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\n", "Total", "Current", "Max");
 256         seq_printf(seq, "     PDID: %10llu %10llu %10llu\n",
 257                         dev->rdev.stats.pd.total, dev->rdev.stats.pd.cur,
 258                         dev->rdev.stats.pd.max);
 259         seq_printf(seq, "      QID: %10llu %10llu %10llu\n",
 260                         dev->rdev.stats.qid.total, dev->rdev.stats.qid.cur,
 261                         dev->rdev.stats.qid.max);
 262         seq_printf(seq, "   TPTMEM: %10llu %10llu %10llu\n",
 263                         dev->rdev.stats.stag.total, dev->rdev.stats.stag.cur,
 264                         dev->rdev.stats.stag.max);
 265         seq_printf(seq, "   PBLMEM: %10llu %10llu %10llu\n",
 266                         dev->rdev.stats.pbl.total, dev->rdev.stats.pbl.cur,
 267                         dev->rdev.stats.pbl.max);
 268         seq_printf(seq, "   RQTMEM: %10llu %10llu %10llu\n",
 269                         dev->rdev.stats.rqt.total, dev->rdev.stats.rqt.cur,
 270                         dev->rdev.stats.rqt.max);
 271         seq_printf(seq, "  OCQPMEM: %10llu %10llu %10llu\n",
 272                         dev->rdev.stats.ocqp.total, dev->rdev.stats.ocqp.cur,
 273                         dev->rdev.stats.ocqp.max);
 274         seq_printf(seq, "  DB FULL: %10llu\n", dev->rdev.stats.db_full);
 275         seq_printf(seq, " DB EMPTY: %10llu\n", dev->rdev.stats.db_empty);
 276         seq_printf(seq, "  DB DROP: %10llu\n", dev->rdev.stats.db_drop);
 277         seq_printf(seq, " DB State: %s Transitions %llu\n",
 278                    db_state_str[dev->db_state],
 279                    dev->rdev.stats.db_state_transitions);
 280         return 0;
 281 }
 282
 283 static int stats_open(struct inode *inode, struct file *file)
 284 {
 285         return single_open(file, stats_show, inode->i_private);
 286 }
 287
 288 static ssize_t stats_clear(struct file *file, const char __user *buf,
 289                 size_t count, loff_t *pos)
 290 {
 291         struct c4iw_dev *dev = ((struct seq_file *)file->private_data)->private;
 292
 293         mutex_lock(&dev->rdev.stats.lock);
 294         dev->rdev.stats.pd.max = 0;
 295         dev->rdev.stats.qid.max = 0;
 296         dev->rdev.stats.stag.max = 0;
 297         dev->rdev.stats.pbl.max = 0;
 298         dev->rdev.stats.rqt.max = 0;
 299         dev->rdev.stats.ocqp.max = 0;
 300         dev->rdev.stats.db_full = 0;
 301         dev->rdev.stats.db_empty = 0;
 302         dev->rdev.stats.db_drop = 0;
 303         dev->rdev.stats.db_state_transitions = 0;
 304         mutex_unlock(&dev->rdev.stats.lock);
 305         return count;
 306 }
 307
 308 static const struct file_operations stats_debugfs_fops = {
 309         .owner   = THIS_MODULE,
 310         .open    = stats_open,
 311         .release = single_release,
 312         .read    = seq_read,
 313         .llseek  = seq_lseek,
 314         .write   = stats_clear,
 315 };
 316
 317 static int setup_debugfs(struct c4iw_dev *devp)
 318 {
 319         struct dentry *de;
 320
 321         if (!devp->debugfs_root)
 322                 return -1;
 323
 324         de = debugfs_create_file("qps", S_IWUSR, devp->debugfs_root,
 325                                  (void *)devp, &qp_debugfs_fops);
 326         if (de && de->d_inode)
 327                 de->d_inode->i_size = 4096;
 328
 329         de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
 330                                  (void *)devp, &stag_debugfs_fops);
 331         if (de && de->d_inode)
 332                 de->d_inode->i_size = 4096;
 333
 334         de = debugfs_create_file("stats", S_IWUSR, devp->debugfs_root,
 335                         (void *)devp, &stats_debugfs_fops);
 336         if (de && de->d_inode)
 337                 de->d_inode->i_size = 4096;
 338
 339         return 0;
 340 }
 341
 342 void c4iw_release_dev_ucontext(struct c4iw_rdev *rdev,
 343                                struct c4iw_dev_ucontext *uctx)
 344 {
 345         struct list_head *pos, *nxt;
 346         struct c4iw_qid_list *entry;
 347
 348         mutex_lock(&uctx->lock);
 349         list_for_each_safe(pos, nxt, &uctx->qpids) {
 350                 entry = list_entry(pos, struct c4iw_qid_list, entry);
 351                 list_del_init(&entry->entry);
 352                 if (!(entry->qid & rdev->qpmask)) {
 353                         c4iw_put_resource(&rdev->resource.qid_fifo, entry->qid,
 354                                         &rdev->resource.qid_fifo_lock);
 355                         mutex_lock(&rdev->stats.lock);
 356                         rdev->stats.qid.cur -= rdev->qpmask + 1;
 357                         mutex_unlock(&rdev->stats.lock);
 358                 }
 359                 kfree(entry);
 360         }
 361
 362         list_for_each_safe(pos, nxt, &uctx->qpids) {
 363                 entry = list_entry(pos, struct c4iw_qid_list, entry);
 364                 list_del_init(&entry->entry);
 365                 kfree(entry);
 366         }
 367         mutex_unlock(&uctx->lock);
 368 }
 369
 370 void c4iw_init_dev_ucontext(struct c4iw_rdev *rdev,
 371                             struct c4iw_dev_ucontext *uctx)
 372 {
 373         INIT_LIST_HEAD(&uctx->qpids);
 374         INIT_LIST_HEAD(&uctx->cqids);
 375         mutex_init(&uctx->lock);
 376 }
 377
 378 /* Caller takes care of locking if needed */
 379 static int c4iw_rdev_open(struct c4iw_rdev *rdev)
 380 {
 381         int err;
 382
 383         c4iw_init_dev_ucontext(rdev, &rdev->uctx);
 384
 385         /*
 386          * qpshift is the number of bits to shift the qpid left in order
 387          * to get the correct address of the doorbell for that qp.
 388          */
 389         rdev->qpshift = PAGE_SHIFT - ilog2(rdev->lldi.udb_density);
 390         rdev->qpmask = rdev->lldi.udb_density - 1;
 391         rdev->cqshift = PAGE_SHIFT - ilog2(rdev->lldi.ucq_density);
 392         rdev->cqmask = rdev->lldi.ucq_density - 1;
 393         PDBG("%s dev %s stag start 0x%0x size 0x%0x num stags %d "
 394              "pbl start 0x%0x size 0x%0x rq start 0x%0x size 0x%0x "
 395              "qp qid start %u size %u cq qid start %u size %u\n",
 396              __func__, pci_name(rdev->lldi.pdev), rdev->lldi.vr->stag.start,
 397              rdev->lldi.vr->stag.size, c4iw_num_stags(rdev),
 398              rdev->lldi.vr->pbl.start,
 399              rdev->lldi.vr->pbl.size, rdev->lldi.vr->rq.start,
 400              rdev->lldi.vr->rq.size,
 401              rdev->lldi.vr->qp.start,
 402              rdev->lldi.vr->qp.size,
 403              rdev->lldi.vr->cq.start,
 404              rdev->lldi.vr->cq.size);
 405         PDBG("udb len 0x%x udb base %p db_reg %p gts_reg %p qpshift %lu "
 406              "qpmask 0x%x cqshift %lu cqmask 0x%x\n",
 407              (unsigned)pci_resource_len(rdev->lldi.pdev, 2),
 408              (void *)pci_resource_start(rdev->lldi.pdev, 2),
 409              rdev->lldi.db_reg,
 410              rdev->lldi.gts_reg,
 411              rdev->qpshift, rdev->qpmask,
 412              rdev->cqshift, rdev->cqmask);
 413
 414         if (c4iw_num_stags(rdev) == 0) {
 415                 err = -EINVAL;
 416                 goto err1;
 417         }
 418
 419         rdev->stats.pd.total = T4_MAX_NUM_PD;
 420         rdev->stats.stag.total = rdev->lldi.vr->stag.size;
 421         rdev->stats.pbl.total = rdev->lldi.vr->pbl.size;
 422         rdev->stats.rqt.total = rdev->lldi.vr->rq.size;
 423         rdev->stats.ocqp.total = rdev->lldi.vr->ocq.size;
 424         rdev->stats.qid.total = rdev->lldi.vr->qp.size;
 425
 426         err = c4iw_init_resource(rdev, c4iw_num_stags(rdev), T4_MAX_NUM_PD);
 427         if (err) {
 428                 printk(KERN_ERR MOD "error %d initializing resources\n", err);
 429                 goto err1;
 430         }
 431         err = c4iw_pblpool_create(rdev);
 432         if (err) {
 433                 printk(KERN_ERR MOD "error %d initializing pbl pool\n", err);
 434                 goto err2;
 435         }
 436         err = c4iw_rqtpool_create(rdev);
 437         if (err) {
 438                 printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
 439                 goto err3;
 440         }
 441         err = c4iw_ocqp_pool_create(rdev);
 442         if (err) {
 443                 printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
 444                 goto err4;
 445         }
 446         return 0;
 447 err4:
 448         c4iw_rqtpool_destroy(rdev);
 449 err3:
 450         c4iw_pblpool_destroy(rdev);
 451 err2:
 452         c4iw_destroy_resource(&rdev->resource);
 453 err1:
 454         return err;
 455 }
 456
 457 static void c4iw_rdev_close(struct c4iw_rdev *rdev)
 458 {
 459         c4iw_pblpool_destroy(rdev);
 460         c4iw_rqtpool_destroy(rdev);
 461         c4iw_destroy_resource(&rdev->resource);
 462 }
 463
 464 static void c4iw_dealloc(struct uld_ctx *ctx)
 465 {
 466         c4iw_rdev_close(&ctx->dev->rdev);
 467         idr_destroy(&ctx->dev->cqidr);
 468         idr_destroy(&ctx->dev->qpidr);
 469         idr_destroy(&ctx->dev->mmidr);
 470         iounmap(ctx->dev->rdev.oc_mw_kva);
 471         ib_dealloc_device(&ctx->dev->ibdev);
 472         ctx->dev = NULL;
 473 }
 474
 475 static void c4iw_remove(struct uld_ctx *ctx)
 476 {
 477         PDBG("%s c4iw_dev %p\n", __func__,  ctx->dev);
 478         c4iw_unregister_device(ctx->dev);
 479         c4iw_dealloc(ctx);
 480 }
 481
 482 static int rdma_supported(const struct cxgb4_lld_info *infop)
 483 {
 484         return infop->vr->stag.size > 0 && infop->vr->pbl.size > 0 &&
 485                infop->vr->rq.size > 0 && infop->vr->qp.size > 0 &&
 486                infop->vr->cq.size > 0 && infop->vr->ocq.size > 0;
 487 }
 488
 489 static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
 490 {
 491         struct c4iw_dev *devp;
 492         int ret;
 493
 494         if (!rdma_supported(infop)) {
 495                 printk(KERN_INFO MOD "%s: RDMA not supported on this device.\n",
 496                        pci_name(infop->pdev));
 497                 return ERR_PTR(-ENOSYS);
 498         }
 499         devp = (struct c4iw_dev *)ib_alloc_device(sizeof(*devp));
 500         if (!devp) {
 501                 printk(KERN_ERR MOD "Cannot allocate ib device\n");
 502                 return ERR_PTR(-ENOMEM);
 503         }
 504         devp->rdev.lldi = *infop;
 505
 506         devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
 507                 (pci_resource_len(devp->rdev.lldi.pdev, 2) -
 508                  roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
 509         devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
 510                                                devp->rdev.lldi.vr->ocq.size);
 511
 512         PDBG(KERN_INFO MOD "ocq memory: "
 513                "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
 514                devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
 515                devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
 516
 517         ret = c4iw_rdev_open(&devp->rdev);
 518         if (ret) {
 519                 printk(KERN_ERR MOD "Unable to open CXIO rdev err %d\n", ret);
 520                 ib_dealloc_device(&devp->ibdev);
 521                 return ERR_PTR(ret);
 522         }
 523
 524         idr_init(&devp->cqidr);
 525         idr_init(&devp->qpidr);
 526         idr_init(&devp->mmidr);
 527         spin_lock_init(&devp->lock);
 528         mutex_init(&devp->rdev.stats.lock);
 529         mutex_init(&devp->db_mutex);
 530
 531         if (c4iw_debugfs_root) {
 532                 devp->debugfs_root = debugfs_create_dir(
 533                                         pci_name(devp->rdev.lldi.pdev),
 534                                         c4iw_debugfs_root);
 535                 setup_debugfs(devp);
 536         }
 537         return devp;
 538 }
 539
 540 static void *c4iw_uld_add(const struct cxgb4_lld_info *infop)
 541 {
 542         struct uld_ctx *ctx;
 543         static int vers_printed;
 544         int i;
 545
 546         if (!vers_printed++)
 547                 printk(KERN_INFO MOD "Chelsio T4 RDMA Driver - version %s\n",
 548                        DRV_VERSION);
 549
 550         ctx = kzalloc(sizeof *ctx, GFP_KERNEL);
 551         if (!ctx) {
 552                 ctx = ERR_PTR(-ENOMEM);
 553                 goto out;
 554         }
 555         ctx->lldi = *infop;
 556
 557         PDBG("%s found device %s nchan %u nrxq %u ntxq %u nports %u\n",
 558              __func__, pci_name(ctx->lldi.pdev),
 559              ctx->lldi.nchan, ctx->lldi.nrxq,
 560              ctx->lldi.ntxq, ctx->lldi.nports);
 561
 562         mutex_lock(&dev_mutex);
 563         list_add_tail(&ctx->entry, &uld_ctx_list);
 564         mutex_unlock(&dev_mutex);
 565
 566         for (i = 0; i < ctx->lldi.nrxq; i++)
 567                 PDBG("rxqid[%u] %u\n", i, ctx->lldi.rxq_ids[i]);
 568 out:
 569         return ctx;
 570 }
 571
 572 static int c4iw_uld_rx_handler(void *handle, const __be64 *rsp,
 573                         const struct pkt_gl *gl)
 574 {
 575         struct uld_ctx *ctx = handle;
 576         struct c4iw_dev *dev = ctx->dev;
 577         struct sk_buff *skb;
 578         const struct cpl_act_establish *rpl;
 579         unsigned int opcode;
 580
 581         if (gl == NULL) {
 582                 /* omit RSS and rsp_ctrl at end of descriptor */
 583                 unsigned int len = 64 - sizeof(struct rsp_ctrl) - 8;
 584
 585                 skb = alloc_skb(256, GFP_ATOMIC);
 586                 if (!skb)
 587                         goto nomem;
 588                 __skb_put(skb, len);
 589                 skb_copy_to_linear_data(skb, &rsp[1], len);
 590         } else if (gl == CXGB4_MSG_AN) {
 591                 const struct rsp_ctrl *rc = (void *)rsp;
 592
 593                 u32 qid = be32_to_cpu(rc->pldbuflen_qid);
 594                 c4iw_ev_handler(dev, qid);
 595                 return 0;
 596         } else {
 597                 skb = cxgb4_pktgl_to_skb(gl, 128, 128);
 598                 if (unlikely(!skb))
 599                         goto nomem;
 600         }
 601
 602         rpl = cplhdr(skb);
 603         opcode = rpl->ot.opcode;
 604
 605         if (c4iw_handlers[opcode])
 606                 c4iw_handlers[opcode](dev, skb);
 607         else
 608                 printk(KERN_INFO "%s no handler opcode 0x%x...\n", __func__,
 609                        opcode);
 610
 611         return 0;
 612 nomem:
 613         return -1;
 614 }
 615
 616 static int c4iw_uld_state_change(void *handle, enum cxgb4_state new_state)
 617 {
 618         struct uld_ctx *ctx = handle;
 619
 620         PDBG("%s new_state %u\n", __func__, new_state);
 621         switch (new_state) {
 622         case CXGB4_STATE_UP:
 623                 printk(KERN_INFO MOD "%s: Up\n", pci_name(ctx->lldi.pdev));
 624                 if (!ctx->dev) {
 625                         int ret;
 626
 627                         ctx->dev = c4iw_alloc(&ctx->lldi);
 628                         if (IS_ERR(ctx->dev)) {
 629                                 printk(KERN_ERR MOD
 630                                        "%s: initialization failed: %ld\n",
 631                                        pci_name(ctx->lldi.pdev),
 632                                        PTR_ERR(ctx->dev));
 633                                 ctx->dev = NULL;
 634                                 break;
 635                         }
 636                         ret = c4iw_register_device(ctx->dev);
 637                         if (ret) {
 638                                 printk(KERN_ERR MOD
 639                                        "%s: RDMA registration failed: %d\n",
 640                                        pci_name(ctx->lldi.pdev), ret);
 641                                 c4iw_dealloc(ctx);
 642                         }
 643                 }
 644                 break;
 645         case CXGB4_STATE_DOWN:
 646                 printk(KERN_INFO MOD "%s: Down\n",
 647                        pci_name(ctx->lldi.pdev));
 648                 if (ctx->dev)
 649                         c4iw_remove(ctx);
 650                 break;
 651         case CXGB4_STATE_START_RECOVERY:
 652                 printk(KERN_INFO MOD "%s: Fatal Error\n",
 653                        pci_name(ctx->lldi.pdev));
 654                 if (ctx->dev) {
 655                         struct ib_event event;
 656
 657                         ctx->dev->rdev.flags |= T4_FATAL_ERROR;
 658                         memset(&event, 0, sizeof event);
 659                         event.event  = IB_EVENT_DEVICE_FATAL;
 660                         event.device = &ctx->dev->ibdev;
 661                         ib_dispatch_event(&event);
 662                         c4iw_remove(ctx);
 663                 }
 664                 break;
 665         case CXGB4_STATE_DETACH:
 666                 printk(KERN_INFO MOD "%s: Detach\n",
 667                        pci_name(ctx->lldi.pdev));
 668                 if (ctx->dev)
 669                         c4iw_remove(ctx);
 670                 break;
 671         }
 672         return 0;
 673 }
 674
 675 static int disable_qp_db(int id, void *p, void *data)
 676 {
 677         struct c4iw_qp *qp = p;
 678
 679         t4_disable_wq_db(&qp->wq);
 680         return 0;
 681 }
 682
 683 static void stop_queues(struct uld_ctx *ctx)
 684 {
 685         spin_lock_irq(&ctx->dev->lock);
 686         if (ctx->dev->db_state == NORMAL) {
 687                 ctx->dev->rdev.stats.db_state_transitions++;
 688                 ctx->dev->db_state = FLOW_CONTROL;
 689                 idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
 690         }
 691         spin_unlock_irq(&ctx->dev->lock);
 692 }
 693
 694 static int enable_qp_db(int id, void *p, void *data)
 695 {
 696         struct c4iw_qp *qp = p;
 697
 698         t4_enable_wq_db(&qp->wq);
 699         return 0;
 700 }
 701
 702 static void resume_queues(struct uld_ctx *ctx)
 703 {
 704         spin_lock_irq(&ctx->dev->lock);
 705         if (ctx->dev->qpcnt <= db_fc_threshold &&
 706             ctx->dev->db_state == FLOW_CONTROL) {
 707                 ctx->dev->db_state = NORMAL;
 708                 ctx->dev->rdev.stats.db_state_transitions++;
 709                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
 710         }
 711         spin_unlock_irq(&ctx->dev->lock);
 712 }
 713
 714 struct qp_list {
 715         unsigned idx;
 716         struct c4iw_qp **qps;
 717 };
 718
 719 static int add_and_ref_qp(int id, void *p, void *data)
 720 {
 721         struct qp_list *qp_listp = data;
 722         struct c4iw_qp *qp = p;
 723
 724         c4iw_qp_add_ref(&qp->ibqp);
 725         qp_listp->qps[qp_listp->idx++] = qp;
 726         return 0;
 727 }
 728
 729 static int count_qps(int id, void *p, void *data)
 730 {
 731         unsigned *countp = data;
 732         (*countp)++;
 733         return 0;
 734 }
 735
 736 static void deref_qps(struct qp_list qp_list)
 737 {
 738         int idx;
 739
 740         for (idx = 0; idx < qp_list.idx; idx++)
 741                 c4iw_qp_rem_ref(&qp_list.qps[idx]->ibqp);
 742 }
 743
 744 static void recover_lost_dbs(struct uld_ctx *ctx, struct qp_list *qp_list)
 745 {
 746         int idx;
 747         int ret;
 748
 749         for (idx = 0; idx < qp_list->idx; idx++) {
 750                 struct c4iw_qp *qp = qp_list->qps[idx];
 751
 752                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
 753                                           qp->wq.sq.qid,
 754                                           t4_sq_host_wq_pidx(&qp->wq),
 755                                           t4_sq_wq_size(&qp->wq));
 756                 if (ret) {
 757                         printk(KERN_ERR MOD "%s: Fatal error - "
 758                                "DB overflow recovery failed - "
 759                                "error syncing SQ qid %u\n",
 760                                pci_name(ctx->lldi.pdev), qp->wq.sq.qid);
 761                         return;
 762                 }
 763
 764                 ret = cxgb4_sync_txq_pidx(qp->rhp->rdev.lldi.ports[0],
 765                                           qp->wq.rq.qid,
 766                                           t4_rq_host_wq_pidx(&qp->wq),
 767                                           t4_rq_wq_size(&qp->wq));
 768
 769                 if (ret) {
 770                         printk(KERN_ERR MOD "%s: Fatal error - "
 771                                "DB overflow recovery failed - "
 772                                "error syncing RQ qid %u\n",
 773                                pci_name(ctx->lldi.pdev), qp->wq.rq.qid);
 774                         return;
 775                 }
 776
 777                 /* Wait for the dbfifo to drain */
 778                 while (cxgb4_dbfifo_count(qp->rhp->rdev.lldi.ports[0], 1) > 0) {
 779                         set_current_state(TASK_UNINTERRUPTIBLE);
 780                         schedule_timeout(usecs_to_jiffies(10));
 781                 }
 782         }
 783 }
 784
 785 static void recover_queues(struct uld_ctx *ctx)
 786 {
 787         int count = 0;
 788         struct qp_list qp_list;
 789         int ret;
 790
 791         /* lock out kernel db ringers */
 792         mutex_lock(&ctx->dev->db_mutex);
 793
 794         /* put all queues in to recovery mode */
 795         spin_lock_irq(&ctx->dev->lock);
 796         ctx->dev->db_state = RECOVERY;
 797         ctx->dev->rdev.stats.db_state_transitions++;
 798         idr_for_each(&ctx->dev->qpidr, disable_qp_db, NULL);
 799         spin_unlock_irq(&ctx->dev->lock);
 800
 801         /* slow everybody down */
 802         set_current_state(TASK_UNINTERRUPTIBLE);
 803         schedule_timeout(usecs_to_jiffies(1000));
 804
 805         /* Wait for the dbfifo to completely drain. */
 806         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
 807                 set_current_state(TASK_UNINTERRUPTIBLE);
 808                 schedule_timeout(usecs_to_jiffies(10));
 809         }
 810
 811         /* flush the SGE contexts */
 812         ret = cxgb4_flush_eq_cache(ctx->dev->rdev.lldi.ports[0]);
 813         if (ret) {
 814                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
 815                        pci_name(ctx->lldi.pdev));
 816                 goto out;
 817         }
 818
 819         /* Count active queues so we can build a list of queues to recover */
 820         spin_lock_irq(&ctx->dev->lock);
 821         idr_for_each(&ctx->dev->qpidr, count_qps, &count);
 822
 823         qp_list.qps = kzalloc(count * sizeof *qp_list.qps, GFP_ATOMIC);
 824         if (!qp_list.qps) {
 825                 printk(KERN_ERR MOD "%s: Fatal error - DB overflow recovery failed\n",
 826                        pci_name(ctx->lldi.pdev));
 827                 spin_unlock_irq(&ctx->dev->lock);
 828                 goto out;
 829         }
 830         qp_list.idx = 0;
 831
 832         /* add and ref each qp so it doesn't get freed */
 833         idr_for_each(&ctx->dev->qpidr, add_and_ref_qp, &qp_list);
 834
 835         spin_unlock_irq(&ctx->dev->lock);
 836
 837         /* now traverse the list in a safe context to recover the db state*/
 838         recover_lost_dbs(ctx, &qp_list);
 839
 840         /* we're almost done!  deref the qps and clean up */
 841         deref_qps(qp_list);
 842         kfree(qp_list.qps);
 843
 844         /* Wait for the dbfifo to completely drain again */
 845         while (cxgb4_dbfifo_count(ctx->dev->rdev.lldi.ports[0], 1) > 0) {
 846                 set_current_state(TASK_UNINTERRUPTIBLE);
 847                 schedule_timeout(usecs_to_jiffies(10));
 848         }
 849
 850         /* resume the queues */
 851         spin_lock_irq(&ctx->dev->lock);
 852         if (ctx->dev->qpcnt > db_fc_threshold)
 853                 ctx->dev->db_state = FLOW_CONTROL;
 854         else {
 855                 ctx->dev->db_state = NORMAL;
 856                 idr_for_each(&ctx->dev->qpidr, enable_qp_db, NULL);
 857         }
 858         ctx->dev->rdev.stats.db_state_transitions++;
 859         spin_unlock_irq(&ctx->dev->lock);
 860
 861 out:
 862         /* start up kernel db ringers again */
 863         mutex_unlock(&ctx->dev->db_mutex);
 864 }
 865
 866 static int c4iw_uld_control(void *handle, enum cxgb4_control control, ...)
 867 {
 868         struct uld_ctx *ctx = handle;
 869
 870         switch (control) {
 871         case CXGB4_CONTROL_DB_FULL:
 872                 stop_queues(ctx);
 873                 mutex_lock(&ctx->dev->rdev.stats.lock);
 874                 ctx->dev->rdev.stats.db_full++;
 875                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 876                 break;
 877         case CXGB4_CONTROL_DB_EMPTY:
 878                 resume_queues(ctx);
 879                 mutex_lock(&ctx->dev->rdev.stats.lock);
 880                 ctx->dev->rdev.stats.db_empty++;
 881                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 882                 break;
 883         case CXGB4_CONTROL_DB_DROP:
 884                 recover_queues(ctx);
 885                 mutex_lock(&ctx->dev->rdev.stats.lock);
 886                 ctx->dev->rdev.stats.db_drop++;
 887                 mutex_unlock(&ctx->dev->rdev.stats.lock);
 888                 break;
 889         default:
 890                 printk(KERN_WARNING MOD "%s: unknown control cmd %u\n",
 891                        pci_name(ctx->lldi.pdev), control);
 892                 break;
 893         }
 894         return 0;
 895 }
 896
 897 static struct cxgb4_uld_info c4iw_uld_info = {
 898         .name = DRV_NAME,
 899         .add = c4iw_uld_add,
 900         .rx_handler = c4iw_uld_rx_handler,
 901         .state_change = c4iw_uld_state_change,
 902         .control = c4iw_uld_control,
 903 };
 904
 905 static int __init c4iw_init_module(void)
 906 {
 907         int err;
 908
 909         err = c4iw_cm_init();
 910         if (err)
 911                 return err;
 912
 913         c4iw_debugfs_root = debugfs_create_dir(DRV_NAME, NULL);
 914         if (!c4iw_debugfs_root)
 915                 printk(KERN_WARNING MOD
 916                        "could not create debugfs entry, continuing\n");
 917
 918         cxgb4_register_uld(CXGB4_ULD_RDMA, &c4iw_uld_info);
 919
 920         return 0;
 921 }
 922
 923 static void __exit c4iw_exit_module(void)
 924 {
 925         struct uld_ctx *ctx, *tmp;
 926
 927         mutex_lock(&dev_mutex);
 928         list_for_each_entry_safe(ctx, tmp, &uld_ctx_list, entry) {
 929                 if (ctx->dev)
 930                         c4iw_remove(ctx);
 931                 kfree(ctx);
 932         }
 933         mutex_unlock(&dev_mutex);
 934         cxgb4_unregister_uld(CXGB4_ULD_RDMA);
 935         c4iw_cm_term();
 936         debugfs_remove_recursive(c4iw_debugfs_root);
 937 }
 938
 939 module_init(c4iw_init_module);
 940 module_exit(c4iw_exit_module);