drivers/infiniband/hw/ipath/ipath_init_chip.c

   1 /*
   2  * Copyright (c) 2006, 2007 QLogic Corporation. 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_p0_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),
 137                                     RCVHQ_RCV_TYPE_EAGER,
 138                                     dd->ipath_port0_skbinfo[e].phys);
 139         }
 140
 141         ret = 0;
 142
 143 bail:
 144         return ret;
 145 }
 146
 147 static int bringup_link(struct ipath_devdata *dd)
 148 {
 149         u64 val, ibc;
 150         int ret = 0;
 151
 152         /* hold IBC in reset */
 153         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
 154         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 155                          dd->ipath_control);
 156
 157         /*
 158          * set initial max size pkt IBC will send, including ICRC; it's the
 159          * PIO buffer size in dwords, less 1; also see ipath_set_mtu()
 160          */
 161         val = (dd->ipath_ibmaxlen >> 2) + 1;
 162         ibc = val << dd->ibcc_mpl_shift;
 163
 164         /* flowcontrolwatermark is in units of KBytes */
 165         ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
 166         /*
 167          * How often flowctrl sent.  More or less in usecs; balance against
 168          * watermark value, so that in theory senders always get a flow
 169          * control update in time to not let the IB link go idle.
 170          */
 171         ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
 172         /* max error tolerance */
 173         ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
 174         /* use "real" buffer space for */
 175         ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
 176         /* IB credit flow control. */
 177         ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
 178         /* initially come up waiting for TS1, without sending anything. */
 179         dd->ipath_ibcctrl = ibc;
 180         /*
 181          * Want to start out with both LINKCMD and LINKINITCMD in NOP
 182          * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
 183          * to stay a NOP. Flag that we are disabled, for the (unlikely)
 184          * case that some recovery path is trying to bring the link up
 185          * before we are ready.
 186          */
 187         ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
 188                 INFINIPATH_IBCC_LINKINITCMD_SHIFT;
 189         dd->ipath_flags |= IPATH_IB_LINK_DISABLED;
 190         ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
 191                    (unsigned long long) ibc);
 192         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
 193
 194         // be sure chip saw it
 195         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 196
 197         ret = dd->ipath_f_bringup_serdes(dd);
 198
 199         if (ret)
 200                 dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
 201                          "not usable\n");
 202         else {
 203                 /* enable IBC */
 204                 dd->ipath_control |= INFINIPATH_C_LINKENABLE;
 205                 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 206                                  dd->ipath_control);
 207         }
 208
 209         return ret;
 210 }
 211
 212 static struct ipath_portdata *create_portdata0(struct ipath_devdata *dd)
 213 {
 214         struct ipath_portdata *pd = NULL;
 215
 216         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 217         if (pd) {
 218                 pd->port_dd = dd;
 219                 pd->port_cnt = 1;
 220                 /* The port 0 pkey table is used by the layer interface. */
 221                 pd->port_pkeys[0] = IPATH_DEFAULT_P_KEY;
 222                 pd->port_seq_cnt = 1;
 223         }
 224         return pd;
 225 }
 226
 227 static int init_chip_first(struct ipath_devdata *dd)
 228 {
 229         struct ipath_portdata *pd;
 230         int ret = 0;
 231         u64 val;
 232
 233         /*
 234          * skip cfgports stuff because we are not allocating memory,
 235          * and we don't want problems if the portcnt changed due to
 236          * cfgports.  We do still check and report a difference, if
 237          * not same (should be impossible).
 238          */
 239         dd->ipath_f_config_ports(dd, ipath_cfgports);
 240         if (!ipath_cfgports)
 241                 dd->ipath_cfgports = dd->ipath_portcnt;
 242         else if (ipath_cfgports <= dd->ipath_portcnt) {
 243                 dd->ipath_cfgports = ipath_cfgports;
 244                 ipath_dbg("Configured to use %u ports out of %u in chip\n",
 245                           dd->ipath_cfgports, ipath_read_kreg32(dd,
 246                           dd->ipath_kregs->kr_portcnt));
 247         } else {
 248                 dd->ipath_cfgports = dd->ipath_portcnt;
 249                 ipath_dbg("Tried to configured to use %u ports; chip "
 250                           "only supports %u\n", ipath_cfgports,
 251                           ipath_read_kreg32(dd,
 252                                   dd->ipath_kregs->kr_portcnt));
 253         }
 254         /*
 255          * Allocate full portcnt array, rather than just cfgports, because
 256          * cleanup iterates across all possible ports.
 257          */
 258         dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_portcnt,
 259                                GFP_KERNEL);
 260
 261         if (!dd->ipath_pd) {
 262                 ipath_dev_err(dd, "Unable to allocate portdata array, "
 263                               "failing\n");
 264                 ret = -ENOMEM;
 265                 goto done;
 266         }
 267
 268         pd = create_portdata0(dd);
 269         if (!pd) {
 270                 ipath_dev_err(dd, "Unable to allocate portdata for port "
 271                               "0, failing\n");
 272                 ret = -ENOMEM;
 273                 goto done;
 274         }
 275         dd->ipath_pd[0] = pd;
 276
 277         dd->ipath_rcvtidcnt =
 278                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 279         dd->ipath_rcvtidbase =
 280                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 281         dd->ipath_rcvegrcnt =
 282                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 283         dd->ipath_rcvegrbase =
 284                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 285         dd->ipath_palign =
 286                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
 287         dd->ipath_piobufbase =
 288                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
 289         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
 290         dd->ipath_piosize2k = val & ~0U;
 291         dd->ipath_piosize4k = val >> 32;
 292         if (dd->ipath_piosize4k == 0 && ipath_mtu4096)
 293                 ipath_mtu4096 = 0; /* 4KB not supported by this chip */
 294         dd->ipath_ibmtu = ipath_mtu4096 ? 4096 : 2048;
 295         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
 296         dd->ipath_piobcnt2k = val & ~0U;
 297         dd->ipath_piobcnt4k = val >> 32;
 298         dd->ipath_pio2kbase =
 299                 (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 300                                  (dd->ipath_piobufbase & 0xffffffff));
 301         if (dd->ipath_piobcnt4k) {
 302                 dd->ipath_pio4kbase = (u32 __iomem *)
 303                         (((char __iomem *) dd->ipath_kregbase) +
 304                          (dd->ipath_piobufbase >> 32));
 305                 /*
 306                  * 4K buffers take 2 pages; we use roundup just to be
 307                  * paranoid; we calculate it once here, rather than on
 308                  * ever buf allocate
 309                  */
 310                 dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
 311                                           dd->ipath_palign);
 312                 ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
 313                           "(%x aligned)\n",
 314                           dd->ipath_piobcnt2k, dd->ipath_piosize2k,
 315                           dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
 316                           dd->ipath_piosize4k, dd->ipath_pio4kbase,
 317                           dd->ipath_4kalign);
 318         }
 319         else ipath_dbg("%u 2k piobufs @ %p\n",
 320                        dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
 321
 322         spin_lock_init(&dd->ipath_tid_lock);
 323         spin_lock_init(&dd->ipath_sendctrl_lock);
 324         spin_lock_init(&dd->ipath_gpio_lock);
 325         spin_lock_init(&dd->ipath_eep_st_lock);
 326         mutex_init(&dd->ipath_eep_lock);
 327
 328 done:
 329         return ret;
 330 }
 331
 332 /**
 333  * init_chip_reset - re-initialize after a reset, or enable
 334  * @dd: the infinipath device
 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 {
 342         u32 rtmp;
 343         int i;
 344
 345         /*
 346          * ensure chip does no sends or receives, tail updates, or
 347          * pioavail updates while we re-initialize
 348          */
 349         dd->ipath_rcvctrl &= ~(1ULL << dd->ipath_r_tailupd_shift);
 350         for (i = 0; i < dd->ipath_portcnt; i++) {
 351                 clear_bit(dd->ipath_r_portenable_shift + i,
 352                           &dd->ipath_rcvctrl);
 353                 clear_bit(dd->ipath_r_intravail_shift + i,
 354                           &dd->ipath_rcvctrl);
 355         }
 356         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 357                 dd->ipath_rcvctrl);
 358
 359         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 360         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, dd->ipath_control);
 361
 362         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
 363         if (rtmp != dd->ipath_rcvtidcnt)
 364                 dev_info(&dd->pcidev->dev, "tidcnt was %u before "
 365                          "reset, now %u, using original\n",
 366                          dd->ipath_rcvtidcnt, rtmp);
 367         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
 368         if (rtmp != dd->ipath_rcvtidbase)
 369                 dev_info(&dd->pcidev->dev, "tidbase was %u before "
 370                          "reset, now %u, using original\n",
 371                          dd->ipath_rcvtidbase, rtmp);
 372         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
 373         if (rtmp != dd->ipath_rcvegrcnt)
 374                 dev_info(&dd->pcidev->dev, "egrcnt was %u before "
 375                          "reset, now %u, using original\n",
 376                          dd->ipath_rcvegrcnt, rtmp);
 377         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
 378         if (rtmp != dd->ipath_rcvegrbase)
 379                 dev_info(&dd->pcidev->dev, "egrbase was %u before "
 380                          "reset, now %u, using original\n",
 381                          dd->ipath_rcvegrbase, rtmp);
 382
 383         return 0;
 384 }
 385
 386 static int init_pioavailregs(struct ipath_devdata *dd)
 387 {
 388         int ret;
 389
 390         dd->ipath_pioavailregs_dma = dma_alloc_coherent(
 391                 &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
 392                 GFP_KERNEL);
 393         if (!dd->ipath_pioavailregs_dma) {
 394                 ipath_dev_err(dd, "failed to allocate PIOavail reg area "
 395                               "in memory\n");
 396                 ret = -ENOMEM;
 397                 goto done;
 398         }
 399
 400         /*
 401          * we really want L2 cache aligned, but for current CPUs of
 402          * interest, they are the same.
 403          */
 404         dd->ipath_statusp = (u64 *)
 405                 ((char *)dd->ipath_pioavailregs_dma +
 406                  ((2 * L1_CACHE_BYTES +
 407                    dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
 408         /* copy the current value now that it's really allocated */
 409         *dd->ipath_statusp = dd->_ipath_status;
 410         /*
 411          * setup buffer to hold freeze msg, accessible to apps,
 412          * following statusp
 413          */
 414         dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
 415         /* and its length */
 416         dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
 417
 418         ret = 0;
 419
 420 done:
 421         return ret;
 422 }
 423
 424 /**
 425  * init_shadow_tids - allocate the shadow TID array
 426  * @dd: the infinipath device
 427  *
 428  * allocate the shadow TID array, so we can ipath_munlock previous
 429  * entries.  It may make more sense to move the pageshadow to the
 430  * port data structure, so we only allocate memory for ports actually
 431  * in use, since we at 8k per port, now.
 432  */
 433 static void init_shadow_tids(struct ipath_devdata *dd)
 434 {
 435         struct page **pages;
 436         dma_addr_t *addrs;
 437
 438         pages = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 439                         sizeof(struct page *));
 440         if (!pages) {
 441                 ipath_dev_err(dd, "failed to allocate shadow page * "
 442                               "array, no expected sends!\n");
 443                 dd->ipath_pageshadow = NULL;
 444                 return;
 445         }
 446
 447         addrs = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 448                         sizeof(dma_addr_t));
 449         if (!addrs) {
 450                 ipath_dev_err(dd, "failed to allocate shadow dma handle "
 451                               "array, no expected sends!\n");
 452                 vfree(dd->ipath_pageshadow);
 453                 dd->ipath_pageshadow = NULL;
 454                 return;
 455         }
 456
 457         memset(pages, 0, dd->ipath_cfgports * dd->ipath_rcvtidcnt *
 458                sizeof(struct page *));
 459
 460         dd->ipath_pageshadow = pages;
 461         dd->ipath_physshadow = addrs;
 462 }
 463
 464 static void enable_chip(struct ipath_devdata *dd, int reinit)
 465 {
 466         u32 val;
 467         u64 rcvmask;
 468         unsigned long flags;
 469         int i;
 470
 471         if (!reinit)
 472                 init_waitqueue_head(&ipath_state_wait);
 473
 474         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 475                          dd->ipath_rcvctrl);
 476
 477         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 478         /* Enable PIO send, and update of PIOavail regs to memory. */
 479         dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
 480                 INFINIPATH_S_PIOBUFAVAILUPD;
 481         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 482         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 483         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 484
 485         /*
 486          * Enable kernel ports' receive and receive interrupt.
 487          * Other ports done as user opens and inits them.
 488          */
 489         rcvmask = 1ULL;
 490         dd->ipath_rcvctrl |= (rcvmask << dd->ipath_r_portenable_shift) |
 491                 (rcvmask << dd->ipath_r_intravail_shift);
 492         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL))
 493                 dd->ipath_rcvctrl |= (1ULL << dd->ipath_r_tailupd_shift);
 494
 495         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 496                          dd->ipath_rcvctrl);
 497
 498         /*
 499          * now ready for use.  this should be cleared whenever we
 500          * detect a reset, or initiate one.
 501          */
 502         dd->ipath_flags |= IPATH_INITTED;
 503
 504         /*
 505          * Init our shadow copies of head from tail values,
 506          * and write head values to match.
 507          */
 508         val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
 509         ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
 510
 511         /* Initialize so we interrupt on next packet received */
 512         ipath_write_ureg(dd, ur_rcvhdrhead,
 513                          dd->ipath_rhdrhead_intr_off |
 514                          dd->ipath_pd[0]->port_head, 0);
 515
 516         /*
 517          * by now pioavail updates to memory should have occurred, so
 518          * copy them into our working/shadow registers; this is in
 519          * case something went wrong with abort, but mostly to get the
 520          * initial values of the generation bit correct.
 521          */
 522         for (i = 0; i < dd->ipath_pioavregs; i++) {
 523                 __le64 pioavail;
 524
 525                 /*
 526                  * Chip Errata bug 6641; even and odd qwords>3 are swapped.
 527                  */
 528                 if (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS))
 529                         pioavail = dd->ipath_pioavailregs_dma[i ^ 1];
 530                 else
 531                         pioavail = dd->ipath_pioavailregs_dma[i];
 532                 dd->ipath_pioavailshadow[i] = le64_to_cpu(pioavail) |
 533                         (~dd->ipath_pioavailkernel[i] <<
 534                         INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT);
 535         }
 536         /* can get counters, stats, etc. */
 537         dd->ipath_flags |= IPATH_PRESENT;
 538 }
 539
 540 static int init_housekeeping(struct ipath_devdata *dd, int reinit)
 541 {
 542         char boardn[32];
 543         int ret = 0;
 544
 545         /*
 546          * have to clear shadow copies of registers at init that are
 547          * not otherwise set here, or all kinds of bizarre things
 548          * happen with driver on chip reset
 549          */
 550         dd->ipath_rcvhdrsize = 0;
 551
 552         /*
 553          * Don't clear ipath_flags as 8bit mode was set before
 554          * entering this func. However, we do set the linkstate to
 555          * unknown, so we can watch for a transition.
 556          * PRESENT is set because we want register reads to work,
 557          * and the kernel infrastructure saw it in config space;
 558          * We clear it if we have failures.
 559          */
 560         dd->ipath_flags |= IPATH_LINKUNK | IPATH_PRESENT;
 561         dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
 562                              IPATH_LINKDOWN | IPATH_LINKINIT);
 563
 564         ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
 565         dd->ipath_revision =
 566                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
 567
 568         /*
 569          * set up fundamental info we need to use the chip; we assume
 570          * if the revision reg and these regs are OK, we don't need to
 571          * special case the rest
 572          */
 573         dd->ipath_sregbase =
 574                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
 575         dd->ipath_cregbase =
 576                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
 577         dd->ipath_uregbase =
 578                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
 579         ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
 580                    "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
 581                    dd->ipath_uregbase, dd->ipath_cregbase);
 582         if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
 583             || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
 584             || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
 585             || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
 586                 ipath_dev_err(dd, "Register read failures from chip, "
 587                               "giving up initialization\n");
 588                 dd->ipath_flags &= ~IPATH_PRESENT;
 589                 ret = -ENODEV;
 590                 goto done;
 591         }
 592
 593
 594         /* clear diagctrl register, in case diags were running and crashed */
 595         ipath_write_kreg (dd, dd->ipath_kregs->kr_hwdiagctrl, 0);
 596
 597         /* clear the initial reset flag, in case first driver load */
 598         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
 599                          INFINIPATH_E_RESET);
 600
 601         ipath_cdbg(VERBOSE, "Revision %llx (PCI %x)\n",
 602                    (unsigned long long) dd->ipath_revision,
 603                    dd->ipath_pcirev);
 604
 605         if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
 606              INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
 607                 ipath_dev_err(dd, "Driver only handles version %d, "
 608                               "chip swversion is %d (%llx), failng\n",
 609                               IPATH_CHIP_SWVERSION,
 610                               (int)(dd->ipath_revision >>
 611                                     INFINIPATH_R_SOFTWARE_SHIFT) &
 612                               INFINIPATH_R_SOFTWARE_MASK,
 613                               (unsigned long long) dd->ipath_revision);
 614                 ret = -ENOSYS;
 615                 goto done;
 616         }
 617         dd->ipath_majrev = (u8) ((dd->ipath_revision >>
 618                                   INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
 619                                  INFINIPATH_R_CHIPREVMAJOR_MASK);
 620         dd->ipath_minrev = (u8) ((dd->ipath_revision >>
 621                                   INFINIPATH_R_CHIPREVMINOR_SHIFT) &
 622                                  INFINIPATH_R_CHIPREVMINOR_MASK);
 623         dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
 624                                     INFINIPATH_R_BOARDID_SHIFT) &
 625                                    INFINIPATH_R_BOARDID_MASK);
 626
 627         ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
 628
 629         snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
 630                  "ChipABI %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
 631                  "SW Compat %u\n",
 632                  IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
 633                  (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
 634                  INFINIPATH_R_ARCH_MASK,
 635                  dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
 636                  (unsigned)(dd->ipath_revision >>
 637                             INFINIPATH_R_SOFTWARE_SHIFT) &
 638                  INFINIPATH_R_SOFTWARE_MASK);
 639
 640         ipath_dbg("%s", dd->ipath_boardversion);
 641
 642         if (ret)
 643                 goto done;
 644
 645         if (reinit)
 646                 ret = init_chip_reset(dd);
 647         else
 648                 ret = init_chip_first(dd);
 649
 650 done:
 651         return ret;
 652 }
 653
 654 /**
 655  * ipath_init_chip - do the actual initialization sequence on the chip
 656  * @dd: the infinipath device
 657  * @reinit: reinitializing, so don't allocate new memory
 658  *
 659  * Do the actual initialization sequence on the chip.  This is done
 660  * both from the init routine called from the PCI infrastructure, and
 661  * when we reset the chip, or detect that it was reset internally,
 662  * or it's administratively re-enabled.
 663  *
 664  * Memory allocation here and in called routines is only done in
 665  * the first case (reinit == 0).  We have to be careful, because even
 666  * without memory allocation, we need to re-write all the chip registers
 667  * TIDs, etc. after the reset or enable has completed.
 668  */
 669 int ipath_init_chip(struct ipath_devdata *dd, int reinit)
 670 {
 671         int ret = 0;
 672         u32 val32, kpiobufs;
 673         u32 piobufs, uports;
 674         u64 val;
 675         struct ipath_portdata *pd;
 676         gfp_t gfp_flags = GFP_USER | __GFP_COMP;
 677         unsigned long flags;
 678
 679         ret = init_housekeeping(dd, reinit);
 680         if (ret)
 681                 goto done;
 682
 683         /*
 684          * we ignore most issues after reporting them, but have to specially
 685          * handle hardware-disabled chips.
 686          */
 687         if (ret == 2) {
 688                 /* unique error, known to ipath_init_one */
 689                 ret = -EPERM;
 690                 goto done;
 691         }
 692
 693         /*
 694          * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
 695          * but then it no longer nicely fits power of two, and since
 696          * we now use routines that backend onto __get_free_pages, the
 697          * rest would be wasted.
 698          */
 699         dd->ipath_rcvhdrcnt = max(dd->ipath_p0_rcvegrcnt, dd->ipath_rcvegrcnt);
 700         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
 701                          dd->ipath_rcvhdrcnt);
 702
 703         /*
 704          * Set up the shadow copies of the piobufavail registers,
 705          * which we compare against the chip registers for now, and
 706          * the in memory DMA'ed copies of the registers.  This has to
 707          * be done early, before we calculate lastport, etc.
 708          */
 709         piobufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
 710         /*
 711          * calc number of pioavail registers, and save it; we have 2
 712          * bits per buffer.
 713          */
 714         dd->ipath_pioavregs = ALIGN(piobufs, sizeof(u64) * BITS_PER_BYTE / 2)
 715                 / (sizeof(u64) * BITS_PER_BYTE / 2);
 716         uports = dd->ipath_cfgports ? dd->ipath_cfgports - 1 : 0;
 717         if (ipath_kpiobufs == 0) {
 718                 /* not set by user (this is default) */
 719                 if (piobufs > 144)
 720                         kpiobufs = 32;
 721                 else
 722                         kpiobufs = 16;
 723         }
 724         else
 725                 kpiobufs = ipath_kpiobufs;
 726
 727         if (kpiobufs + (uports * IPATH_MIN_USER_PORT_BUFCNT) > piobufs) {
 728                 int i = (int) piobufs -
 729                         (int) (uports * IPATH_MIN_USER_PORT_BUFCNT);
 730                 if (i < 1)
 731                         i = 1;
 732                 dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs of "
 733                          "%d for kernel leaves too few for %d user ports "
 734                          "(%d each); using %u\n", kpiobufs,
 735                          piobufs, uports, IPATH_MIN_USER_PORT_BUFCNT, i);
 736                 /*
 737                  * shouldn't change ipath_kpiobufs, because could be
 738                  * different for different devices...
 739                  */
 740                 kpiobufs = i;
 741         }
 742         dd->ipath_lastport_piobuf = piobufs - kpiobufs;
 743         dd->ipath_pbufsport =
 744                 uports ? dd->ipath_lastport_piobuf / uports : 0;
 745         val32 = dd->ipath_lastport_piobuf - (dd->ipath_pbufsport * uports);
 746         if (val32 > 0) {
 747                 ipath_dbg("allocating %u pbufs/port leaves %u unused, "
 748                           "add to kernel\n", dd->ipath_pbufsport, val32);
 749                 dd->ipath_lastport_piobuf -= val32;
 750                 kpiobufs += val32;
 751                 ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
 752                           dd->ipath_pbufsport, val32);
 753         }
 754         dd->ipath_lastpioindex = 0;
 755         dd->ipath_lastpioindexl = dd->ipath_piobcnt2k;
 756         ipath_chg_pioavailkernel(dd, 0, piobufs, 1);
 757         ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
 758                    "each for %u user ports\n", kpiobufs,
 759                    piobufs, dd->ipath_pbufsport, uports);
 760
 761         dd->ipath_f_early_init(dd);
 762         /*
 763          * Cancel any possible active sends from early driver load.
 764          * Follows early_init because some chips have to initialize
 765          * PIO buffers in early_init to avoid false parity errors.
 766          */
 767         ipath_cancel_sends(dd, 0);
 768
 769         /*
 770          * Early_init sets rcvhdrentsize and rcvhdrsize, so this must be
 771          * done after early_init.
 772          */
 773         dd->ipath_hdrqlast =
 774                 dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
 775         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
 776                          dd->ipath_rcvhdrentsize);
 777         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
 778                          dd->ipath_rcvhdrsize);
 779
 780         if (!reinit) {
 781                 ret = init_pioavailregs(dd);
 782                 init_shadow_tids(dd);
 783                 if (ret)
 784                         goto done;
 785         }
 786
 787         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
 788                          dd->ipath_pioavailregs_phys);
 789         /*
 790          * this is to detect s/w errors, which the h/w works around by
 791          * ignoring the low 6 bits of address, if it wasn't aligned.
 792          */
 793         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
 794         if (val != dd->ipath_pioavailregs_phys) {
 795                 ipath_dev_err(dd, "Catastrophic software error, "
 796                               "SendPIOAvailAddr written as %lx, "
 797                               "read back as %llx\n",
 798                               (unsigned long) dd->ipath_pioavailregs_phys,
 799                               (unsigned long long) val);
 800                 ret = -EINVAL;
 801                 goto done;
 802         }
 803
 804         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
 805
 806         /*
 807          * make sure we are not in freeze, and PIO send enabled, so
 808          * writes to pbc happen
 809          */
 810         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
 811         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 812                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 813         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
 814
 815         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 816         dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE;
 817         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 818         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 819         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 820
 821         /*
 822          * before error clears, since we expect serdes pll errors during
 823          * this, the first time after reset
 824          */
 825         if (bringup_link(dd)) {
 826                 dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
 827                 ret = -ENETDOWN;
 828                 goto done;
 829         }
 830
 831         /*
 832          * clear any "expected" hwerrs from reset and/or initialization
 833          * clear any that aren't enabled (at least this once), and then
 834          * set the enable mask
 835          */
 836         dd->ipath_f_init_hwerrors(dd);
 837         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 838                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
 839         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 840                          dd->ipath_hwerrmask);
 841
 842         /* clear all */
 843         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
 844         /* enable errors that are masked, at least this first time. */
 845         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 846                          ~dd->ipath_maskederrs);
 847         dd->ipath_maskederrs = 0; /* don't re-enable ignored in timer */
 848         dd->ipath_errormask =
 849                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
 850         /* clear any interrupts up to this point (ints still not enabled) */
 851         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
 852
 853         dd->ipath_f_tidtemplate(dd);
 854
 855         /*
 856          * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
 857          * re-init, the simplest way to handle this is to free
 858          * existing, and re-allocate.
 859          * Need to re-create rest of port 0 portdata as well.
 860          */
 861         pd = dd->ipath_pd[0];
 862         if (reinit) {
 863                 struct ipath_portdata *npd;
 864
 865                 /*
 866                  * Alloc and init new ipath_portdata for port0,
 867                  * Then free old pd. Could lead to fragmentation, but also
 868                  * makes later support for hot-swap easier.
 869                  */
 870                 npd = create_portdata0(dd);
 871                 if (npd) {
 872                         ipath_free_pddata(dd, pd);
 873                         dd->ipath_pd[0] = npd;
 874                         pd = npd;
 875                 } else {
 876                         ipath_dev_err(dd, "Unable to allocate portdata"
 877                                       " for port 0, failing\n");
 878                         ret = -ENOMEM;
 879                         goto done;
 880                 }
 881         }
 882         ret = ipath_create_rcvhdrq(dd, pd);
 883         if (!ret)
 884                 ret = create_port0_egr(dd);
 885         if (ret) {
 886                 ipath_dev_err(dd, "failed to allocate kernel port's "
 887                               "rcvhdrq and/or egr bufs\n");
 888                 goto done;
 889         }
 890         else
 891                 enable_chip(dd, reinit);
 892
 893         if (!reinit) {
 894                 /*
 895                  * Used when we close a port, for DMA already in flight
 896                  * at close.
 897                  */
 898                 dd->ipath_dummy_hdrq = dma_alloc_coherent(
 899                         &dd->pcidev->dev, dd->ipath_pd[0]->port_rcvhdrq_size,
 900                         &dd->ipath_dummy_hdrq_phys,
 901                         gfp_flags);
 902                 if (!dd->ipath_dummy_hdrq) {
 903                         dev_info(&dd->pcidev->dev,
 904                                 "Couldn't allocate 0x%lx bytes for dummy hdrq\n",
 905                                 dd->ipath_pd[0]->port_rcvhdrq_size);
 906                         /* fallback to just 0'ing */
 907                         dd->ipath_dummy_hdrq_phys = 0UL;
 908                 }
 909         }
 910
 911         /*
 912          * cause retrigger of pending interrupts ignored during init,
 913          * even if we had errors
 914          */
 915         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 916
 917         if (!dd->ipath_stats_timer_active) {
 918                 /*
 919                  * first init, or after an admin disable/enable
 920                  * set up stats retrieval timer, even if we had errors
 921                  * in last portion of setup
 922                  */
 923                 init_timer(&dd->ipath_stats_timer);
 924                 dd->ipath_stats_timer.function = ipath_get_faststats;
 925                 dd->ipath_stats_timer.data = (unsigned long) dd;
 926                 /* every 5 seconds; */
 927                 dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
 928                 /* takes ~16 seconds to overflow at full IB 4x bandwdith */
 929                 add_timer(&dd->ipath_stats_timer);
 930                 dd->ipath_stats_timer_active = 1;
 931         }
 932
 933         /* Set up HoL state */
 934         init_timer(&dd->ipath_hol_timer);
 935         dd->ipath_hol_timer.function = ipath_hol_event;
 936         dd->ipath_hol_timer.data = (unsigned long)dd;
 937         dd->ipath_hol_state = IPATH_HOL_UP;
 938
 939 done:
 940         if (!ret) {
 941                 *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
 942                 if (!dd->ipath_f_intrsetup(dd)) {
 943                         /* now we can enable all interrupts from the chip */
 944                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
 945                                          -1LL);
 946                         /* force re-interrupt of any pending interrupts. */
 947                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
 948                                          0ULL);
 949                         /* chip is usable; mark it as initialized */
 950                         *dd->ipath_statusp |= IPATH_STATUS_INITTED;
 951                 } else
 952                         ipath_dev_err(dd, "No interrupts enabled, couldn't "
 953                                       "setup interrupt address\n");
 954
 955                 if (dd->ipath_cfgports > ipath_stats.sps_nports)
 956                         /*
 957                          * sps_nports is a global, so, we set it to
 958                          * the highest number of ports of any of the
 959                          * chips we find; we never decrement it, at
 960                          * least for now.  Since this might have changed
 961                          * over disable/enable or prior to reset, always
 962                          * do the check and potentially adjust.
 963                          */
 964                         ipath_stats.sps_nports = dd->ipath_cfgports;
 965         } else
 966                 ipath_dbg("Failed (%d) to initialize chip\n", ret);
 967
 968         /* if ret is non-zero, we probably should do some cleanup
 969            here... */
 970         return ret;
 971 }
 972
 973 static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
 974 {
 975         struct ipath_devdata *dd;
 976         unsigned long flags;
 977         unsigned short val;
 978         int ret;
 979
 980         ret = ipath_parse_ushort(str, &val);
 981
 982         spin_lock_irqsave(&ipath_devs_lock, flags);
 983
 984         if (ret < 0)
 985                 goto bail;
 986
 987         if (val == 0) {
 988                 ret = -EINVAL;
 989                 goto bail;
 990         }
 991
 992         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
 993                 if (dd->ipath_kregbase)
 994                         continue;
 995                 if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
 996                            (dd->ipath_cfgports *
 997                             IPATH_MIN_USER_PORT_BUFCNT)))
 998                 {
 999                         ipath_dev_err(
1000                                 dd,
1001                                 "Allocating %d PIO bufs for kernel leaves "
1002                                 "too few for %d user ports (%d each)\n",
1003                                 val, dd->ipath_cfgports - 1,
1004                                 IPATH_MIN_USER_PORT_BUFCNT);
1005                         ret = -EINVAL;
1006                         goto bail;
1007                 }
1008                 dd->ipath_lastport_piobuf =
1009                         dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
1010         }
1011
1012         ipath_kpiobufs = val;
1013         ret = 0;
1014 bail:
1015         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1016
1017         return ret;
1018 }