIPATH_S_PIOBUFAVAILUPD;
 500         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 501                          dd->ipath_sendctrl);
 502
 503         /*
 504          * enable port 0 receive, and receive interrupt.  other ports
 505          * done as user opens and inits them.
 506          */
 507         dd->ipath_rcvctrl = INFINIPATH_R_TAILUPD |
 508                 (1ULL << INFINIPATH_R_PORTENABLE_SHIFT) |
 509                 (1ULL << INFINIPATH_R_INTRAVAIL_SHIFT);
 510         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 511                          dd->ipath_rcvctrl);
 512
 513         /*
 514          * now ready for use.  this should be cleared whenever we
 515          * detect a reset, or initiate one.
 516          */
 517         dd->ipath_flags |= IPATH_INITTED;
 518
 519         /*
 520          * init our shadow copies of head from tail values, and write
 521          * head values to match.
 522          */
 523         val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
 524         (void)ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
 525         dd->ipath_port0head = ipath_read_ureg32(dd, ur_rcvhdrtail, 0);
 526
 527         /* Initialize so we interrupt on next packet received */
 528         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
 529                                dd->ipath_rhdrhead_intr_off |
 530                                dd->ipath_port0head, 0);
 531
 532         /*
 533          * by now pioavail updates to memory should have occurred, so
 534          * copy them into our working/shadow registers; this is in
 535          * case something went wrong with abort, but mostly to get the
 536          * initial values of the generation bit correct.
 537          */
 538         for (i = 0; i < dd->ipath_pioavregs; i++) {
 539                 __le64 val;
 540
 541                 /*
 542                  * Chip Errata bug 6641; even and odd qwords>3 are swapped.
 543                  */
 544                 if (i > 3) {
 545                         if (i & 1)
 546                                 val = dd->ipath_pioavailregs_dma[i - 1];
 547                         else
 548                                 val = dd->ipath_pioavailregs_dma[i + 1];
 549                 }
 550                 else
 551                         val = dd->ipath_pioavailregs_dma[i];
 552                 dd->ipath_pioavailshadow[i] = le64_to_cpu(val);
 553         }
 554         /* can get counters, stats, etc. */
 555         dd->ipath_flags |= IPATH_PRESENT;
 556 }
 557
 558 static int init_housekeeping(struct ipath_devdata *dd,
 559                              struct ipath_portdata **pdp, int reinit)
 560 {
 561         char boardn[32];
 562         int ret = 0;
 563
 564         /*
 565          * have to clear shadow copies of registers at init that are
 566          * not otherwise set here, or all kinds of bizarre things
 567          * happen with driver on chip reset
 568          */
 569         dd->ipath_rcvhdrsize = 0;
 570
 571         /*
 572          * Don't clear ipath_flags as 8bit mode was set before
 573          * entering this func. However, we do set the linkstate to
 574          * unknown, so we can watch for a transition.
 575          * PRESENT is set because we want register reads to work,
 576          * and the kernel infrastructure saw it in config space;
 577          * We clear it if we have failures.
 578          */
 579         dd->ipath_flags |= IPATH_LINKUNK | IPATH_PRESENT;
 580         dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
 581                              IPATH_LINKDOWN | IPATH_LINKINIT);
 582
 583         ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
 584         dd->ipath_revision =
 585                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
 586
 587         /*
 588          * set up fundamental info we need to use the chip; we assume
 589          * if the revision reg and these regs are OK, we don't need to
 590          * special case the rest
 591          */
 592         dd->ipath_sregbase =
 593                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
 594         dd->ipath_cregbase =
 595                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
 596         dd->ipath_uregbase =
 597                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
 598         ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
 599                    "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
 600                    dd->ipath_uregbase, dd->ipath_cregbase);
 601         if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
 602             || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
 603             || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
 604             || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
 605                 ipath_dev_err(dd, "Register read failures from chip, "
 606                               "giving up initialization\n");
 607                 dd->ipath_flags &= ~IPATH_PRESENT;
 608                 ret = -ENODEV;
 609                 goto done;
 610         }
 611
 612
 613         /* clear diagctrl register, in case diags were running and crashed */
 614         ipath_write_kreg (dd, dd->ipath_kregs->kr_hwdiagctrl, 0);
 615
 616         /* clear the initial reset flag, in case first driver load */
 617         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 618                          INFINIPATH_E_RESET);
 619
 620         if (reinit)
 621                 ret = init_chip_reset(dd, pdp);
 622         else
 623                 ret = init_chip_first(dd, pdp);
 624
 625         if (ret)
 626                 goto done;
 627
 628         ipath_cdbg(VERBOSE, "Revision %llx (PCI %x), %u ports, %u tids, "
 629                    "%u egrtids\n", (unsigned long long) dd->ipath_revision,
 630                    dd->ipath_pcirev, dd->ipath_portcnt, dd->ipath_rcvtidcnt,
 631                    dd->ipath_rcvegrcnt);
 632
 633         if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
 634              INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
 635                 ipath_dev_err(dd, "Driver only handles version %d, "
 636                               "chip swversion is %d (%llx), failng\n",
 637                               IPATH_CHIP_SWVERSION,
 638                               (int)(dd->ipath_revision >>
 639                                     INFINIPATH_R_SOFTWARE_SHIFT) &
 640                               INFINIPATH_R_SOFTWARE_MASK,
 641                               (unsigned long long) dd->ipath_revision);
 642                 ret = -ENOSYS;
 643                 goto done;
 644         }
 645         dd->ipath_majrev = (u8) ((dd->ipath_revision >>
 646                                   INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
 647                                  INFINIPATH_R_CHIPREVMAJOR_MASK);
 648         dd->ipath_minrev = (u8) ((dd->ipath_revision >>
 649                                   INFINIPATH_R_CHIPREVMINOR_SHIFT) &
 650                                  INFINIPATH_R_CHIPREVMINOR_MASK);
 651         dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
 652                                     INFINIPATH_R_BOARDID_SHIFT) &
 653                                    INFINIPATH_R_BOARDID_MASK);
 654
 655         ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
 656
 657         snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
 658                  "Driver %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
 659                  "SW Compat %u\n",
 660                  IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
 661                  (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
 662                  INFINIPATH_R_ARCH_MASK,
 663                  dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
 664                  (unsigned)(dd->ipath_revision >>
 665                             INFINIPATH_R_SOFTWARE_SHIFT) &
 666                  INFINIPATH_R_SOFTWARE_MASK);
 667
 668         ipath_dbg("%s", dd->ipath_boardversion);
 669
 670 done:
 671         return ret;
 672 }
 673
 674
 675 /**
 676  * ipath_init_chip - do the actual initialization sequence on the chip
 677  * @dd: the infinipath device
 678  * @reinit: reinitializing, so don't allocate new memory
 679  *
 680  * Do the actual initialization sequence on the chip.  This is done
 681  * both from the init routine called from the PCI infrastructure, and
 682  * when we reset the chip, or detect that it was reset internally,
 683  * or it's administratively re-enabled.
 684  *
 685  * Memory allocation here and in called routines is only done in
 686  * the first case (reinit == 0).  We have to be careful, because even
 687  * without memory allocation, we need to re-write all the chip registers
 688  * TIDs, etc. after the reset or enable has completed.
 689  */
 690 int ipath_init_chip(struct ipath_devdata *dd, int reinit)
 691 {
 692         int ret = 0, i;
 693         u32 val32, kpiobufs;
 694         u32 piobufs, uports;
 695         u64 val;
 696         struct ipath_portdata *pd = NULL; /* keep gcc4 happy */
 697         gfp_t gfp_flags = GFP_USER | __GFP_COMP;
 698
 699         ret = init_housekeeping(dd, &pd, reinit);
 700         if (ret)
 701                 goto done;
 702
 703         /*
 704          * we ignore most issues after reporting them, but have to specially
 705          * handle hardware-disabled chips.
 706          */
 707         if (ret == 2) {
 708                 /* unique error, known to ipath_init_one */
 709                 ret = -EPERM;
 710                 goto done;
 711         }
 712
 713         /*
 714          * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
 715          * but then it no longer nicely fits power of two, and since
 716          * we now use routines that backend onto __get_free_pages, the
 717          * rest would be wasted.
 718          */
 719         dd->ipath_rcvhdrcnt = dd->ipath_rcvegrcnt;
 720         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
 721                          dd->ipath_rcvhdrcnt);
 722
 723         /*
 724          * Set up the shadow copies of the piobufavail registers,
 725          * which we compare against the chip registers for now, and
 726          * the in memory DMA'ed copies of the registers.  This has to
 727          * be done early, before we calculate lastport, etc.
 728          */
 729         piobufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
 730         /*
 731          * calc number of pioavail registers, and save it; we have 2
 732          * bits per buffer.
 733          */
 734         dd->ipath_pioavregs = ALIGN(piobufs, sizeof(u64) * BITS_PER_BYTE / 2)
 735                 / (sizeof(u64) * BITS_PER_BYTE / 2);
 736         uports = dd->ipath_cfgports ? dd->ipath_cfgports - 1 : 0;
 737         if (ipath_kpiobufs == 0) {
 738                 /* not set by user (this is default) */
 739                 if (piobufs >= (uports * IPATH_MIN_USER_PORT_BUFCNT) + 32)
 740                         kpiobufs = 32;
 741                 else
 742                         kpiobufs = 16;
 743         }
 744         else
 745                 kpiobufs = ipath_kpiobufs;
 746
 747         if (kpiobufs + (uports * IPATH_MIN_USER_PORT_BUFCNT) > piobufs) {
 748                 i = (int) piobufs -
 749                         (int) (uports * IPATH_MIN_USER_PORT_BUFCNT);
 750                 if (i < 0)
 751                         i = 0;
 752                 dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs of "
 753                          "%d for kernel leaves too few for %d user ports "
 754                          "(%d each); using %u\n", kpiobufs,
 755                          piobufs, uports, IPATH_MIN_USER_PORT_BUFCNT, i);
 756                 /*
 757                  * shouldn't change ipath_kpiobufs, because could be
 758                  * different for different devices...
 759                  */
 760                 kpiobufs = i;
 761         }
 762         dd->ipath_lastport_piobuf = piobufs - kpiobufs;
 763         dd->ipath_pbufsport =
 764                 uports ? dd->ipath_lastport_piobuf / uports : 0;
 765         val32 = dd->ipath_lastport_piobuf - (dd->ipath_pbufsport * uports);
 766         if (val32 > 0) {
 767                 ipath_dbg("allocating %u pbufs/port leaves %u unused, "
 768                           "add to kernel\n", dd->ipath_pbufsport, val32);
 769                 dd->ipath_lastport_piobuf -= val32;
 770                 ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
 771                           dd->ipath_pbufsport, val32);
 772         }
 773         dd->ipath_lastpioindex = dd->ipath_lastport_piobuf;
 774         ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
 775                    "each for %u user ports\n", kpiobufs,
 776                    piobufs, dd->ipath_pbufsport, uports);
 777
 778         dd->ipath_f_early_init(dd);
 779
 780         /* early_init sets rcvhdrentsize and rcvhdrsize, so this must be
 781          * done after early_init */
 782         dd->ipath_hdrqlast =
 783                 dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
 784         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
 785                          dd->ipath_rcvhdrentsize);
 786         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
 787                          dd->ipath_rcvhdrsize);
 788
 789         if (!reinit) {
 790                 ret = init_pioavailregs(dd);
 791                 init_shadow_tids(dd);
 792                 if (ret)
 793                         goto done;
 794         }
 795
 796         (void)ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
 797                                dd->ipath_pioavailregs_phys);
 798         /*
 799          * this is to detect s/w errors, which the h/w works around by
 800          * ignoring the low 6 bits of address, if it wasn't aligned.
 801          */
 802         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
 803         if (val != dd->ipath_pioavailregs_phys) {
 804                 ipath_dev_err(dd, "Catastrophic software error, "
 805                               "SendPIOAvailAddr written as %lx, "
 806                               "read back as %llx\n",
 807                               (unsigned long) dd->ipath_pioavailregs_phys,
 808                               (unsigned long long) val);
 809                 ret = -EINVAL;
 810                 goto done;
 811         }
 812
 813         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
 814
 815         /*
 816          * make sure we are not in freeze, and PIO send enabled, so
 817          * writes to pbc happen
 818          */
 819         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
 820         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 821                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 822         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
 823         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 824                          INFINIPATH_S_PIOENABLE);
 825
 826         /*
 827          * before error clears, since we expect serdes pll errors during
 828          * this, the first time after reset
 829          */
 830         if (bringup_link(dd)) {
 831                 dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
 832                 ret = -ENETDOWN;
 833                 goto done;
 834         }
 835
 836         /*
 837          * clear any "expected" hwerrs from reset and/or initialization
 838          * clear any that aren't enabled (at least this once), and then
 839          * set the enable mask
 840          */
 841         dd->ipath_f_init_hwerrors(dd);
 842         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 843                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 844         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 845                          dd->ipath_hwerrmask);
 846
 847         dd->ipath_maskederrs = dd->ipath_ignorederrs;
 848         /* clear all */
 849         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
 850         /* enable errors that are masked, at least this first time. */
 851         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 852                          ~dd->ipath_maskederrs);
 853         /* clear any interrups up to this point (ints still not enabled) */
 854         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
 855
 856         /*
 857          * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
 858          * re-init, the simplest way to handle this is to free
 859          * existing, and re-allocate.
 860          * Need to re-create rest of port 0 portdata as well.
 861          */
 862         if (reinit) {
 863                 /* Alloc and init new ipath_portdata for port0,
 864                  * Then free old pd. Could lead to fragmentation, but also
 865                  * makes later support for hot-swap easier.
 866                  */
 867                 struct ipath_portdata *npd;
 868                 npd = create_portdata0(dd);
 869                 if (npd) {
 870                         ipath_free_pddata(dd, pd);
 871                         dd->ipath_pd[0] = pd = npd;
 872                 } else {
 873                         ipath_dev_err(dd, "Unable to allocate portdata for"
 874                                       "  port 0, failing\n");
 875                         ret = -ENOMEM;
 876                         goto done;
 877                 }
 878         }
 879         dd->ipath_f_tidtemplate(dd);
 880         ret = ipath_create_rcvhdrq(dd, pd);
 881         if (!ret) {
 882                 dd->ipath_hdrqtailptr =
 883                         (volatile __le64 *)pd->port_rcvhdrtail_kvaddr;
 884                 ret = create_port0_egr(dd);
 885         }
 886         if (ret)
 887                 ipath_dev_err(dd, "failed to allocate port 0 (kernel) "
 888                               "rcvhdrq and/or egr bufs\n");
 889         else
 890                 enable_chip(dd, pd, reinit);
 891
 892
 893         if (!ret && !reinit) {
 894             /* used when we close a port, for DMA already in flight at close */
 895                 dd->ipath_dummy_hdrq = dma_alloc_coherent(
 896                         &dd->pcidev->dev, pd->port_rcvhdrq_size,
 897                         &dd->ipath_dummy_hdrq_phys,
 898                         gfp_flags);
 899                 if (!dd->ipath_dummy_hdrq ) {
 900                         dev_info(&dd->pcidev->dev,
 901                                 "Couldn't allocate 0x%lx bytes for dummy hdrq\n",
 902                                 pd->port_rcvhdrq_size);
 903                         /* fallback to just 0'ing */
 904                         dd->ipath_dummy_hdrq_phys = 0UL;
 905                 }
 906         }
 907
 908         /*
 909          * cause retrigger of pending interrupts ignored during init,
 910          * even if we had errors
 911          */
 912         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 913
 914         if(!dd->ipath_stats_timer_active) {
 915                 /*
 916                  * first init, or after an admin disable/enable
 917                  * set up stats retrieval timer, even if we had errors
 918                  * in last portion of setup
 919                  */
 920                 init_timer(&dd->ipath_stats_timer);
 921                 dd->ipath_stats_timer.function = ipath_get_faststats;
 922                 dd->ipath_stats_timer.data = (unsigned long) dd;
 923                 /* every 5 seconds; */
 924                 dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
 925                 /* takes ~16 seconds to overflow at full IB 4x bandwdith */
 926                 add_timer(&dd->ipath_stats_timer);
 927                 dd->ipath_stats_timer_active = 1;
 928         }
 929
 930 done:
 931         if (!ret) {
 932                 *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
 933                 if (!dd->ipath_f_intrsetup(dd)) {
 934                         /* now we can enable all interrupts from the chip */
 935                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 936                                          -1LL);
 937                         /* force re-interrupt of any pending interrupts. */
 938                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
 939                                          0ULL);
 940                         /* chip is usable; mark it as initialized */
 941                         *dd->ipath_statusp |= IPATH_STATUS_INITTED;
 942                 } else
 943                         ipath_dev_err(dd, "No interrupts enabled, couldn't "
 944                                       "setup interrupt address\n");
 945
 946                 if (dd->ipath_cfgports > ipath_stats.sps_nports)
 947                         /*
 948                          * sps_nports is a global, so, we set it to
 949                          * the highest number of ports of any of the
 950                          * chips we find; we never decrement it, at
 951                          * least for now.  Since this might have changed
 952                          * over disable/enable or prior to reset, always
 953                          * do the check and potentially adjust.
 954                          */
 955                         ipath_stats.sps_nports = dd->ipath_cfgports;
 956         } else
 957                 ipath_dbg("Failed (%d) to initialize chip\n", ret);
 958
 959         /* if ret is non-zero, we probably should do some cleanup
 960            here... */
 961         return ret;
 962 }
 963
 964 static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
 965 {
 966         struct ipath_devdata *dd;
 967         unsigned long flags;
 968         unsigned short val;
 969         int ret;
 970
 971         ret = ipath_parse_ushort(str, &val);
 972
 973         spin_lock_irqsave(&ipath_devs_lock, flags);
 974
 975         if (ret < 0)
 976                 goto bail;
 977
 978         if (val == 0) {
 979                 ret = -EINVAL;
 980                 goto bail;
 981         }
 982
 983         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 984                 if (dd->ipath_kregbase)
 985                         continue;
 986                 if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
 987                            (dd->ipath_cfgports *
 988                             IPATH_MIN_USER_PORT_BUFCNT)))
 989                 {
 990                         ipath_dev_err(
 991                                 dd,
 992                                 "Allocating %d PIO bufs for kernel leaves "
 993                                 "too few for %d user ports (%d each)\n",
 994                                 val, dd->ipath_cfgports - 1,
 995                                 IPATH_MIN_USER_PORT_BUFCNT);
 996                         ret = -EINVAL;
 997                         goto bail;
 998                 }
 999                 dd->ipath_lastport_piobuf =
1000                         dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
1001         }
1002
1003         ipath_kpiobufs = val;
1004         ret = 0;
1005 bail:
1006         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1007
1008         return ret;
1009 }