SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_user_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 #include <linux/mm.h>
  33 #include <linux/types.h>
  34 #include <linux/device.h>
  35 #include <linux/dmapool.h>
  36 #include <linux/slab.h>
  37 #include <linux/list.h>
  38 #include <linux/highmem.h>
  39 #include <linux/io.h>
  40 #include <linux/uio.h>
  41 #include <linux/rbtree.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/delay.h>
  44
  45 #include "ipath_kernel.h"
  46 #include "ipath_user_sdma.h"
  47
  48 /* minimum size of header */
  49 #define IPATH_USER_SDMA_MIN_HEADER_LENGTH       64
  50 /* expected size of headers (for dma_pool) */
  51 #define IPATH_USER_SDMA_EXP_HEADER_LENGTH       64
  52 /* length mask in PBC (lower 11 bits) */
  53 #define IPATH_PBC_LENGTH_MASK                   ((1 << 11) - 1)
  54
  55 struct ipath_user_sdma_pkt {
  56         u8 naddr;               /* dimension of addr (1..3) ... */
  57         u32 counter;            /* sdma pkts queued counter for this entry */
  58         u64 added;              /* global descq number of entries */
  59
  60         struct {
  61                 u32 offset;                     /* offset for kvaddr, addr */
  62                 u32 length;                     /* length in page */
  63                 u8  put_page;                   /* should we put_page? */
  64                 u8  dma_mapped;                 /* is page dma_mapped? */
  65                 struct page *page;              /* may be NULL (coherent mem) */
  66                 void *kvaddr;                   /* FIXME: only for pio hack */
  67                 dma_addr_t addr;
  68         } addr[4];   /* max pages, any more and we coalesce */
  69         struct list_head list;  /* list element */
  70 };
  71
  72 struct ipath_user_sdma_queue {
  73         /*
  74          * pkts sent to dma engine are queued on this
  75          * list head.  the type of the elements of this
  76          * list are struct ipath_user_sdma_pkt...
  77          */
  78         struct list_head sent;
  79
  80         /* headers with expected length are allocated from here... */
  81         char header_cache_name[64];
  82         struct dma_pool *header_cache;
  83
  84         /* packets are allocated from the slab cache... */
  85         char pkt_slab_name[64];
  86         struct kmem_cache *pkt_slab;
  87
  88         /* as packets go on the queued queue, they are counted... */
  89         u32 counter;
  90         u32 sent_counter;
  91
  92         /* dma page table */
  93         struct rb_root dma_pages_root;
  94
  95         /* protect everything above... */
  96         struct mutex lock;
  97 };
  98
  99 struct ipath_user_sdma_queue *
 100 ipath_user_sdma_queue_create(struct device *dev, int unit, int port, int sport)
 101 {
 102         struct ipath_user_sdma_queue *pq =
 103                 kmalloc(sizeof(struct ipath_user_sdma_queue), GFP_KERNEL);
 104
 105         if (!pq)
 106                 goto done;
 107
 108         pq->counter = 0;
 109         pq->sent_counter = 0;
 110         INIT_LIST_HEAD(&pq->sent);
 111
 112         mutex_init(&pq->lock);
 113
 114         snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
 115                  "ipath-user-sdma-pkts-%u-%02u.%02u", unit, port, sport);
 116         pq->pkt_slab = kmem_cache_create(pq->pkt_slab_name,
 117                                          sizeof(struct ipath_user_sdma_pkt),
 118                                          0, 0, NULL);
 119
 120         if (!pq->pkt_slab)
 121                 goto err_kfree;
 122
 123         snprintf(pq->header_cache_name, sizeof(pq->header_cache_name),
 124                  "ipath-user-sdma-headers-%u-%02u.%02u", unit, port, sport);
 125         pq->header_cache = dma_pool_create(pq->header_cache_name,
 126                                            dev,
 127                                            IPATH_USER_SDMA_EXP_HEADER_LENGTH,
 128                                            4, 0);
 129         if (!pq->header_cache)
 130                 goto err_slab;
 131
 132         pq->dma_pages_root = RB_ROOT;
 133
 134         goto done;
 135
 136 err_slab:
 137         kmem_cache_destroy(pq->pkt_slab);
 138 err_kfree:
 139         kfree(pq);
 140         pq = NULL;
 141
 142 done:
 143         return pq;
 144 }
 145
 146 static void ipath_user_sdma_init_frag(struct ipath_user_sdma_pkt *pkt,
 147                                       int i, size_t offset, size_t len,
 148                                       int put_page, int dma_mapped,
 149                                       struct page *page,
 150                                       void *kvaddr, dma_addr_t dma_addr)
 151 {
 152         pkt->addr[i].offset = offset;
 153         pkt->addr[i].length = len;
 154         pkt->addr[i].put_page = put_page;
 155         pkt->addr[i].dma_mapped = dma_mapped;
 156         pkt->addr[i].page = page;
 157         pkt->addr[i].kvaddr = kvaddr;
 158         pkt->addr[i].addr = dma_addr;
 159 }
 160
 161 static void ipath_user_sdma_init_header(struct ipath_user_sdma_pkt *pkt,
 162                                         u32 counter, size_t offset,
 163                                         size_t len, int dma_mapped,
 164                                         struct page *page,
 165                                         void *kvaddr, dma_addr_t dma_addr)
 166 {
 167         pkt->naddr = 1;
 168         pkt->counter = counter;
 169         ipath_user_sdma_init_frag(pkt, 0, offset, len, 0, dma_mapped, page,
 170                                   kvaddr, dma_addr);
 171 }
 172
 173 /* we've too many pages in the iovec, coalesce to a single page */
 174 static int ipath_user_sdma_coalesce(const struct ipath_devdata *dd,
 175                                     struct ipath_user_sdma_pkt *pkt,
 176                                     const struct iovec *iov,
 177                                     unsigned long niov) {
 178         int ret = 0;
 179         struct page *page = alloc_page(GFP_KERNEL);
 180         void *mpage_save;
 181         char *mpage;
 182         int i;
 183         int len = 0;
 184         dma_addr_t dma_addr;
 185
 186         if (!page) {
 187                 ret = -ENOMEM;
 188                 goto done;
 189         }
 190
 191         mpage = kmap(page);
 192         mpage_save = mpage;
 193         for (i = 0; i < niov; i++) {
 194                 int cfur;
 195
 196                 cfur = copy_from_user(mpage,
 197                                       iov[i].iov_base, iov[i].iov_len);
 198                 if (cfur) {
 199                         ret = -EFAULT;
 200                         goto free_unmap;
 201                 }
 202
 203                 mpage += iov[i].iov_len;
 204                 len += iov[i].iov_len;
 205         }
 206
 207         dma_addr = dma_map_page(&dd->pcidev->dev, page, 0, len,
 208                                 DMA_TO_DEVICE);
 209         if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
 210                 ret = -ENOMEM;
 211                 goto free_unmap;
 212         }
 213
 214         ipath_user_sdma_init_frag(pkt, 1, 0, len, 0, 1, page, mpage_save,
 215                                   dma_addr);
 216         pkt->naddr = 2;
 217
 218         goto done;
 219
 220 free_unmap:
 221         kunmap(page);
 222         __free_page(page);
 223 done:
 224         return ret;
 225 }
 226
 227 /* how many pages in this iovec element? */
 228 static int ipath_user_sdma_num_pages(const struct iovec *iov)
 229 {
 230         const unsigned long addr  = (unsigned long) iov->iov_base;
 231         const unsigned long  len  = iov->iov_len;
 232         const unsigned long spage = addr & PAGE_MASK;
 233         const unsigned long epage = (addr + len - 1) & PAGE_MASK;
 234
 235         return 1 + ((epage - spage) >> PAGE_SHIFT);
 236 }
 237
 238 /* truncate length to page boundry */
 239 static int ipath_user_sdma_page_length(unsigned long addr, unsigned long len)
 240 {
 241         const unsigned long offset = addr & ~PAGE_MASK;
 242
 243         return ((offset + len) > PAGE_SIZE) ? (PAGE_SIZE - offset) : len;
 244 }
 245
 246 static void ipath_user_sdma_free_pkt_frag(struct device *dev,
 247                                           struct ipath_user_sdma_queue *pq,
 248                                           struct ipath_user_sdma_pkt *pkt,
 249                                           int frag)
 250 {
 251         const int i = frag;
 252
 253         if (pkt->addr[i].page) {
 254                 if (pkt->addr[i].dma_mapped)
 255                         dma_unmap_page(dev,
 256                                        pkt->addr[i].addr,
 257                                        pkt->addr[i].length,
 258                                        DMA_TO_DEVICE);
 259
 260                 if (pkt->addr[i].kvaddr)
 261                         kunmap(pkt->addr[i].page);
 262
 263                 if (pkt->addr[i].put_page)
 264                         put_page(pkt->addr[i].page);
 265                 else
 266                         __free_page(pkt->addr[i].page);
 267         } else if (pkt->addr[i].kvaddr)
 268                 /* free coherent mem from cache... */
 269                 dma_pool_free(pq->header_cache,
 270                               pkt->addr[i].kvaddr, pkt->addr[i].addr);
 271 }
 272
 273 /* return number of pages pinned... */
 274 static int ipath_user_sdma_pin_pages(const struct ipath_devdata *dd,
 275                                      struct ipath_user_sdma_pkt *pkt,
 276                                      unsigned long addr, int tlen, int npages)
 277 {
 278         struct page *pages[2];
 279         int j;
 280         int ret;
 281
 282         ret = get_user_pages(current, current->mm, addr,
 283                              npages, 0, 1, pages, NULL);
 284
 285         if (ret != npages) {
 286                 int i;
 287
 288                 for (i = 0; i < ret; i++)
 289                         put_page(pages[i]);
 290
 291                 ret = -ENOMEM;
 292                 goto done;
 293         }
 294
 295         for (j = 0; j < npages; j++) {
 296                 /* map the pages... */
 297                 const int flen =
 298                         ipath_user_sdma_page_length(addr, tlen);
 299                 dma_addr_t dma_addr =
 300                         dma_map_page(&dd->pcidev->dev,
 301                                      pages[j], 0, flen, DMA_TO_DEVICE);
 302                 unsigned long fofs = addr & ~PAGE_MASK;
 303
 304                 if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
 305                         ret = -ENOMEM;
 306                         goto done;
 307                 }
 308
 309                 ipath_user_sdma_init_frag(pkt, pkt->naddr, fofs, flen, 1, 1,
 310                                           pages[j], kmap(pages[j]),
 311                                           dma_addr);
 312
 313                 pkt->naddr++;
 314                 addr += flen;
 315                 tlen -= flen;
 316         }
 317
 318 done:
 319         return ret;
 320 }
 321
 322 static int ipath_user_sdma_pin_pkt(const struct ipath_devdata *dd,
 323                                    struct ipath_user_sdma_queue *pq,
 324                                    struct ipath_user_sdma_pkt *pkt,
 325                                    const struct iovec *iov,
 326                                    unsigned long niov)
 327 {
 328         int ret = 0;
 329         unsigned long idx;
 330
 331         for (idx = 0; idx < niov; idx++) {
 332                 const int npages = ipath_user_sdma_num_pages(iov + idx);
 333                 const unsigned long addr = (unsigned long) iov[idx].iov_base;
 334
 335                 ret = ipath_user_sdma_pin_pages(dd, pkt,
 336                                                 addr, iov[idx].iov_len,
 337                                                 npages);
 338                 if (ret < 0)
 339                         goto free_pkt;
 340         }
 341
 342         goto done;
 343
 344 free_pkt:
 345         for (idx = 0; idx < pkt->naddr; idx++)
 346                 ipath_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
 347
 348 done:
 349         return ret;
 350 }
 351
 352 static int ipath_user_sdma_init_payload(const struct ipath_devdata *dd,
 353                                         struct ipath_user_sdma_queue *pq,
 354                                         struct ipath_user_sdma_pkt *pkt,
 355                                         const struct iovec *iov,
 356                                         unsigned long niov, int npages)
 357 {
 358         int ret = 0;
 359
 360         if (npages >= ARRAY_SIZE(pkt->addr))
 361                 ret = ipath_user_sdma_coalesce(dd, pkt, iov, niov);
 362         else
 363                 ret = ipath_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
 364
 365         return ret;
 366 }
 367
 368 /* free a packet list -- return counter value of last packet */
 369 static void ipath_user_sdma_free_pkt_list(struct device *dev,
 370                                           struct ipath_user_sdma_queue *pq,
 371                                           struct list_head *list)
 372 {
 373         struct ipath_user_sdma_pkt *pkt, *pkt_next;
 374
 375         list_for_each_entry_safe(pkt, pkt_next, list, list) {
 376                 int i;
 377
 378                 for (i = 0; i < pkt->naddr; i++)
 379                         ipath_user_sdma_free_pkt_frag(dev, pq, pkt, i);
 380
 381                 kmem_cache_free(pq->pkt_slab, pkt);
 382         }
 383 }
 384
 385 /*
 386  * copy headers, coalesce etc -- pq->lock must be held
 387  *
 388  * we queue all the packets to list, returning the
 389  * number of bytes total.  list must be empty initially,
 390  * as, if there is an error we clean it...
 391  */
 392 static int ipath_user_sdma_queue_pkts(const struct ipath_devdata *dd,
 393                                       struct ipath_user_sdma_queue *pq,
 394                                       struct list_head *list,
 395                                       const struct iovec *iov,
 396                                       unsigned long niov,
 397                                       int maxpkts)
 398 {
 399         unsigned long idx = 0;
 400         int ret = 0;
 401         int npkts = 0;
 402         struct page *page = NULL;
 403         __le32 *pbc;
 404         dma_addr_t dma_addr;
 405         struct ipath_user_sdma_pkt *pkt = NULL;
 406         size_t len;
 407         size_t nw;
 408         u32 counter = pq->counter;
 409         int dma_mapped = 0;
 410
 411         while (idx < niov && npkts < maxpkts) {
 412                 const unsigned long addr = (unsigned long) iov[idx].iov_base;
 413                 const unsigned long idx_save = idx;
 414                 unsigned pktnw;
 415                 unsigned pktnwc;
 416                 int nfrags = 0;
 417                 int npages = 0;
 418                 int cfur;
 419
 420                 dma_mapped = 0;
 421                 len = iov[idx].iov_len;
 422                 nw = len >> 2;
 423                 page = NULL;
 424
 425                 pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
 426                 if (!pkt) {
 427                         ret = -ENOMEM;
 428                         goto free_list;
 429                 }
 430
 431                 if (len < IPATH_USER_SDMA_MIN_HEADER_LENGTH ||
 432                     len > PAGE_SIZE || len & 3 || addr & 3) {
 433                         ret = -EINVAL;
 434                         goto free_pkt;
 435                 }
 436
 437                 if (len == IPATH_USER_SDMA_EXP_HEADER_LENGTH)
 438                         pbc = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
 439                                              &dma_addr);
 440                 else
 441                         pbc = NULL;
 442
 443                 if (!pbc) {
 444                         page = alloc_page(GFP_KERNEL);
 445                         if (!page) {
 446                                 ret = -ENOMEM;
 447                                 goto free_pkt;
 448                         }
 449                         pbc = kmap(page);
 450                 }
 451
 452                 cfur = copy_from_user(pbc, iov[idx].iov_base, len);
 453                 if (cfur) {
 454                         ret = -EFAULT;
 455                         goto free_pbc;
 456                 }
 457
 458                 /*
 459                  * this assignment is a bit strange.  it's because the
 460                  * the pbc counts the number of 32 bit words in the full
 461                  * packet _except_ the first word of the pbc itself...
 462                  */
 463                 pktnwc = nw - 1;
 464
 465                 /*
 466                  * pktnw computation yields the number of 32 bit words
 467                  * that the caller has indicated in the PBC.  note that
 468                  * this is one less than the total number of words that
 469                  * goes to the send DMA engine as the first 32 bit word
 470                  * of the PBC itself is not counted.  Armed with this count,
 471                  * we can verify that the packet is consistent with the
 472                  * iovec lengths.
 473                  */
 474                 pktnw = le32_to_cpu(*pbc) & IPATH_PBC_LENGTH_MASK;
 475                 if (pktnw < pktnwc || pktnw > pktnwc + (PAGE_SIZE >> 2)) {
 476                         ret = -EINVAL;
 477                         goto free_pbc;
 478                 }
 479
 480
 481                 idx++;
 482                 while (pktnwc < pktnw && idx < niov) {
 483                         const size_t slen = iov[idx].iov_len;
 484                         const unsigned long faddr =
 485                                 (unsigned long) iov[idx].iov_base;
 486
 487                         if (slen & 3 || faddr & 3 || !slen ||
 488                             slen > PAGE_SIZE) {
 489                                 ret = -EINVAL;
 490                                 goto free_pbc;
 491                         }
 492
 493                         npages++;
 494                         if ((faddr & PAGE_MASK) !=
 495                             ((faddr + slen - 1) & PAGE_MASK))
 496                                 npages++;
 497
 498                         pktnwc += slen >> 2;
 499                         idx++;
 500                         nfrags++;
 501                 }
 502
 503                 if (pktnwc != pktnw) {
 504                         ret = -EINVAL;
 505                         goto free_pbc;
 506                 }
 507
 508                 if (page) {
 509                         dma_addr = dma_map_page(&dd->pcidev->dev,
 510                                                 page, 0, len, DMA_TO_DEVICE);
 511                         if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
 512                                 ret = -ENOMEM;
 513                                 goto free_pbc;
 514                         }
 515
 516                         dma_mapped = 1;
 517                 }
 518
 519                 ipath_user_sdma_init_header(pkt, counter, 0, len, dma_mapped,
 520                                             page, pbc, dma_addr);
 521
 522                 if (nfrags) {
 523                         ret = ipath_user_sdma_init_payload(dd, pq, pkt,
 524                                                            iov + idx_save + 1,
 525                                                            nfrags, npages);
 526                         if (ret < 0)
 527                                 goto free_pbc_dma;
 528                 }
 529
 530                 counter++;
 531                 npkts++;
 532
 533                 list_add_tail(&pkt->list, list);
 534         }
 535
 536         ret = idx;
 537         goto done;
 538
 539 free_pbc_dma:
 540         if (dma_mapped)
 541                 dma_unmap_page(&dd->pcidev->dev, dma_addr, len, DMA_TO_DEVICE);
 542 free_pbc:
 543         if (page) {
 544                 kunmap(page);
 545                 __free_page(page);
 546         } else
 547                 dma_pool_free(pq->header_cache, pbc, dma_addr);
 548 free_pkt:
 549         kmem_cache_free(pq->pkt_slab, pkt);
 550 free_list:
 551         ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
 552 done:
 553         return ret;
 554 }
 555
 556 static void ipath_user_sdma_set_complete_counter(struct ipath_user_sdma_queue *pq,
 557                                                  u32 c)
 558 {
 559         pq->sent_counter = c;
 560 }
 561
 562 /* try to clean out queue -- needs pq->lock */
 563 static int ipath_user_sdma_queue_clean(const struct ipath_devdata *dd,
 564                                        struct ipath_user_sdma_queue *pq)
 565 {
 566         struct list_head free_list;
 567         struct ipath_user_sdma_pkt *pkt;
 568         struct ipath_user_sdma_pkt *pkt_prev;
 569         int ret = 0;
 570
 571         INIT_LIST_HEAD(&free_list);
 572
 573         list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
 574                 s64 descd = dd->ipath_sdma_descq_removed - pkt->added;
 575
 576                 if (descd < 0)
 577                         break;
 578
 579                 list_move_tail(&pkt->list, &free_list);
 580
 581                 /* one more packet cleaned */
 582                 ret++;
 583         }
 584
 585         if (!list_empty(&free_list)) {
 586                 u32 counter;
 587
 588                 pkt = list_entry(free_list.prev,
 589                                  struct ipath_user_sdma_pkt, list);
 590                 counter = pkt->counter;
 591
 592                 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
 593                 ipath_user_sdma_set_complete_counter(pq, counter);
 594         }
 595
 596         return ret;
 597 }
 598
 599 void ipath_user_sdma_queue_destroy(struct ipath_user_sdma_queue *pq)
 600 {
 601         if (!pq)
 602                 return;
 603
 604         kmem_cache_destroy(pq->pkt_slab);
 605         dma_pool_destroy(pq->header_cache);
 606         kfree(pq);
 607 }
 608
 609 /* clean descriptor queue, returns > 0 if some elements cleaned */
 610 static int ipath_user_sdma_hwqueue_clean(struct ipath_devdata *dd)
 611 {
 612         int ret;
 613         unsigned long flags;
 614
 615         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 616         ret = ipath_sdma_make_progress(dd);
 617         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 618
 619         return ret;
 620 }
 621
 622 /* we're in close, drain packets so that we can cleanup successfully... */
 623 void ipath_user_sdma_queue_drain(struct ipath_devdata *dd,
 624                                  struct ipath_user_sdma_queue *pq)
 625 {
 626         int i;
 627
 628         if (!pq)
 629                 return;
 630
 631         for (i = 0; i < 100; i++) {
 632                 mutex_lock(&pq->lock);
 633                 if (list_empty(&pq->sent)) {
 634                         mutex_unlock(&pq->lock);
 635                         break;
 636                 }
 637                 ipath_user_sdma_hwqueue_clean(dd);
 638                 ipath_user_sdma_queue_clean(dd, pq);
 639                 mutex_unlock(&pq->lock);
 640                 msleep(10);
 641         }
 642
 643         if (!list_empty(&pq->sent)) {
 644                 struct list_head free_list;
 645
 646                 printk(KERN_INFO "drain: lists not empty: forcing!\n");
 647                 INIT_LIST_HEAD(&free_list);
 648                 mutex_lock(&pq->lock);
 649                 list_splice_init(&pq->sent, &free_list);
 650                 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
 651                 mutex_unlock(&pq->lock);
 652         }
 653 }
 654
 655 static inline __le64 ipath_sdma_make_desc0(struct ipath_devdata *dd,
 656                                            u64 addr, u64 dwlen, u64 dwoffset)
 657 {
 658         return cpu_to_le64(/* SDmaPhyAddr[31:0] */
 659                            ((addr & 0xfffffffcULL) << 32) |
 660                            /* SDmaGeneration[1:0] */
 661                            ((dd->ipath_sdma_generation & 3ULL) << 30) |
 662                            /* SDmaDwordCount[10:0] */
 663                            ((dwlen & 0x7ffULL) << 16) |
 664                            /* SDmaBufOffset[12:2] */
 665                            (dwoffset & 0x7ffULL));
 666 }
 667
 668 static inline __le64 ipath_sdma_make_first_desc0(__le64 descq)
 669 {
 670         return descq | cpu_to_le64(1ULL << 12);
 671 }
 672
 673 static inline __le64 ipath_sdma_make_last_desc0(__le64 descq)
 674 {
 675                                               /* last */  /* dma head */
 676         return descq | cpu_to_le64(1ULL << 11 | 1ULL << 13);
 677 }
 678
 679 static inline __le64 ipath_sdma_make_desc1(u64 addr)
 680 {
 681         /* SDmaPhyAddr[47:32] */
 682         return cpu_to_le64(addr >> 32);
 683 }
 684
 685 static void ipath_user_sdma_send_frag(struct ipath_devdata *dd,
 686                                       struct ipath_user_sdma_pkt *pkt, int idx,
 687                                       unsigned ofs, u16 tail)
 688 {
 689         const u64 addr = (u64) pkt->addr[idx].addr +
 690                 (u64) pkt->addr[idx].offset;
 691         const u64 dwlen = (u64) pkt->addr[idx].length / 4;
 692         __le64 *descqp;
 693         __le64 descq0;
 694
 695         descqp = &dd->ipath_sdma_descq[tail].qw[0];
 696
 697         descq0 = ipath_sdma_make_desc0(dd, addr, dwlen, ofs);
 698         if (idx == 0)
 699                 descq0 = ipath_sdma_make_first_desc0(descq0);
 700         if (idx == pkt->naddr - 1)
 701                 descq0 = ipath_sdma_make_last_desc0(descq0);
 702
 703         descqp[0] = descq0;
 704         descqp[1] = ipath_sdma_make_desc1(addr);
 705 }
 706
 707 /* pq->lock must be held, get packets on the wire... */
 708 static int ipath_user_sdma_push_pkts(struct ipath_devdata *dd,
 709                                      struct ipath_user_sdma_queue *pq,
 710                                      struct list_head *pktlist)
 711 {
 712         int ret = 0;
 713         unsigned long flags;
 714         u16 tail;
 715
 716         if (list_empty(pktlist))
 717                 return 0;
 718
 719         if (unlikely(!(dd->ipath_flags & IPATH_LINKACTIVE)))
 720                 return -ECOMM;
 721
 722         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 723
 724         if (unlikely(dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK)) {
 725                 ret = -ECOMM;
 726                 goto unlock;
 727         }
 728
 729         tail = dd->ipath_sdma_descq_tail;
 730         while (!list_empty(pktlist)) {
 731                 struct ipath_user_sdma_pkt *pkt =
 732                         list_entry(pktlist->next, struct ipath_user_sdma_pkt,
 733                                    list);
 734                 int i;
 735                 unsigned ofs = 0;
 736                 u16 dtail = tail;
 737
 738                 if (pkt->naddr > ipath_sdma_descq_freecnt(dd))
 739                         goto unlock_check_tail;
 740
 741                 for (i = 0; i < pkt->naddr; i++) {
 742                         ipath_user_sdma_send_frag(dd, pkt, i, ofs, tail);
 743                         ofs += pkt->addr[i].length >> 2;
 744
 745                         if (++tail == dd->ipath_sdma_descq_cnt) {
 746                                 tail = 0;
 747                                 ++dd->ipath_sdma_generation;
 748                         }
 749                 }
 750
 751                 if ((ofs<<2) > dd->ipath_ibmaxlen) {
 752                         ipath_dbg("packet size %X > ibmax %X, fail\n",
 753                                 ofs<<2, dd->ipath_ibmaxlen);
 754                         ret = -EMSGSIZE;
 755                         goto unlock;
 756                 }
 757
 758                 /*
 759                  * if the packet is >= 2KB mtu equivalent, we have to use
 760                  * the large buffers, and have to mark each descriptor as
 761                  * part of a large buffer packet.
 762                  */
 763                 if (ofs >= IPATH_SMALLBUF_DWORDS) {
 764                         for (i = 0; i < pkt->naddr; i++) {
 765                                 dd->ipath_sdma_descq[dtail].qw[0] |=
 766                                         cpu_to_le64(1ULL << 14);
 767                                 if (++dtail == dd->ipath_sdma_descq_cnt)
 768                                         dtail = 0;
 769                         }
 770                 }
 771
 772                 dd->ipath_sdma_descq_added += pkt->naddr;
 773                 pkt->added = dd->ipath_sdma_descq_added;
 774                 list_move_tail(&pkt->list, &pq->sent);
 775                 ret++;
 776         }
 777
 778 unlock_check_tail:
 779         /* advance the tail on the chip if necessary */
 780         if (dd->ipath_sdma_descq_tail != tail) {
 781                 wmb();
 782                 ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
 783                 dd->ipath_sdma_descq_tail = tail;
 784         }
 785
 786 unlock:
 787         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 788
 789         return ret;
 790 }
 791
 792 int ipath_user_sdma_writev(struct ipath_devdata *dd,
 793                            struct ipath_user_sdma_queue *pq,
 794                            const struct iovec *iov,
 795                            unsigned long dim)
 796 {
 797         int ret = 0;
 798         struct list_head list;
 799         int npkts = 0;
 800
 801         INIT_LIST_HEAD(&list);
 802
 803         mutex_lock(&pq->lock);
 804
 805         if (dd->ipath_sdma_descq_added != dd->ipath_sdma_descq_removed) {
 806                 ipath_user_sdma_hwqueue_clean(dd);
 807                 ipath_user_sdma_queue_clean(dd, pq);
 808         }
 809
 810         while (dim) {
 811                 const int mxp = 8;
 812
 813                 down_write(&current->mm->mmap_sem);
 814                 ret = ipath_user_sdma_queue_pkts(dd, pq, &list, iov, dim, mxp);
 815                 up_write(&current->mm->mmap_sem);
 816
 817                 if (ret <= 0)
 818                         goto done_unlock;
 819                 else {
 820                         dim -= ret;
 821                         iov += ret;
 822                 }
 823
 824                 /* force packets onto the sdma hw queue... */
 825                 if (!list_empty(&list)) {
 826                         /*
 827                          * lazily clean hw queue.  the 4 is a guess of about
 828                          * how many sdma descriptors a packet will take (it
 829                          * doesn't have to be perfect).
 830                          */
 831                         if (ipath_sdma_descq_freecnt(dd) < ret * 4) {
 832                                 ipath_user_sdma_hwqueue_clean(dd);
 833                                 ipath_user_sdma_queue_clean(dd, pq);
 834                         }
 835
 836                         ret = ipath_user_sdma_push_pkts(dd, pq, &list);
 837                         if (ret < 0)
 838                                 goto done_unlock;
 839                         else {
 840                                 npkts += ret;
 841                                 pq->counter += ret;
 842
 843                                 if (!list_empty(&list))
 844                                         goto done_unlock;
 845                         }
 846                 }
 847         }
 848
 849 done_unlock:
 850         if (!list_empty(&list))
 851                 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &list);
 852         mutex_unlock(&pq->lock);
 853
 854         return (ret < 0) ? ret : npkts;
 855 }
 856
 857 int ipath_user_sdma_make_progress(struct ipath_devdata *dd,
 858                                   struct ipath_user_sdma_queue *pq)
 859 {
 860         int ret = 0;
 861
 862         mutex_lock(&pq->lock);
 863         ipath_user_sdma_hwqueue_clean(dd);
 864         ret = ipath_user_sdma_queue_clean(dd, pq);
 865         mutex_unlock(&pq->lock);
 866
 867         return ret;
 868 }
 869
 870 u32 ipath_user_sdma_complete_counter(const struct ipath_user_sdma_queue *pq)
 871 {
 872         return pq->sent_counter;
 873 }
 874
 875 u32 ipath_user_sdma_inflight_counter(struct ipath_user_sdma_queue *pq)
 876 {
 877         return pq->counter;
 878 }
 879