drivers/infiniband/hw/qib/qib_user_sdma.c

   1 /*
   2  * Copyright (c) 2007, 2008, 2009 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 "qib.h"
  46 #include "qib_user_sdma.h"
  47
  48 /* minimum size of header */
  49 #define QIB_USER_SDMA_MIN_HEADER_LENGTH 64
  50 /* expected size of headers (for dma_pool) */
  51 #define QIB_USER_SDMA_EXP_HEADER_LENGTH 64
  52 /* attempt to drain the queue for 5secs */
  53 #define QIB_USER_SDMA_DRAIN_TIMEOUT 500
  54
  55 /*
  56  * track how many times a process open this driver.
  57  */
  58 static struct rb_root qib_user_sdma_rb_root = RB_ROOT;
  59
  60 struct qib_user_sdma_rb_node {
  61         struct rb_node node;
  62         int refcount;
  63         pid_t pid;
  64 };
  65
  66 struct qib_user_sdma_pkt {
  67         struct list_head list;  /* list element */
  68
  69         u8  tiddma;             /* if this is NEW tid-sdma */
  70         u8  largepkt;           /* this is large pkt from kmalloc */
  71         u16 frag_size;          /* frag size used by PSM */
  72         u16 index;              /* last header index or push index */
  73         u16 naddr;              /* dimension of addr (1..3) ... */
  74         u16 addrlimit;          /* addr array size */
  75         u16 tidsmidx;           /* current tidsm index */
  76         u16 tidsmcount;         /* tidsm array item count */
  77         u16 payload_size;       /* payload size so far for header */
  78         u32 bytes_togo;         /* bytes for processing */
  79         u32 counter;            /* sdma pkts queued counter for this entry */
  80         struct qib_tid_session_member *tidsm;   /* tid session member array */
  81         struct qib_user_sdma_queue *pq; /* which pq this pkt belongs to */
  82         u64 added;              /* global descq number of entries */
  83
  84         struct {
  85                 u16 offset;                     /* offset for kvaddr, addr */
  86                 u16 length;                     /* length in page */
  87                 u16 first_desc;                 /* first desc */
  88                 u16 last_desc;                  /* last desc */
  89                 u16 put_page;                   /* should we put_page? */
  90                 u16 dma_mapped;                 /* is page dma_mapped? */
  91                 u16 dma_length;                 /* for dma_unmap_page() */
  92                 u16 padding;
  93                 struct page *page;              /* may be NULL (coherent mem) */
  94                 void *kvaddr;                   /* FIXME: only for pio hack */
  95                 dma_addr_t addr;
  96         } addr[4];   /* max pages, any more and we coalesce */
  97 };
  98
  99 struct qib_user_sdma_queue {
 100         /*
 101          * pkts sent to dma engine are queued on this
 102          * list head.  the type of the elements of this
 103          * list are struct qib_user_sdma_pkt...
 104          */
 105         struct list_head sent;
 106
 107         /*
 108          * Because above list will be accessed by both process and
 109          * signal handler, we need a spinlock for it.
 110          */
 111         spinlock_t sent_lock ____cacheline_aligned_in_smp;
 112
 113         /* headers with expected length are allocated from here... */
 114         char header_cache_name[64];
 115         struct dma_pool *header_cache;
 116
 117         /* packets are allocated from the slab cache... */
 118         char pkt_slab_name[64];
 119         struct kmem_cache *pkt_slab;
 120
 121         /* as packets go on the queued queue, they are counted... */
 122         u32 counter;
 123         u32 sent_counter;
 124         /* pending packets, not sending yet */
 125         u32 num_pending;
 126         /* sending packets, not complete yet */
 127         u32 num_sending;
 128         /* global descq number of entry of last sending packet */
 129         u64 added;
 130
 131         /* dma page table */
 132         struct rb_root dma_pages_root;
 133
 134         struct qib_user_sdma_rb_node *sdma_rb_node;
 135
 136         /* protect everything above... */
 137         struct mutex lock;
 138 };
 139
 140 static struct qib_user_sdma_rb_node *
 141 qib_user_sdma_rb_search(struct rb_root *root, pid_t pid)
 142 {
 143         struct qib_user_sdma_rb_node *sdma_rb_node;
 144         struct rb_node *node = root->rb_node;
 145
 146         while (node) {
 147                 sdma_rb_node = container_of(node,
 148                         struct qib_user_sdma_rb_node, node);
 149                 if (pid < sdma_rb_node->pid)
 150                         node = node->rb_left;
 151                 else if (pid > sdma_rb_node->pid)
 152                         node = node->rb_right;
 153                 else
 154                         return sdma_rb_node;
 155         }
 156         return NULL;
 157 }
 158
 159 static int
 160 qib_user_sdma_rb_insert(struct rb_root *root, struct qib_user_sdma_rb_node *new)
 161 {
 162         struct rb_node **node = &(root->rb_node);
 163         struct rb_node *parent = NULL;
 164         struct qib_user_sdma_rb_node *got;
 165
 166         while (*node) {
 167                 got = container_of(*node, struct qib_user_sdma_rb_node, node);
 168                 parent = *node;
 169                 if (new->pid < got->pid)
 170                         node = &((*node)->rb_left);
 171                 else if (new->pid > got->pid)
 172                         node = &((*node)->rb_right);
 173                 else
 174                         return 0;
 175         }
 176
 177         rb_link_node(&new->node, parent, node);
 178         rb_insert_color(&new->node, root);
 179         return 1;
 180 }
 181
 182 struct qib_user_sdma_queue *
 183 qib_user_sdma_queue_create(struct device *dev, int unit, int ctxt, int sctxt)
 184 {
 185         struct qib_user_sdma_queue *pq =
 186                 kmalloc(sizeof(struct qib_user_sdma_queue), GFP_KERNEL);
 187         struct qib_user_sdma_rb_node *sdma_rb_node;
 188
 189         if (!pq)
 190                 goto done;
 191
 192         pq->counter = 0;
 193         pq->sent_counter = 0;
 194         pq->num_pending = 0;
 195         pq->num_sending = 0;
 196         pq->added = 0;
 197         pq->sdma_rb_node = NULL;
 198
 199         INIT_LIST_HEAD(&pq->sent);
 200         spin_lock_init(&pq->sent_lock);
 201         mutex_init(&pq->lock);
 202
 203         snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
 204                  "qib-user-sdma-pkts-%u-%02u.%02u", unit, ctxt, sctxt);
 205         pq->pkt_slab = kmem_cache_create(pq->pkt_slab_name,
 206                                          sizeof(struct qib_user_sdma_pkt),
 207                                          0, 0, NULL);
 208
 209         if (!pq->pkt_slab)
 210                 goto err_kfree;
 211
 212         snprintf(pq->header_cache_name, sizeof(pq->header_cache_name),
 213                  "qib-user-sdma-headers-%u-%02u.%02u", unit, ctxt, sctxt);
 214         pq->header_cache = dma_pool_create(pq->header_cache_name,
 215                                            dev,
 216                                            QIB_USER_SDMA_EXP_HEADER_LENGTH,
 217                                            4, 0);
 218         if (!pq->header_cache)
 219                 goto err_slab;
 220
 221         pq->dma_pages_root = RB_ROOT;
 222
 223         sdma_rb_node = qib_user_sdma_rb_search(&qib_user_sdma_rb_root,
 224                                         current->pid);
 225         if (sdma_rb_node) {
 226                 sdma_rb_node->refcount++;
 227         } else {
 228                 int ret;
 229                 sdma_rb_node = kmalloc(sizeof(
 230                         struct qib_user_sdma_rb_node), GFP_KERNEL);
 231                 if (!sdma_rb_node)
 232                         goto err_rb;
 233
 234                 sdma_rb_node->refcount = 1;
 235                 sdma_rb_node->pid = current->pid;
 236
 237                 ret = qib_user_sdma_rb_insert(&qib_user_sdma_rb_root,
 238                                         sdma_rb_node);
 239                 BUG_ON(ret == 0);
 240         }
 241         pq->sdma_rb_node = sdma_rb_node;
 242
 243         goto done;
 244
 245 err_rb:
 246         dma_pool_destroy(pq->header_cache);
 247 err_slab:
 248         kmem_cache_destroy(pq->pkt_slab);
 249 err_kfree:
 250         kfree(pq);
 251         pq = NULL;
 252
 253 done:
 254         return pq;
 255 }
 256
 257 static void qib_user_sdma_init_frag(struct qib_user_sdma_pkt *pkt,
 258                                     int i, u16 offset, u16 len,
 259                                     u16 first_desc, u16 last_desc,
 260                                     u16 put_page, u16 dma_mapped,
 261                                     struct page *page, void *kvaddr,
 262                                     dma_addr_t dma_addr, u16 dma_length)
 263 {
 264         pkt->addr[i].offset = offset;
 265         pkt->addr[i].length = len;
 266         pkt->addr[i].first_desc = first_desc;
 267         pkt->addr[i].last_desc = last_desc;
 268         pkt->addr[i].put_page = put_page;
 269         pkt->addr[i].dma_mapped = dma_mapped;
 270         pkt->addr[i].page = page;
 271         pkt->addr[i].kvaddr = kvaddr;
 272         pkt->addr[i].addr = dma_addr;
 273         pkt->addr[i].dma_length = dma_length;
 274 }
 275
 276 static void *qib_user_sdma_alloc_header(struct qib_user_sdma_queue *pq,
 277                                 size_t len, dma_addr_t *dma_addr)
 278 {
 279         void *hdr;
 280
 281         if (len == QIB_USER_SDMA_EXP_HEADER_LENGTH)
 282                 hdr = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
 283                                              dma_addr);
 284         else
 285                 hdr = NULL;
 286
 287         if (!hdr) {
 288                 hdr = kmalloc(len, GFP_KERNEL);
 289                 if (!hdr)
 290                         return NULL;
 291
 292                 *dma_addr = 0;
 293         }
 294
 295         return hdr;
 296 }
 297
 298 static int qib_user_sdma_page_to_frags(const struct qib_devdata *dd,
 299                                        struct qib_user_sdma_queue *pq,
 300                                        struct qib_user_sdma_pkt *pkt,
 301                                        struct page *page, u16 put,
 302                                        u16 offset, u16 len, void *kvaddr)
 303 {
 304         __le16 *pbc16;
 305         void *pbcvaddr;
 306         struct qib_message_header *hdr;
 307         u16 newlen, pbclen, lastdesc, dma_mapped;
 308         u32 vcto;
 309         union qib_seqnum seqnum;
 310         dma_addr_t pbcdaddr;
 311         dma_addr_t dma_addr =
 312                 dma_map_page(&dd->pcidev->dev,
 313                         page, offset, len, DMA_TO_DEVICE);
 314         int ret = 0;
 315
 316         if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
 317                 /*
 318                  * dma mapping error, pkt has not managed
 319                  * this page yet, return the page here so
 320                  * the caller can ignore this page.
 321                  */
 322                 if (put) {
 323                         put_page(page);
 324                 } else {
 325                         /* coalesce case */
 326                         kunmap(page);
 327                         __free_page(page);
 328                 }
 329                 ret = -ENOMEM;
 330                 goto done;
 331         }
 332         offset = 0;
 333         dma_mapped = 1;
 334
 335
 336 next_fragment:
 337
 338         /*
 339          * In tid-sdma, the transfer length is restricted by
 340          * receiver side current tid page length.
 341          */
 342         if (pkt->tiddma && len > pkt->tidsm[pkt->tidsmidx].length)
 343                 newlen = pkt->tidsm[pkt->tidsmidx].length;
 344         else
 345                 newlen = len;
 346
 347         /*
 348          * Then the transfer length is restricted by MTU.
 349          * the last descriptor flag is determined by:
 350          * 1. the current packet is at frag size length.
 351          * 2. the current tid page is done if tid-sdma.
 352          * 3. there is no more byte togo if sdma.
 353          */
 354         lastdesc = 0;
 355         if ((pkt->payload_size + newlen) >= pkt->frag_size) {
 356                 newlen = pkt->frag_size - pkt->payload_size;
 357                 lastdesc = 1;
 358         } else if (pkt->tiddma) {
 359                 if (newlen == pkt->tidsm[pkt->tidsmidx].length)
 360                         lastdesc = 1;
 361         } else {
 362                 if (newlen == pkt->bytes_togo)
 363                         lastdesc = 1;
 364         }
 365
 366         /* fill the next fragment in this page */
 367         qib_user_sdma_init_frag(pkt, pkt->naddr, /* index */
 368                 offset, newlen,         /* offset, len */
 369                 0, lastdesc,            /* first last desc */
 370                 put, dma_mapped,        /* put page, dma mapped */
 371                 page, kvaddr,           /* struct page, virt addr */
 372                 dma_addr, len);         /* dma addr, dma length */
 373         pkt->bytes_togo -= newlen;
 374         pkt->payload_size += newlen;
 375         pkt->naddr++;
 376         if (pkt->naddr == pkt->addrlimit) {
 377                 ret = -EFAULT;
 378                 goto done;
 379         }
 380
 381         /* If there is no more byte togo. (lastdesc==1) */
 382         if (pkt->bytes_togo == 0) {
 383                 /* The packet is done, header is not dma mapped yet.
 384                  * it should be from kmalloc */
 385                 if (!pkt->addr[pkt->index].addr) {
 386                         pkt->addr[pkt->index].addr =
 387                                 dma_map_single(&dd->pcidev->dev,
 388                                         pkt->addr[pkt->index].kvaddr,
 389                                         pkt->addr[pkt->index].dma_length,
 390                                         DMA_TO_DEVICE);
 391                         if (dma_mapping_error(&dd->pcidev->dev,
 392                                         pkt->addr[pkt->index].addr)) {
 393                                 ret = -ENOMEM;
 394                                 goto done;
 395                         }
 396                         pkt->addr[pkt->index].dma_mapped = 1;
 397                 }
 398
 399                 goto done;
 400         }
 401
 402         /* If tid-sdma, advance tid info. */
 403         if (pkt->tiddma) {
 404                 pkt->tidsm[pkt->tidsmidx].length -= newlen;
 405                 if (pkt->tidsm[pkt->tidsmidx].length) {
 406                         pkt->tidsm[pkt->tidsmidx].offset += newlen;
 407                 } else {
 408                         pkt->tidsmidx++;
 409                         if (pkt->tidsmidx == pkt->tidsmcount) {
 410                                 ret = -EFAULT;
 411                                 goto done;
 412                         }
 413                 }
 414         }
 415
 416         /*
 417          * If this is NOT the last descriptor. (newlen==len)
 418          * the current packet is not done yet, but the current
 419          * send side page is done.
 420          */
 421         if (lastdesc == 0)
 422                 goto done;
 423
 424         /*
 425          * If running this driver under PSM with message size
 426          * fitting into one transfer unit, it is not possible
 427          * to pass this line. otherwise, it is a buggggg.
 428          */
 429
 430         /*
 431          * Since the current packet is done, and there are more
 432          * bytes togo, we need to create a new sdma header, copying
 433          * from previous sdma header and modify both.
 434          */
 435         pbclen = pkt->addr[pkt->index].length;
 436         pbcvaddr = qib_user_sdma_alloc_header(pq, pbclen, &pbcdaddr);
 437         if (!pbcvaddr) {
 438                 ret = -ENOMEM;
 439                 goto done;
 440         }
 441         /* Copy the previous sdma header to new sdma header */
 442         pbc16 = (__le16 *)pkt->addr[pkt->index].kvaddr;
 443         memcpy(pbcvaddr, pbc16, pbclen);
 444
 445         /* Modify the previous sdma header */
 446         hdr = (struct qib_message_header *)&pbc16[4];
 447
 448         /* New pbc length */
 449         pbc16[0] = cpu_to_le16(le16_to_cpu(pbc16[0])-(pkt->bytes_togo>>2));
 450
 451         /* New packet length */
 452         hdr->lrh[2] = cpu_to_be16(le16_to_cpu(pbc16[0]));
 453
 454         if (pkt->tiddma) {
 455                 /* turn on the header suppression */
 456                 hdr->iph.pkt_flags =
 457                         cpu_to_le16(le16_to_cpu(hdr->iph.pkt_flags)|0x2);
 458                 /* turn off ACK_REQ: 0x04 and EXPECTED_DONE: 0x20 */
 459                 hdr->flags &= ~(0x04|0x20);
 460         } else {
 461                 /* turn off extra bytes: 20-21 bits */
 462                 hdr->bth[0] = cpu_to_be32(be32_to_cpu(hdr->bth[0])&0xFFCFFFFF);
 463                 /* turn off ACK_REQ: 0x04 */
 464                 hdr->flags &= ~(0x04);
 465         }
 466
 467         /* New kdeth checksum */
 468         vcto = le32_to_cpu(hdr->iph.ver_ctxt_tid_offset);
 469         hdr->iph.chksum = cpu_to_le16(QIB_LRH_BTH +
 470                 be16_to_cpu(hdr->lrh[2]) -
 471                 ((vcto>>16)&0xFFFF) - (vcto&0xFFFF) -
 472                 le16_to_cpu(hdr->iph.pkt_flags));
 473
 474         /* The packet is done, header is not dma mapped yet.
 475          * it should be from kmalloc */
 476         if (!pkt->addr[pkt->index].addr) {
 477                 pkt->addr[pkt->index].addr =
 478                         dma_map_single(&dd->pcidev->dev,
 479                                 pkt->addr[pkt->index].kvaddr,
 480                                 pkt->addr[pkt->index].dma_length,
 481                                 DMA_TO_DEVICE);
 482                 if (dma_mapping_error(&dd->pcidev->dev,
 483                                 pkt->addr[pkt->index].addr)) {
 484                         ret = -ENOMEM;
 485                         goto done;
 486                 }
 487                 pkt->addr[pkt->index].dma_mapped = 1;
 488         }
 489
 490         /* Modify the new sdma header */
 491         pbc16 = (__le16 *)pbcvaddr;
 492         hdr = (struct qib_message_header *)&pbc16[4];
 493
 494         /* New pbc length */
 495         pbc16[0] = cpu_to_le16(le16_to_cpu(pbc16[0])-(pkt->payload_size>>2));
 496
 497         /* New packet length */
 498         hdr->lrh[2] = cpu_to_be16(le16_to_cpu(pbc16[0]));
 499
 500         if (pkt->tiddma) {
 501                 /* Set new tid and offset for new sdma header */
 502                 hdr->iph.ver_ctxt_tid_offset = cpu_to_le32(
 503                         (le32_to_cpu(hdr->iph.ver_ctxt_tid_offset)&0xFF000000) +
 504                         (pkt->tidsm[pkt->tidsmidx].tid<<QLOGIC_IB_I_TID_SHIFT) +
 505                         (pkt->tidsm[pkt->tidsmidx].offset>>2));
 506         } else {
 507                 /* Middle protocol new packet offset */
 508                 hdr->uwords[2] += pkt->payload_size;
 509         }
 510
 511         /* New kdeth checksum */
 512         vcto = le32_to_cpu(hdr->iph.ver_ctxt_tid_offset);
 513         hdr->iph.chksum = cpu_to_le16(QIB_LRH_BTH +
 514                 be16_to_cpu(hdr->lrh[2]) -
 515                 ((vcto>>16)&0xFFFF) - (vcto&0xFFFF) -
 516                 le16_to_cpu(hdr->iph.pkt_flags));
 517
 518         /* Next sequence number in new sdma header */
 519         seqnum.val = be32_to_cpu(hdr->bth[2]);
 520         if (pkt->tiddma)
 521                 seqnum.seq++;
 522         else
 523                 seqnum.pkt++;
 524         hdr->bth[2] = cpu_to_be32(seqnum.val);
 525
 526         /* Init new sdma header. */
 527         qib_user_sdma_init_frag(pkt, pkt->naddr, /* index */
 528                 0, pbclen,              /* offset, len */
 529                 1, 0,                   /* first last desc */
 530                 0, 0,                   /* put page, dma mapped */
 531                 NULL, pbcvaddr,         /* struct page, virt addr */
 532                 pbcdaddr, pbclen);      /* dma addr, dma length */
 533         pkt->index = pkt->naddr;
 534         pkt->payload_size = 0;
 535         pkt->naddr++;
 536         if (pkt->naddr == pkt->addrlimit) {
 537                 ret = -EFAULT;
 538                 goto done;
 539         }
 540
 541         /* Prepare for next fragment in this page */
 542         if (newlen != len) {
 543                 if (dma_mapped) {
 544                         put = 0;
 545                         dma_mapped = 0;
 546                         page = NULL;
 547                         kvaddr = NULL;
 548                 }
 549                 len -= newlen;
 550                 offset += newlen;
 551
 552                 goto next_fragment;
 553         }
 554
 555 done:
 556         return ret;
 557 }
 558
 559 /* we've too many pages in the iovec, coalesce to a single page */
 560 static int qib_user_sdma_coalesce(const struct qib_devdata *dd,
 561                                   struct qib_user_sdma_queue *pq,
 562                                   struct qib_user_sdma_pkt *pkt,
 563                                   const struct iovec *iov,
 564                                   unsigned long niov)
 565 {
 566         int ret = 0;
 567         struct page *page = alloc_page(GFP_KERNEL);
 568         void *mpage_save;
 569         char *mpage;
 570         int i;
 571         int len = 0;
 572
 573         if (!page) {
 574                 ret = -ENOMEM;
 575                 goto done;
 576         }
 577
 578         mpage = kmap(page);
 579         mpage_save = mpage;
 580         for (i = 0; i < niov; i++) {
 581                 int cfur;
 582
 583                 cfur = copy_from_user(mpage,
 584                                       iov[i].iov_base, iov[i].iov_len);
 585                 if (cfur) {
 586                         ret = -EFAULT;
 587                         goto free_unmap;
 588                 }
 589
 590                 mpage += iov[i].iov_len;
 591                 len += iov[i].iov_len;
 592         }
 593
 594         ret = qib_user_sdma_page_to_frags(dd, pq, pkt,
 595                         page, 0, 0, len, mpage_save);
 596         goto done;
 597
 598 free_unmap:
 599         kunmap(page);
 600         __free_page(page);
 601 done:
 602         return ret;
 603 }
 604
 605 /*
 606  * How many pages in this iovec element?
 607  */
 608 static int qib_user_sdma_num_pages(const struct iovec *iov)
 609 {
 610         const unsigned long addr  = (unsigned long) iov->iov_base;
 611         const unsigned long  len  = iov->iov_len;
 612         const unsigned long spage = addr & PAGE_MASK;
 613         const unsigned long epage = (addr + len - 1) & PAGE_MASK;
 614
 615         return 1 + ((epage - spage) >> PAGE_SHIFT);
 616 }
 617
 618 static void qib_user_sdma_free_pkt_frag(struct device *dev,
 619                                         struct qib_user_sdma_queue *pq,
 620                                         struct qib_user_sdma_pkt *pkt,
 621                                         int frag)
 622 {
 623         const int i = frag;
 624
 625         if (pkt->addr[i].page) {
 626                 /* only user data has page */
 627                 if (pkt->addr[i].dma_mapped)
 628                         dma_unmap_page(dev,
 629                                        pkt->addr[i].addr,
 630                                        pkt->addr[i].dma_length,
 631                                        DMA_TO_DEVICE);
 632
 633                 if (pkt->addr[i].kvaddr)
 634                         kunmap(pkt->addr[i].page);
 635
 636                 if (pkt->addr[i].put_page)
 637                         put_page(pkt->addr[i].page);
 638                 else
 639                         __free_page(pkt->addr[i].page);
 640         } else if (pkt->addr[i].kvaddr) {
 641                 /* for headers */
 642                 if (pkt->addr[i].dma_mapped) {
 643                         /* from kmalloc & dma mapped */
 644                         dma_unmap_single(dev,
 645                                        pkt->addr[i].addr,
 646                                        pkt->addr[i].dma_length,
 647                                        DMA_TO_DEVICE);
 648                         kfree(pkt->addr[i].kvaddr);
 649                 } else if (pkt->addr[i].addr) {
 650                         /* free coherent mem from cache... */
 651                         dma_pool_free(pq->header_cache,
 652                               pkt->addr[i].kvaddr, pkt->addr[i].addr);
 653                 } else {
 654                         /* from kmalloc but not dma mapped */
 655                         kfree(pkt->addr[i].kvaddr);
 656                 }
 657         }
 658 }
 659
 660 /* return number of pages pinned... */
 661 static int qib_user_sdma_pin_pages(const struct qib_devdata *dd,
 662                                    struct qib_user_sdma_queue *pq,
 663                                    struct qib_user_sdma_pkt *pkt,
 664                                    unsigned long addr, int tlen, int npages)
 665 {
 666         struct page *pages[8];
 667         int i, j;
 668         int ret = 0;
 669
 670         while (npages) {
 671                 if (npages > 8)
 672                         j = 8;
 673                 else
 674                         j = npages;
 675
 676                 ret = get_user_pages_fast(addr, j, 0, pages);
 677                 if (ret != j) {
 678                         i = 0;
 679                         j = ret;
 680                         ret = -ENOMEM;
 681                         goto free_pages;
 682                 }
 683
 684                 for (i = 0; i < j; i++) {
 685                         /* map the pages... */
 686                         unsigned long fofs = addr & ~PAGE_MASK;
 687                         int flen = ((fofs + tlen) > PAGE_SIZE) ?
 688                                 (PAGE_SIZE - fofs) : tlen;
 689
 690                         ret = qib_user_sdma_page_to_frags(dd, pq, pkt,
 691                                 pages[i], 1, fofs, flen, NULL);
 692                         if (ret < 0) {
 693                                 /* current page has beed taken
 694                                  * care of inside above call.
 695                                  */
 696                                 i++;
 697                                 goto free_pages;
 698                         }
 699
 700                         addr += flen;
 701                         tlen -= flen;
 702                 }
 703
 704                 npages -= j;
 705         }
 706
 707         goto done;
 708
 709         /* if error, return all pages not managed by pkt */
 710 free_pages:
 711         while (i < j)
 712                 put_page(pages[i++]);
 713
 714 done:
 715         return ret;
 716 }
 717
 718 static int qib_user_sdma_pin_pkt(const struct qib_devdata *dd,
 719                                  struct qib_user_sdma_queue *pq,
 720                                  struct qib_user_sdma_pkt *pkt,
 721                                  const struct iovec *iov,
 722                                  unsigned long niov)
 723 {
 724         int ret = 0;
 725         unsigned long idx;
 726
 727         for (idx = 0; idx < niov; idx++) {
 728                 const int npages = qib_user_sdma_num_pages(iov + idx);
 729                 const unsigned long addr = (unsigned long) iov[idx].iov_base;
 730
 731                 ret = qib_user_sdma_pin_pages(dd, pq, pkt, addr,
 732                                               iov[idx].iov_len, npages);
 733                 if (ret < 0)
 734                         goto free_pkt;
 735         }
 736
 737         goto done;
 738
 739 free_pkt:
 740         /* we need to ignore the first entry here */
 741         for (idx = 1; idx < pkt->naddr; idx++)
 742                 qib_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
 743
 744         /* need to dma unmap the first entry, this is to restore to
 745          * the original state so that caller can free the memory in
 746          * error condition. Caller does not know if dma mapped or not*/
 747         if (pkt->addr[0].dma_mapped) {
 748                 dma_unmap_single(&dd->pcidev->dev,
 749                        pkt->addr[0].addr,
 750                        pkt->addr[0].dma_length,
 751                        DMA_TO_DEVICE);
 752                 pkt->addr[0].addr = 0;
 753                 pkt->addr[0].dma_mapped = 0;
 754         }
 755
 756 done:
 757         return ret;
 758 }
 759
 760 static int qib_user_sdma_init_payload(const struct qib_devdata *dd,
 761                                       struct qib_user_sdma_queue *pq,
 762                                       struct qib_user_sdma_pkt *pkt,
 763                                       const struct iovec *iov,
 764                                       unsigned long niov, int npages)
 765 {
 766         int ret = 0;
 767
 768         if (pkt->frag_size == pkt->bytes_togo &&
 769                         npages >= ARRAY_SIZE(pkt->addr))
 770                 ret = qib_user_sdma_coalesce(dd, pq, pkt, iov, niov);
 771         else
 772                 ret = qib_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
 773
 774         return ret;
 775 }
 776
 777 /* free a packet list -- return counter value of last packet */
 778 static void qib_user_sdma_free_pkt_list(struct device *dev,
 779                                         struct qib_user_sdma_queue *pq,
 780                                         struct list_head *list)
 781 {
 782         struct qib_user_sdma_pkt *pkt, *pkt_next;
 783
 784         list_for_each_entry_safe(pkt, pkt_next, list, list) {
 785                 int i;
 786
 787                 for (i = 0; i < pkt->naddr; i++)
 788                         qib_user_sdma_free_pkt_frag(dev, pq, pkt, i);
 789
 790                 if (pkt->largepkt)
 791                         kfree(pkt);
 792                 else
 793                         kmem_cache_free(pq->pkt_slab, pkt);
 794         }
 795         INIT_LIST_HEAD(list);
 796 }
 797
 798 /*
 799  * copy headers, coalesce etc -- pq->lock must be held
 800  *
 801  * we queue all the packets to list, returning the
 802  * number of bytes total.  list must be empty initially,
 803  * as, if there is an error we clean it...
 804  */
 805 static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
 806                                     struct qib_pportdata *ppd,
 807                                     struct qib_user_sdma_queue *pq,
 808                                     const struct iovec *iov,
 809                                     unsigned long niov,
 810                                     struct list_head *list,
 811                                     int *maxpkts, int *ndesc)
 812 {
 813         unsigned long idx = 0;
 814         int ret = 0;
 815         int npkts = 0;
 816         __le32 *pbc;
 817         dma_addr_t dma_addr;
 818         struct qib_user_sdma_pkt *pkt = NULL;
 819         size_t len;
 820         size_t nw;
 821         u32 counter = pq->counter;
 822         u16 frag_size;
 823
 824         while (idx < niov && npkts < *maxpkts) {
 825                 const unsigned long addr = (unsigned long) iov[idx].iov_base;
 826                 const unsigned long idx_save = idx;
 827                 unsigned pktnw;
 828                 unsigned pktnwc;
 829                 int nfrags = 0;
 830                 int npages = 0;
 831                 int bytes_togo = 0;
 832                 int tiddma = 0;
 833                 int cfur;
 834
 835                 len = iov[idx].iov_len;
 836                 nw = len >> 2;
 837
 838                 if (len < QIB_USER_SDMA_MIN_HEADER_LENGTH ||
 839                     len > PAGE_SIZE || len & 3 || addr & 3) {
 840                         ret = -EINVAL;
 841                         goto free_list;
 842                 }
 843
 844                 pbc = qib_user_sdma_alloc_header(pq, len, &dma_addr);
 845                 if (!pbc) {
 846                         ret = -ENOMEM;
 847                         goto free_list;
 848                 }
 849
 850                 cfur = copy_from_user(pbc, iov[idx].iov_base, len);
 851                 if (cfur) {
 852                         ret = -EFAULT;
 853                         goto free_pbc;
 854                 }
 855
 856                 /*
 857                  * This assignment is a bit strange.  it's because the
 858                  * the pbc counts the number of 32 bit words in the full
 859                  * packet _except_ the first word of the pbc itself...
 860                  */
 861                 pktnwc = nw - 1;
 862
 863                 /*
 864                  * pktnw computation yields the number of 32 bit words
 865                  * that the caller has indicated in the PBC.  note that
 866                  * this is one less than the total number of words that
 867                  * goes to the send DMA engine as the first 32 bit word
 868                  * of the PBC itself is not counted.  Armed with this count,
 869                  * we can verify that the packet is consistent with the
 870                  * iovec lengths.
 871                  */
 872                 pktnw = le32_to_cpu(*pbc) & 0xFFFF;
 873                 if (pktnw < pktnwc) {
 874                         ret = -EINVAL;
 875                         goto free_pbc;
 876                 }
 877
 878                 idx++;
 879                 while (pktnwc < pktnw && idx < niov) {
 880                         const size_t slen = iov[idx].iov_len;
 881                         const unsigned long faddr =
 882                                 (unsigned long) iov[idx].iov_base;
 883
 884                         if (slen & 3 || faddr & 3 || !slen) {
 885                                 ret = -EINVAL;
 886                                 goto free_pbc;
 887                         }
 888
 889                         npages += qib_user_sdma_num_pages(&iov[idx]);
 890
 891                         bytes_togo += slen;
 892                         pktnwc += slen >> 2;
 893                         idx++;
 894                         nfrags++;
 895                 }
 896
 897                 if (pktnwc != pktnw) {
 898                         ret = -EINVAL;
 899                         goto free_pbc;
 900                 }
 901
 902                 frag_size = ((le32_to_cpu(*pbc))>>16) & 0xFFFF;
 903                 if (((frag_size ? frag_size : bytes_togo) + len) >
 904                                                 ppd->ibmaxlen) {
 905                         ret = -EINVAL;
 906                         goto free_pbc;
 907                 }
 908
 909                 if (frag_size) {
 910                         int pktsize, tidsmsize, n;
 911
 912                         n = npages*((2*PAGE_SIZE/frag_size)+1);
 913                         pktsize = sizeof(*pkt) + sizeof(pkt->addr[0])*n;
 914
 915                         /*
 916                          * Determine if this is tid-sdma or just sdma.
 917                          */
 918                         tiddma = (((le32_to_cpu(pbc[7])>>
 919                                 QLOGIC_IB_I_TID_SHIFT)&
 920                                 QLOGIC_IB_I_TID_MASK) !=
 921                                 QLOGIC_IB_I_TID_MASK);
 922
 923                         if (tiddma)
 924                                 tidsmsize = iov[idx].iov_len;
 925                         else
 926                                 tidsmsize = 0;
 927
 928                         pkt = kmalloc(pktsize+tidsmsize, GFP_KERNEL);
 929                         if (!pkt) {
 930                                 ret = -ENOMEM;
 931                                 goto free_pbc;
 932                         }
 933                         pkt->largepkt = 1;
 934                         pkt->frag_size = frag_size;
 935                         pkt->addrlimit = n + ARRAY_SIZE(pkt->addr);
 936
 937                         if (tiddma) {
 938                                 char *tidsm = (char *)pkt + pktsize;
 939                                 cfur = copy_from_user(tidsm,
 940                                         iov[idx].iov_base, tidsmsize);
 941                                 if (cfur) {
 942                                         ret = -EFAULT;
 943                                         goto free_pkt;
 944                                 }
 945                                 pkt->tidsm =
 946                                         (struct qib_tid_session_member *)tidsm;
 947                                 pkt->tidsmcount = tidsmsize/
 948                                         sizeof(struct qib_tid_session_member);
 949                                 pkt->tidsmidx = 0;
 950                                 idx++;
 951                         }
 952
 953                         /*
 954                          * pbc 'fill1' field is borrowed to pass frag size,
 955                          * we need to clear it after picking frag size, the
 956                          * hardware requires this field to be zero.
 957                          */
 958                         *pbc = cpu_to_le32(le32_to_cpu(*pbc) & 0x0000FFFF);
 959                 } else {
 960                         pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
 961                         if (!pkt) {
 962                                 ret = -ENOMEM;
 963                                 goto free_pbc;
 964                         }
 965                         pkt->largepkt = 0;
 966                         pkt->frag_size = bytes_togo;
 967                         pkt->addrlimit = ARRAY_SIZE(pkt->addr);
 968                 }
 969                 pkt->bytes_togo = bytes_togo;
 970                 pkt->payload_size = 0;
 971                 pkt->counter = counter;
 972                 pkt->tiddma = tiddma;
 973
 974                 /* setup the first header */
 975                 qib_user_sdma_init_frag(pkt, 0, /* index */
 976                         0, len,         /* offset, len */
 977                         1, 0,           /* first last desc */
 978                         0, 0,           /* put page, dma mapped */
 979                         NULL, pbc,      /* struct page, virt addr */
 980                         dma_addr, len); /* dma addr, dma length */
 981                 pkt->index = 0;
 982                 pkt->naddr = 1;
 983
 984                 if (nfrags) {
 985                         ret = qib_user_sdma_init_payload(dd, pq, pkt,
 986                                                          iov + idx_save + 1,
 987                                                          nfrags, npages);
 988                         if (ret < 0)
 989                                 goto free_pkt;
 990                 } else {
 991                         /* since there is no payload, mark the
 992                          * header as the last desc. */
 993                         pkt->addr[0].last_desc = 1;
 994
 995                         if (dma_addr == 0) {
 996                                 /*
 997                                  * the header is not dma mapped yet.
 998                                  * it should be from kmalloc.
 999                                  */
1000                                 dma_addr = dma_map_single(&dd->pcidev->dev,
1001                                         pbc, len, DMA_TO_DEVICE);
1002                                 if (dma_mapping_error(&dd->pcidev->dev,
1003                                                                 dma_addr)) {
1004                                         ret = -ENOMEM;
1005                                         goto free_pkt;
1006                                 }
1007                                 pkt->addr[0].addr = dma_addr;
1008                                 pkt->addr[0].dma_mapped = 1;
1009                         }
1010                 }
1011
1012                 counter++;
1013                 npkts++;
1014                 pkt->pq = pq;
1015                 pkt->index = 0; /* reset index for push on hw */
1016                 *ndesc += pkt->naddr;
1017
1018                 list_add_tail(&pkt->list, list);
1019         }
1020
1021         *maxpkts = npkts;
1022         ret = idx;
1023         goto done;
1024
1025 free_pkt:
1026         if (pkt->largepkt)
1027                 kfree(pkt);
1028         else
1029                 kmem_cache_free(pq->pkt_slab, pkt);
1030 free_pbc:
1031         if (dma_addr)
1032                 dma_pool_free(pq->header_cache, pbc, dma_addr);
1033         else
1034                 kfree(pbc);
1035 free_list:
1036         qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
1037 done:
1038         return ret;
1039 }
1040
1041 static void qib_user_sdma_set_complete_counter(struct qib_user_sdma_queue *pq,
1042                                                u32 c)
1043 {
1044         pq->sent_counter = c;
1045 }
1046
1047 /* try to clean out queue -- needs pq->lock */
1048 static int qib_user_sdma_queue_clean(struct qib_pportdata *ppd,
1049                                      struct qib_user_sdma_queue *pq)
1050 {
1051         struct qib_devdata *dd = ppd->dd;
1052         struct list_head free_list;
1053         struct qib_user_sdma_pkt *pkt;
1054         struct qib_user_sdma_pkt *pkt_prev;
1055         unsigned long flags;
1056         int ret = 0;
1057
1058         if (!pq->num_sending)
1059                 return 0;
1060
1061         INIT_LIST_HEAD(&free_list);
1062
1063         /*
1064          * We need this spin lock here because interrupt handler
1065          * might modify this list in qib_user_sdma_send_desc(), also
1066          * we can not get interrupted, otherwise it is a deadlock.
1067          */
1068         spin_lock_irqsave(&pq->sent_lock, flags);
1069         list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
1070                 s64 descd = ppd->sdma_descq_removed - pkt->added;
1071
1072                 if (descd < 0)
1073                         break;
1074
1075                 list_move_tail(&pkt->list, &free_list);
1076
1077                 /* one more packet cleaned */
1078                 ret++;
1079                 pq->num_sending--;
1080         }
1081         spin_unlock_irqrestore(&pq->sent_lock, flags);
1082
1083         if (!list_empty(&free_list)) {
1084                 u32 counter;
1085
1086                 pkt = list_entry(free_list.prev,
1087                                  struct qib_user_sdma_pkt, list);
1088                 counter = pkt->counter;
1089
1090                 qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
1091                 qib_user_sdma_set_complete_counter(pq, counter);
1092         }
1093
1094         return ret;
1095 }
1096
1097 void qib_user_sdma_queue_destroy(struct qib_user_sdma_queue *pq)
1098 {
1099         if (!pq)
1100                 return;
1101
1102         pq->sdma_rb_node->refcount--;
1103         if (pq->sdma_rb_node->refcount == 0) {
1104                 rb_erase(&pq->sdma_rb_node->node, &qib_user_sdma_rb_root);
1105                 kfree(pq->sdma_rb_node);
1106         }
1107         dma_pool_destroy(pq->header_cache);
1108         kmem_cache_destroy(pq->pkt_slab);
1109         kfree(pq);
1110 }
1111
1112 /* clean descriptor queue, returns > 0 if some elements cleaned */
1113 static int qib_user_sdma_hwqueue_clean(struct qib_pportdata *ppd)
1114 {
1115         int ret;
1116         unsigned long flags;
1117
1118         spin_lock_irqsave(&ppd->sdma_lock, flags);
1119         ret = qib_sdma_make_progress(ppd);
1120         spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1121
1122         return ret;
1123 }
1124
1125 /* we're in close, drain packets so that we can cleanup successfully... */
1126 void qib_user_sdma_queue_drain(struct qib_pportdata *ppd,
1127                                struct qib_user_sdma_queue *pq)
1128 {
1129         struct qib_devdata *dd = ppd->dd;
1130         unsigned long flags;
1131         int i;
1132
1133         if (!pq)
1134                 return;
1135
1136         for (i = 0; i < QIB_USER_SDMA_DRAIN_TIMEOUT; i++) {
1137                 mutex_lock(&pq->lock);
1138                 if (!pq->num_pending && !pq->num_sending) {
1139                         mutex_unlock(&pq->lock);
1140                         break;
1141                 }
1142                 qib_user_sdma_hwqueue_clean(ppd);
1143                 qib_user_sdma_queue_clean(ppd, pq);
1144                 mutex_unlock(&pq->lock);
1145                 msleep(10);
1146         }
1147
1148         if (pq->num_pending || pq->num_sending) {
1149                 struct qib_user_sdma_pkt *pkt;
1150                 struct qib_user_sdma_pkt *pkt_prev;
1151                 struct list_head free_list;
1152
1153                 mutex_lock(&pq->lock);
1154                 spin_lock_irqsave(&ppd->sdma_lock, flags);
1155                 /*
1156                  * Since we hold sdma_lock, it is safe without sent_lock.
1157                  */
1158                 if (pq->num_pending) {
1159                         list_for_each_entry_safe(pkt, pkt_prev,
1160                                         &ppd->sdma_userpending, list) {
1161                                 if (pkt->pq == pq) {
1162                                         list_move_tail(&pkt->list, &pq->sent);
1163                                         pq->num_pending--;
1164                                         pq->num_sending++;
1165                                 }
1166                         }
1167                 }
1168                 spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1169
1170                 qib_dev_err(dd, "user sdma lists not empty: forcing!\n");
1171                 INIT_LIST_HEAD(&free_list);
1172                 list_splice_init(&pq->sent, &free_list);
1173                 pq->num_sending = 0;
1174                 qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
1175                 mutex_unlock(&pq->lock);
1176         }
1177 }
1178
1179 static inline __le64 qib_sdma_make_desc0(u8 gen,
1180                                          u64 addr, u64 dwlen, u64 dwoffset)
1181 {
1182         return cpu_to_le64(/* SDmaPhyAddr[31:0] */
1183                            ((addr & 0xfffffffcULL) << 32) |
1184                            /* SDmaGeneration[1:0] */
1185                            ((gen & 3ULL) << 30) |
1186                            /* SDmaDwordCount[10:0] */
1187                            ((dwlen & 0x7ffULL) << 16) |
1188                            /* SDmaBufOffset[12:2] */
1189                            (dwoffset & 0x7ffULL));
1190 }
1191
1192 static inline __le64 qib_sdma_make_first_desc0(__le64 descq)
1193 {
1194         return descq | cpu_to_le64(1ULL << 12);
1195 }
1196
1197 static inline __le64 qib_sdma_make_last_desc0(__le64 descq)
1198 {
1199                                               /* last */  /* dma head */
1200         return descq | cpu_to_le64(1ULL << 11 | 1ULL << 13);
1201 }
1202
1203 static inline __le64 qib_sdma_make_desc1(u64 addr)
1204 {
1205         /* SDmaPhyAddr[47:32] */
1206         return cpu_to_le64(addr >> 32);
1207 }
1208
1209 static void qib_user_sdma_send_frag(struct qib_pportdata *ppd,
1210                                     struct qib_user_sdma_pkt *pkt, int idx,
1211                                     unsigned ofs, u16 tail, u8 gen)
1212 {
1213         const u64 addr = (u64) pkt->addr[idx].addr +
1214                 (u64) pkt->addr[idx].offset;
1215         const u64 dwlen = (u64) pkt->addr[idx].length / 4;
1216         __le64 *descqp;
1217         __le64 descq0;
1218
1219         descqp = &ppd->sdma_descq[tail].qw[0];
1220
1221         descq0 = qib_sdma_make_desc0(gen, addr, dwlen, ofs);
1222         if (pkt->addr[idx].first_desc)
1223                 descq0 = qib_sdma_make_first_desc0(descq0);
1224         if (pkt->addr[idx].last_desc) {
1225                 descq0 = qib_sdma_make_last_desc0(descq0);
1226                 if (ppd->sdma_intrequest) {
1227                         descq0 |= cpu_to_le64(1ULL << 15);
1228                         ppd->sdma_intrequest = 0;
1229                 }
1230         }
1231
1232         descqp[0] = descq0;
1233         descqp[1] = qib_sdma_make_desc1(addr);
1234 }
1235
1236 void qib_user_sdma_send_desc(struct qib_pportdata *ppd,
1237                                 struct list_head *pktlist)
1238 {
1239         struct qib_devdata *dd = ppd->dd;
1240         u16 nfree, nsent;
1241         u16 tail, tail_c;
1242         u8 gen, gen_c;
1243
1244         nfree = qib_sdma_descq_freecnt(ppd);
1245         if (!nfree)
1246                 return;
1247
1248 retry:
1249         nsent = 0;
1250         tail_c = tail = ppd->sdma_descq_tail;
1251         gen_c = gen = ppd->sdma_generation;
1252         while (!list_empty(pktlist)) {
1253                 struct qib_user_sdma_pkt *pkt =
1254                         list_entry(pktlist->next, struct qib_user_sdma_pkt,
1255                                    list);
1256                 int i, j, c = 0;
1257                 unsigned ofs = 0;
1258                 u16 dtail = tail;
1259
1260                 for (i = pkt->index; i < pkt->naddr && nfree; i++) {
1261                         qib_user_sdma_send_frag(ppd, pkt, i, ofs, tail, gen);
1262                         ofs += pkt->addr[i].length >> 2;
1263
1264                         if (++tail == ppd->sdma_descq_cnt) {
1265                                 tail = 0;
1266                                 ++gen;
1267                                 ppd->sdma_intrequest = 1;
1268                         } else if (tail == (ppd->sdma_descq_cnt>>1)) {
1269                                 ppd->sdma_intrequest = 1;
1270                         }
1271                         nfree--;
1272                         if (pkt->addr[i].last_desc == 0)
1273                                 continue;
1274
1275                         /*
1276                          * If the packet is >= 2KB mtu equivalent, we
1277                          * have to use the large buffers, and have to
1278                          * mark each descriptor as part of a large
1279                          * buffer packet.
1280                          */
1281                         if (ofs > dd->piosize2kmax_dwords) {
1282                                 for (j = pkt->index; j <= i; j++) {
1283                                         ppd->sdma_descq[dtail].qw[0] |=
1284                                                 cpu_to_le64(1ULL << 14);
1285                                         if (++dtail == ppd->sdma_descq_cnt)
1286                                                 dtail = 0;
1287                                 }
1288                         }
1289                         c += i + 1 - pkt->index;
1290                         pkt->index = i + 1; /* index for next first */
1291                         tail_c = dtail = tail;
1292                         gen_c = gen;
1293                         ofs = 0;  /* reset for next packet */
1294                 }
1295
1296                 ppd->sdma_descq_added += c;
1297                 nsent += c;
1298                 if (pkt->index == pkt->naddr) {
1299                         pkt->added = ppd->sdma_descq_added;
1300                         pkt->pq->added = pkt->added;
1301                         pkt->pq->num_pending--;
1302                         spin_lock(&pkt->pq->sent_lock);
1303                         pkt->pq->num_sending++;
1304                         list_move_tail(&pkt->list, &pkt->pq->sent);
1305                         spin_unlock(&pkt->pq->sent_lock);
1306                 }
1307                 if (!nfree || (nsent<<2) > ppd->sdma_descq_cnt)
1308                         break;
1309         }
1310
1311         /* advance the tail on the chip if necessary */
1312         if (ppd->sdma_descq_tail != tail_c) {
1313                 ppd->sdma_generation = gen_c;
1314                 dd->f_sdma_update_tail(ppd, tail_c);
1315         }
1316
1317         if (nfree && !list_empty(pktlist))
1318                 goto retry;
1319
1320         return;
1321 }
1322
1323 /* pq->lock must be held, get packets on the wire... */
1324 static int qib_user_sdma_push_pkts(struct qib_pportdata *ppd,
1325                                  struct qib_user_sdma_queue *pq,
1326                                  struct list_head *pktlist, int count)
1327 {
1328         unsigned long flags;
1329
1330         if (unlikely(!(ppd->lflags & QIBL_LINKACTIVE)))
1331                 return -ECOMM;
1332
1333         /* non-blocking mode */
1334         if (pq->sdma_rb_node->refcount > 1) {
1335                 spin_lock_irqsave(&ppd->sdma_lock, flags);
1336                 if (unlikely(!__qib_sdma_running(ppd))) {
1337                         spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1338                         return -ECOMM;
1339                 }
1340                 pq->num_pending += count;
1341                 list_splice_tail_init(pktlist, &ppd->sdma_userpending);
1342                 qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
1343                 spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1344                 return 0;
1345         }
1346
1347         /* In this case, descriptors from this process are not
1348          * linked to ppd pending queue, interrupt handler
1349          * won't update this process, it is OK to directly
1350          * modify without sdma lock.
1351          */
1352
1353
1354         pq->num_pending += count;
1355         /*
1356          * Blocking mode for single rail process, we must
1357          * release/regain sdma_lock to give other process
1358          * chance to make progress. This is important for
1359          * performance.
1360          */
1361         do {
1362                 spin_lock_irqsave(&ppd->sdma_lock, flags);
1363                 if (unlikely(!__qib_sdma_running(ppd))) {
1364                         spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1365                         return -ECOMM;
1366                 }
1367                 qib_user_sdma_send_desc(ppd, pktlist);
1368                 if (!list_empty(pktlist))
1369                         qib_sdma_make_progress(ppd);
1370                 spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1371         } while (!list_empty(pktlist));
1372
1373         return 0;
1374 }
1375
1376 int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
1377                          struct qib_user_sdma_queue *pq,
1378                          const struct iovec *iov,
1379                          unsigned long dim)
1380 {
1381         struct qib_devdata *dd = rcd->dd;
1382         struct qib_pportdata *ppd = rcd->ppd;
1383         int ret = 0;
1384         struct list_head list;
1385         int npkts = 0;
1386
1387         INIT_LIST_HEAD(&list);
1388
1389         mutex_lock(&pq->lock);
1390
1391         /* why not -ECOMM like qib_user_sdma_push_pkts() below? */
1392         if (!qib_sdma_running(ppd))
1393                 goto done_unlock;
1394
1395         /* if I have packets not complete yet */
1396         if (pq->added > ppd->sdma_descq_removed)
1397                 qib_user_sdma_hwqueue_clean(ppd);
1398         /* if I have complete packets to be freed */
1399         if (pq->num_sending)
1400                 qib_user_sdma_queue_clean(ppd, pq);
1401
1402         while (dim) {
1403                 int mxp = 1;
1404                 int ndesc = 0;
1405
1406                 ret = qib_user_sdma_queue_pkts(dd, ppd, pq,
1407                                 iov, dim, &list, &mxp, &ndesc);
1408                 if (ret < 0)
1409                         goto done_unlock;
1410                 else {
1411                         dim -= ret;
1412                         iov += ret;
1413                 }
1414
1415                 /* force packets onto the sdma hw queue... */
1416                 if (!list_empty(&list)) {
1417                         /*
1418                          * Lazily clean hw queue.
1419                          */
1420                         if (qib_sdma_descq_freecnt(ppd) < ndesc) {
1421                                 qib_user_sdma_hwqueue_clean(ppd);
1422                                 if (pq->num_sending)
1423                                         qib_user_sdma_queue_clean(ppd, pq);
1424                         }
1425
1426                         ret = qib_user_sdma_push_pkts(ppd, pq, &list, mxp);
1427                         if (ret < 0)
1428                                 goto done_unlock;
1429                         else {
1430                                 npkts += mxp;
1431                                 pq->counter += mxp;
1432                         }
1433                 }
1434         }
1435
1436 done_unlock:
1437         if (!list_empty(&list))
1438                 qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &list);
1439         mutex_unlock(&pq->lock);
1440
1441         return (ret < 0) ? ret : npkts;
1442 }
1443
1444 int qib_user_sdma_make_progress(struct qib_pportdata *ppd,
1445                                 struct qib_user_sdma_queue *pq)
1446 {
1447         int ret = 0;
1448
1449         mutex_lock(&pq->lock);
1450         qib_user_sdma_hwqueue_clean(ppd);
1451         ret = qib_user_sdma_queue_clean(ppd, pq);
1452         mutex_unlock(&pq->lock);
1453
1454         return ret;
1455 }
1456
1457 u32 qib_user_sdma_complete_counter(const struct qib_user_sdma_queue *pq)
1458 {
1459         return pq ? pq->sent_counter : 0;
1460 }
1461
1462 u32 qib_user_sdma_inflight_counter(struct qib_user_sdma_queue *pq)
1463 {
1464         return pq ? pq->counter : 0;
1465 }