err.no Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_sdma.c

   1 /*
   2  * Copyright (c) 2007, 2008 QLogic Corporation. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/spinlock.h>
  34
  35 #include "ipath_kernel.h"
  36 #include "ipath_verbs.h"
  37 #include "ipath_common.h"
  38
  39 #define SDMA_DESCQ_SZ PAGE_SIZE /* 256 entries per 4KB page */
  40
  41 static void vl15_watchdog_enq(struct ipath_devdata *dd)
  42 {
  43         /* ipath_sdma_lock must already be held */
  44         if (atomic_inc_return(&dd->ipath_sdma_vl15_count) == 1) {
  45                 unsigned long interval = (HZ + 19) / 20;
  46                 dd->ipath_sdma_vl15_timer.expires = jiffies + interval;
  47                 add_timer(&dd->ipath_sdma_vl15_timer);
  48         }
  49 }
  50
  51 static void vl15_watchdog_deq(struct ipath_devdata *dd)
  52 {
  53         /* ipath_sdma_lock must already be held */
  54         if (atomic_dec_return(&dd->ipath_sdma_vl15_count) != 0) {
  55                 unsigned long interval = (HZ + 19) / 20;
  56                 mod_timer(&dd->ipath_sdma_vl15_timer, jiffies + interval);
  57         } else {
  58                 del_timer(&dd->ipath_sdma_vl15_timer);
  59         }
  60 }
  61
  62 static void vl15_watchdog_timeout(unsigned long opaque)
  63 {
  64         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
  65
  66         if (atomic_read(&dd->ipath_sdma_vl15_count) != 0) {
  67                 ipath_dbg("vl15 watchdog timeout - clearing\n");
  68                 ipath_cancel_sends(dd, 1);
  69                 ipath_hol_down(dd);
  70         } else {
  71                 ipath_dbg("vl15 watchdog timeout - "
  72                           "condition already cleared\n");
  73         }
  74 }
  75
  76 static void unmap_desc(struct ipath_devdata *dd, unsigned head)
  77 {
  78         __le64 *descqp = &dd->ipath_sdma_descq[head].qw[0];
  79         u64 desc[2];
  80         dma_addr_t addr;
  81         size_t len;
  82
  83         desc[0] = le64_to_cpu(descqp[0]);
  84         desc[1] = le64_to_cpu(descqp[1]);
  85
  86         addr = (desc[1] << 32) | (desc[0] >> 32);
  87         len = (desc[0] >> 14) & (0x7ffULL << 2);
  88         dma_unmap_single(&dd->pcidev->dev, addr, len, DMA_TO_DEVICE);
  89 }
  90
  91 /*
  92  * ipath_sdma_lock should be locked before calling this.
  93  */
  94 int ipath_sdma_make_progress(struct ipath_devdata *dd)
  95 {
  96         struct list_head *lp = NULL;
  97         struct ipath_sdma_txreq *txp = NULL;
  98         u16 dmahead;
  99         u16 start_idx = 0;
 100         int progress = 0;
 101
 102         if (!list_empty(&dd->ipath_sdma_activelist)) {
 103                 lp = dd->ipath_sdma_activelist.next;
 104                 txp = list_entry(lp, struct ipath_sdma_txreq, list);
 105                 start_idx = txp->start_idx;
 106         }
 107
 108         /*
 109          * Read the SDMA head register in order to know that the
 110          * interrupt clear has been written to the chip.
 111          * Otherwise, we may not get an interrupt for the last
 112          * descriptor in the queue.
 113          */
 114         dmahead = (u16)ipath_read_kreg32(dd, dd->ipath_kregs->kr_senddmahead);
 115         /* sanity check return value for error handling (chip reset, etc.) */
 116         if (dmahead >= dd->ipath_sdma_descq_cnt)
 117                 goto done;
 118
 119         while (dd->ipath_sdma_descq_head != dmahead) {
 120                 if (txp && txp->flags & IPATH_SDMA_TXREQ_F_FREEDESC &&
 121                     dd->ipath_sdma_descq_head == start_idx) {
 122                         unmap_desc(dd, dd->ipath_sdma_descq_head);
 123                         start_idx++;
 124                         if (start_idx == dd->ipath_sdma_descq_cnt)
 125                                 start_idx = 0;
 126                 }
 127
 128                 /* increment free count and head */
 129                 dd->ipath_sdma_descq_removed++;
 130                 if (++dd->ipath_sdma_descq_head == dd->ipath_sdma_descq_cnt)
 131                         dd->ipath_sdma_descq_head = 0;
 132
 133                 if (txp && txp->next_descq_idx == dd->ipath_sdma_descq_head) {
 134                         /* move to notify list */
 135                         if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
 136                                 vl15_watchdog_deq(dd);
 137                         list_move_tail(lp, &dd->ipath_sdma_notifylist);
 138                         if (!list_empty(&dd->ipath_sdma_activelist)) {
 139                                 lp = dd->ipath_sdma_activelist.next;
 140                                 txp = list_entry(lp, struct ipath_sdma_txreq,
 141                                                  list);
 142                                 start_idx = txp->start_idx;
 143                         } else {
 144                                 lp = NULL;
 145                                 txp = NULL;
 146                         }
 147                 }
 148                 progress = 1;
 149         }
 150
 151         if (progress)
 152                 tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
 153
 154 done:
 155         return progress;
 156 }
 157
 158 static void ipath_sdma_notify(struct ipath_devdata *dd, struct list_head *list)
 159 {
 160         struct ipath_sdma_txreq *txp, *txp_next;
 161
 162         list_for_each_entry_safe(txp, txp_next, list, list) {
 163                 list_del_init(&txp->list);
 164
 165                 if (txp->callback)
 166                         (*txp->callback)(txp->callback_cookie,
 167                                          txp->callback_status);
 168         }
 169 }
 170
 171 static void sdma_notify_taskbody(struct ipath_devdata *dd)
 172 {
 173         unsigned long flags;
 174         struct list_head list;
 175
 176         INIT_LIST_HEAD(&list);
 177
 178         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 179
 180         list_splice_init(&dd->ipath_sdma_notifylist, &list);
 181
 182         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 183
 184         ipath_sdma_notify(dd, &list);
 185
 186         /*
 187          * The IB verbs layer needs to see the callback before getting
 188          * the call to ipath_ib_piobufavail() because the callback
 189          * handles releasing resources the next send will need.
 190          * Otherwise, we could do these calls in
 191          * ipath_sdma_make_progress().
 192          */
 193         ipath_ib_piobufavail(dd->verbs_dev);
 194 }
 195
 196 static void sdma_notify_task(unsigned long opaque)
 197 {
 198         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
 199
 200         if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 201                 sdma_notify_taskbody(dd);
 202 }
 203
 204 static void dump_sdma_state(struct ipath_devdata *dd)
 205 {
 206         unsigned long reg;
 207
 208         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmastatus);
 209         ipath_cdbg(VERBOSE, "kr_senddmastatus: 0x%016lx\n", reg);
 210
 211         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendctrl);
 212         ipath_cdbg(VERBOSE, "kr_sendctrl: 0x%016lx\n", reg);
 213
 214         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask0);
 215         ipath_cdbg(VERBOSE, "kr_senddmabufmask0: 0x%016lx\n", reg);
 216
 217         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask1);
 218         ipath_cdbg(VERBOSE, "kr_senddmabufmask1: 0x%016lx\n", reg);
 219
 220         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask2);
 221         ipath_cdbg(VERBOSE, "kr_senddmabufmask2: 0x%016lx\n", reg);
 222
 223         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmatail);
 224         ipath_cdbg(VERBOSE, "kr_senddmatail: 0x%016lx\n", reg);
 225
 226         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmahead);
 227         ipath_cdbg(VERBOSE, "kr_senddmahead: 0x%016lx\n", reg);
 228 }
 229
 230 static void sdma_abort_task(unsigned long opaque)
 231 {
 232         struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
 233         u64 status;
 234         unsigned long flags;
 235
 236         if (test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 237                 return;
 238
 239         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 240
 241         status = dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK;
 242
 243         /* nothing to do */
 244         if (status == IPATH_SDMA_ABORT_NONE)
 245                 goto unlock;
 246
 247         /* ipath_sdma_abort() is done, waiting for interrupt */
 248         if (status == IPATH_SDMA_ABORT_DISARMED) {
 249                 if (jiffies < dd->ipath_sdma_abort_intr_timeout)
 250                         goto resched_noprint;
 251                 /* give up, intr got lost somewhere */
 252                 ipath_dbg("give up waiting for SDMADISABLED intr\n");
 253                 __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 254                 status = IPATH_SDMA_ABORT_ABORTED;
 255         }
 256
 257         /* everything is stopped, time to clean up and restart */
 258         if (status == IPATH_SDMA_ABORT_ABORTED) {
 259                 struct ipath_sdma_txreq *txp, *txpnext;
 260                 u64 hwstatus;
 261                 int notify = 0;
 262
 263                 hwstatus = ipath_read_kreg64(dd,
 264                                 dd->ipath_kregs->kr_senddmastatus);
 265
 266                 if (/* ScoreBoardDrainInProg */
 267                     test_bit(63, &hwstatus) ||
 268                     /* AbortInProg */
 269                     test_bit(62, &hwstatus) ||
 270                     /* InternalSDmaEnable */
 271                     test_bit(61, &hwstatus) ||
 272                     /* ScbEmpty */
 273                     !test_bit(30, &hwstatus)) {
 274                         if (dd->ipath_sdma_reset_wait > 0) {
 275                                 /* not done shutting down sdma */
 276                                 --dd->ipath_sdma_reset_wait;
 277                                 goto resched;
 278                         }
 279                         ipath_cdbg(VERBOSE, "gave up waiting for quiescent "
 280                                 "status after SDMA reset, continuing\n");
 281                         dump_sdma_state(dd);
 282                 }
 283
 284                 /* dequeue all "sent" requests */
 285                 list_for_each_entry_safe(txp, txpnext,
 286                                          &dd->ipath_sdma_activelist, list) {
 287                         txp->callback_status = IPATH_SDMA_TXREQ_S_ABORTED;
 288                         if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
 289                                 vl15_watchdog_deq(dd);
 290                         list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
 291                         notify = 1;
 292                 }
 293                 if (notify)
 294                         tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
 295
 296                 /* reset our notion of head and tail */
 297                 dd->ipath_sdma_descq_tail = 0;
 298                 dd->ipath_sdma_descq_head = 0;
 299                 dd->ipath_sdma_head_dma[0] = 0;
 300                 dd->ipath_sdma_generation = 0;
 301                 dd->ipath_sdma_descq_removed = dd->ipath_sdma_descq_added;
 302
 303                 /* Reset SendDmaLenGen */
 304                 ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen,
 305                         (u64) dd->ipath_sdma_descq_cnt | (1ULL << 18));
 306
 307                 /* done with sdma state for a bit */
 308                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 309
 310                 /*
 311                  * Don't restart sdma here. Wait until link is up to ACTIVE.
 312                  * VL15 MADs used to bring the link up use PIO, and multiple
 313                  * link transitions otherwise cause the sdma engine to be
 314                  * stopped and started multiple times.
 315                  * The disable is done here, including the shadow, so the
 316                  * state is kept consistent.
 317                  * See ipath_restart_sdma() for the actual starting of sdma.
 318                  */
 319                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 320                 dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
 321                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 322                                  dd->ipath_sendctrl);
 323                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 324                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 325
 326                 /* make sure I see next message */
 327                 dd->ipath_sdma_abort_jiffies = 0;
 328
 329                 goto done;
 330         }
 331
 332 resched:
 333         /*
 334          * for now, keep spinning
 335          * JAG - this is bad to just have default be a loop without
 336          * state change
 337          */
 338         if (jiffies > dd->ipath_sdma_abort_jiffies) {
 339                 ipath_dbg("looping with status 0x%016llx\n",
 340                           dd->ipath_sdma_status);
 341                 dd->ipath_sdma_abort_jiffies = jiffies + 5 * HZ;
 342         }
 343 resched_noprint:
 344         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 345         if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 346                 tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
 347         return;
 348
 349 unlock:
 350         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 351 done:
 352         return;
 353 }
 354
 355 /*
 356  * This is called from interrupt context.
 357  */
 358 void ipath_sdma_intr(struct ipath_devdata *dd)
 359 {
 360         unsigned long flags;
 361
 362         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 363
 364         (void) ipath_sdma_make_progress(dd);
 365
 366         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 367 }
 368
 369 static int alloc_sdma(struct ipath_devdata *dd)
 370 {
 371         int ret = 0;
 372
 373         /* Allocate memory for SendDMA descriptor FIFO */
 374         dd->ipath_sdma_descq = dma_alloc_coherent(&dd->pcidev->dev,
 375                 SDMA_DESCQ_SZ, &dd->ipath_sdma_descq_phys, GFP_KERNEL);
 376
 377         if (!dd->ipath_sdma_descq) {
 378                 ipath_dev_err(dd, "failed to allocate SendDMA descriptor "
 379                         "FIFO memory\n");
 380                 ret = -ENOMEM;
 381                 goto done;
 382         }
 383
 384         dd->ipath_sdma_descq_cnt =
 385                 SDMA_DESCQ_SZ / sizeof(struct ipath_sdma_desc);
 386
 387         /* Allocate memory for DMA of head register to memory */
 388         dd->ipath_sdma_head_dma = dma_alloc_coherent(&dd->pcidev->dev,
 389                 PAGE_SIZE, &dd->ipath_sdma_head_phys, GFP_KERNEL);
 390         if (!dd->ipath_sdma_head_dma) {
 391                 ipath_dev_err(dd, "failed to allocate SendDMA head memory\n");
 392                 ret = -ENOMEM;
 393                 goto cleanup_descq;
 394         }
 395         dd->ipath_sdma_head_dma[0] = 0;
 396
 397         init_timer(&dd->ipath_sdma_vl15_timer);
 398         dd->ipath_sdma_vl15_timer.function = vl15_watchdog_timeout;
 399         dd->ipath_sdma_vl15_timer.data = (unsigned long)dd;
 400         atomic_set(&dd->ipath_sdma_vl15_count, 0);
 401
 402         goto done;
 403
 404 cleanup_descq:
 405         dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
 406                 (void *)dd->ipath_sdma_descq, dd->ipath_sdma_descq_phys);
 407         dd->ipath_sdma_descq = NULL;
 408         dd->ipath_sdma_descq_phys = 0;
 409 done:
 410         return ret;
 411 }
 412
 413 int setup_sdma(struct ipath_devdata *dd)
 414 {
 415         int ret = 0;
 416         unsigned i, n;
 417         u64 tmp64;
 418         u64 senddmabufmask[3] = { 0 };
 419         unsigned long flags;
 420
 421         ret = alloc_sdma(dd);
 422         if (ret)
 423                 goto done;
 424
 425         if (!dd->ipath_sdma_descq) {
 426                 ipath_dev_err(dd, "SendDMA memory not allocated\n");
 427                 goto done;
 428         }
 429
 430         dd->ipath_sdma_status = 0;
 431         dd->ipath_sdma_abort_jiffies = 0;
 432         dd->ipath_sdma_generation = 0;
 433         dd->ipath_sdma_descq_tail = 0;
 434         dd->ipath_sdma_descq_head = 0;
 435         dd->ipath_sdma_descq_removed = 0;
 436         dd->ipath_sdma_descq_added = 0;
 437
 438         /* Set SendDmaBase */
 439         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase,
 440                          dd->ipath_sdma_descq_phys);
 441         /* Set SendDmaLenGen */
 442         tmp64 = dd->ipath_sdma_descq_cnt;
 443         tmp64 |= 1<<18; /* enable generation checking */
 444         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, tmp64);
 445         /* Set SendDmaTail */
 446         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail,
 447                          dd->ipath_sdma_descq_tail);
 448         /* Set SendDmaHeadAddr */
 449         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr,
 450                          dd->ipath_sdma_head_phys);
 451
 452         /*
 453          * Reserve all the former "kernel" piobufs, using high number range
 454          * so we get as many 4K buffers as possible
 455          */
 456         n = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
 457         i = dd->ipath_lastport_piobuf + dd->ipath_pioreserved;
 458         ipath_chg_pioavailkernel(dd, i, n - i , 0);
 459         for (; i < n; ++i) {
 460                 unsigned word = i / 64;
 461                 unsigned bit = i & 63;
 462                 BUG_ON(word >= 3);
 463                 senddmabufmask[word] |= 1ULL << bit;
 464         }
 465         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0,
 466                          senddmabufmask[0]);
 467         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1,
 468                          senddmabufmask[1]);
 469         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2,
 470                          senddmabufmask[2]);
 471
 472         INIT_LIST_HEAD(&dd->ipath_sdma_activelist);
 473         INIT_LIST_HEAD(&dd->ipath_sdma_notifylist);
 474
 475         tasklet_init(&dd->ipath_sdma_notify_task, sdma_notify_task,
 476                      (unsigned long) dd);
 477         tasklet_init(&dd->ipath_sdma_abort_task, sdma_abort_task,
 478                      (unsigned long) dd);
 479
 480         /*
 481          * No use to turn on SDMA here, as link is probably not ACTIVE
 482          * Just mark it RUNNING and enable the interrupt, and let the
 483          * ipath_restart_sdma() on link transition to ACTIVE actually
 484          * enable it.
 485          */
 486         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 487         dd->ipath_sendctrl |= INFINIPATH_S_SDMAINTENABLE;
 488         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 489         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 490         __set_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
 491         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 492
 493 done:
 494         return ret;
 495 }
 496
 497 void teardown_sdma(struct ipath_devdata *dd)
 498 {
 499         struct ipath_sdma_txreq *txp, *txpnext;
 500         unsigned long flags;
 501         dma_addr_t sdma_head_phys = 0;
 502         dma_addr_t sdma_descq_phys = 0;
 503         void *sdma_descq = NULL;
 504         void *sdma_head_dma = NULL;
 505
 506         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 507         __clear_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
 508         __set_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
 509         __set_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status);
 510         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 511
 512         tasklet_kill(&dd->ipath_sdma_abort_task);
 513         tasklet_kill(&dd->ipath_sdma_notify_task);
 514
 515         /* turn off sdma */
 516         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 517         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
 518         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 519                 dd->ipath_sendctrl);
 520         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 521         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 522
 523         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 524         /* dequeue all "sent" requests */
 525         list_for_each_entry_safe(txp, txpnext, &dd->ipath_sdma_activelist,
 526                                  list) {
 527                 txp->callback_status = IPATH_SDMA_TXREQ_S_SHUTDOWN;
 528                 if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
 529                         vl15_watchdog_deq(dd);
 530                 list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
 531         }
 532         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 533
 534         sdma_notify_taskbody(dd);
 535
 536         del_timer_sync(&dd->ipath_sdma_vl15_timer);
 537
 538         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 539
 540         dd->ipath_sdma_abort_jiffies = 0;
 541
 542         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase, 0);
 543         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, 0);
 544         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, 0);
 545         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr, 0);
 546         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0, 0);
 547         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1, 0);
 548         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2, 0);
 549
 550         if (dd->ipath_sdma_head_dma) {
 551                 sdma_head_dma = (void *) dd->ipath_sdma_head_dma;
 552                 sdma_head_phys = dd->ipath_sdma_head_phys;
 553                 dd->ipath_sdma_head_dma = NULL;
 554                 dd->ipath_sdma_head_phys = 0;
 555         }
 556
 557         if (dd->ipath_sdma_descq) {
 558                 sdma_descq = dd->ipath_sdma_descq;
 559                 sdma_descq_phys = dd->ipath_sdma_descq_phys;
 560                 dd->ipath_sdma_descq = NULL;
 561                 dd->ipath_sdma_descq_phys = 0;
 562         }
 563
 564         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 565
 566         if (sdma_head_dma)
 567                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
 568                                   sdma_head_dma, sdma_head_phys);
 569
 570         if (sdma_descq)
 571                 dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
 572                                   sdma_descq, sdma_descq_phys);
 573 }
 574
 575 /*
 576  * [Re]start SDMA, if we use it, and it's not already OK.
 577  * This is called on transition to link ACTIVE, either the first or
 578  * subsequent times.
 579  */
 580 void ipath_restart_sdma(struct ipath_devdata *dd)
 581 {
 582         unsigned long flags;
 583         int needed = 1;
 584
 585         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA))
 586                 goto bail;
 587
 588         /*
 589          * First, make sure we should, which is to say,
 590          * check that we are "RUNNING" (not in teardown)
 591          * and not "SHUTDOWN"
 592          */
 593         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 594         if (!test_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status)
 595                 || test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 596                         needed = 0;
 597         else {
 598                 __clear_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
 599                 __clear_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
 600                 __clear_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
 601         }
 602         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 603         if (!needed) {
 604                 ipath_dbg("invalid attempt to restart SDMA, status 0x%016llx\n",
 605                         dd->ipath_sdma_status);
 606                 goto bail;
 607         }
 608         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 609         /*
 610          * First clear, just to be safe. Enable is only done
 611          * in chip on 0->1 transition
 612          */
 613         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
 614         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 615         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 616         dd->ipath_sendctrl |= INFINIPATH_S_SDMAENABLE;
 617         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
 618         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 619         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 620
 621 bail:
 622         return;
 623 }
 624
 625 static inline void make_sdma_desc(struct ipath_devdata *dd,
 626         u64 *sdmadesc, u64 addr, u64 dwlen, u64 dwoffset)
 627 {
 628         WARN_ON(addr & 3);
 629         /* SDmaPhyAddr[47:32] */
 630         sdmadesc[1] = addr >> 32;
 631         /* SDmaPhyAddr[31:0] */
 632         sdmadesc[0] = (addr & 0xfffffffcULL) << 32;
 633         /* SDmaGeneration[1:0] */
 634         sdmadesc[0] |= (dd->ipath_sdma_generation & 3ULL) << 30;
 635         /* SDmaDwordCount[10:0] */
 636         sdmadesc[0] |= (dwlen & 0x7ffULL) << 16;
 637         /* SDmaBufOffset[12:2] */
 638         sdmadesc[0] |= dwoffset & 0x7ffULL;
 639 }
 640
 641 /*
 642  * This function queues one IB packet onto the send DMA queue per call.
 643  * The caller is responsible for checking:
 644  * 1) The number of send DMA descriptor entries is less than the size of
 645  *    the descriptor queue.
 646  * 2) The IB SGE addresses and lengths are 32-bit aligned
 647  *    (except possibly the last SGE's length)
 648  * 3) The SGE addresses are suitable for passing to dma_map_single().
 649  */
 650 int ipath_sdma_verbs_send(struct ipath_devdata *dd,
 651         struct ipath_sge_state *ss, u32 dwords,
 652         struct ipath_verbs_txreq *tx)
 653 {
 654
 655         unsigned long flags;
 656         struct ipath_sge *sge;
 657         int ret = 0;
 658         u16 tail;
 659         __le64 *descqp;
 660         u64 sdmadesc[2];
 661         u32 dwoffset;
 662         dma_addr_t addr;
 663
 664         if ((tx->map_len + (dwords<<2)) > dd->ipath_ibmaxlen) {
 665                 ipath_dbg("packet size %X > ibmax %X, fail\n",
 666                         tx->map_len + (dwords<<2), dd->ipath_ibmaxlen);
 667                 ret = -EMSGSIZE;
 668                 goto fail;
 669         }
 670
 671         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 672
 673 retry:
 674         if (unlikely(test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status))) {
 675                 ret = -EBUSY;
 676                 goto unlock;
 677         }
 678
 679         if (tx->txreq.sg_count > ipath_sdma_descq_freecnt(dd)) {
 680                 if (ipath_sdma_make_progress(dd))
 681                         goto retry;
 682                 ret = -ENOBUFS;
 683                 goto unlock;
 684         }
 685
 686         addr = dma_map_single(&dd->pcidev->dev, tx->txreq.map_addr,
 687                               tx->map_len, DMA_TO_DEVICE);
 688         if (dma_mapping_error(addr)) {
 689                 ret = -EIO;
 690                 goto unlock;
 691         }
 692
 693         dwoffset = tx->map_len >> 2;
 694         make_sdma_desc(dd, sdmadesc, (u64) addr, dwoffset, 0);
 695
 696         /* SDmaFirstDesc */
 697         sdmadesc[0] |= 1ULL << 12;
 698         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
 699                 sdmadesc[0] |= 1ULL << 14;      /* SDmaUseLargeBuf */
 700
 701         /* write to the descq */
 702         tail = dd->ipath_sdma_descq_tail;
 703         descqp = &dd->ipath_sdma_descq[tail].qw[0];
 704         *descqp++ = cpu_to_le64(sdmadesc[0]);
 705         *descqp++ = cpu_to_le64(sdmadesc[1]);
 706
 707         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEDESC)
 708                 tx->txreq.start_idx = tail;
 709
 710         /* increment the tail */
 711         if (++tail == dd->ipath_sdma_descq_cnt) {
 712                 tail = 0;
 713                 descqp = &dd->ipath_sdma_descq[0].qw[0];
 714                 ++dd->ipath_sdma_generation;
 715         }
 716
 717         sge = &ss->sge;
 718         while (dwords) {
 719                 u32 dw;
 720                 u32 len;
 721
 722                 len = dwords << 2;
 723                 if (len > sge->length)
 724                         len = sge->length;
 725                 if (len > sge->sge_length)
 726                         len = sge->sge_length;
 727                 BUG_ON(len == 0);
 728                 dw = (len + 3) >> 2;
 729                 addr = dma_map_single(&dd->pcidev->dev, sge->vaddr, dw << 2,
 730                                       DMA_TO_DEVICE);
 731                 make_sdma_desc(dd, sdmadesc, (u64) addr, dw, dwoffset);
 732                 /* SDmaUseLargeBuf has to be set in every descriptor */
 733                 if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
 734                         sdmadesc[0] |= 1ULL << 14;
 735                 /* write to the descq */
 736                 *descqp++ = cpu_to_le64(sdmadesc[0]);
 737                 *descqp++ = cpu_to_le64(sdmadesc[1]);
 738
 739                 /* increment the tail */
 740                 if (++tail == dd->ipath_sdma_descq_cnt) {
 741                         tail = 0;
 742                         descqp = &dd->ipath_sdma_descq[0].qw[0];
 743                         ++dd->ipath_sdma_generation;
 744                 }
 745                 sge->vaddr += len;
 746                 sge->length -= len;
 747                 sge->sge_length -= len;
 748                 if (sge->sge_length == 0) {
 749                         if (--ss->num_sge)
 750                                 *sge = *ss->sg_list++;
 751                 } else if (sge->length == 0 && sge->mr != NULL) {
 752                         if (++sge->n >= IPATH_SEGSZ) {
 753                                 if (++sge->m >= sge->mr->mapsz)
 754                                         break;
 755                                 sge->n = 0;
 756                         }
 757                         sge->vaddr =
 758                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 759                         sge->length =
 760                                 sge->mr->map[sge->m]->segs[sge->n].length;
 761                 }
 762
 763                 dwoffset += dw;
 764                 dwords -= dw;
 765         }
 766
 767         if (!tail)
 768                 descqp = &dd->ipath_sdma_descq[dd->ipath_sdma_descq_cnt].qw[0];
 769         descqp -= 2;
 770         /* SDmaLastDesc */
 771         descqp[0] |= __constant_cpu_to_le64(1ULL << 11);
 772         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_INTREQ) {
 773                 /* SDmaIntReq */
 774                 descqp[0] |= __constant_cpu_to_le64(1ULL << 15);
 775         }
 776
 777         /* Commit writes to memory and advance the tail on the chip */
 778         wmb();
 779         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
 780
 781         tx->txreq.next_descq_idx = tail;
 782         tx->txreq.callback_status = IPATH_SDMA_TXREQ_S_OK;
 783         dd->ipath_sdma_descq_tail = tail;
 784         dd->ipath_sdma_descq_added += tx->txreq.sg_count;
 785         list_add_tail(&tx->txreq.list, &dd->ipath_sdma_activelist);
 786         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_VL15)
 787                 vl15_watchdog_enq(dd);
 788
 789 unlock:
 790         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 791 fail:
 792         return ret;
 793 }