Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / infiniband / hw / cxgb3 / iwch_provider.c
diff --git a/kernel/drivers/infiniband/hw/cxgb3/iwch_provider.c b/kernel/drivers/infiniband/hw/cxgb3/iwch_provider.c
new file mode 100644 (file)
index 0000000..811b24a
--- /dev/null
@@ -0,0 +1,1467 @@
+/*
+ * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/device.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/delay.h>
+#include <linux/errno.h>
+#include <linux/list.h>
+#include <linux/sched.h>
+#include <linux/spinlock.h>
+#include <linux/ethtool.h>
+#include <linux/rtnetlink.h>
+#include <linux/inetdevice.h>
+#include <linux/slab.h>
+
+#include <asm/io.h>
+#include <asm/irq.h>
+#include <asm/byteorder.h>
+
+#include <rdma/iw_cm.h>
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_user_verbs.h>
+
+#include "cxio_hal.h"
+#include "iwch.h"
+#include "iwch_provider.h"
+#include "iwch_cm.h"
+#include "iwch_user.h"
+#include "common.h"
+
+static struct ib_ah *iwch_ah_create(struct ib_pd *pd,
+                                   struct ib_ah_attr *ah_attr)
+{
+       return ERR_PTR(-ENOSYS);
+}
+
+static int iwch_ah_destroy(struct ib_ah *ah)
+{
+       return -ENOSYS;
+}
+
+static int iwch_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+       return -ENOSYS;
+}
+
+static int iwch_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+       return -ENOSYS;
+}
+
+static int iwch_process_mad(struct ib_device *ibdev,
+                           int mad_flags,
+                           u8 port_num,
+                           struct ib_wc *in_wc,
+                           struct ib_grh *in_grh,
+                           struct ib_mad *in_mad, struct ib_mad *out_mad)
+{
+       return -ENOSYS;
+}
+
+static int iwch_dealloc_ucontext(struct ib_ucontext *context)
+{
+       struct iwch_dev *rhp = to_iwch_dev(context->device);
+       struct iwch_ucontext *ucontext = to_iwch_ucontext(context);
+       struct iwch_mm_entry *mm, *tmp;
+
+       PDBG("%s context %p\n", __func__, context);
+       list_for_each_entry_safe(mm, tmp, &ucontext->mmaps, entry)
+               kfree(mm);
+       cxio_release_ucontext(&rhp->rdev, &ucontext->uctx);
+       kfree(ucontext);
+       return 0;
+}
+
+static struct ib_ucontext *iwch_alloc_ucontext(struct ib_device *ibdev,
+                                       struct ib_udata *udata)
+{
+       struct iwch_ucontext *context;
+       struct iwch_dev *rhp = to_iwch_dev(ibdev);
+
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+       context = kzalloc(sizeof(*context), GFP_KERNEL);
+       if (!context)
+               return ERR_PTR(-ENOMEM);
+       cxio_init_ucontext(&rhp->rdev, &context->uctx);
+       INIT_LIST_HEAD(&context->mmaps);
+       spin_lock_init(&context->mmap_lock);
+       return &context->ibucontext;
+}
+
+static int iwch_destroy_cq(struct ib_cq *ib_cq)
+{
+       struct iwch_cq *chp;
+
+       PDBG("%s ib_cq %p\n", __func__, ib_cq);
+       chp = to_iwch_cq(ib_cq);
+
+       remove_handle(chp->rhp, &chp->rhp->cqidr, chp->cq.cqid);
+       atomic_dec(&chp->refcnt);
+       wait_event(chp->wait, !atomic_read(&chp->refcnt));
+
+       cxio_destroy_cq(&chp->rhp->rdev, &chp->cq);
+       kfree(chp);
+       return 0;
+}
+
+static struct ib_cq *iwch_create_cq(struct ib_device *ibdev, int entries, int vector,
+                            struct ib_ucontext *ib_context,
+                            struct ib_udata *udata)
+{
+       struct iwch_dev *rhp;
+       struct iwch_cq *chp;
+       struct iwch_create_cq_resp uresp;
+       struct iwch_create_cq_req ureq;
+       struct iwch_ucontext *ucontext = NULL;
+       static int warned;
+       size_t resplen;
+
+       PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
+       rhp = to_iwch_dev(ibdev);
+       chp = kzalloc(sizeof(*chp), GFP_KERNEL);
+       if (!chp)
+               return ERR_PTR(-ENOMEM);
+
+       if (ib_context) {
+               ucontext = to_iwch_ucontext(ib_context);
+               if (!t3a_device(rhp)) {
+                       if (ib_copy_from_udata(&ureq, udata, sizeof (ureq))) {
+                               kfree(chp);
+                               return ERR_PTR(-EFAULT);
+                       }
+                       chp->user_rptr_addr = (u32 __user *)(unsigned long)ureq.user_rptr_addr;
+               }
+       }
+
+       if (t3a_device(rhp)) {
+
+               /*
+                * T3A: Add some fluff to handle extra CQEs inserted
+                * for various errors.
+                * Additional CQE possibilities:
+                *      TERMINATE,
+                *      incoming RDMA WRITE Failures
+                *      incoming RDMA READ REQUEST FAILUREs
+                * NOTE: We cannot ensure the CQ won't overflow.
+                */
+               entries += 16;
+       }
+       entries = roundup_pow_of_two(entries);
+       chp->cq.size_log2 = ilog2(entries);
+
+       if (cxio_create_cq(&rhp->rdev, &chp->cq, !ucontext)) {
+               kfree(chp);
+               return ERR_PTR(-ENOMEM);
+       }
+       chp->rhp = rhp;
+       chp->ibcq.cqe = 1 << chp->cq.size_log2;
+       spin_lock_init(&chp->lock);
+       spin_lock_init(&chp->comp_handler_lock);
+       atomic_set(&chp->refcnt, 1);
+       init_waitqueue_head(&chp->wait);
+       if (insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid)) {
+               cxio_destroy_cq(&chp->rhp->rdev, &chp->cq);
+               kfree(chp);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       if (ucontext) {
+               struct iwch_mm_entry *mm;
+
+               mm = kmalloc(sizeof *mm, GFP_KERNEL);
+               if (!mm) {
+                       iwch_destroy_cq(&chp->ibcq);
+                       return ERR_PTR(-ENOMEM);
+               }
+               uresp.cqid = chp->cq.cqid;
+               uresp.size_log2 = chp->cq.size_log2;
+               spin_lock(&ucontext->mmap_lock);
+               uresp.key = ucontext->key;
+               ucontext->key += PAGE_SIZE;
+               spin_unlock(&ucontext->mmap_lock);
+               mm->key = uresp.key;
+               mm->addr = virt_to_phys(chp->cq.queue);
+               if (udata->outlen < sizeof uresp) {
+                       if (!warned++)
+                               printk(KERN_WARNING MOD "Warning - "
+                                      "downlevel libcxgb3 (non-fatal).\n");
+                       mm->len = PAGE_ALIGN((1UL << uresp.size_log2) *
+                                            sizeof(struct t3_cqe));
+                       resplen = sizeof(struct iwch_create_cq_resp_v0);
+               } else {
+                       mm->len = PAGE_ALIGN(((1UL << uresp.size_log2) + 1) *
+                                            sizeof(struct t3_cqe));
+                       uresp.memsize = mm->len;
+                       uresp.reserved = 0;
+                       resplen = sizeof uresp;
+               }
+               if (ib_copy_to_udata(udata, &uresp, resplen)) {
+                       kfree(mm);
+                       iwch_destroy_cq(&chp->ibcq);
+                       return ERR_PTR(-EFAULT);
+               }
+               insert_mmap(ucontext, mm);
+       }
+       PDBG("created cqid 0x%0x chp %p size 0x%0x, dma_addr 0x%0llx\n",
+            chp->cq.cqid, chp, (1 << chp->cq.size_log2),
+            (unsigned long long) chp->cq.dma_addr);
+       return &chp->ibcq;
+}
+
+static int iwch_resize_cq(struct ib_cq *cq, int cqe, struct ib_udata *udata)
+{
+#ifdef notyet
+       struct iwch_cq *chp = to_iwch_cq(cq);
+       struct t3_cq oldcq, newcq;
+       int ret;
+
+       PDBG("%s ib_cq %p cqe %d\n", __func__, cq, cqe);
+
+       /* We don't downsize... */
+       if (cqe <= cq->cqe)
+               return 0;
+
+       /* create new t3_cq with new size */
+       cqe = roundup_pow_of_two(cqe+1);
+       newcq.size_log2 = ilog2(cqe);
+
+       /* Dont allow resize to less than the current wce count */
+       if (cqe < Q_COUNT(chp->cq.rptr, chp->cq.wptr)) {
+               return -ENOMEM;
+       }
+
+       /* Quiesce all QPs using this CQ */
+       ret = iwch_quiesce_qps(chp);
+       if (ret) {
+               return ret;
+       }
+
+       ret = cxio_create_cq(&chp->rhp->rdev, &newcq);
+       if (ret) {
+               return ret;
+       }
+
+       /* copy CQEs */
+       memcpy(newcq.queue, chp->cq.queue, (1 << chp->cq.size_log2) *
+                                       sizeof(struct t3_cqe));
+
+       /* old iwch_qp gets new t3_cq but keeps old cqid */
+       oldcq = chp->cq;
+       chp->cq = newcq;
+       chp->cq.cqid = oldcq.cqid;
+
+       /* resize new t3_cq to update the HW context */
+       ret = cxio_resize_cq(&chp->rhp->rdev, &chp->cq);
+       if (ret) {
+               chp->cq = oldcq;
+               return ret;
+       }
+       chp->ibcq.cqe = (1<<chp->cq.size_log2) - 1;
+
+       /* destroy old t3_cq */
+       oldcq.cqid = newcq.cqid;
+       ret = cxio_destroy_cq(&chp->rhp->rdev, &oldcq);
+       if (ret) {
+               printk(KERN_ERR MOD "%s - cxio_destroy_cq failed %d\n",
+                       __func__, ret);
+       }
+
+       /* add user hooks here */
+
+       /* resume qps */
+       ret = iwch_resume_qps(chp);
+       return ret;
+#else
+       return -ENOSYS;
+#endif
+}
+
+static int iwch_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
+{
+       struct iwch_dev *rhp;
+       struct iwch_cq *chp;
+       enum t3_cq_opcode cq_op;
+       int err;
+       unsigned long flag;
+       u32 rptr;
+
+       chp = to_iwch_cq(ibcq);
+       rhp = chp->rhp;
+       if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
+               cq_op = CQ_ARM_SE;
+       else
+               cq_op = CQ_ARM_AN;
+       if (chp->user_rptr_addr) {
+               if (get_user(rptr, chp->user_rptr_addr))
+                       return -EFAULT;
+               spin_lock_irqsave(&chp->lock, flag);
+               chp->cq.rptr = rptr;
+       } else
+               spin_lock_irqsave(&chp->lock, flag);
+       PDBG("%s rptr 0x%x\n", __func__, chp->cq.rptr);
+       err = cxio_hal_cq_op(&rhp->rdev, &chp->cq, cq_op, 0);
+       spin_unlock_irqrestore(&chp->lock, flag);
+       if (err < 0)
+               printk(KERN_ERR MOD "Error %d rearming CQID 0x%x\n", err,
+                      chp->cq.cqid);
+       if (err > 0 && !(flags & IB_CQ_REPORT_MISSED_EVENTS))
+               err = 0;
+       return err;
+}
+
+static int iwch_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
+{
+       int len = vma->vm_end - vma->vm_start;
+       u32 key = vma->vm_pgoff << PAGE_SHIFT;
+       struct cxio_rdev *rdev_p;
+       int ret = 0;
+       struct iwch_mm_entry *mm;
+       struct iwch_ucontext *ucontext;
+       u64 addr;
+
+       PDBG("%s pgoff 0x%lx key 0x%x len %d\n", __func__, vma->vm_pgoff,
+            key, len);
+
+       if (vma->vm_start & (PAGE_SIZE-1)) {
+               return -EINVAL;
+       }
+
+       rdev_p = &(to_iwch_dev(context->device)->rdev);
+       ucontext = to_iwch_ucontext(context);
+
+       mm = remove_mmap(ucontext, key, len);
+       if (!mm)
+               return -EINVAL;
+       addr = mm->addr;
+       kfree(mm);
+
+       if ((addr >= rdev_p->rnic_info.udbell_physbase) &&
+           (addr < (rdev_p->rnic_info.udbell_physbase +
+                      rdev_p->rnic_info.udbell_len))) {
+
+               /*
+                * Map T3 DB register.
+                */
+               if (vma->vm_flags & VM_READ) {
+                       return -EPERM;
+               }
+
+               vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+               vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
+               vma->vm_flags &= ~VM_MAYREAD;
+               ret = io_remap_pfn_range(vma, vma->vm_start,
+                                        addr >> PAGE_SHIFT,
+                                        len, vma->vm_page_prot);
+       } else {
+
+               /*
+                * Map WQ or CQ contig dma memory...
+                */
+               ret = remap_pfn_range(vma, vma->vm_start,
+                                     addr >> PAGE_SHIFT,
+                                     len, vma->vm_page_prot);
+       }
+
+       return ret;
+}
+
+static int iwch_deallocate_pd(struct ib_pd *pd)
+{
+       struct iwch_dev *rhp;
+       struct iwch_pd *php;
+
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+       PDBG("%s ibpd %p pdid 0x%x\n", __func__, pd, php->pdid);
+       cxio_hal_put_pdid(rhp->rdev.rscp, php->pdid);
+       kfree(php);
+       return 0;
+}
+
+static struct ib_pd *iwch_allocate_pd(struct ib_device *ibdev,
+                              struct ib_ucontext *context,
+                              struct ib_udata *udata)
+{
+       struct iwch_pd *php;
+       u32 pdid;
+       struct iwch_dev *rhp;
+
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+       rhp = (struct iwch_dev *) ibdev;
+       pdid = cxio_hal_get_pdid(rhp->rdev.rscp);
+       if (!pdid)
+               return ERR_PTR(-EINVAL);
+       php = kzalloc(sizeof(*php), GFP_KERNEL);
+       if (!php) {
+               cxio_hal_put_pdid(rhp->rdev.rscp, pdid);
+               return ERR_PTR(-ENOMEM);
+       }
+       php->pdid = pdid;
+       php->rhp = rhp;
+       if (context) {
+               if (ib_copy_to_udata(udata, &php->pdid, sizeof (__u32))) {
+                       iwch_deallocate_pd(&php->ibpd);
+                       return ERR_PTR(-EFAULT);
+               }
+       }
+       PDBG("%s pdid 0x%0x ptr 0x%p\n", __func__, pdid, php);
+       return &php->ibpd;
+}
+
+static int iwch_dereg_mr(struct ib_mr *ib_mr)
+{
+       struct iwch_dev *rhp;
+       struct iwch_mr *mhp;
+       u32 mmid;
+
+       PDBG("%s ib_mr %p\n", __func__, ib_mr);
+       /* There can be no memory windows */
+       if (atomic_read(&ib_mr->usecnt))
+               return -EINVAL;
+
+       mhp = to_iwch_mr(ib_mr);
+       rhp = mhp->rhp;
+       mmid = mhp->attr.stag >> 8;
+       cxio_dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
+                      mhp->attr.pbl_addr);
+       iwch_free_pbl(mhp);
+       remove_handle(rhp, &rhp->mmidr, mmid);
+       if (mhp->kva)
+               kfree((void *) (unsigned long) mhp->kva);
+       if (mhp->umem)
+               ib_umem_release(mhp->umem);
+       PDBG("%s mmid 0x%x ptr %p\n", __func__, mmid, mhp);
+       kfree(mhp);
+       return 0;
+}
+
+static struct ib_mr *iwch_register_phys_mem(struct ib_pd *pd,
+                                       struct ib_phys_buf *buffer_list,
+                                       int num_phys_buf,
+                                       int acc,
+                                       u64 *iova_start)
+{
+       __be64 *page_list;
+       int shift;
+       u64 total_size;
+       int npages;
+       struct iwch_dev *rhp;
+       struct iwch_pd *php;
+       struct iwch_mr *mhp;
+       int ret;
+
+       PDBG("%s ib_pd %p\n", __func__, pd);
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+
+       mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
+       if (!mhp)
+               return ERR_PTR(-ENOMEM);
+
+       mhp->rhp = rhp;
+
+       /* First check that we have enough alignment */
+       if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK)) {
+               ret = -EINVAL;
+               goto err;
+       }
+
+       if (num_phys_buf > 1 &&
+           ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK)) {
+               ret = -EINVAL;
+               goto err;
+       }
+
+       ret = build_phys_page_list(buffer_list, num_phys_buf, iova_start,
+                                  &total_size, &npages, &shift, &page_list);
+       if (ret)
+               goto err;
+
+       ret = iwch_alloc_pbl(mhp, npages);
+       if (ret) {
+               kfree(page_list);
+               goto err_pbl;
+       }
+
+       ret = iwch_write_pbl(mhp, page_list, npages, 0);
+       kfree(page_list);
+       if (ret)
+               goto err_pbl;
+
+       mhp->attr.pdid = php->pdid;
+       mhp->attr.zbva = 0;
+
+       mhp->attr.perms = iwch_ib_to_tpt_access(acc);
+       mhp->attr.va_fbo = *iova_start;
+       mhp->attr.page_size = shift - 12;
+
+       mhp->attr.len = (u32) total_size;
+       mhp->attr.pbl_size = npages;
+       ret = iwch_register_mem(rhp, php, mhp, shift);
+       if (ret)
+               goto err_pbl;
+
+       return &mhp->ibmr;
+
+err_pbl:
+       iwch_free_pbl(mhp);
+
+err:
+       kfree(mhp);
+       return ERR_PTR(ret);
+
+}
+
+static int iwch_reregister_phys_mem(struct ib_mr *mr,
+                                    int mr_rereg_mask,
+                                    struct ib_pd *pd,
+                                    struct ib_phys_buf *buffer_list,
+                                    int num_phys_buf,
+                                    int acc, u64 * iova_start)
+{
+
+       struct iwch_mr mh, *mhp;
+       struct iwch_pd *php;
+       struct iwch_dev *rhp;
+       __be64 *page_list = NULL;
+       int shift = 0;
+       u64 total_size;
+       int npages = 0;
+       int ret;
+
+       PDBG("%s ib_mr %p ib_pd %p\n", __func__, mr, pd);
+
+       /* There can be no memory windows */
+       if (atomic_read(&mr->usecnt))
+               return -EINVAL;
+
+       mhp = to_iwch_mr(mr);
+       rhp = mhp->rhp;
+       php = to_iwch_pd(mr->pd);
+
+       /* make sure we are on the same adapter */
+       if (rhp != php->rhp)
+               return -EINVAL;
+
+       memcpy(&mh, mhp, sizeof *mhp);
+
+       if (mr_rereg_mask & IB_MR_REREG_PD)
+               php = to_iwch_pd(pd);
+       if (mr_rereg_mask & IB_MR_REREG_ACCESS)
+               mh.attr.perms = iwch_ib_to_tpt_access(acc);
+       if (mr_rereg_mask & IB_MR_REREG_TRANS) {
+               ret = build_phys_page_list(buffer_list, num_phys_buf,
+                                          iova_start,
+                                          &total_size, &npages,
+                                          &shift, &page_list);
+               if (ret)
+                       return ret;
+       }
+
+       ret = iwch_reregister_mem(rhp, php, &mh, shift, npages);
+       kfree(page_list);
+       if (ret) {
+               return ret;
+       }
+       if (mr_rereg_mask & IB_MR_REREG_PD)
+               mhp->attr.pdid = php->pdid;
+       if (mr_rereg_mask & IB_MR_REREG_ACCESS)
+               mhp->attr.perms = iwch_ib_to_tpt_access(acc);
+       if (mr_rereg_mask & IB_MR_REREG_TRANS) {
+               mhp->attr.zbva = 0;
+               mhp->attr.va_fbo = *iova_start;
+               mhp->attr.page_size = shift - 12;
+               mhp->attr.len = (u32) total_size;
+               mhp->attr.pbl_size = npages;
+       }
+
+       return 0;
+}
+
+
+static struct ib_mr *iwch_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+                                     u64 virt, int acc, struct ib_udata *udata)
+{
+       __be64 *pages;
+       int shift, n, len;
+       int i, k, entry;
+       int err = 0;
+       struct iwch_dev *rhp;
+       struct iwch_pd *php;
+       struct iwch_mr *mhp;
+       struct iwch_reg_user_mr_resp uresp;
+       struct scatterlist *sg;
+       PDBG("%s ib_pd %p\n", __func__, pd);
+
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+       mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
+       if (!mhp)
+               return ERR_PTR(-ENOMEM);
+
+       mhp->rhp = rhp;
+
+       mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
+       if (IS_ERR(mhp->umem)) {
+               err = PTR_ERR(mhp->umem);
+               kfree(mhp);
+               return ERR_PTR(err);
+       }
+
+       shift = ffs(mhp->umem->page_size) - 1;
+
+       n = mhp->umem->nmap;
+
+       err = iwch_alloc_pbl(mhp, n);
+       if (err)
+               goto err;
+
+       pages = (__be64 *) __get_free_page(GFP_KERNEL);
+       if (!pages) {
+               err = -ENOMEM;
+               goto err_pbl;
+       }
+
+       i = n = 0;
+
+       for_each_sg(mhp->umem->sg_head.sgl, sg, mhp->umem->nmap, entry) {
+                       len = sg_dma_len(sg) >> shift;
+                       for (k = 0; k < len; ++k) {
+                               pages[i++] = cpu_to_be64(sg_dma_address(sg) +
+                                       mhp->umem->page_size * k);
+                               if (i == PAGE_SIZE / sizeof *pages) {
+                                       err = iwch_write_pbl(mhp, pages, i, n);
+                                       if (err)
+                                               goto pbl_done;
+                                       n += i;
+                                       i = 0;
+                               }
+                       }
+       }
+
+       if (i)
+               err = iwch_write_pbl(mhp, pages, i, n);
+
+pbl_done:
+       free_page((unsigned long) pages);
+       if (err)
+               goto err_pbl;
+
+       mhp->attr.pdid = php->pdid;
+       mhp->attr.zbva = 0;
+       mhp->attr.perms = iwch_ib_to_tpt_access(acc);
+       mhp->attr.va_fbo = virt;
+       mhp->attr.page_size = shift - 12;
+       mhp->attr.len = (u32) length;
+
+       err = iwch_register_mem(rhp, php, mhp, shift);
+       if (err)
+               goto err_pbl;
+
+       if (udata && !t3a_device(rhp)) {
+               uresp.pbl_addr = (mhp->attr.pbl_addr -
+                                rhp->rdev.rnic_info.pbl_base) >> 3;
+               PDBG("%s user resp pbl_addr 0x%x\n", __func__,
+                    uresp.pbl_addr);
+
+               if (ib_copy_to_udata(udata, &uresp, sizeof (uresp))) {
+                       iwch_dereg_mr(&mhp->ibmr);
+                       err = -EFAULT;
+                       goto err;
+               }
+       }
+
+       return &mhp->ibmr;
+
+err_pbl:
+       iwch_free_pbl(mhp);
+
+err:
+       ib_umem_release(mhp->umem);
+       kfree(mhp);
+       return ERR_PTR(err);
+}
+
+static struct ib_mr *iwch_get_dma_mr(struct ib_pd *pd, int acc)
+{
+       struct ib_phys_buf bl;
+       u64 kva;
+       struct ib_mr *ibmr;
+
+       PDBG("%s ib_pd %p\n", __func__, pd);
+
+       /*
+        * T3 only supports 32 bits of size.
+        */
+       bl.size = 0xffffffff;
+       bl.addr = 0;
+       kva = 0;
+       ibmr = iwch_register_phys_mem(pd, &bl, 1, acc, &kva);
+       return ibmr;
+}
+
+static struct ib_mw *iwch_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
+{
+       struct iwch_dev *rhp;
+       struct iwch_pd *php;
+       struct iwch_mw *mhp;
+       u32 mmid;
+       u32 stag = 0;
+       int ret;
+
+       if (type != IB_MW_TYPE_1)
+               return ERR_PTR(-EINVAL);
+
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+       mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
+       if (!mhp)
+               return ERR_PTR(-ENOMEM);
+       ret = cxio_allocate_window(&rhp->rdev, &stag, php->pdid);
+       if (ret) {
+               kfree(mhp);
+               return ERR_PTR(ret);
+       }
+       mhp->rhp = rhp;
+       mhp->attr.pdid = php->pdid;
+       mhp->attr.type = TPT_MW;
+       mhp->attr.stag = stag;
+       mmid = (stag) >> 8;
+       mhp->ibmw.rkey = stag;
+       if (insert_handle(rhp, &rhp->mmidr, mhp, mmid)) {
+               cxio_deallocate_window(&rhp->rdev, mhp->attr.stag);
+               kfree(mhp);
+               return ERR_PTR(-ENOMEM);
+       }
+       PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
+       return &(mhp->ibmw);
+}
+
+static int iwch_dealloc_mw(struct ib_mw *mw)
+{
+       struct iwch_dev *rhp;
+       struct iwch_mw *mhp;
+       u32 mmid;
+
+       mhp = to_iwch_mw(mw);
+       rhp = mhp->rhp;
+       mmid = (mw->rkey) >> 8;
+       cxio_deallocate_window(&rhp->rdev, mhp->attr.stag);
+       remove_handle(rhp, &rhp->mmidr, mmid);
+       PDBG("%s ib_mw %p mmid 0x%x ptr %p\n", __func__, mw, mmid, mhp);
+       kfree(mhp);
+       return 0;
+}
+
+static struct ib_mr *iwch_alloc_fast_reg_mr(struct ib_pd *pd, int pbl_depth)
+{
+       struct iwch_dev *rhp;
+       struct iwch_pd *php;
+       struct iwch_mr *mhp;
+       u32 mmid;
+       u32 stag = 0;
+       int ret = 0;
+
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+       mhp = kzalloc(sizeof(*mhp), GFP_KERNEL);
+       if (!mhp)
+               goto err;
+
+       mhp->rhp = rhp;
+       ret = iwch_alloc_pbl(mhp, pbl_depth);
+       if (ret)
+               goto err1;
+       mhp->attr.pbl_size = pbl_depth;
+       ret = cxio_allocate_stag(&rhp->rdev, &stag, php->pdid,
+                                mhp->attr.pbl_size, mhp->attr.pbl_addr);
+       if (ret)
+               goto err2;
+       mhp->attr.pdid = php->pdid;
+       mhp->attr.type = TPT_NON_SHARED_MR;
+       mhp->attr.stag = stag;
+       mhp->attr.state = 1;
+       mmid = (stag) >> 8;
+       mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
+       if (insert_handle(rhp, &rhp->mmidr, mhp, mmid))
+               goto err3;
+
+       PDBG("%s mmid 0x%x mhp %p stag 0x%x\n", __func__, mmid, mhp, stag);
+       return &(mhp->ibmr);
+err3:
+       cxio_dereg_mem(&rhp->rdev, stag, mhp->attr.pbl_size,
+                      mhp->attr.pbl_addr);
+err2:
+       iwch_free_pbl(mhp);
+err1:
+       kfree(mhp);
+err:
+       return ERR_PTR(ret);
+}
+
+static struct ib_fast_reg_page_list *iwch_alloc_fastreg_pbl(
+                                       struct ib_device *device,
+                                       int page_list_len)
+{
+       struct ib_fast_reg_page_list *page_list;
+
+       page_list = kmalloc(sizeof *page_list + page_list_len * sizeof(u64),
+                           GFP_KERNEL);
+       if (!page_list)
+               return ERR_PTR(-ENOMEM);
+
+       page_list->page_list = (u64 *)(page_list + 1);
+       page_list->max_page_list_len = page_list_len;
+
+       return page_list;
+}
+
+static void iwch_free_fastreg_pbl(struct ib_fast_reg_page_list *page_list)
+{
+       kfree(page_list);
+}
+
+static int iwch_destroy_qp(struct ib_qp *ib_qp)
+{
+       struct iwch_dev *rhp;
+       struct iwch_qp *qhp;
+       struct iwch_qp_attributes attrs;
+       struct iwch_ucontext *ucontext;
+
+       qhp = to_iwch_qp(ib_qp);
+       rhp = qhp->rhp;
+
+       attrs.next_state = IWCH_QP_STATE_ERROR;
+       iwch_modify_qp(rhp, qhp, IWCH_QP_ATTR_NEXT_STATE, &attrs, 0);
+       wait_event(qhp->wait, !qhp->ep);
+
+       remove_handle(rhp, &rhp->qpidr, qhp->wq.qpid);
+
+       atomic_dec(&qhp->refcnt);
+       wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
+
+       ucontext = ib_qp->uobject ? to_iwch_ucontext(ib_qp->uobject->context)
+                                 : NULL;
+       cxio_destroy_qp(&rhp->rdev, &qhp->wq,
+                       ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
+
+       PDBG("%s ib_qp %p qpid 0x%0x qhp %p\n", __func__,
+            ib_qp, qhp->wq.qpid, qhp);
+       kfree(qhp);
+       return 0;
+}
+
+static struct ib_qp *iwch_create_qp(struct ib_pd *pd,
+                            struct ib_qp_init_attr *attrs,
+                            struct ib_udata *udata)
+{
+       struct iwch_dev *rhp;
+       struct iwch_qp *qhp;
+       struct iwch_pd *php;
+       struct iwch_cq *schp;
+       struct iwch_cq *rchp;
+       struct iwch_create_qp_resp uresp;
+       int wqsize, sqsize, rqsize;
+       struct iwch_ucontext *ucontext;
+
+       PDBG("%s ib_pd %p\n", __func__, pd);
+       if (attrs->qp_type != IB_QPT_RC)
+               return ERR_PTR(-EINVAL);
+       php = to_iwch_pd(pd);
+       rhp = php->rhp;
+       schp = get_chp(rhp, ((struct iwch_cq *) attrs->send_cq)->cq.cqid);
+       rchp = get_chp(rhp, ((struct iwch_cq *) attrs->recv_cq)->cq.cqid);
+       if (!schp || !rchp)
+               return ERR_PTR(-EINVAL);
+
+       /* The RQT size must be # of entries + 1 rounded up to a power of two */
+       rqsize = roundup_pow_of_two(attrs->cap.max_recv_wr);
+       if (rqsize == attrs->cap.max_recv_wr)
+               rqsize = roundup_pow_of_two(attrs->cap.max_recv_wr+1);
+
+       /* T3 doesn't support RQT depth < 16 */
+       if (rqsize < 16)
+               rqsize = 16;
+
+       if (rqsize > T3_MAX_RQ_SIZE)
+               return ERR_PTR(-EINVAL);
+
+       if (attrs->cap.max_inline_data > T3_MAX_INLINE)
+               return ERR_PTR(-EINVAL);
+
+       /*
+        * NOTE: The SQ and total WQ sizes don't need to be
+        * a power of two.  However, all the code assumes
+        * they are. EG: Q_FREECNT() and friends.
+        */
+       sqsize = roundup_pow_of_two(attrs->cap.max_send_wr);
+       wqsize = roundup_pow_of_two(rqsize + sqsize);
+
+       /*
+        * Kernel users need more wq space for fastreg WRs which can take
+        * 2 WR fragments.
+        */
+       ucontext = pd->uobject ? to_iwch_ucontext(pd->uobject->context) : NULL;
+       if (!ucontext && wqsize < (rqsize + (2 * sqsize)))
+               wqsize = roundup_pow_of_two(rqsize +
+                               roundup_pow_of_two(attrs->cap.max_send_wr * 2));
+       PDBG("%s wqsize %d sqsize %d rqsize %d\n", __func__,
+            wqsize, sqsize, rqsize);
+       qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
+       if (!qhp)
+               return ERR_PTR(-ENOMEM);
+       qhp->wq.size_log2 = ilog2(wqsize);
+       qhp->wq.rq_size_log2 = ilog2(rqsize);
+       qhp->wq.sq_size_log2 = ilog2(sqsize);
+       if (cxio_create_qp(&rhp->rdev, !udata, &qhp->wq,
+                          ucontext ? &ucontext->uctx : &rhp->rdev.uctx)) {
+               kfree(qhp);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       attrs->cap.max_recv_wr = rqsize - 1;
+       attrs->cap.max_send_wr = sqsize;
+       attrs->cap.max_inline_data = T3_MAX_INLINE;
+
+       qhp->rhp = rhp;
+       qhp->attr.pd = php->pdid;
+       qhp->attr.scq = ((struct iwch_cq *) attrs->send_cq)->cq.cqid;
+       qhp->attr.rcq = ((struct iwch_cq *) attrs->recv_cq)->cq.cqid;
+       qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
+       qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
+       qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
+       qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
+       qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
+       qhp->attr.state = IWCH_QP_STATE_IDLE;
+       qhp->attr.next_state = IWCH_QP_STATE_IDLE;
+
+       /*
+        * XXX - These don't get passed in from the openib user
+        * at create time.  The CM sets them via a QP modify.
+        * Need to fix...  I think the CM should
+        */
+       qhp->attr.enable_rdma_read = 1;
+       qhp->attr.enable_rdma_write = 1;
+       qhp->attr.enable_bind = 1;
+       qhp->attr.max_ord = 1;
+       qhp->attr.max_ird = 1;
+
+       spin_lock_init(&qhp->lock);
+       init_waitqueue_head(&qhp->wait);
+       atomic_set(&qhp->refcnt, 1);
+
+       if (insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.qpid)) {
+               cxio_destroy_qp(&rhp->rdev, &qhp->wq,
+                       ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
+               kfree(qhp);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       if (udata) {
+
+               struct iwch_mm_entry *mm1, *mm2;
+
+               mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
+               if (!mm1) {
+                       iwch_destroy_qp(&qhp->ibqp);
+                       return ERR_PTR(-ENOMEM);
+               }
+
+               mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
+               if (!mm2) {
+                       kfree(mm1);
+                       iwch_destroy_qp(&qhp->ibqp);
+                       return ERR_PTR(-ENOMEM);
+               }
+
+               uresp.qpid = qhp->wq.qpid;
+               uresp.size_log2 = qhp->wq.size_log2;
+               uresp.sq_size_log2 = qhp->wq.sq_size_log2;
+               uresp.rq_size_log2 = qhp->wq.rq_size_log2;
+               spin_lock(&ucontext->mmap_lock);
+               uresp.key = ucontext->key;
+               ucontext->key += PAGE_SIZE;
+               uresp.db_key = ucontext->key;
+               ucontext->key += PAGE_SIZE;
+               spin_unlock(&ucontext->mmap_lock);
+               if (ib_copy_to_udata(udata, &uresp, sizeof (uresp))) {
+                       kfree(mm1);
+                       kfree(mm2);
+                       iwch_destroy_qp(&qhp->ibqp);
+                       return ERR_PTR(-EFAULT);
+               }
+               mm1->key = uresp.key;
+               mm1->addr = virt_to_phys(qhp->wq.queue);
+               mm1->len = PAGE_ALIGN(wqsize * sizeof (union t3_wr));
+               insert_mmap(ucontext, mm1);
+               mm2->key = uresp.db_key;
+               mm2->addr = qhp->wq.udb & PAGE_MASK;
+               mm2->len = PAGE_SIZE;
+               insert_mmap(ucontext, mm2);
+       }
+       qhp->ibqp.qp_num = qhp->wq.qpid;
+       init_timer(&(qhp->timer));
+       PDBG("%s sq_num_entries %d, rq_num_entries %d "
+            "qpid 0x%0x qhp %p dma_addr 0x%llx size %d rq_addr 0x%x\n",
+            __func__, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
+            qhp->wq.qpid, qhp, (unsigned long long) qhp->wq.dma_addr,
+            1 << qhp->wq.size_log2, qhp->wq.rq_addr);
+       return &qhp->ibqp;
+}
+
+static int iwch_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+                     int attr_mask, struct ib_udata *udata)
+{
+       struct iwch_dev *rhp;
+       struct iwch_qp *qhp;
+       enum iwch_qp_attr_mask mask = 0;
+       struct iwch_qp_attributes attrs;
+
+       PDBG("%s ib_qp %p\n", __func__, ibqp);
+
+       /* iwarp does not support the RTR state */
+       if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
+               attr_mask &= ~IB_QP_STATE;
+
+       /* Make sure we still have something left to do */
+       if (!attr_mask)
+               return 0;
+
+       memset(&attrs, 0, sizeof attrs);
+       qhp = to_iwch_qp(ibqp);
+       rhp = qhp->rhp;
+
+       attrs.next_state = iwch_convert_state(attr->qp_state);
+       attrs.enable_rdma_read = (attr->qp_access_flags &
+                              IB_ACCESS_REMOTE_READ) ?  1 : 0;
+       attrs.enable_rdma_write = (attr->qp_access_flags &
+                               IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
+       attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
+
+
+       mask |= (attr_mask & IB_QP_STATE) ? IWCH_QP_ATTR_NEXT_STATE : 0;
+       mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
+                       (IWCH_QP_ATTR_ENABLE_RDMA_READ |
+                        IWCH_QP_ATTR_ENABLE_RDMA_WRITE |
+                        IWCH_QP_ATTR_ENABLE_RDMA_BIND) : 0;
+
+       return iwch_modify_qp(rhp, qhp, mask, &attrs, 0);
+}
+
+void iwch_qp_add_ref(struct ib_qp *qp)
+{
+       PDBG("%s ib_qp %p\n", __func__, qp);
+       atomic_inc(&(to_iwch_qp(qp)->refcnt));
+}
+
+void iwch_qp_rem_ref(struct ib_qp *qp)
+{
+       PDBG("%s ib_qp %p\n", __func__, qp);
+       if (atomic_dec_and_test(&(to_iwch_qp(qp)->refcnt)))
+               wake_up(&(to_iwch_qp(qp)->wait));
+}
+
+static struct ib_qp *iwch_get_qp(struct ib_device *dev, int qpn)
+{
+       PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
+       return (struct ib_qp *)get_qhp(to_iwch_dev(dev), qpn);
+}
+
+
+static int iwch_query_pkey(struct ib_device *ibdev,
+                          u8 port, u16 index, u16 * pkey)
+{
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+       *pkey = 0;
+       return 0;
+}
+
+static int iwch_query_gid(struct ib_device *ibdev, u8 port,
+                         int index, union ib_gid *gid)
+{
+       struct iwch_dev *dev;
+
+       PDBG("%s ibdev %p, port %d, index %d, gid %p\n",
+              __func__, ibdev, port, index, gid);
+       dev = to_iwch_dev(ibdev);
+       BUG_ON(port == 0 || port > 2);
+       memset(&(gid->raw[0]), 0, sizeof(gid->raw));
+       memcpy(&(gid->raw[0]), dev->rdev.port_info.lldevs[port-1]->dev_addr, 6);
+       return 0;
+}
+
+static u64 fw_vers_string_to_u64(struct iwch_dev *iwch_dev)
+{
+       struct ethtool_drvinfo info;
+       struct net_device *lldev = iwch_dev->rdev.t3cdev_p->lldev;
+       char *cp, *next;
+       unsigned fw_maj, fw_min, fw_mic;
+
+       lldev->ethtool_ops->get_drvinfo(lldev, &info);
+
+       next = info.fw_version + 1;
+       cp = strsep(&next, ".");
+       sscanf(cp, "%i", &fw_maj);
+       cp = strsep(&next, ".");
+       sscanf(cp, "%i", &fw_min);
+       cp = strsep(&next, ".");
+       sscanf(cp, "%i", &fw_mic);
+
+       return (((u64)fw_maj & 0xffff) << 32) | ((fw_min & 0xffff) << 16) |
+              (fw_mic & 0xffff);
+}
+
+static int iwch_query_device(struct ib_device *ibdev,
+                            struct ib_device_attr *props)
+{
+
+       struct iwch_dev *dev;
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+
+       dev = to_iwch_dev(ibdev);
+       memset(props, 0, sizeof *props);
+       memcpy(&props->sys_image_guid, dev->rdev.t3cdev_p->lldev->dev_addr, 6);
+       props->hw_ver = dev->rdev.t3cdev_p->type;
+       props->fw_ver = fw_vers_string_to_u64(dev);
+       props->device_cap_flags = dev->device_cap_flags;
+       props->page_size_cap = dev->attr.mem_pgsizes_bitmask;
+       props->vendor_id = (u32)dev->rdev.rnic_info.pdev->vendor;
+       props->vendor_part_id = (u32)dev->rdev.rnic_info.pdev->device;
+       props->max_mr_size = dev->attr.max_mr_size;
+       props->max_qp = dev->attr.max_qps;
+       props->max_qp_wr = dev->attr.max_wrs;
+       props->max_sge = dev->attr.max_sge_per_wr;
+       props->max_sge_rd = 1;
+       props->max_qp_rd_atom = dev->attr.max_rdma_reads_per_qp;
+       props->max_qp_init_rd_atom = dev->attr.max_rdma_reads_per_qp;
+       props->max_cq = dev->attr.max_cqs;
+       props->max_cqe = dev->attr.max_cqes_per_cq;
+       props->max_mr = dev->attr.max_mem_regs;
+       props->max_pd = dev->attr.max_pds;
+       props->local_ca_ack_delay = 0;
+       props->max_fast_reg_page_list_len = T3_MAX_FASTREG_DEPTH;
+
+       return 0;
+}
+
+static int iwch_query_port(struct ib_device *ibdev,
+                          u8 port, struct ib_port_attr *props)
+{
+       struct iwch_dev *dev;
+       struct net_device *netdev;
+       struct in_device *inetdev;
+
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+
+       dev = to_iwch_dev(ibdev);
+       netdev = dev->rdev.port_info.lldevs[port-1];
+
+       memset(props, 0, sizeof(struct ib_port_attr));
+       props->max_mtu = IB_MTU_4096;
+       if (netdev->mtu >= 4096)
+               props->active_mtu = IB_MTU_4096;
+       else if (netdev->mtu >= 2048)
+               props->active_mtu = IB_MTU_2048;
+       else if (netdev->mtu >= 1024)
+               props->active_mtu = IB_MTU_1024;
+       else if (netdev->mtu >= 512)
+               props->active_mtu = IB_MTU_512;
+       else
+               props->active_mtu = IB_MTU_256;
+
+       if (!netif_carrier_ok(netdev))
+               props->state = IB_PORT_DOWN;
+       else {
+               inetdev = in_dev_get(netdev);
+               if (inetdev) {
+                       if (inetdev->ifa_list)
+                               props->state = IB_PORT_ACTIVE;
+                       else
+                               props->state = IB_PORT_INIT;
+                       in_dev_put(inetdev);
+               } else
+                       props->state = IB_PORT_INIT;
+       }
+
+       props->port_cap_flags =
+           IB_PORT_CM_SUP |
+           IB_PORT_SNMP_TUNNEL_SUP |
+           IB_PORT_REINIT_SUP |
+           IB_PORT_DEVICE_MGMT_SUP |
+           IB_PORT_VENDOR_CLASS_SUP | IB_PORT_BOOT_MGMT_SUP;
+       props->gid_tbl_len = 1;
+       props->pkey_tbl_len = 1;
+       props->active_width = 2;
+       props->active_speed = IB_SPEED_DDR;
+       props->max_msg_sz = -1;
+
+       return 0;
+}
+
+static ssize_t show_rev(struct device *dev, struct device_attribute *attr,
+                       char *buf)
+{
+       struct iwch_dev *iwch_dev = container_of(dev, struct iwch_dev,
+                                                ibdev.dev);
+       PDBG("%s dev 0x%p\n", __func__, dev);
+       return sprintf(buf, "%d\n", iwch_dev->rdev.t3cdev_p->type);
+}
+
+static ssize_t show_fw_ver(struct device *dev, struct device_attribute *attr, char *buf)
+{
+       struct iwch_dev *iwch_dev = container_of(dev, struct iwch_dev,
+                                                ibdev.dev);
+       struct ethtool_drvinfo info;
+       struct net_device *lldev = iwch_dev->rdev.t3cdev_p->lldev;
+
+       PDBG("%s dev 0x%p\n", __func__, dev);
+       lldev->ethtool_ops->get_drvinfo(lldev, &info);
+       return sprintf(buf, "%s\n", info.fw_version);
+}
+
+static ssize_t show_hca(struct device *dev, struct device_attribute *attr,
+                       char *buf)
+{
+       struct iwch_dev *iwch_dev = container_of(dev, struct iwch_dev,
+                                                ibdev.dev);
+       struct ethtool_drvinfo info;
+       struct net_device *lldev = iwch_dev->rdev.t3cdev_p->lldev;
+
+       PDBG("%s dev 0x%p\n", __func__, dev);
+       lldev->ethtool_ops->get_drvinfo(lldev, &info);
+       return sprintf(buf, "%s\n", info.driver);
+}
+
+static ssize_t show_board(struct device *dev, struct device_attribute *attr,
+                         char *buf)
+{
+       struct iwch_dev *iwch_dev = container_of(dev, struct iwch_dev,
+                                                ibdev.dev);
+       PDBG("%s dev 0x%p\n", __func__, dev);
+       return sprintf(buf, "%x.%x\n", iwch_dev->rdev.rnic_info.pdev->vendor,
+                      iwch_dev->rdev.rnic_info.pdev->device);
+}
+
+static int iwch_get_mib(struct ib_device *ibdev,
+                       union rdma_protocol_stats *stats)
+{
+       struct iwch_dev *dev;
+       struct tp_mib_stats m;
+       int ret;
+
+       PDBG("%s ibdev %p\n", __func__, ibdev);
+       dev = to_iwch_dev(ibdev);
+       ret = dev->rdev.t3cdev_p->ctl(dev->rdev.t3cdev_p, RDMA_GET_MIB, &m);
+       if (ret)
+               return -ENOSYS;
+
+       memset(stats, 0, sizeof *stats);
+       stats->iw.ipInReceives = ((u64) m.ipInReceive_hi << 32) +
+                               m.ipInReceive_lo;
+       stats->iw.ipInHdrErrors = ((u64) m.ipInHdrErrors_hi << 32) +
+                                 m.ipInHdrErrors_lo;
+       stats->iw.ipInAddrErrors = ((u64) m.ipInAddrErrors_hi << 32) +
+                                  m.ipInAddrErrors_lo;
+       stats->iw.ipInUnknownProtos = ((u64) m.ipInUnknownProtos_hi << 32) +
+                                     m.ipInUnknownProtos_lo;
+       stats->iw.ipInDiscards = ((u64) m.ipInDiscards_hi << 32) +
+                                m.ipInDiscards_lo;
+       stats->iw.ipInDelivers = ((u64) m.ipInDelivers_hi << 32) +
+                                m.ipInDelivers_lo;
+       stats->iw.ipOutRequests = ((u64) m.ipOutRequests_hi << 32) +
+                                 m.ipOutRequests_lo;
+       stats->iw.ipOutDiscards = ((u64) m.ipOutDiscards_hi << 32) +
+                                 m.ipOutDiscards_lo;
+       stats->iw.ipOutNoRoutes = ((u64) m.ipOutNoRoutes_hi << 32) +
+                                 m.ipOutNoRoutes_lo;
+       stats->iw.ipReasmTimeout = (u64) m.ipReasmTimeout;
+       stats->iw.ipReasmReqds = (u64) m.ipReasmReqds;
+       stats->iw.ipReasmOKs = (u64) m.ipReasmOKs;
+       stats->iw.ipReasmFails = (u64) m.ipReasmFails;
+       stats->iw.tcpActiveOpens = (u64) m.tcpActiveOpens;
+       stats->iw.tcpPassiveOpens = (u64) m.tcpPassiveOpens;
+       stats->iw.tcpAttemptFails = (u64) m.tcpAttemptFails;
+       stats->iw.tcpEstabResets = (u64) m.tcpEstabResets;
+       stats->iw.tcpOutRsts = (u64) m.tcpOutRsts;
+       stats->iw.tcpCurrEstab = (u64) m.tcpCurrEstab;
+       stats->iw.tcpInSegs = ((u64) m.tcpInSegs_hi << 32) +
+                             m.tcpInSegs_lo;
+       stats->iw.tcpOutSegs = ((u64) m.tcpOutSegs_hi << 32) +
+                              m.tcpOutSegs_lo;
+       stats->iw.tcpRetransSegs = ((u64) m.tcpRetransSeg_hi << 32) +
+                                 m.tcpRetransSeg_lo;
+       stats->iw.tcpInErrs = ((u64) m.tcpInErrs_hi << 32) +
+                             m.tcpInErrs_lo;
+       stats->iw.tcpRtoMin = (u64) m.tcpRtoMin;
+       stats->iw.tcpRtoMax = (u64) m.tcpRtoMax;
+       return 0;
+}
+
+static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
+static DEVICE_ATTR(fw_ver, S_IRUGO, show_fw_ver, NULL);
+static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
+static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
+
+static struct device_attribute *iwch_class_attributes[] = {
+       &dev_attr_hw_rev,
+       &dev_attr_fw_ver,
+       &dev_attr_hca_type,
+       &dev_attr_board_id,
+};
+
+int iwch_register_device(struct iwch_dev *dev)
+{
+       int ret;
+       int i;
+
+       PDBG("%s iwch_dev %p\n", __func__, dev);
+       strlcpy(dev->ibdev.name, "cxgb3_%d", IB_DEVICE_NAME_MAX);
+       memset(&dev->ibdev.node_guid, 0, sizeof(dev->ibdev.node_guid));
+       memcpy(&dev->ibdev.node_guid, dev->rdev.t3cdev_p->lldev->dev_addr, 6);
+       dev->ibdev.owner = THIS_MODULE;
+       dev->device_cap_flags = IB_DEVICE_LOCAL_DMA_LKEY |
+                               IB_DEVICE_MEM_WINDOW |
+                               IB_DEVICE_MEM_MGT_EXTENSIONS;
+
+       /* cxgb3 supports STag 0. */
+       dev->ibdev.local_dma_lkey = 0;
+
+       dev->ibdev.uverbs_cmd_mask =
+           (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
+           (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
+           (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
+           (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
+           (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
+           (1ull << IB_USER_VERBS_CMD_REG_MR) |
+           (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
+           (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
+           (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
+           (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
+           (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
+           (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
+           (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
+           (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
+           (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
+           (1ull << IB_USER_VERBS_CMD_POST_SEND) |
+           (1ull << IB_USER_VERBS_CMD_POST_RECV);
+       dev->ibdev.node_type = RDMA_NODE_RNIC;
+       memcpy(dev->ibdev.node_desc, IWCH_NODE_DESC, sizeof(IWCH_NODE_DESC));
+       dev->ibdev.phys_port_cnt = dev->rdev.port_info.nports;
+       dev->ibdev.num_comp_vectors = 1;
+       dev->ibdev.dma_device = &(dev->rdev.rnic_info.pdev->dev);
+       dev->ibdev.query_device = iwch_query_device;
+       dev->ibdev.query_port = iwch_query_port;
+       dev->ibdev.query_pkey = iwch_query_pkey;
+       dev->ibdev.query_gid = iwch_query_gid;
+       dev->ibdev.alloc_ucontext = iwch_alloc_ucontext;
+       dev->ibdev.dealloc_ucontext = iwch_dealloc_ucontext;
+       dev->ibdev.mmap = iwch_mmap;
+       dev->ibdev.alloc_pd = iwch_allocate_pd;
+       dev->ibdev.dealloc_pd = iwch_deallocate_pd;
+       dev->ibdev.create_ah = iwch_ah_create;
+       dev->ibdev.destroy_ah = iwch_ah_destroy;
+       dev->ibdev.create_qp = iwch_create_qp;
+       dev->ibdev.modify_qp = iwch_ib_modify_qp;
+       dev->ibdev.destroy_qp = iwch_destroy_qp;
+       dev->ibdev.create_cq = iwch_create_cq;
+       dev->ibdev.destroy_cq = iwch_destroy_cq;
+       dev->ibdev.resize_cq = iwch_resize_cq;
+       dev->ibdev.poll_cq = iwch_poll_cq;
+       dev->ibdev.get_dma_mr = iwch_get_dma_mr;
+       dev->ibdev.reg_phys_mr = iwch_register_phys_mem;
+       dev->ibdev.rereg_phys_mr = iwch_reregister_phys_mem;
+       dev->ibdev.reg_user_mr = iwch_reg_user_mr;
+       dev->ibdev.dereg_mr = iwch_dereg_mr;
+       dev->ibdev.alloc_mw = iwch_alloc_mw;
+       dev->ibdev.bind_mw = iwch_bind_mw;
+       dev->ibdev.dealloc_mw = iwch_dealloc_mw;
+       dev->ibdev.alloc_fast_reg_mr = iwch_alloc_fast_reg_mr;
+       dev->ibdev.alloc_fast_reg_page_list = iwch_alloc_fastreg_pbl;
+       dev->ibdev.free_fast_reg_page_list = iwch_free_fastreg_pbl;
+       dev->ibdev.attach_mcast = iwch_multicast_attach;
+       dev->ibdev.detach_mcast = iwch_multicast_detach;
+       dev->ibdev.process_mad = iwch_process_mad;
+       dev->ibdev.req_notify_cq = iwch_arm_cq;
+       dev->ibdev.post_send = iwch_post_send;
+       dev->ibdev.post_recv = iwch_post_receive;
+       dev->ibdev.get_protocol_stats = iwch_get_mib;
+       dev->ibdev.uverbs_abi_ver = IWCH_UVERBS_ABI_VERSION;
+
+       dev->ibdev.iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL);
+       if (!dev->ibdev.iwcm)
+               return -ENOMEM;
+
+       dev->ibdev.iwcm->connect = iwch_connect;
+       dev->ibdev.iwcm->accept = iwch_accept_cr;
+       dev->ibdev.iwcm->reject = iwch_reject_cr;
+       dev->ibdev.iwcm->create_listen = iwch_create_listen;
+       dev->ibdev.iwcm->destroy_listen = iwch_destroy_listen;
+       dev->ibdev.iwcm->add_ref = iwch_qp_add_ref;
+       dev->ibdev.iwcm->rem_ref = iwch_qp_rem_ref;
+       dev->ibdev.iwcm->get_qp = iwch_get_qp;
+
+       ret = ib_register_device(&dev->ibdev, NULL);
+       if (ret)
+               goto bail1;
+
+       for (i = 0; i < ARRAY_SIZE(iwch_class_attributes); ++i) {
+               ret = device_create_file(&dev->ibdev.dev,
+                                        iwch_class_attributes[i]);
+               if (ret) {
+                       goto bail2;
+               }
+       }
+       return 0;
+bail2:
+       ib_unregister_device(&dev->ibdev);
+bail1:
+       kfree(dev->ibdev.iwcm);
+       return ret;
+}
+
+void iwch_unregister_device(struct iwch_dev *dev)
+{
+       int i;
+
+       PDBG("%s iwch_dev %p\n", __func__, dev);
+       for (i = 0; i < ARRAY_SIZE(iwch_class_attributes); ++i)
+               device_remove_file(&dev->ibdev.dev,
+                                  iwch_class_attributes[i]);
+       ib_unregister_device(&dev->ibdev);
+       kfree(dev->ibdev.iwcm);
+       return;
+}