drivers/infiniband/hw/hfi1/init.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #include <linux/pci.h>
  49 #include <linux/netdevice.h>
  50 #include <linux/vmalloc.h>
  51 #include <linux/delay.h>
  52 #include <linux/idr.h>
  53 #include <linux/module.h>
  54 #include <linux/printk.h>
  55 #include <linux/hrtimer.h>
  56 #include <rdma/rdma_vt.h>
  57
  58 #include "hfi.h"
  59 #include "device.h"
  60 #include "common.h"
  61 #include "trace.h"
  62 #include "mad.h"
  63 #include "sdma.h"
  64 #include "debugfs.h"
  65 #include "verbs.h"
  66 #include "aspm.h"
  67
  68 #undef pr_fmt
  69 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  70
  71 /*
  72  * min buffers we want to have per context, after driver
  73  */
  74 #define HFI1_MIN_USER_CTXT_BUFCNT 7
  75
  76 #define HFI1_MIN_HDRQ_EGRBUF_CNT 2
  77 #define HFI1_MAX_HDRQ_EGRBUF_CNT 16352
  78 #define HFI1_MIN_EAGER_BUFFER_SIZE (4 * 1024) /* 4KB */
  79 #define HFI1_MAX_EAGER_BUFFER_SIZE (256 * 1024) /* 256KB */
  80
  81 /*
  82  * Number of user receive contexts we are configured to use (to allow for more
  83  * pio buffers per ctxt, etc.)  Zero means use one user context per CPU.
  84  */
  85 int num_user_contexts = -1;
  86 module_param_named(num_user_contexts, num_user_contexts, uint, S_IRUGO);
  87 MODULE_PARM_DESC(
  88         num_user_contexts, "Set max number of user contexts to use");
  89
  90 uint krcvqs[RXE_NUM_DATA_VL];
  91 int krcvqsset;
  92 module_param_array(krcvqs, uint, &krcvqsset, S_IRUGO);
  93 MODULE_PARM_DESC(krcvqs, "Array of the number of non-control kernel receive queues by VL");
  94
  95 /* computed based on above array */
  96 unsigned n_krcvqs;
  97
  98 static unsigned hfi1_rcvarr_split = 25;
  99 module_param_named(rcvarr_split, hfi1_rcvarr_split, uint, S_IRUGO);
 100 MODULE_PARM_DESC(rcvarr_split, "Percent of context's RcvArray entries used for Eager buffers");
 101
 102 static uint eager_buffer_size = (2 << 20); /* 2MB */
 103 module_param(eager_buffer_size, uint, S_IRUGO);
 104 MODULE_PARM_DESC(eager_buffer_size, "Size of the eager buffers, default: 2MB");
 105
 106 static uint rcvhdrcnt = 2048; /* 2x the max eager buffer count */
 107 module_param_named(rcvhdrcnt, rcvhdrcnt, uint, S_IRUGO);
 108 MODULE_PARM_DESC(rcvhdrcnt, "Receive header queue count (default 2048)");
 109
 110 static uint hfi1_hdrq_entsize = 32;
 111 module_param_named(hdrq_entsize, hfi1_hdrq_entsize, uint, S_IRUGO);
 112 MODULE_PARM_DESC(hdrq_entsize, "Size of header queue entries: 2 - 8B, 16 - 64B (default), 32 - 128B");
 113
 114 unsigned int user_credit_return_threshold = 33; /* default is 33% */
 115 module_param(user_credit_return_threshold, uint, S_IRUGO);
 116 MODULE_PARM_DESC(user_credit_return_threshold, "Credit return threshold for user send contexts, return when unreturned credits passes this many blocks (in percent of allocated blocks, 0 is off)");
 117
 118 static inline u64 encode_rcv_header_entry_size(u16);
 119
 120 static struct idr hfi1_unit_table;
 121 u32 hfi1_cpulist_count;
 122 unsigned long *hfi1_cpulist;
 123
 124 /*
 125  * Common code for creating the receive context array.
 126  */
 127 int hfi1_create_ctxts(struct hfi1_devdata *dd)
 128 {
 129         unsigned i;
 130         int ret;
 131
 132         /* Control context has to be always 0 */
 133         BUILD_BUG_ON(HFI1_CTRL_CTXT != 0);
 134
 135         dd->rcd = kzalloc_node(dd->num_rcv_contexts * sizeof(*dd->rcd),
 136                                GFP_KERNEL, dd->node);
 137         if (!dd->rcd)
 138                 goto nomem;
 139
 140         /* create one or more kernel contexts */
 141         for (i = 0; i < dd->first_user_ctxt; ++i) {
 142                 struct hfi1_pportdata *ppd;
 143                 struct hfi1_ctxtdata *rcd;
 144
 145                 ppd = dd->pport + (i % dd->num_pports);
 146                 rcd = hfi1_create_ctxtdata(ppd, i, dd->node);
 147                 if (!rcd) {
 148                         dd_dev_err(dd,
 149                                    "Unable to allocate kernel receive context, failing\n");
 150                         goto nomem;
 151                 }
 152                 /*
 153                  * Set up the kernel context flags here and now because they
 154                  * use default values for all receive side memories.  User
 155                  * contexts will be handled as they are created.
 156                  */
 157                 rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
 158                         HFI1_CAP_KGET(NODROP_RHQ_FULL) |
 159                         HFI1_CAP_KGET(NODROP_EGR_FULL) |
 160                         HFI1_CAP_KGET(DMA_RTAIL);
 161
 162                 /* Control context must use DMA_RTAIL */
 163                 if (rcd->ctxt == HFI1_CTRL_CTXT)
 164                         rcd->flags |= HFI1_CAP_DMA_RTAIL;
 165                 rcd->seq_cnt = 1;
 166
 167                 rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
 168                 if (!rcd->sc) {
 169                         dd_dev_err(dd,
 170                                    "Unable to allocate kernel send context, failing\n");
 171                         dd->rcd[rcd->ctxt] = NULL;
 172                         hfi1_free_ctxtdata(dd, rcd);
 173                         goto nomem;
 174                 }
 175
 176                 ret = hfi1_init_ctxt(rcd->sc);
 177                 if (ret < 0) {
 178                         dd_dev_err(dd,
 179                                    "Failed to setup kernel receive context, failing\n");
 180                         sc_free(rcd->sc);
 181                         dd->rcd[rcd->ctxt] = NULL;
 182                         hfi1_free_ctxtdata(dd, rcd);
 183                         ret = -EFAULT;
 184                         goto bail;
 185                 }
 186         }
 187
 188         /*
 189          * Initialize aspm, to be done after gen3 transition and setting up
 190          * contexts and before enabling interrupts
 191          */
 192         aspm_init(dd);
 193
 194         return 0;
 195 nomem:
 196         ret = -ENOMEM;
 197 bail:
 198         kfree(dd->rcd);
 199         dd->rcd = NULL;
 200         return ret;
 201 }
 202
 203 /*
 204  * Common code for user and kernel context setup.
 205  */
 206 struct hfi1_ctxtdata *hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, u32 ctxt,
 207                                            int numa)
 208 {
 209         struct hfi1_devdata *dd = ppd->dd;
 210         struct hfi1_ctxtdata *rcd;
 211         unsigned kctxt_ngroups = 0;
 212         u32 base;
 213
 214         if (dd->rcv_entries.nctxt_extra >
 215             dd->num_rcv_contexts - dd->first_user_ctxt)
 216                 kctxt_ngroups = (dd->rcv_entries.nctxt_extra -
 217                                  (dd->num_rcv_contexts - dd->first_user_ctxt));
 218         rcd = kzalloc(sizeof(*rcd), GFP_KERNEL);
 219         if (rcd) {
 220                 u32 rcvtids, max_entries;
 221
 222                 hfi1_cdbg(PROC, "setting up context %u\n", ctxt);
 223
 224                 INIT_LIST_HEAD(&rcd->qp_wait_list);
 225                 rcd->ppd = ppd;
 226                 rcd->dd = dd;
 227                 rcd->cnt = 1;
 228                 rcd->ctxt = ctxt;
 229                 dd->rcd[ctxt] = rcd;
 230                 rcd->numa_id = numa;
 231                 rcd->rcv_array_groups = dd->rcv_entries.ngroups;
 232
 233                 mutex_init(&rcd->exp_lock);
 234
 235                 /*
 236                  * Calculate the context's RcvArray entry starting point.
 237                  * We do this here because we have to take into account all
 238                  * the RcvArray entries that previous context would have
 239                  * taken and we have to account for any extra groups
 240                  * assigned to the kernel or user contexts.
 241                  */
 242                 if (ctxt < dd->first_user_ctxt) {
 243                         if (ctxt < kctxt_ngroups) {
 244                                 base = ctxt * (dd->rcv_entries.ngroups + 1);
 245                                 rcd->rcv_array_groups++;
 246                         } else
 247                                 base = kctxt_ngroups +
 248                                         (ctxt * dd->rcv_entries.ngroups);
 249                 } else {
 250                         u16 ct = ctxt - dd->first_user_ctxt;
 251
 252                         base = ((dd->n_krcv_queues * dd->rcv_entries.ngroups) +
 253                                 kctxt_ngroups);
 254                         if (ct < dd->rcv_entries.nctxt_extra) {
 255                                 base += ct * (dd->rcv_entries.ngroups + 1);
 256                                 rcd->rcv_array_groups++;
 257                         } else
 258                                 base += dd->rcv_entries.nctxt_extra +
 259                                         (ct * dd->rcv_entries.ngroups);
 260                 }
 261                 rcd->eager_base = base * dd->rcv_entries.group_size;
 262
 263                 /* Validate and initialize Rcv Hdr Q variables */
 264                 if (rcvhdrcnt % HDRQ_INCREMENT) {
 265                         dd_dev_err(dd,
 266                                    "ctxt%u: header queue count %d must be divisible by %lu\n",
 267                                    rcd->ctxt, rcvhdrcnt, HDRQ_INCREMENT);
 268                         goto bail;
 269                 }
 270                 rcd->rcvhdrq_cnt = rcvhdrcnt;
 271                 rcd->rcvhdrqentsize = hfi1_hdrq_entsize;
 272                 /*
 273                  * Simple Eager buffer allocation: we have already pre-allocated
 274                  * the number of RcvArray entry groups. Each ctxtdata structure
 275                  * holds the number of groups for that context.
 276                  *
 277                  * To follow CSR requirements and maintain cacheline alignment,
 278                  * make sure all sizes and bases are multiples of group_size.
 279                  *
 280                  * The expected entry count is what is left after assigning
 281                  * eager.
 282                  */
 283                 max_entries = rcd->rcv_array_groups *
 284                         dd->rcv_entries.group_size;
 285                 rcvtids = ((max_entries * hfi1_rcvarr_split) / 100);
 286                 rcd->egrbufs.count = round_down(rcvtids,
 287                                                 dd->rcv_entries.group_size);
 288                 if (rcd->egrbufs.count > MAX_EAGER_ENTRIES) {
 289                         dd_dev_err(dd, "ctxt%u: requested too many RcvArray entries.\n",
 290                                    rcd->ctxt);
 291                         rcd->egrbufs.count = MAX_EAGER_ENTRIES;
 292                 }
 293                 hfi1_cdbg(PROC,
 294                           "ctxt%u: max Eager buffer RcvArray entries: %u\n",
 295                           rcd->ctxt, rcd->egrbufs.count);
 296
 297                 /*
 298                  * Allocate array that will hold the eager buffer accounting
 299                  * data.
 300                  * This will allocate the maximum possible buffer count based
 301                  * on the value of the RcvArray split parameter.
 302                  * The resulting value will be rounded down to the closest
 303                  * multiple of dd->rcv_entries.group_size.
 304                  */
 305                 rcd->egrbufs.buffers = kcalloc(rcd->egrbufs.count,
 306                                                sizeof(*rcd->egrbufs.buffers),
 307                                                GFP_KERNEL);
 308                 if (!rcd->egrbufs.buffers)
 309                         goto bail;
 310                 rcd->egrbufs.rcvtids = kcalloc(rcd->egrbufs.count,
 311                                                sizeof(*rcd->egrbufs.rcvtids),
 312                                                GFP_KERNEL);
 313                 if (!rcd->egrbufs.rcvtids)
 314                         goto bail;
 315                 rcd->egrbufs.size = eager_buffer_size;
 316                 /*
 317                  * The size of the buffers programmed into the RcvArray
 318                  * entries needs to be big enough to handle the highest
 319                  * MTU supported.
 320                  */
 321                 if (rcd->egrbufs.size < hfi1_max_mtu) {
 322                         rcd->egrbufs.size = __roundup_pow_of_two(hfi1_max_mtu);
 323                         hfi1_cdbg(PROC,
 324                                   "ctxt%u: eager bufs size too small. Adjusting to %zu\n",
 325                                     rcd->ctxt, rcd->egrbufs.size);
 326                 }
 327                 rcd->egrbufs.rcvtid_size = HFI1_MAX_EAGER_BUFFER_SIZE;
 328
 329                 if (ctxt < dd->first_user_ctxt) { /* N/A for PSM contexts */
 330                         rcd->opstats = kzalloc(sizeof(*rcd->opstats),
 331                                 GFP_KERNEL);
 332                         if (!rcd->opstats)
 333                                 goto bail;
 334                 }
 335         }
 336         return rcd;
 337 bail:
 338         kfree(rcd->egrbufs.rcvtids);
 339         kfree(rcd->egrbufs.buffers);
 340         kfree(rcd);
 341         return NULL;
 342 }
 343
 344 /*
 345  * Convert a receive header entry size that to the encoding used in the CSR.
 346  *
 347  * Return a zero if the given size is invalid.
 348  */
 349 static inline u64 encode_rcv_header_entry_size(u16 size)
 350 {
 351         /* there are only 3 valid receive header entry sizes */
 352         if (size == 2)
 353                 return 1;
 354         if (size == 16)
 355                 return 2;
 356         else if (size == 32)
 357                 return 4;
 358         return 0; /* invalid */
 359 }
 360
 361 /*
 362  * Select the largest ccti value over all SLs to determine the intra-
 363  * packet gap for the link.
 364  *
 365  * called with cca_timer_lock held (to protect access to cca_timer
 366  * array), and rcu_read_lock() (to protect access to cc_state).
 367  */
 368 void set_link_ipg(struct hfi1_pportdata *ppd)
 369 {
 370         struct hfi1_devdata *dd = ppd->dd;
 371         struct cc_state *cc_state;
 372         int i;
 373         u16 cce, ccti_limit, max_ccti = 0;
 374         u16 shift, mult;
 375         u64 src;
 376         u32 current_egress_rate; /* Mbits /sec */
 377         u32 max_pkt_time;
 378         /*
 379          * max_pkt_time is the maximum packet egress time in units
 380          * of the fabric clock period 1/(805 MHz).
 381          */
 382
 383         cc_state = get_cc_state(ppd);
 384
 385         if (!cc_state)
 386                 /*
 387                  * This should _never_ happen - rcu_read_lock() is held,
 388                  * and set_link_ipg() should not be called if cc_state
 389                  * is NULL.
 390                  */
 391                 return;
 392
 393         for (i = 0; i < OPA_MAX_SLS; i++) {
 394                 u16 ccti = ppd->cca_timer[i].ccti;
 395
 396                 if (ccti > max_ccti)
 397                         max_ccti = ccti;
 398         }
 399
 400         ccti_limit = cc_state->cct.ccti_limit;
 401         if (max_ccti > ccti_limit)
 402                 max_ccti = ccti_limit;
 403
 404         cce = cc_state->cct.entries[max_ccti].entry;
 405         shift = (cce & 0xc000) >> 14;
 406         mult = (cce & 0x3fff);
 407
 408         current_egress_rate = active_egress_rate(ppd);
 409
 410         max_pkt_time = egress_cycles(ppd->ibmaxlen, current_egress_rate);
 411
 412         src = (max_pkt_time >> shift) * mult;
 413
 414         src &= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SMASK;
 415         src <<= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SHIFT;
 416
 417         write_csr(dd, SEND_STATIC_RATE_CONTROL, src);
 418 }
 419
 420 static enum hrtimer_restart cca_timer_fn(struct hrtimer *t)
 421 {
 422         struct cca_timer *cca_timer;
 423         struct hfi1_pportdata *ppd;
 424         int sl;
 425         u16 ccti_timer, ccti_min;
 426         struct cc_state *cc_state;
 427         unsigned long flags;
 428         enum hrtimer_restart ret = HRTIMER_NORESTART;
 429
 430         cca_timer = container_of(t, struct cca_timer, hrtimer);
 431         ppd = cca_timer->ppd;
 432         sl = cca_timer->sl;
 433
 434         rcu_read_lock();
 435
 436         cc_state = get_cc_state(ppd);
 437
 438         if (!cc_state) {
 439                 rcu_read_unlock();
 440                 return HRTIMER_NORESTART;
 441         }
 442
 443         /*
 444          * 1) decrement ccti for SL
 445          * 2) calculate IPG for link (set_link_ipg())
 446          * 3) restart timer, unless ccti is at min value
 447          */
 448
 449         ccti_min = cc_state->cong_setting.entries[sl].ccti_min;
 450         ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
 451
 452         spin_lock_irqsave(&ppd->cca_timer_lock, flags);
 453
 454         if (cca_timer->ccti > ccti_min) {
 455                 cca_timer->ccti--;
 456                 set_link_ipg(ppd);
 457         }
 458
 459         if (cca_timer->ccti > ccti_min) {
 460                 unsigned long nsec = 1024 * ccti_timer;
 461                 /* ccti_timer is in units of 1.024 usec */
 462                 hrtimer_forward_now(t, ns_to_ktime(nsec));
 463                 ret = HRTIMER_RESTART;
 464         }
 465
 466         spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
 467         rcu_read_unlock();
 468         return ret;
 469 }
 470
 471 /*
 472  * Common code for initializing the physical port structure.
 473  */
 474 void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
 475                          struct hfi1_devdata *dd, u8 hw_pidx, u8 port)
 476 {
 477         int i, size;
 478         uint default_pkey_idx;
 479
 480         ppd->dd = dd;
 481         ppd->hw_pidx = hw_pidx;
 482         ppd->port = port; /* IB port number, not index */
 483
 484         default_pkey_idx = 1;
 485
 486         ppd->pkeys[default_pkey_idx] = DEFAULT_P_KEY;
 487         if (loopback) {
 488                 hfi1_early_err(&pdev->dev,
 489                                "Faking data partition 0x8001 in idx %u\n",
 490                                !default_pkey_idx);
 491                 ppd->pkeys[!default_pkey_idx] = 0x8001;
 492         }
 493
 494         INIT_WORK(&ppd->link_vc_work, handle_verify_cap);
 495         INIT_WORK(&ppd->link_up_work, handle_link_up);
 496         INIT_WORK(&ppd->link_down_work, handle_link_down);
 497         INIT_WORK(&ppd->freeze_work, handle_freeze);
 498         INIT_WORK(&ppd->link_downgrade_work, handle_link_downgrade);
 499         INIT_WORK(&ppd->sma_message_work, handle_sma_message);
 500         INIT_WORK(&ppd->link_bounce_work, handle_link_bounce);
 501         INIT_WORK(&ppd->linkstate_active_work, receive_interrupt_work);
 502         INIT_WORK(&ppd->qsfp_info.qsfp_work, qsfp_event);
 503
 504         mutex_init(&ppd->hls_lock);
 505         spin_lock_init(&ppd->sdma_alllock);
 506         spin_lock_init(&ppd->qsfp_info.qsfp_lock);
 507
 508         ppd->qsfp_info.ppd = ppd;
 509         ppd->sm_trap_qp = 0x0;
 510         ppd->sa_qp = 0x1;
 511
 512         ppd->hfi1_wq = NULL;
 513
 514         spin_lock_init(&ppd->cca_timer_lock);
 515
 516         for (i = 0; i < OPA_MAX_SLS; i++) {
 517                 hrtimer_init(&ppd->cca_timer[i].hrtimer, CLOCK_MONOTONIC,
 518                              HRTIMER_MODE_REL);
 519                 ppd->cca_timer[i].ppd = ppd;
 520                 ppd->cca_timer[i].sl = i;
 521                 ppd->cca_timer[i].ccti = 0;
 522                 ppd->cca_timer[i].hrtimer.function = cca_timer_fn;
 523         }
 524
 525         ppd->cc_max_table_entries = IB_CC_TABLE_CAP_DEFAULT;
 526
 527         spin_lock_init(&ppd->cc_state_lock);
 528         spin_lock_init(&ppd->cc_log_lock);
 529         size = sizeof(struct cc_state);
 530         RCU_INIT_POINTER(ppd->cc_state, kzalloc(size, GFP_KERNEL));
 531         if (!rcu_dereference(ppd->cc_state))
 532                 goto bail;
 533         return;
 534
 535 bail:
 536
 537         hfi1_early_err(&pdev->dev,
 538                        "Congestion Control Agent disabled for port %d\n", port);
 539 }
 540
 541 /*
 542  * Do initialization for device that is only needed on
 543  * first detect, not on resets.
 544  */
 545 static int loadtime_init(struct hfi1_devdata *dd)
 546 {
 547         return 0;
 548 }
 549
 550 /**
 551  * init_after_reset - re-initialize after a reset
 552  * @dd: the hfi1_ib device
 553  *
 554  * sanity check at least some of the values after reset, and
 555  * ensure no receive or transmit (explicitly, in case reset
 556  * failed
 557  */
 558 static int init_after_reset(struct hfi1_devdata *dd)
 559 {
 560         int i;
 561
 562         /*
 563          * Ensure chip does no sends or receives, tail updates, or
 564          * pioavail updates while we re-initialize.  This is mostly
 565          * for the driver data structures, not chip registers.
 566          */
 567         for (i = 0; i < dd->num_rcv_contexts; i++)
 568                 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
 569                                   HFI1_RCVCTRL_INTRAVAIL_DIS |
 570                                   HFI1_RCVCTRL_TAILUPD_DIS, i);
 571         pio_send_control(dd, PSC_GLOBAL_DISABLE);
 572         for (i = 0; i < dd->num_send_contexts; i++)
 573                 sc_disable(dd->send_contexts[i].sc);
 574
 575         return 0;
 576 }
 577
 578 static void enable_chip(struct hfi1_devdata *dd)
 579 {
 580         u32 rcvmask;
 581         u32 i;
 582
 583         /* enable PIO send */
 584         pio_send_control(dd, PSC_GLOBAL_ENABLE);
 585
 586         /*
 587          * Enable kernel ctxts' receive and receive interrupt.
 588          * Other ctxts done as user opens and initializes them.
 589          */
 590         for (i = 0; i < dd->first_user_ctxt; ++i) {
 591                 rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
 592                 rcvmask |= HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, DMA_RTAIL) ?
 593                         HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
 594                 if (!HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, MULTI_PKT_EGR))
 595                         rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
 596                 if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_RHQ_FULL))
 597                         rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
 598                 if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_EGR_FULL))
 599                         rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
 600                 hfi1_rcvctrl(dd, rcvmask, i);
 601                 sc_enable(dd->rcd[i]->sc);
 602         }
 603 }
 604
 605 /**
 606  * create_workqueues - create per port workqueues
 607  * @dd: the hfi1_ib device
 608  */
 609 static int create_workqueues(struct hfi1_devdata *dd)
 610 {
 611         int pidx;
 612         struct hfi1_pportdata *ppd;
 613
 614         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 615                 ppd = dd->pport + pidx;
 616                 if (!ppd->hfi1_wq) {
 617                         ppd->hfi1_wq =
 618                                 alloc_workqueue(
 619                                     "hfi%d_%d",
 620                                     WQ_SYSFS | WQ_HIGHPRI | WQ_CPU_INTENSIVE,
 621                                     dd->num_sdma,
 622                                     dd->unit, pidx);
 623                         if (!ppd->hfi1_wq)
 624                                 goto wq_error;
 625                 }
 626         }
 627         return 0;
 628 wq_error:
 629         pr_err("alloc_workqueue failed for port %d\n", pidx + 1);
 630         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 631                 ppd = dd->pport + pidx;
 632                 if (ppd->hfi1_wq) {
 633                         destroy_workqueue(ppd->hfi1_wq);
 634                         ppd->hfi1_wq = NULL;
 635                 }
 636         }
 637         return -ENOMEM;
 638 }
 639
 640 /**
 641  * hfi1_init - do the actual initialization sequence on the chip
 642  * @dd: the hfi1_ib device
 643  * @reinit: re-initializing, so don't allocate new memory
 644  *
 645  * Do the actual initialization sequence on the chip.  This is done
 646  * both from the init routine called from the PCI infrastructure, and
 647  * when we reset the chip, or detect that it was reset internally,
 648  * or it's administratively re-enabled.
 649  *
 650  * Memory allocation here and in called routines is only done in
 651  * the first case (reinit == 0).  We have to be careful, because even
 652  * without memory allocation, we need to re-write all the chip registers
 653  * TIDs, etc. after the reset or enable has completed.
 654  */
 655 int hfi1_init(struct hfi1_devdata *dd, int reinit)
 656 {
 657         int ret = 0, pidx, lastfail = 0;
 658         unsigned i, len;
 659         struct hfi1_ctxtdata *rcd;
 660         struct hfi1_pportdata *ppd;
 661
 662         /* Set up recv low level handlers */
 663         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EXPECTED] =
 664                                                 kdeth_process_expected;
 665         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EAGER] =
 666                                                 kdeth_process_eager;
 667         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_IB] = process_receive_ib;
 668         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_ERROR] =
 669                                                 process_receive_error;
 670         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_BYPASS] =
 671                                                 process_receive_bypass;
 672         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID5] =
 673                                                 process_receive_invalid;
 674         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID6] =
 675                                                 process_receive_invalid;
 676         dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID7] =
 677                                                 process_receive_invalid;
 678         dd->rhf_rcv_function_map = dd->normal_rhf_rcv_functions;
 679
 680         /* Set up send low level handlers */
 681         dd->process_pio_send = hfi1_verbs_send_pio;
 682         dd->process_dma_send = hfi1_verbs_send_dma;
 683         dd->pio_inline_send = pio_copy;
 684
 685         if (is_ax(dd)) {
 686                 atomic_set(&dd->drop_packet, DROP_PACKET_ON);
 687                 dd->do_drop = 1;
 688         } else {
 689                 atomic_set(&dd->drop_packet, DROP_PACKET_OFF);
 690                 dd->do_drop = 0;
 691         }
 692
 693         /* make sure the link is not "up" */
 694         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 695                 ppd = dd->pport + pidx;
 696                 ppd->linkup = 0;
 697         }
 698
 699         if (reinit)
 700                 ret = init_after_reset(dd);
 701         else
 702                 ret = loadtime_init(dd);
 703         if (ret)
 704                 goto done;
 705
 706         /* allocate dummy tail memory for all receive contexts */
 707         dd->rcvhdrtail_dummy_kvaddr = dma_zalloc_coherent(
 708                 &dd->pcidev->dev, sizeof(u64),
 709                 &dd->rcvhdrtail_dummy_physaddr,
 710                 GFP_KERNEL);
 711
 712         if (!dd->rcvhdrtail_dummy_kvaddr) {
 713                 dd_dev_err(dd, "cannot allocate dummy tail memory\n");
 714                 ret = -ENOMEM;
 715                 goto done;
 716         }
 717
 718         /* dd->rcd can be NULL if early initialization failed */
 719         for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
 720                 /*
 721                  * Set up the (kernel) rcvhdr queue and egr TIDs.  If doing
 722                  * re-init, the simplest way to handle this is to free
 723                  * existing, and re-allocate.
 724                  * Need to re-create rest of ctxt 0 ctxtdata as well.
 725                  */
 726                 rcd = dd->rcd[i];
 727                 if (!rcd)
 728                         continue;
 729
 730                 rcd->do_interrupt = &handle_receive_interrupt;
 731
 732                 lastfail = hfi1_create_rcvhdrq(dd, rcd);
 733                 if (!lastfail)
 734                         lastfail = hfi1_setup_eagerbufs(rcd);
 735                 if (lastfail) {
 736                         dd_dev_err(dd,
 737                                    "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
 738                         ret = lastfail;
 739                 }
 740         }
 741
 742         /* Allocate enough memory for user event notification. */
 743         len = PAGE_ALIGN(dd->chip_rcv_contexts * HFI1_MAX_SHARED_CTXTS *
 744                          sizeof(*dd->events));
 745         dd->events = vmalloc_user(len);
 746         if (!dd->events)
 747                 dd_dev_err(dd, "Failed to allocate user events page\n");
 748         /*
 749          * Allocate a page for device and port status.
 750          * Page will be shared amongst all user processes.
 751          */
 752         dd->status = vmalloc_user(PAGE_SIZE);
 753         if (!dd->status)
 754                 dd_dev_err(dd, "Failed to allocate dev status page\n");
 755         else
 756                 dd->freezelen = PAGE_SIZE - (sizeof(*dd->status) -
 757                                              sizeof(dd->status->freezemsg));
 758         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 759                 ppd = dd->pport + pidx;
 760                 if (dd->status)
 761                         /* Currently, we only have one port */
 762                         ppd->statusp = &dd->status->port;
 763
 764                 set_mtu(ppd);
 765         }
 766
 767         /* enable chip even if we have an error, so we can debug cause */
 768         enable_chip(dd);
 769
 770 done:
 771         /*
 772          * Set status even if port serdes is not initialized
 773          * so that diags will work.
 774          */
 775         if (dd->status)
 776                 dd->status->dev |= HFI1_STATUS_CHIP_PRESENT |
 777                         HFI1_STATUS_INITTED;
 778         if (!ret) {
 779                 /* enable all interrupts from the chip */
 780                 set_intr_state(dd, 1);
 781
 782                 /* chip is OK for user apps; mark it as initialized */
 783                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 784                         ppd = dd->pport + pidx;
 785
 786                         /*
 787                          * start the serdes - must be after interrupts are
 788                          * enabled so we are notified when the link goes up
 789                          */
 790                         lastfail = bringup_serdes(ppd);
 791                         if (lastfail)
 792                                 dd_dev_info(dd,
 793                                             "Failed to bring up port %u\n",
 794                                             ppd->port);
 795
 796                         /*
 797                          * Set status even if port serdes is not initialized
 798                          * so that diags will work.
 799                          */
 800                         if (ppd->statusp)
 801                                 *ppd->statusp |= HFI1_STATUS_CHIP_PRESENT |
 802                                                         HFI1_STATUS_INITTED;
 803                         if (!ppd->link_speed_enabled)
 804                                 continue;
 805                 }
 806         }
 807
 808         /* if ret is non-zero, we probably should do some cleanup here... */
 809         return ret;
 810 }
 811
 812 static inline struct hfi1_devdata *__hfi1_lookup(int unit)
 813 {
 814         return idr_find(&hfi1_unit_table, unit);
 815 }
 816
 817 struct hfi1_devdata *hfi1_lookup(int unit)
 818 {
 819         struct hfi1_devdata *dd;
 820         unsigned long flags;
 821
 822         spin_lock_irqsave(&hfi1_devs_lock, flags);
 823         dd = __hfi1_lookup(unit);
 824         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
 825
 826         return dd;
 827 }
 828
 829 /*
 830  * Stop the timers during unit shutdown, or after an error late
 831  * in initialization.
 832  */
 833 static void stop_timers(struct hfi1_devdata *dd)
 834 {
 835         struct hfi1_pportdata *ppd;
 836         int pidx;
 837
 838         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 839                 ppd = dd->pport + pidx;
 840                 if (ppd->led_override_timer.data) {
 841                         del_timer_sync(&ppd->led_override_timer);
 842                         atomic_set(&ppd->led_override_timer_active, 0);
 843                 }
 844         }
 845 }
 846
 847 /**
 848  * shutdown_device - shut down a device
 849  * @dd: the hfi1_ib device
 850  *
 851  * This is called to make the device quiet when we are about to
 852  * unload the driver, and also when the device is administratively
 853  * disabled.   It does not free any data structures.
 854  * Everything it does has to be setup again by hfi1_init(dd, 1)
 855  */
 856 static void shutdown_device(struct hfi1_devdata *dd)
 857 {
 858         struct hfi1_pportdata *ppd;
 859         unsigned pidx;
 860         int i;
 861
 862         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 863                 ppd = dd->pport + pidx;
 864
 865                 ppd->linkup = 0;
 866                 if (ppd->statusp)
 867                         *ppd->statusp &= ~(HFI1_STATUS_IB_CONF |
 868                                            HFI1_STATUS_IB_READY);
 869         }
 870         dd->flags &= ~HFI1_INITTED;
 871
 872         /* mask interrupts, but not errors */
 873         set_intr_state(dd, 0);
 874
 875         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 876                 ppd = dd->pport + pidx;
 877                 for (i = 0; i < dd->num_rcv_contexts; i++)
 878                         hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
 879                                           HFI1_RCVCTRL_CTXT_DIS |
 880                                           HFI1_RCVCTRL_INTRAVAIL_DIS |
 881                                           HFI1_RCVCTRL_PKEY_DIS |
 882                                           HFI1_RCVCTRL_ONE_PKT_EGR_DIS, i);
 883                 /*
 884                  * Gracefully stop all sends allowing any in progress to
 885                  * trickle out first.
 886                  */
 887                 for (i = 0; i < dd->num_send_contexts; i++)
 888                         sc_flush(dd->send_contexts[i].sc);
 889         }
 890
 891         /*
 892          * Enough for anything that's going to trickle out to have actually
 893          * done so.
 894          */
 895         udelay(20);
 896
 897         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 898                 ppd = dd->pport + pidx;
 899
 900                 /* disable all contexts */
 901                 for (i = 0; i < dd->num_send_contexts; i++)
 902                         sc_disable(dd->send_contexts[i].sc);
 903                 /* disable the send device */
 904                 pio_send_control(dd, PSC_GLOBAL_DISABLE);
 905
 906                 shutdown_led_override(ppd);
 907
 908                 /*
 909                  * Clear SerdesEnable.
 910                  * We can't count on interrupts since we are stopping.
 911                  */
 912                 hfi1_quiet_serdes(ppd);
 913
 914                 if (ppd->hfi1_wq) {
 915                         destroy_workqueue(ppd->hfi1_wq);
 916                         ppd->hfi1_wq = NULL;
 917                 }
 918         }
 919         sdma_exit(dd);
 920 }
 921
 922 /**
 923  * hfi1_free_ctxtdata - free a context's allocated data
 924  * @dd: the hfi1_ib device
 925  * @rcd: the ctxtdata structure
 926  *
 927  * free up any allocated data for a context
 928  * This should not touch anything that would affect a simultaneous
 929  * re-allocation of context data, because it is called after hfi1_mutex
 930  * is released (and can be called from reinit as well).
 931  * It should never change any chip state, or global driver state.
 932  */
 933 void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
 934 {
 935         unsigned e;
 936
 937         if (!rcd)
 938                 return;
 939
 940         if (rcd->rcvhdrq) {
 941                 dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
 942                                   rcd->rcvhdrq, rcd->rcvhdrq_phys);
 943                 rcd->rcvhdrq = NULL;
 944                 if (rcd->rcvhdrtail_kvaddr) {
 945                         dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
 946                                           (void *)rcd->rcvhdrtail_kvaddr,
 947                                           rcd->rcvhdrqtailaddr_phys);
 948                         rcd->rcvhdrtail_kvaddr = NULL;
 949                 }
 950         }
 951
 952         /* all the RcvArray entries should have been cleared by now */
 953         kfree(rcd->egrbufs.rcvtids);
 954
 955         for (e = 0; e < rcd->egrbufs.alloced; e++) {
 956                 if (rcd->egrbufs.buffers[e].phys)
 957                         dma_free_coherent(&dd->pcidev->dev,
 958                                           rcd->egrbufs.buffers[e].len,
 959                                           rcd->egrbufs.buffers[e].addr,
 960                                           rcd->egrbufs.buffers[e].phys);
 961         }
 962         kfree(rcd->egrbufs.buffers);
 963
 964         sc_free(rcd->sc);
 965         vfree(rcd->user_event_mask);
 966         vfree(rcd->subctxt_uregbase);
 967         vfree(rcd->subctxt_rcvegrbuf);
 968         vfree(rcd->subctxt_rcvhdr_base);
 969         kfree(rcd->opstats);
 970         kfree(rcd);
 971 }
 972
 973 /*
 974  * Release our hold on the shared asic data.  If we are the last one,
 975  * free the structure.  Must be holding hfi1_devs_lock.
 976  */
 977 static void release_asic_data(struct hfi1_devdata *dd)
 978 {
 979         int other;
 980
 981         if (!dd->asic_data)
 982                 return;
 983         dd->asic_data->dds[dd->hfi1_id] = NULL;
 984         other = dd->hfi1_id ? 0 : 1;
 985         if (!dd->asic_data->dds[other]) {
 986                 /* we are the last holder, free it */
 987                 kfree(dd->asic_data);
 988         }
 989         dd->asic_data = NULL;
 990 }
 991
 992 static void __hfi1_free_devdata(struct kobject *kobj)
 993 {
 994         struct hfi1_devdata *dd =
 995                 container_of(kobj, struct hfi1_devdata, kobj);
 996         unsigned long flags;
 997
 998         spin_lock_irqsave(&hfi1_devs_lock, flags);
 999         idr_remove(&hfi1_unit_table, dd->unit);
1000         list_del(&dd->list);
1001         release_asic_data(dd);
1002         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1003         free_platform_config(dd);
1004         rcu_barrier(); /* wait for rcu callbacks to complete */
1005         free_percpu(dd->int_counter);
1006         free_percpu(dd->rcv_limit);
1007         hfi1_dev_affinity_free(dd);
1008         free_percpu(dd->send_schedule);
1009         rvt_dealloc_device(&dd->verbs_dev.rdi);
1010 }
1011
1012 static struct kobj_type hfi1_devdata_type = {
1013         .release = __hfi1_free_devdata,
1014 };
1015
1016 void hfi1_free_devdata(struct hfi1_devdata *dd)
1017 {
1018         kobject_put(&dd->kobj);
1019 }
1020
1021 /*
1022  * Allocate our primary per-unit data structure.  Must be done via verbs
1023  * allocator, because the verbs cleanup process both does cleanup and
1024  * free of the data structure.
1025  * "extra" is for chip-specific data.
1026  *
1027  * Use the idr mechanism to get a unit number for this unit.
1028  */
1029 struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev, size_t extra)
1030 {
1031         unsigned long flags;
1032         struct hfi1_devdata *dd;
1033         int ret, nports;
1034
1035         /* extra is * number of ports */
1036         nports = extra / sizeof(struct hfi1_pportdata);
1037
1038         dd = (struct hfi1_devdata *)rvt_alloc_device(sizeof(*dd) + extra,
1039                                                      nports);
1040         if (!dd)
1041                 return ERR_PTR(-ENOMEM);
1042         dd->num_pports = nports;
1043         dd->pport = (struct hfi1_pportdata *)(dd + 1);
1044
1045         INIT_LIST_HEAD(&dd->list);
1046         idr_preload(GFP_KERNEL);
1047         spin_lock_irqsave(&hfi1_devs_lock, flags);
1048
1049         ret = idr_alloc(&hfi1_unit_table, dd, 0, 0, GFP_NOWAIT);
1050         if (ret >= 0) {
1051                 dd->unit = ret;
1052                 list_add(&dd->list, &hfi1_dev_list);
1053         }
1054
1055         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1056         idr_preload_end();
1057
1058         if (ret < 0) {
1059                 hfi1_early_err(&pdev->dev,
1060                                "Could not allocate unit ID: error %d\n", -ret);
1061                 goto bail;
1062         }
1063         /*
1064          * Initialize all locks for the device. This needs to be as early as
1065          * possible so locks are usable.
1066          */
1067         spin_lock_init(&dd->sc_lock);
1068         spin_lock_init(&dd->sendctrl_lock);
1069         spin_lock_init(&dd->rcvctrl_lock);
1070         spin_lock_init(&dd->uctxt_lock);
1071         spin_lock_init(&dd->hfi1_diag_trans_lock);
1072         spin_lock_init(&dd->sc_init_lock);
1073         spin_lock_init(&dd->dc8051_lock);
1074         spin_lock_init(&dd->dc8051_memlock);
1075         seqlock_init(&dd->sc2vl_lock);
1076         spin_lock_init(&dd->sde_map_lock);
1077         spin_lock_init(&dd->pio_map_lock);
1078         init_waitqueue_head(&dd->event_queue);
1079
1080         dd->int_counter = alloc_percpu(u64);
1081         if (!dd->int_counter) {
1082                 ret = -ENOMEM;
1083                 hfi1_early_err(&pdev->dev,
1084                                "Could not allocate per-cpu int_counter\n");
1085                 goto bail;
1086         }
1087
1088         dd->rcv_limit = alloc_percpu(u64);
1089         if (!dd->rcv_limit) {
1090                 ret = -ENOMEM;
1091                 hfi1_early_err(&pdev->dev,
1092                                "Could not allocate per-cpu rcv_limit\n");
1093                 goto bail;
1094         }
1095
1096         dd->send_schedule = alloc_percpu(u64);
1097         if (!dd->send_schedule) {
1098                 ret = -ENOMEM;
1099                 hfi1_early_err(&pdev->dev,
1100                                "Could not allocate per-cpu int_counter\n");
1101                 goto bail;
1102         }
1103
1104         if (!hfi1_cpulist_count) {
1105                 u32 count = num_online_cpus();
1106
1107                 hfi1_cpulist = kcalloc(BITS_TO_LONGS(count), sizeof(long),
1108                                        GFP_KERNEL);
1109                 if (hfi1_cpulist)
1110                         hfi1_cpulist_count = count;
1111                 else
1112                         hfi1_early_err(
1113                         &pdev->dev,
1114                         "Could not alloc cpulist info, cpu affinity might be wrong\n");
1115         }
1116         kobject_init(&dd->kobj, &hfi1_devdata_type);
1117         return dd;
1118
1119 bail:
1120         if (!list_empty(&dd->list))
1121                 list_del_init(&dd->list);
1122         rvt_dealloc_device(&dd->verbs_dev.rdi);
1123         return ERR_PTR(ret);
1124 }
1125
1126 /*
1127  * Called from freeze mode handlers, and from PCI error
1128  * reporting code.  Should be paranoid about state of
1129  * system and data structures.
1130  */
1131 void hfi1_disable_after_error(struct hfi1_devdata *dd)
1132 {
1133         if (dd->flags & HFI1_INITTED) {
1134                 u32 pidx;
1135
1136                 dd->flags &= ~HFI1_INITTED;
1137                 if (dd->pport)
1138                         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1139                                 struct hfi1_pportdata *ppd;
1140
1141                                 ppd = dd->pport + pidx;
1142                                 if (dd->flags & HFI1_PRESENT)
1143                                         set_link_state(ppd, HLS_DN_DISABLE);
1144
1145                                 if (ppd->statusp)
1146                                         *ppd->statusp &= ~HFI1_STATUS_IB_READY;
1147                         }
1148         }
1149
1150         /*
1151          * Mark as having had an error for driver, and also
1152          * for /sys and status word mapped to user programs.
1153          * This marks unit as not usable, until reset.
1154          */
1155         if (dd->status)
1156                 dd->status->dev |= HFI1_STATUS_HWERROR;
1157 }
1158
1159 static void remove_one(struct pci_dev *);
1160 static int init_one(struct pci_dev *, const struct pci_device_id *);
1161
1162 #define DRIVER_LOAD_MSG "Intel " DRIVER_NAME " loaded: "
1163 #define PFX DRIVER_NAME ": "
1164
1165 static const struct pci_device_id hfi1_pci_tbl[] = {
1166         { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL0) },
1167         { PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL1) },
1168         { 0, }
1169 };
1170
1171 MODULE_DEVICE_TABLE(pci, hfi1_pci_tbl);
1172
1173 static struct pci_driver hfi1_pci_driver = {
1174         .name = DRIVER_NAME,
1175         .probe = init_one,
1176         .remove = remove_one,
1177         .id_table = hfi1_pci_tbl,
1178         .err_handler = &hfi1_pci_err_handler,
1179 };
1180
1181 static void __init compute_krcvqs(void)
1182 {
1183         int i;
1184
1185         for (i = 0; i < krcvqsset; i++)
1186                 n_krcvqs += krcvqs[i];
1187 }
1188
1189 /*
1190  * Do all the generic driver unit- and chip-independent memory
1191  * allocation and initialization.
1192  */
1193 static int __init hfi1_mod_init(void)
1194 {
1195         int ret;
1196
1197         ret = dev_init();
1198         if (ret)
1199                 goto bail;
1200
1201         /* validate max MTU before any devices start */
1202         if (!valid_opa_max_mtu(hfi1_max_mtu)) {
1203                 pr_err("Invalid max_mtu 0x%x, using 0x%x instead\n",
1204                        hfi1_max_mtu, HFI1_DEFAULT_MAX_MTU);
1205                 hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
1206         }
1207         /* valid CUs run from 1-128 in powers of 2 */
1208         if (hfi1_cu > 128 || !is_power_of_2(hfi1_cu))
1209                 hfi1_cu = 1;
1210         /* valid credit return threshold is 0-100, variable is unsigned */
1211         if (user_credit_return_threshold > 100)
1212                 user_credit_return_threshold = 100;
1213
1214         compute_krcvqs();
1215         /*
1216          * sanitize receive interrupt count, time must wait until after
1217          * the hardware type is known
1218          */
1219         if (rcv_intr_count > RCV_HDR_HEAD_COUNTER_MASK)
1220                 rcv_intr_count = RCV_HDR_HEAD_COUNTER_MASK;
1221         /* reject invalid combinations */
1222         if (rcv_intr_count == 0 && rcv_intr_timeout == 0) {
1223                 pr_err("Invalid mode: both receive interrupt count and available timeout are zero - setting interrupt count to 1\n");
1224                 rcv_intr_count = 1;
1225         }
1226         if (rcv_intr_count > 1 && rcv_intr_timeout == 0) {
1227                 /*
1228                  * Avoid indefinite packet delivery by requiring a timeout
1229                  * if count is > 1.
1230                  */
1231                 pr_err("Invalid mode: receive interrupt count greater than 1 and available timeout is zero - setting available timeout to 1\n");
1232                 rcv_intr_timeout = 1;
1233         }
1234         if (rcv_intr_dynamic && !(rcv_intr_count > 1 && rcv_intr_timeout > 0)) {
1235                 /*
1236                  * The dynamic algorithm expects a non-zero timeout
1237                  * and a count > 1.
1238                  */
1239                 pr_err("Invalid mode: dynamic receive interrupt mitigation with invalid count and timeout - turning dynamic off\n");
1240                 rcv_intr_dynamic = 0;
1241         }
1242
1243         /* sanitize link CRC options */
1244         link_crc_mask &= SUPPORTED_CRCS;
1245
1246         /*
1247          * These must be called before the driver is registered with
1248          * the PCI subsystem.
1249          */
1250         idr_init(&hfi1_unit_table);
1251
1252         hfi1_dbg_init();
1253         ret = hfi1_wss_init();
1254         if (ret < 0)
1255                 goto bail_wss;
1256         ret = pci_register_driver(&hfi1_pci_driver);
1257         if (ret < 0) {
1258                 pr_err("Unable to register driver: error %d\n", -ret);
1259                 goto bail_dev;
1260         }
1261         goto bail; /* all OK */
1262
1263 bail_dev:
1264         hfi1_wss_exit();
1265 bail_wss:
1266         hfi1_dbg_exit();
1267         idr_destroy(&hfi1_unit_table);
1268         dev_cleanup();
1269 bail:
1270         return ret;
1271 }
1272
1273 module_init(hfi1_mod_init);
1274
1275 /*
1276  * Do the non-unit driver cleanup, memory free, etc. at unload.
1277  */
1278 static void __exit hfi1_mod_cleanup(void)
1279 {
1280         pci_unregister_driver(&hfi1_pci_driver);
1281         hfi1_wss_exit();
1282         hfi1_dbg_exit();
1283         hfi1_cpulist_count = 0;
1284         kfree(hfi1_cpulist);
1285
1286         idr_destroy(&hfi1_unit_table);
1287         dispose_firmware();     /* asymmetric with obtain_firmware() */
1288         dev_cleanup();
1289 }
1290
1291 module_exit(hfi1_mod_cleanup);
1292
1293 /* this can only be called after a successful initialization */
1294 static void cleanup_device_data(struct hfi1_devdata *dd)
1295 {
1296         int ctxt;
1297         int pidx;
1298         struct hfi1_ctxtdata **tmp;
1299         unsigned long flags;
1300
1301         /* users can't do anything more with chip */
1302         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1303                 struct hfi1_pportdata *ppd = &dd->pport[pidx];
1304                 struct cc_state *cc_state;
1305                 int i;
1306
1307                 if (ppd->statusp)
1308                         *ppd->statusp &= ~HFI1_STATUS_CHIP_PRESENT;
1309
1310                 for (i = 0; i < OPA_MAX_SLS; i++)
1311                         hrtimer_cancel(&ppd->cca_timer[i].hrtimer);
1312
1313                 spin_lock(&ppd->cc_state_lock);
1314                 cc_state = get_cc_state(ppd);
1315                 RCU_INIT_POINTER(ppd->cc_state, NULL);
1316                 spin_unlock(&ppd->cc_state_lock);
1317
1318                 if (cc_state)
1319                         call_rcu(&cc_state->rcu, cc_state_reclaim);
1320         }
1321
1322         free_credit_return(dd);
1323
1324         /*
1325          * Free any resources still in use (usually just kernel contexts)
1326          * at unload; we do for ctxtcnt, because that's what we allocate.
1327          * We acquire lock to be really paranoid that rcd isn't being
1328          * accessed from some interrupt-related code (that should not happen,
1329          * but best to be sure).
1330          */
1331         spin_lock_irqsave(&dd->uctxt_lock, flags);
1332         tmp = dd->rcd;
1333         dd->rcd = NULL;
1334         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1335
1336         if (dd->rcvhdrtail_dummy_kvaddr) {
1337                 dma_free_coherent(&dd->pcidev->dev, sizeof(u64),
1338                                   (void *)dd->rcvhdrtail_dummy_kvaddr,
1339                                   dd->rcvhdrtail_dummy_physaddr);
1340                 dd->rcvhdrtail_dummy_kvaddr = NULL;
1341         }
1342
1343         for (ctxt = 0; tmp && ctxt < dd->num_rcv_contexts; ctxt++) {
1344                 struct hfi1_ctxtdata *rcd = tmp[ctxt];
1345
1346                 tmp[ctxt] = NULL; /* debugging paranoia */
1347                 if (rcd) {
1348                         hfi1_clear_tids(rcd);
1349                         hfi1_free_ctxtdata(dd, rcd);
1350                 }
1351         }
1352         kfree(tmp);
1353         free_pio_map(dd);
1354         /* must follow rcv context free - need to remove rcv's hooks */
1355         for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
1356                 sc_free(dd->send_contexts[ctxt].sc);
1357         dd->num_send_contexts = 0;
1358         kfree(dd->send_contexts);
1359         dd->send_contexts = NULL;
1360         kfree(dd->hw_to_sw);
1361         dd->hw_to_sw = NULL;
1362         kfree(dd->boardname);
1363         vfree(dd->events);
1364         vfree(dd->status);
1365 }
1366
1367 /*
1368  * Clean up on unit shutdown, or error during unit load after
1369  * successful initialization.
1370  */
1371 static void postinit_cleanup(struct hfi1_devdata *dd)
1372 {
1373         hfi1_start_cleanup(dd);
1374
1375         hfi1_pcie_ddcleanup(dd);
1376         hfi1_pcie_cleanup(dd->pcidev);
1377
1378         cleanup_device_data(dd);
1379
1380         hfi1_free_devdata(dd);
1381 }
1382
1383 static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1384 {
1385         int ret = 0, j, pidx, initfail;
1386         struct hfi1_devdata *dd = NULL;
1387         struct hfi1_pportdata *ppd;
1388
1389         /* First, lock the non-writable module parameters */
1390         HFI1_CAP_LOCK();
1391
1392         /* Validate some global module parameters */
1393         if (rcvhdrcnt <= HFI1_MIN_HDRQ_EGRBUF_CNT) {
1394                 hfi1_early_err(&pdev->dev, "Header queue  count too small\n");
1395                 ret = -EINVAL;
1396                 goto bail;
1397         }
1398         if (rcvhdrcnt > HFI1_MAX_HDRQ_EGRBUF_CNT) {
1399                 hfi1_early_err(&pdev->dev,
1400                                "Receive header queue count cannot be greater than %u\n",
1401                                HFI1_MAX_HDRQ_EGRBUF_CNT);
1402                 ret = -EINVAL;
1403                 goto bail;
1404         }
1405         /* use the encoding function as a sanitization check */
1406         if (!encode_rcv_header_entry_size(hfi1_hdrq_entsize)) {
1407                 hfi1_early_err(&pdev->dev, "Invalid HdrQ Entry size %u\n",
1408                                hfi1_hdrq_entsize);
1409                 ret = -EINVAL;
1410                 goto bail;
1411         }
1412
1413         /* The receive eager buffer size must be set before the receive
1414          * contexts are created.
1415          *
1416          * Set the eager buffer size.  Validate that it falls in a range
1417          * allowed by the hardware - all powers of 2 between the min and
1418          * max.  The maximum valid MTU is within the eager buffer range
1419          * so we do not need to cap the max_mtu by an eager buffer size
1420          * setting.
1421          */
1422         if (eager_buffer_size) {
1423                 if (!is_power_of_2(eager_buffer_size))
1424                         eager_buffer_size =
1425                                 roundup_pow_of_two(eager_buffer_size);
1426                 eager_buffer_size =
1427                         clamp_val(eager_buffer_size,
1428                                   MIN_EAGER_BUFFER * 8,
1429                                   MAX_EAGER_BUFFER_TOTAL);
1430                 hfi1_early_info(&pdev->dev, "Eager buffer size %u\n",
1431                                 eager_buffer_size);
1432         } else {
1433                 hfi1_early_err(&pdev->dev, "Invalid Eager buffer size of 0\n");
1434                 ret = -EINVAL;
1435                 goto bail;
1436         }
1437
1438         /* restrict value of hfi1_rcvarr_split */
1439         hfi1_rcvarr_split = clamp_val(hfi1_rcvarr_split, 0, 100);
1440
1441         ret = hfi1_pcie_init(pdev, ent);
1442         if (ret)
1443                 goto bail;
1444
1445         /*
1446          * Do device-specific initialization, function table setup, dd
1447          * allocation, etc.
1448          */
1449         switch (ent->device) {
1450         case PCI_DEVICE_ID_INTEL0:
1451         case PCI_DEVICE_ID_INTEL1:
1452                 dd = hfi1_init_dd(pdev, ent);
1453                 break;
1454         default:
1455                 hfi1_early_err(&pdev->dev,
1456                                "Failing on unknown Intel deviceid 0x%x\n",
1457                                ent->device);
1458                 ret = -ENODEV;
1459         }
1460
1461         if (IS_ERR(dd))
1462                 ret = PTR_ERR(dd);
1463         if (ret)
1464                 goto clean_bail; /* error already printed */
1465
1466         ret = create_workqueues(dd);
1467         if (ret)
1468                 goto clean_bail;
1469
1470         /* do the generic initialization */
1471         initfail = hfi1_init(dd, 0);
1472
1473         ret = hfi1_register_ib_device(dd);
1474
1475         /*
1476          * Now ready for use.  this should be cleared whenever we
1477          * detect a reset, or initiate one.  If earlier failure,
1478          * we still create devices, so diags, etc. can be used
1479          * to determine cause of problem.
1480          */
1481         if (!initfail && !ret) {
1482                 dd->flags |= HFI1_INITTED;
1483                 /* create debufs files after init and ib register */
1484                 hfi1_dbg_ibdev_init(&dd->verbs_dev);
1485         }
1486
1487         j = hfi1_device_create(dd);
1488         if (j)
1489                 dd_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1490
1491         if (initfail || ret) {
1492                 stop_timers(dd);
1493                 flush_workqueue(ib_wq);
1494                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1495                         hfi1_quiet_serdes(dd->pport + pidx);
1496                         ppd = dd->pport + pidx;
1497                         if (ppd->hfi1_wq) {
1498                                 destroy_workqueue(ppd->hfi1_wq);
1499                                 ppd->hfi1_wq = NULL;
1500                         }
1501                 }
1502                 if (!j)
1503                         hfi1_device_remove(dd);
1504                 if (!ret)
1505                         hfi1_unregister_ib_device(dd);
1506                 postinit_cleanup(dd);
1507                 if (initfail)
1508                         ret = initfail;
1509                 goto bail;      /* everything already cleaned */
1510         }
1511
1512         sdma_start(dd);
1513
1514         return 0;
1515
1516 clean_bail:
1517         hfi1_pcie_cleanup(pdev);
1518 bail:
1519         return ret;
1520 }
1521
1522 static void remove_one(struct pci_dev *pdev)
1523 {
1524         struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1525
1526         /* close debugfs files before ib unregister */
1527         hfi1_dbg_ibdev_exit(&dd->verbs_dev);
1528         /* unregister from IB core */
1529         hfi1_unregister_ib_device(dd);
1530
1531         /*
1532          * Disable the IB link, disable interrupts on the device,
1533          * clear dma engines, etc.
1534          */
1535         shutdown_device(dd);
1536
1537         stop_timers(dd);
1538
1539         /* wait until all of our (qsfp) queue_work() calls complete */
1540         flush_workqueue(ib_wq);
1541
1542         hfi1_device_remove(dd);
1543
1544         postinit_cleanup(dd);
1545 }
1546
1547 /**
1548  * hfi1_create_rcvhdrq - create a receive header queue
1549  * @dd: the hfi1_ib device
1550  * @rcd: the context data
1551  *
1552  * This must be contiguous memory (from an i/o perspective), and must be
1553  * DMA'able (which means for some systems, it will go through an IOMMU,
1554  * or be forced into a low address range).
1555  */
1556 int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1557 {
1558         unsigned amt;
1559         u64 reg;
1560
1561         if (!rcd->rcvhdrq) {
1562                 dma_addr_t phys_hdrqtail;
1563                 gfp_t gfp_flags;
1564
1565                 /*
1566                  * rcvhdrqentsize is in DWs, so we have to convert to bytes
1567                  * (* sizeof(u32)).
1568                  */
1569                 amt = PAGE_ALIGN(rcd->rcvhdrq_cnt * rcd->rcvhdrqentsize *
1570                                  sizeof(u32));
1571
1572                 gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1573                         GFP_USER : GFP_KERNEL;
1574                 rcd->rcvhdrq = dma_zalloc_coherent(
1575                         &dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1576                         gfp_flags | __GFP_COMP);
1577
1578                 if (!rcd->rcvhdrq) {
1579                         dd_dev_err(dd,
1580                                    "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1581                                    amt, rcd->ctxt);
1582                         goto bail;
1583                 }
1584
1585                 if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
1586                         rcd->rcvhdrtail_kvaddr = dma_zalloc_coherent(
1587                                 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1588                                 gfp_flags);
1589                         if (!rcd->rcvhdrtail_kvaddr)
1590                                 goto bail_free;
1591                         rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1592                 }
1593
1594                 rcd->rcvhdrq_size = amt;
1595         }
1596         /*
1597          * These values are per-context:
1598          *      RcvHdrCnt
1599          *      RcvHdrEntSize
1600          *      RcvHdrSize
1601          */
1602         reg = ((u64)(rcd->rcvhdrq_cnt >> HDRQ_SIZE_SHIFT)
1603                         & RCV_HDR_CNT_CNT_MASK)
1604                 << RCV_HDR_CNT_CNT_SHIFT;
1605         write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_CNT, reg);
1606         reg = (encode_rcv_header_entry_size(rcd->rcvhdrqentsize)
1607                         & RCV_HDR_ENT_SIZE_ENT_SIZE_MASK)
1608                 << RCV_HDR_ENT_SIZE_ENT_SIZE_SHIFT;
1609         write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_ENT_SIZE, reg);
1610         reg = (dd->rcvhdrsize & RCV_HDR_SIZE_HDR_SIZE_MASK)
1611                 << RCV_HDR_SIZE_HDR_SIZE_SHIFT;
1612         write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_SIZE, reg);
1613
1614         /*
1615          * Program dummy tail address for every receive context
1616          * before enabling any receive context
1617          */
1618         write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_TAIL_ADDR,
1619                         dd->rcvhdrtail_dummy_physaddr);
1620
1621         return 0;
1622
1623 bail_free:
1624         dd_dev_err(dd,
1625                    "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1626                    rcd->ctxt);
1627         vfree(rcd->user_event_mask);
1628         rcd->user_event_mask = NULL;
1629         dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1630                           rcd->rcvhdrq_phys);
1631         rcd->rcvhdrq = NULL;
1632 bail:
1633         return -ENOMEM;
1634 }
1635
1636 /**
1637  * allocate eager buffers, both kernel and user contexts.
1638  * @rcd: the context we are setting up.
1639  *
1640  * Allocate the eager TID buffers and program them into hip.
1641  * They are no longer completely contiguous, we do multiple allocation
1642  * calls.  Otherwise we get the OOM code involved, by asking for too
1643  * much per call, with disastrous results on some kernels.
1644  */
1645 int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd)
1646 {
1647         struct hfi1_devdata *dd = rcd->dd;
1648         u32 max_entries, egrtop, alloced_bytes = 0, idx = 0;
1649         gfp_t gfp_flags;
1650         u16 order;
1651         int ret = 0;
1652         u16 round_mtu = roundup_pow_of_two(hfi1_max_mtu);
1653
1654         /*
1655          * GFP_USER, but without GFP_FS, so buffer cache can be
1656          * coalesced (we hope); otherwise, even at order 4,
1657          * heavy filesystem activity makes these fail, and we can
1658          * use compound pages.
1659          */
1660         gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1661
1662         /*
1663          * The minimum size of the eager buffers is a groups of MTU-sized
1664          * buffers.
1665          * The global eager_buffer_size parameter is checked against the
1666          * theoretical lower limit of the value. Here, we check against the
1667          * MTU.
1668          */
1669         if (rcd->egrbufs.size < (round_mtu * dd->rcv_entries.group_size))
1670                 rcd->egrbufs.size = round_mtu * dd->rcv_entries.group_size;
1671         /*
1672          * If using one-pkt-per-egr-buffer, lower the eager buffer
1673          * size to the max MTU (page-aligned).
1674          */
1675         if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
1676                 rcd->egrbufs.rcvtid_size = round_mtu;
1677
1678         /*
1679          * Eager buffers sizes of 1MB or less require smaller TID sizes
1680          * to satisfy the "multiple of 8 RcvArray entries" requirement.
1681          */
1682         if (rcd->egrbufs.size <= (1 << 20))
1683                 rcd->egrbufs.rcvtid_size = max((unsigned long)round_mtu,
1684                         rounddown_pow_of_two(rcd->egrbufs.size / 8));
1685
1686         while (alloced_bytes < rcd->egrbufs.size &&
1687                rcd->egrbufs.alloced < rcd->egrbufs.count) {
1688                 rcd->egrbufs.buffers[idx].addr =
1689                         dma_zalloc_coherent(&dd->pcidev->dev,
1690                                             rcd->egrbufs.rcvtid_size,
1691                                             &rcd->egrbufs.buffers[idx].phys,
1692                                             gfp_flags);
1693                 if (rcd->egrbufs.buffers[idx].addr) {
1694                         rcd->egrbufs.buffers[idx].len =
1695                                 rcd->egrbufs.rcvtid_size;
1696                         rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].addr =
1697                                 rcd->egrbufs.buffers[idx].addr;
1698                         rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].phys =
1699                                 rcd->egrbufs.buffers[idx].phys;
1700                         rcd->egrbufs.alloced++;
1701                         alloced_bytes += rcd->egrbufs.rcvtid_size;
1702                         idx++;
1703                 } else {
1704                         u32 new_size, i, j;
1705                         u64 offset = 0;
1706
1707                         /*
1708                          * Fail the eager buffer allocation if:
1709                          *   - we are already using the lowest acceptable size
1710                          *   - we are using one-pkt-per-egr-buffer (this implies
1711                          *     that we are accepting only one size)
1712                          */
1713                         if (rcd->egrbufs.rcvtid_size == round_mtu ||
1714                             !HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR)) {
1715                                 dd_dev_err(dd, "ctxt%u: Failed to allocate eager buffers\n",
1716                                            rcd->ctxt);
1717                                 goto bail_rcvegrbuf_phys;
1718                         }
1719
1720                         new_size = rcd->egrbufs.rcvtid_size / 2;
1721
1722                         /*
1723                          * If the first attempt to allocate memory failed, don't
1724                          * fail everything but continue with the next lower
1725                          * size.
1726                          */
1727                         if (idx == 0) {
1728                                 rcd->egrbufs.rcvtid_size = new_size;
1729                                 continue;
1730                         }
1731
1732                         /*
1733                          * Re-partition already allocated buffers to a smaller
1734                          * size.
1735                          */
1736                         rcd->egrbufs.alloced = 0;
1737                         for (i = 0, j = 0, offset = 0; j < idx; i++) {
1738                                 if (i >= rcd->egrbufs.count)
1739                                         break;
1740                                 rcd->egrbufs.rcvtids[i].phys =
1741                                         rcd->egrbufs.buffers[j].phys + offset;
1742                                 rcd->egrbufs.rcvtids[i].addr =
1743                                         rcd->egrbufs.buffers[j].addr + offset;
1744                                 rcd->egrbufs.alloced++;
1745                                 if ((rcd->egrbufs.buffers[j].phys + offset +
1746                                      new_size) ==
1747                                     (rcd->egrbufs.buffers[j].phys +
1748                                      rcd->egrbufs.buffers[j].len)) {
1749                                         j++;
1750                                         offset = 0;
1751                                 } else {
1752                                         offset += new_size;
1753                                 }
1754                         }
1755                         rcd->egrbufs.rcvtid_size = new_size;
1756                 }
1757         }
1758         rcd->egrbufs.numbufs = idx;
1759         rcd->egrbufs.size = alloced_bytes;
1760
1761         hfi1_cdbg(PROC,
1762                   "ctxt%u: Alloced %u rcv tid entries @ %uKB, total %zuKB\n",
1763                   rcd->ctxt, rcd->egrbufs.alloced, rcd->egrbufs.rcvtid_size,
1764                   rcd->egrbufs.size);
1765
1766         /*
1767          * Set the contexts rcv array head update threshold to the closest
1768          * power of 2 (so we can use a mask instead of modulo) below half
1769          * the allocated entries.
1770          */
1771         rcd->egrbufs.threshold =
1772                 rounddown_pow_of_two(rcd->egrbufs.alloced / 2);
1773         /*
1774          * Compute the expected RcvArray entry base. This is done after
1775          * allocating the eager buffers in order to maximize the
1776          * expected RcvArray entries for the context.
1777          */
1778         max_entries = rcd->rcv_array_groups * dd->rcv_entries.group_size;
1779         egrtop = roundup(rcd->egrbufs.alloced, dd->rcv_entries.group_size);
1780         rcd->expected_count = max_entries - egrtop;
1781         if (rcd->expected_count > MAX_TID_PAIR_ENTRIES * 2)
1782                 rcd->expected_count = MAX_TID_PAIR_ENTRIES * 2;
1783
1784         rcd->expected_base = rcd->eager_base + egrtop;
1785         hfi1_cdbg(PROC, "ctxt%u: eager:%u, exp:%u, egrbase:%u, expbase:%u\n",
1786                   rcd->ctxt, rcd->egrbufs.alloced, rcd->expected_count,
1787                   rcd->eager_base, rcd->expected_base);
1788
1789         if (!hfi1_rcvbuf_validate(rcd->egrbufs.rcvtid_size, PT_EAGER, &order)) {
1790                 hfi1_cdbg(PROC,
1791                           "ctxt%u: current Eager buffer size is invalid %u\n",
1792                           rcd->ctxt, rcd->egrbufs.rcvtid_size);
1793                 ret = -EINVAL;
1794                 goto bail;
1795         }
1796
1797         for (idx = 0; idx < rcd->egrbufs.alloced; idx++) {
1798                 hfi1_put_tid(dd, rcd->eager_base + idx, PT_EAGER,
1799                              rcd->egrbufs.rcvtids[idx].phys, order);
1800                 cond_resched();
1801         }
1802         goto bail;
1803
1804 bail_rcvegrbuf_phys:
1805         for (idx = 0; idx < rcd->egrbufs.alloced &&
1806              rcd->egrbufs.buffers[idx].addr;
1807              idx++) {
1808                 dma_free_coherent(&dd->pcidev->dev,
1809                                   rcd->egrbufs.buffers[idx].len,
1810                                   rcd->egrbufs.buffers[idx].addr,
1811                                   rcd->egrbufs.buffers[idx].phys);
1812                 rcd->egrbufs.buffers[idx].addr = NULL;
1813                 rcd->egrbufs.buffers[idx].phys = 0;
1814                 rcd->egrbufs.buffers[idx].len = 0;
1815         }
1816 bail:
1817         return ret;
1818 }