drivers/infiniband/hw/cxgb4/device.c

   1 /*
   2  * Copyright (c) 2009-2010 Chelsio, Inc. 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/module.h>
  33 #include <linux/moduleparam.h>
  34 #include <linux/debugfs.h>
  35
  36 #include <rdma/ib_verbs.h>
  37
  38 #include "iw_cxgb4.h"
  39
  40 #define DRV_VERSION "0.1"
  41
  42 MODULE_AUTHOR("Steve Wise");
  43 MODULE_DESCRIPTION("Chelsio T4 RDMA Driver");
  44 MODULE_LICENSE("Dual BSD/GPL");
  45 MODULE_VERSION(DRV_VERSION);
  46
  47 struct uld_ctx {
  48         struct list_head entry;
  49         struct cxgb4_lld_info lldi;
  50         struct c4iw_dev *dev;
  51 };
  52
  53 static LIST_HEAD(uld_ctx_list);
  54 static DEFINE_MUTEX(dev_mutex);
  55
  56 static struct dentry *c4iw_debugfs_root;
  57
  58 struct c4iw_debugfs_data {
  59         struct c4iw_dev *devp;
  60         char *buf;
  61         int bufsize;
  62         int pos;
  63 };
  64
  65 static int count_idrs(int id, void *p, void *data)
  66 {
  67         int *countp = data;
  68
  69         *countp = *countp + 1;
  70         return 0;
  71 }
  72
  73 static ssize_t debugfs_read(struct file *file, char __user *buf, size_t count,
  74                             loff_t *ppos)
  75 {
  76         struct c4iw_debugfs_data *d = file->private_data;
  77
  78         return simple_read_from_buffer(buf, count, ppos, d->buf, d->pos);
  79 }
  80
  81 static int dump_qp(int id, void *p, void *data)
  82 {
  83         struct c4iw_qp *qp = p;
  84         struct c4iw_debugfs_data *qpd = data;
  85         int space;
  86         int cc;
  87
  88         if (id != qp->wq.sq.qid)
  89                 return 0;
  90
  91         space = qpd->bufsize - qpd->pos - 1;
  92         if (space == 0)
  93                 return 1;
  94
  95         if (qp->ep)
  96                 cc = snprintf(qpd->buf + qpd->pos, space,
  97                              "qp sq id %u rq id %u state %u onchip %u "
  98                              "ep tid %u state %u %pI4:%u->%pI4:%u\n",
  99                              qp->wq.sq.qid, qp->wq.rq.qid, (int)qp->attr.state,
 100                              qp->wq.sq.flags & T4_SQ_ONCHIP,
 101                              qp->ep->hwtid, (int)qp->ep->com.state,
 102                              &qp->ep->com.local_addr.sin_addr.s_addr,
 103                              ntohs(qp->ep->com.local_addr.sin_port),
 104                              &qp->ep->com.remote_addr.sin_addr.s_addr,
 105                              ntohs(qp->ep->com.remote_addr.sin_port));
 106         else
 107                 cc = snprintf(qpd->buf + qpd->pos, space,
 108                              "qp sq id %u rq id %u state %u onchip %u\n",
 109                               qp->wq.sq.qid, qp->wq.rq.qid,
 110                               (int)qp->attr.state,
 111                               qp->wq.sq.flags & T4_SQ_ONCHIP);
 112         if (cc < space)
 113                 qpd->pos += cc;
 114         return 0;
 115 }
 116
 117 static int qp_release(struct inode *inode, struct file *file)
 118 {
 119         struct c4iw_debugfs_data *qpd = file->private_data;
 120         if (!qpd) {
 121                 printk(KERN_INFO "%s null qpd?\n", __func__);
 122                 return 0;
 123         }
 124         vfree(qpd->buf);
 125         kfree(qpd);
 126         return 0;
 127 }
 128
 129 static int qp_open(struct inode *inode, struct file *file)
 130 {
 131         struct c4iw_debugfs_data *qpd;
 132         int ret = 0;
 133         int count = 1;
 134
 135         qpd = kmalloc(sizeof *qpd, GFP_KERNEL);
 136         if (!qpd) {
 137                 ret = -ENOMEM;
 138                 goto out;
 139         }
 140         qpd->devp = inode->i_private;
 141         qpd->pos = 0;
 142
 143         spin_lock_irq(&qpd->devp->lock);
 144         idr_for_each(&qpd->devp->qpidr, count_idrs, &count);
 145         spin_unlock_irq(&qpd->devp->lock);
 146
 147         qpd->bufsize = count * 128;
 148         qpd->buf = vmalloc(qpd->bufsize);
 149         if (!qpd->buf) {
 150                 ret = -ENOMEM;
 151                 goto err1;
 152         }
 153
 154         spin_lock_irq(&qpd->devp->lock);
 155         idr_for_each(&qpd->devp->qpidr, dump_qp, qpd);
 156         spin_unlock_irq(&qpd->devp->lock);
 157
 158         qpd->buf[qpd->pos++] = 0;
 159         file->private_data = qpd;
 160         goto out;
 161 err1:
 162         kfree(qpd);
 163 out:
 164         return ret;
 165 }
 166
 167 static const struct file_operations qp_debugfs_fops = {
 168         .owner   = THIS_MODULE,
 169         .open    = qp_open,
 170         .release = qp_release,
 171         .read    = debugfs_read,
 172         .llseek  = default_llseek,
 173 };
 174
 175 static int dump_stag(int id, void *p, void *data)
 176 {
 177         struct c4iw_debugfs_data *stagd = data;
 178         int space;
 179         int cc;
 180
 181         space = stagd->bufsize - stagd->pos - 1;
 182         if (space == 0)
 183                 return 1;
 184
 185         cc = snprintf(stagd->buf + stagd->pos, space, "0x%x\n", id<<8);
 186         if (cc < space)
 187                 stagd->pos += cc;
 188         return 0;
 189 }
 190
 191 static int stag_release(struct inode *inode, struct file *file)
 192 {
 193         struct c4iw_debugfs_data *stagd = file->private_data;
 194         if (!stagd) {
 195                 printk(KERN_INFO "%s null stagd?\n", __func__);
 196                 return 0;
 197         }
 198         kfree(stagd->buf);
 199         kfree(stagd);
 200         return 0;
 201 }
 202
 203 static int stag_open(struct inode *inode, struct file *file)
 204 {
 205         struct c4iw_debugfs_data *stagd;
 206         int ret = 0;
 207         int count = 1;
 208
 209         stagd = kmalloc(sizeof *stagd, GFP_KERNEL);
 210         if (!stagd) {
 211                 ret = -ENOMEM;
 212                 goto out;
 213         }
 214         stagd->devp = inode->i_private;
 215         stagd->pos = 0;
 216
 217         spin_lock_irq(&stagd->devp->lock);
 218         idr_for_each(&stagd->devp->mmidr, count_idrs, &count);
 219         spin_unlock_irq(&stagd->devp->lock);
 220
 221         stagd->bufsize = count * sizeof("0x12345678\n");
 222         stagd->buf = kmalloc(stagd->bufsize, GFP_KERNEL);
 223         if (!stagd->buf) {
 224                 ret = -ENOMEM;
 225                 goto err1;
 226         }
 227
 228         spin_lock_irq(&stagd->devp->lock);
 229         idr_for_each(&stagd->devp->mmidr, dump_stag, stagd);
 230         spin_unlock_irq(&stagd->devp->lock);
 231
 232         stagd->buf[stagd->pos++] = 0;
 233         file->private_data = stagd;
 234         goto out;
 235 err1:
 236         kfree(stagd);
 237 out:
 238         return ret;
 239 }
 240
 241 static const struct file_operations stag_debugfs_fops = {
 242         .owner   = THIS_MODULE,
 243         .open    = stag_open,
 244         .release = stag_release,
 245         .read    = debugfs_read,
 246         .llseek  = default_llseek,
 247 };
 248
 249 static char *db_state_str[] = {"NORMAL", "FLOW_CONTROL", "RECOVERY"};
 250
 251 static int stats_show(struct seq_file *seq, void *v)
 252 {
 253         struct c4iw_dev *dev = seq->private;
 254
 255         seq_printf(seq, "   Object: %10s %10s %10s %10s\n", "Total", "Current",
 256                    "Max", "Fail");
 257         seq_printf(seq, "     PDID: %10llu %10llu %10llu %10llu\n",
 258                         dev->rdev.stats.pd.total, dev->rdev.stats.pd.cur,
 259                         dev->rdev.stats.pd.max, dev->rdev.stats.pd.fail);
 260         seq_printf(seq, "      QID: %10llu %10llu %10llu %10llu\n",
 261                         dev->rdev.stats.qid.total, dev->rdev.stats.qid.cur,
 262                         dev->rdev.stats.qid.max, dev->rdev.stats.qid.fail);
 263         seq_printf(seq, "   TPTMEM: %10llu %10llu %10llu %10llu\n",
 264                         dev->rdev.stats.stag.total, dev->rdev.stats.stag.cur,
 265                         dev->rdev.stats.stag.max, dev->rdev.stats.stag.fail);
 266         seq_printf(seq, "   PBLMEM: %10llu %10llu %10llu %10llu\n",
 267                         dev->rdev.stats.pbl.total, dev->rdev.stats.pbl.cur,
 268                         dev->rdev.stats.pbl.max, dev->rdev.stats.pbl.fail);
 269         seq_printf(seq, "   RQTMEM: %10llu %10llu %10llu %10llu\n",
 270                         dev->rdev.stats.rqt.total, dev->rdev.stats.rqt.cur,
 271                         dev->rdev.stats.rqt.max, dev->rdev.stats.rqt.fail);
 272         seq_printf(seq, "  OCQPMEM: %10llu %10llu %10llu %10llu\n",
 273                         dev->rdev.stats.ocqp.total, dev->rdev.stats.ocqp.cur,
 274                         dev->rdev.stats.ocqp.max, dev->rdev.stats.ocqp.fail);
 275         seq_printf(seq, "  DB FULL: %10llu\n", dev->rdev.stats.db_full);
 276         seq_printf(seq, " DB EMPTY: %10llu\n", dev->rdev.stats.db_empty);
 277         seq_printf(seq, "  DB DROP: %10llu\n", dev->rdev.stats.db_drop);
 278         seq_printf(seq, " DB State: %s Transitions %llu\n",
 279                    db_state_str[dev->db_state],
 280                    dev->rdev.stats.db_state_transitions);
 281         return 0;
 282 }
 283
 284 static int stats_open(struct inode *inode, struct file *file)
 285 {
 286         return single_open(file, stats_show, inode->i_private);
 287 }
 288
 289 static ssize_t stats_clear(struct file *file, const char __user *buf,
 290                 size_t count, loff_t *pos)
 291 {
 292         struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
 293
 294         mutex_lock(&dev->rdev.stats.lock);
 295         dev->rdev.stats.pd.max = 0;
 296         dev->rdev.stats.pd.fail = 0;
 297         dev->rdev.stats.qid.max = 0;
 298         dev->rdev.stats.qid.fail = 0;
 299         dev->rdev.stats.stag.max = 0;
 300         dev->rdev.stats.stag.fail = 0;
 301         dev->rdev.stats.pbl.max = 0;
 302         dev->rdev.stats.pbl.fail = 0;
 303         dev->rdev.stats.rqt.max = 0;
 304         dev->rdev.stats.rqt.fail = 0;
 305         dev->rdev.stats.ocqp.max = 0;
 306         dev->rdev.stats.ocqp.fail = 0;
 307         dev->rdev.stats.db_full = 0;
 308         dev->rdev.stats.db_empty = 0;
 309         dev->rdev.stats.db_drop = 0;
 310         dev->rdev.stats.db_state_transitions = 0;
 311         mutex_unlock(&dev->rdev.stats.lock);
 312         return count;
 313 }
 314
 315 static const struct file_operations stats_debugfs_fops = {
 316         .owner   = THIS_MODULE,
 317         .open    = stats_open,
 318         .release = single_release,
 319         .read    = seq_read,
 320         .llseek  = seq_lseek,
 321         .write   = stats_clear,
 322 };
 323
 324 static int setup_debugfs(struct c4iw_dev *devp)
 325 {
 326         struct dentry *de;
 327
 328         if (!devp->debugfs_root)
 329                 return -1;
 330
 331         de = debugfs_create_file("qps", S_IWUSR, devp->debugfs_root,
 332                                  (void *)devp, &qp_debugfs_fops);
 333         if (de && de->d_inode)
 334                 de->d_inode->i_size = 4096;
 335
 336         de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
 337                                  (void *)devp, &stag_debugfs_fops);
 338         if (de && de->d_inode)
 339                 de->d_inode->i_size = 4096;
 340
 341         de = debugfs_create_file("stats", S_IWUSR, devp->debugfs_root,
 342                         (void *)devp, &stats_debugfs_fops);
 343         if (de && de->d_inode)
 344                 de->d_inode->i_size = 4096;
 345
 346         return 0;
 347 }
 348
 349 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
 350                                struct c4iw_dev_ucontext *uctx)
 351 {
 352         struct list_head *pos, *nxt;
 353         struct c4iw_qid_list *entry;
 354
 355         mutex_lock(&uctx->lock);
 356         list_for_each_safe(pos, nxt, &uctx->qpids) {
 357                 entry = list_entry(pos, struct c4iw_qid_list, entry);
 358                 list_del_init(&entry->entry);
 359                 if (!(entry->qid & rdev->qpmask)) {
 360                         c4iw_put_resource(&rdev->resource.qid_table,
 361                                           entry->qid);
 362                         mutex_lock(&rdev->stats.lock);
 363                         rdev->stats.qid.cur -= rdev->qpmask + 1;
 364                         mutex_unlock(&rdev->stats.lock);
 365                 }
 366                 kfree(entry);
 367         }
 368
 369         list_for_each_safe(pos, nxt, &uctx->qpids) {
 370                 entry = list_entry(pos, struct c4iw_qid_list, entry);
 371                 list_del_init(&entry->entry);
 372                 kfree(entry);
 373         }
 374         mutex_unlock(&uctx->lock);
 375 }
 376
 377 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
 378                             struct c4iw_dev_ucontext *uctx)
 379 {
 380         INIT_LIST_HEAD(&uctx->qpids);
 381         INIT_LIST_HEAD(&uctx->cqids);
 382         mutex_init(&uctx->lock);
 383 }
 384
 385 /* Caller takes care of locking if needed */
 386 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
 387 {
 388         int err;
 389
 390         c4iw_init_dev_ucontext(rdev, &rdev->uctx);
 391
 392         /*
 393          * qpshift is the number of bits to shift the qpid left in order
 394          * to get the correct address of the doorbell for that qp.
 395          */
 396         rdev->qpshift = PAGE_SHIFT - ilog2(rdev->lldi.udb_density);
 397         rdev->qpmask = rdev->lldi.udb_density - 1;
 398         rdev->cqshift = PAGE_SHIFT - ilog2(rdev->lldi.ucq_density);
 399         rdev->cqmask = rdev->lldi.ucq_density - 1;
 400         PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
 401              "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
 402              "qp qid start %u size %u cq qid start %u size %u\n",
 403              __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
 404              rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
 405              rdev->lldi.vr->pbl.start,
 406              rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
 407              rdev->lldi.vr->rq.size,
 408              rdev->lldi.vr->qp.start,
 409              rdev->lldi.vr->qp.size,
 410              rdev->lldi.vr->cq.start,
 411              rdev->lldi.vr->cq.size);
 412         PDBG("udb len 0x%x udb base %p db_reg %p gts_reg %p qpshift %lu "
 413              "qpmask 0x%x cqshift %lu cqmask 0x%x\n",
 414              (unsigned)pci_resource_len(rdev->lldi.pdev, 2),
 415              (void *)pci_resource_start(rdev->lldi.pdev, 2),
 416              rdev->lldi.db_reg,
 417              rdev->lldi.gts_reg,
 418              rdev->qpshift, rdev->qpmask,
 419              rdev->cqshift, rdev->cqmask);
 420
 421         if (c4iw_num_stags(rdev) == 0) {
 422                 err = -EINVAL;
 423                 goto err1;
 424         }
 425
 426         rdev->stats.pd.total = T4_MAX_NUM_PD;
 427         rdev->stats.stag.total = rdev->lldi.vr->stag.size;
 428         rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
 429         rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
 430         rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
 431         rdev->stats.qid.total = rdev->lldi.vr->qp.size;
 432
 433         err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
 434         if (err) {
 435                 printk(KERN_ERR MOD "error %d initializing resources\n", err);
 436                 goto err1;
 437         }
 438         err = c4iw_pblpool_create(rdev);
 439         if (err) {
 440                 printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
 441                 goto err2;
 442         }
 443         err = c4iw_rqtpool_create(rdev);
 444         if (err) {
 445                 printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
 446                 goto err3;
 447         }
 448         err = c4iw_ocqp_pool_create(rdev);
 449         if (err) {
 450                 printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
 451                 goto err4;
 452         }
 453         return 0;
 454 err4:
 455         c4iw_rqtpool_destroy(rdev);
 456 err3:
 457         c4iw_pblpool_destroy(rdev);
 458 err2:
 459         c4iw_destroy_resource(&rdev->resource);
 460 err1:
 461         return err;
 462 }
 463
 464 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
 465 {
 466         c4iw_pblpool_destroy(rdev);
 467         c4iw_rqtpool_destroy(rdev);
 468         c4iw_destroy_resource(&rdev->resource);
 469 }
 470
 471 static void c4iw_dealloc(struct uld_ctx *ctx)
 472 {
 473         c4iw_rdev_close(&ctx->dev->rdev);
 474         idr_destroy(&ctx->dev->cqidr);
 475         idr_destroy(&ctx->dev->qpidr);
 476         idr_destroy(&ctx->dev->mmidr);
 477         iounmap(ctx->dev->rdev.oc_mw_kva);
 478         ib_dealloc_device(&ctx->dev->ibdev);
 479         ctx->dev = NULL;
 480 }
 481
 482 static void c4iw_remove(struct uld_ctx *ctx)
 483 {
 484         PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
 485         c4iw_unregister_device(ctx->dev);
 486         c4iw_dealloc(ctx);
 487 }
 488
 489 static int rdma_supported(const struct cxgb4_lld_info *infop)
 490 {
 491         return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
 492                infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
 493                infop->vr->cq.size > 0 && infop->vr->ocq.size > 0;
 494 }
 495
 496 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
 497 {
 498         struct c4iw_dev *devp;
 499         int ret;
 500
 501         if (!rdma_supported(infop)) {
 502                 printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
 503                        pci_name(infop->pdev));
 504                 return ERR_PTR(-ENOSYS);
 505         }
 506         devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
 507         if (!devp) {
 508                 printk(KERN_ERR MOD "Cannot allocate ib device\n");
 509                 return ERR_PTR(-ENOMEM);
 510         }
 511         devp->rdev.lldi = *infop;
 512
 513         devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
 514                 (pci_resource_len(devp->rdev.lldi.pdev, 2) -
 515                  roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
 516         devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
 517                                                devp->rdev.lldi.vr->ocq.size);
 518
 519         PDBG(KERN_INFO MOD "ocq memory: "
 520                "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
 521                devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
 522                devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
 523
 524         ret = c4iw_rdev_open(&devp->rdev);
 525         if (ret) {
 526                 printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
 527                 ib_dealloc_device(&devp->ibdev);
 528                 return ERR_PTR(ret);
 529         }
 530
 531         idr_init(&devp->cqidr);
 532         idr_init(&devp->qpidr);
 533         idr_init(&devp->mmidr);
 534         spin_lock_init(&devp->lock);
 535         mutex_init(&devp->rdev.stats.lock);
 536         mutex_init(&devp->db_mutex);
 537
 538         if (c4iw_debugfs_root) {
 539                 devp->debugfs_root = debugfs_create_dir(
 540                                         pci_name(devp->rdev.lldi.pdev),
 541                                         c4iw_debugfs_root);
 542                 setup_debugfs(devp);
 543         }
 544         return devp;
 545 }
 546
 547 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
 548 {
 549         struct uld_ctx *ctx;
 550         static int vers_printed;
 551         int i;
 552
 553         if (!vers_printed++)
 554                 printk(KERN_INFO MOD "Chelsio T4 RDMA Driver - version %s\n",
 555                        DRV_VERSION);
 556
 557         ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
 558         if (!ctx) {
 559                 ctx = ERR_PTR(-ENOMEM);
 560                 goto out;
 561         }
 562         ctx->lldi = *infop;
 563
 564         PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
 565              __func__, pci_name(ctx->lldi.pdev),
 566              ctx->lldi.nchan, ctx->lldi.nrxq,
 567              ctx->lldi.ntxq, ctx->lldi.nports);
 568
 569         mutex_lock(&dev_mutex);
 570         list_add_tail(&ctx->entry, &uld_ctx_list);
 571         mutex_unlock(&dev_mutex);
 572
 573         for (i = 0; i < ctx->lldi.nrxq; i++)
 574                 PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
 575 out:
 576         return ctx;
 577 }
 578
 579 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
 580                         const struct pkt_gl *gl)
 581 {
 582         struct uld_ctx *ctx = handle;
 583         struct c4iw_dev *dev = ctx->dev;
 584         struct sk_buff *skb;
 585         const struct cpl_act_establish *rpl;
 586         unsigned int opcode;
 587
 588         if (gl == NULL) {
 589                 /* omit RSS and rsp_ctrl at end of descriptor */
 590                 unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
 591
 592                 skb = alloc_skb(256, GFP_ATOMIC);
 593                 if (!skb)
 594                         goto nomem;
 595                 __skb_put(skb, len);
 596                 skb_copy_to_linear_data(skb, &rsp[1], len);
 597         } else if (gl == CXGB4_MSG_AN) {
 598                 const struct rsp_ctrl *rc = (void *)rsp;
 599
 600                 u32 qid = be32_to_cpu(rc->pldbuflen_qid);
 601                 c4iw_ev_handler(dev, qid);
 602                 return 0;
 603         } else {
 604                 skb = cxgb4_pktgl_to_skb(gl, 128, 128);
 605                 if (unlikely(!skb))
 606                         goto nomem;
 607         }
 608
 609         rpl = cplhdr(skb);
 610         opcode = rpl->ot.opcode;
 611
 612         if (c4iw_handlers[opcode])
 613                 c4iw_handlers[opcode](dev, skb);
 614         else
 615                 printk(KERN_INFO "%s no handler opcode 0x%x...\n", __func__,
 616                        opcode);
 617
 618         return 0;
 619 nomem:
 620         return -1;
 621 }
 622
 623 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
 624 {
 625         struct uld_ctx *ctx = handle;
 626
 627         PDBG("%s new_state %u\n", __func__, new_state);
 628         switch (new_state) {
 629         case CXGB4_STATE_UP:
 630                 printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
 631                 if (!ctx->dev) {
 632                         int ret;
 633
 634                         ctx->dev = c4iw_alloc(&ctx->lldi);
 635                         if (IS_ERR(ctx->dev)) {
 636                                 printk(KERN_ERR MOD
 637                                        "%s: initialization failed: %ld\n",
 638                                        pci_name(ctx->lldi.pdev),
 639                                        PTR_ERR(ctx->dev));
 640                                 ctx->dev = NULL;
 641                                 break;
 642                         }
 643                         ret = c4iw_register_device(ctx->dev);
 644                         if (ret) {
 645                                 printk(KERN_ERR MOD
 646                                        "%s: RDMA registration failed: %d\n",
 647                                        pci_name(ctx->lldi.pdev), ret);
 648                                 c4iw_dealloc(ctx);
 649                         }
 650                 }
 651                 break;
 652         case CXGB4_STATE_DOWN:
 653                 printk(KERN_INFO MOD "%s: Down\n",
 654                        pci_name(ctx->lldi.pdev));
 655                 if (ctx->dev)
 656                         c4iw_remove(ctx);
 657                 break;
 658         case CXGB4_STATE_START_RECOVERY:
 659                 printk(KERN_INFO MOD "%s: Fatal Error\n",
 660                        pci_name(ctx->lldi.pdev));
 661                 if (ctx->dev) {
 662                         struct ib_event event;
 663
 664                         ctx->dev->rdev.flags |= T4_FATAL_ERROR;
 665                         memset(&event, 0, sizeof event);
 666                         event.event  = IB_EVENT_DEVICE_FATAL;
 667                         event.device = &ctx->dev->ibdev;
 668                         ib_dispatch_event(&event);
 669                         c4iw_remove(ctx);
 670                 }
 671                 break;
 672         case CXGB4_STATE_DETACH:
 673                 printk(KERN_INFO MOD "%s: Detach\n",
 674                        pci_name(ctx->lldi.pdev));
 675                 if (ctx->dev)
 676                         c4iw_remove(ctx);
 677                 break;
 678         }
 679         return 0;
 680 }
 681
 682 static int disable_qp_db(int id, void *p, void *data)
 683 {
 684         struct c4iw_qp *qp = p;
 685
 686         t4_disable_wq_db(&qp->wq);
 687         return 0;
 688 }
 689
 690 static void stop_queues(struct uld_ctx *ctx)
 691 {
 692         spin_lock_irq(&ctx->dev->lock);
 693         if (ctx->dev->db_state == NORMAL) {
 694                 ctx->dev->rdev.stats.db_state_transitions++;
 695                 ctx->dev->db_state = FLOW_CONTROL;
 696                 idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
 697         }
 698         spin_unlock_irq(&ctx->dev->lock);
 699 }
 700
 701 static int enable_qp_db(int id, void *p, void *data)
 702 {
 703         struct c4iw_qp *qp = p;
 704
 705         t4_enable_wq_db(&qp->wq);
 706         return 0;
 707 }
 708
 709 static void resume_queues(struct uld_ctx *ctx)
 710 {
 711         spin_lock_irq(&ctx->dev->lock);
 712         if (ctx->dev->qpcnt <= db_fc_threshold &&
 713             ctx->dev->db_state == FLOW_CONTROL) {
 714                 ctx->dev->db_state = NORMAL;
 715                 ctx->dev->rdev.stats.db_state_transitions++;
 716                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
 717         }
 718         spin_unlock_irq(&ctx->dev->lock);
 719 }
 720
 721 struct qp_list {
 722         unsigned idx;
 723         struct c4iw_qp **qps;
 724 };
 725
 726 static int add_and_ref_qp(int id, void *p, void *data)
 727 {
 728         struct qp_list *qp_listp = data;
 729         struct c4iw_qp *qp = p;
 730
 731         c4iw_qp_add_ref(&qp->ibqp);
 732         qp_listp->qps[qp_listp->idx++] = qp;
 733         return 0;
 734 }
 735
 736 static int count_qps(int id, void *p, void *data)
 737 {
 738         unsigned *countp = data;
 739         (*countp)++;
 740         return 0;
 741 }
 742
 743 static void deref_qps(struct qp_list qp_list)
 744 {
 745         int idx;
 746
 747         for (idx = 0; idx < qp_list.idx; idx++)
 748                 c4iw_qp_rem_ref(&qp_list.qps[idx]->ibqp);
 749 }
 750
 751 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
 752 {
 753         int idx;
 754         int ret;
 755
 756         for (idx = 0; idx < qp_list->idx; idx++) {
 757                 struct c4iw_qp *qp = qp_list->qps[idx];
 758
 759                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
 760                                           qp->wq.sq.qid,
 761                                           t4_sq_host_wq_pidx(&qp->wq),
 762                                           t4_sq_wq_size(&qp->wq));
 763                 if (ret) {
 764                         printk(KERN_ERR MOD "%s: Fatal error - "
 765                                "DB overflow recovery failed - "
 766                                "error syncing SQ qid %u\n",
 767                                pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
 768                         return;
 769                 }
 770
 771                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
 772                                           qp->wq.rq.qid,
 773                                           t4_rq_host_wq_pidx(&qp->wq),
 774                                           t4_rq_wq_size(&qp->wq));
 775
 776                 if (ret) {
 777                         printk(KERN_ERR MOD "%s: Fatal error - "
 778                                "DB overflow recovery failed - "
 779                                "error syncing RQ qid %u\n",
 780                                pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
 781                         return;
 782                 }
 783
 784                 /* Wait for the dbfifo to drain */
 785                 while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
 786                         set_current_state(TASK_UNINTERRUPTIBLE);
 787                         schedule_timeout(usecs_to_jiffies(10));
 788                 }
 789         }
 790 }
 791
 792 static void recover_queues(struct uld_ctx *ctx)
 793 {
 794         int count = 0;
 795         struct qp_list qp_list;
 796         int ret;
 797
 798         /* lock out kernel db ringers */
 799         mutex_lock(&ctx->dev->db_mutex);
 800
 801         /* put all queues in to recovery mode */
 802         spin_lock_irq(&ctx->dev->lock);
 803         ctx->dev->db_state = RECOVERY;
 804         ctx->dev->rdev.stats.db_state_transitions++;
 805         idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
 806         spin_unlock_irq(&ctx->dev->lock);
 807
 808         /* slow everybody down */
 809         set_current_state(TASK_UNINTERRUPTIBLE);
 810         schedule_timeout(usecs_to_jiffies(1000));
 811
 812         /* Wait for the dbfifo to completely drain. */
 813         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
 814                 set_current_state(TASK_UNINTERRUPTIBLE);
 815                 schedule_timeout(usecs_to_jiffies(10));
 816         }
 817
 818         /* flush the SGE contexts */
 819         ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
 820         if (ret) {
 821                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
 822                        pci_name(ctx->lldi.pdev));
 823                 goto out;
 824         }
 825
 826         /* Count active queues so we can build a list of queues to recover */
 827         spin_lock_irq(&ctx->dev->lock);
 828         idr_for_each(&ctx->dev->qpidr, count_qps, &count);
 829
 830         qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
 831         if (!qp_list.qps) {
 832                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
 833                        pci_name(ctx->lldi.pdev));
 834                 spin_unlock_irq(&ctx->dev->lock);
 835                 goto out;
 836         }
 837         qp_list.idx = 0;
 838
 839         /* add and ref each qp so it doesn't get freed */
 840         idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
 841
 842         spin_unlock_irq(&ctx->dev->lock);
 843
 844         /* now traverse the list in a safe context to recover the db state*/
 845         recover_lost_dbs(ctx, &qp_list);
 846
 847         /* we're almost done!  deref the qps and clean up */
 848         deref_qps(qp_list);
 849         kfree(qp_list.qps);
 850
 851         /* Wait for the dbfifo to completely drain again */
 852         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
 853                 set_current_state(TASK_UNINTERRUPTIBLE);
 854                 schedule_timeout(usecs_to_jiffies(10));
 855         }
 856
 857         /* resume the queues */
 858         spin_lock_irq(&ctx->dev->lock);
 859         if (ctx->dev->qpcnt > db_fc_threshold)
 860                 ctx->dev->db_state = FLOW_CONTROL;
 861         else {
 862                 ctx->dev->db_state = NORMAL;
 863                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
 864         }
 865         ctx->dev->rdev.stats.db_state_transitions++;
 866         spin_unlock_irq(&ctx->dev->lock);
 867
 868 out:
 869         /* start up kernel db ringers again */
 870         mutex_unlock(&ctx->dev->db_mutex);
 871 }
 872
 873 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
 874 {
 875         struct uld_ctx *ctx = handle;
 876
 877         switch (control) {
 878         case CXGB4_CONTROL_DB_FULL:
 879                 stop_queues(ctx);
 880                 mutex_lock(&ctx->dev->rdev.stats.lock);
 881                 ctx->dev->rdev.stats.db_full++;
 882                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 883                 break;
 884         case CXGB4_CONTROL_DB_EMPTY:
 885                 resume_queues(ctx);
 886                 mutex_lock(&ctx->dev->rdev.stats.lock);
 887                 ctx->dev->rdev.stats.db_empty++;
 888                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 889                 break;
 890         case CXGB4_CONTROL_DB_DROP:
 891                 recover_queues(ctx);
 892                 mutex_lock(&ctx->dev->rdev.stats.lock);
 893                 ctx->dev->rdev.stats.db_drop++;
 894                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 895                 break;
 896         default:
 897                 printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
 898                        pci_name(ctx->lldi.pdev), control);
 899                 break;
 900         }
 901         return 0;
 902 }
 903
 904 static struct cxgb4_uld_info c4iw_uld_info = {
 905         .name = DRV_NAME,
 906         .add = c4iw_uld_add,
 907         .rx_handler = c4iw_uld_rx_handler,
 908         .state_change = c4iw_uld_state_change,
 909         .control = c4iw_uld_control,
 910 };
 911
 912 static int __init c4iw_init_module(void)
 913 {
 914         int err;
 915
 916         err = c4iw_cm_init();
 917         if (err)
 918                 return err;
 919
 920         c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
 921         if (!c4iw_debugfs_root)
 922                 printk(KERN_WARNING MOD
 923                        "could not create debugfs entry, continuing\n");
 924
 925         cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
 926
 927         return 0;
 928 }
 929
 930 static void __exit c4iw_exit_module(void)
 931 {
 932         struct uld_ctx *ctx, *tmp;
 933
 934         mutex_lock(&dev_mutex);
 935         list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
 936                 if (ctx->dev)
 937                         c4iw_remove(ctx);
 938                 kfree(ctx);
 939         }
 940         mutex_unlock(&dev_mutex);
 941         cxgb4_unregister_uld(CXGB4_ULD_RDMA);
 942         c4iw_cm_term();
 943         debugfs_remove_recursive(c4iw_debugfs_root);
 944 }
 945
 946 module_init(c4iw_init_module);
 947 module_exit(c4iw_exit_module);