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