| 1 | // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause |
| 2 | /* |
| 3 | * Copyright(c) 2016 Intel Corporation. |
| 4 | */ |
| 5 | |
| 6 | #include <linux/err.h> |
| 7 | #include <linux/slab.h> |
| 8 | #include <linux/vmalloc.h> |
| 9 | #include <rdma/uverbs_ioctl.h> |
| 10 | |
| 11 | #include "srq.h" |
| 12 | #include "vt.h" |
| 13 | #include "qp.h" |
| 14 | /** |
| 15 | * rvt_driver_srq_init - init srq resources on a per driver basis |
| 16 | * @rdi: rvt dev structure |
| 17 | * |
| 18 | * Do any initialization needed when a driver registers with rdmavt. |
| 19 | */ |
| 20 | void rvt_driver_srq_init(struct rvt_dev_info *rdi) |
| 21 | { |
| 22 | spin_lock_init(&rdi->n_srqs_lock); |
| 23 | rdi->n_srqs_allocated = 0; |
| 24 | } |
| 25 | |
| 26 | /** |
| 27 | * rvt_create_srq - create a shared receive queue |
| 28 | * @ibsrq: the protection domain of the SRQ to create |
| 29 | * @srq_init_attr: the attributes of the SRQ |
| 30 | * @udata: data from libibverbs when creating a user SRQ |
| 31 | * |
| 32 | * Return: 0 on success |
| 33 | */ |
| 34 | int rvt_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *srq_init_attr, |
| 35 | struct ib_udata *udata) |
| 36 | { |
| 37 | struct rvt_dev_info *dev = ib_to_rvt(ibdev: ibsrq->device); |
| 38 | struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); |
| 39 | u32 sz; |
| 40 | int ret; |
| 41 | |
| 42 | if (srq_init_attr->srq_type != IB_SRQT_BASIC) |
| 43 | return -EOPNOTSUPP; |
| 44 | |
| 45 | if (srq_init_attr->attr.max_sge == 0 || |
| 46 | srq_init_attr->attr.max_sge > dev->dparms.props.max_srq_sge || |
| 47 | srq_init_attr->attr.max_wr == 0 || |
| 48 | srq_init_attr->attr.max_wr > dev->dparms.props.max_srq_wr) |
| 49 | return -EINVAL; |
| 50 | |
| 51 | /* |
| 52 | * Need to use vmalloc() if we want to support large #s of entries. |
| 53 | */ |
| 54 | srq->rq.size = srq_init_attr->attr.max_wr + 1; |
| 55 | srq->rq.max_sge = srq_init_attr->attr.max_sge; |
| 56 | sz = sizeof(struct ib_sge) * srq->rq.max_sge + |
| 57 | sizeof(struct rvt_rwqe); |
| 58 | if (rvt_alloc_rq(rq: &srq->rq, size: srq->rq.size * sz, |
| 59 | node: dev->dparms.node, udata)) { |
| 60 | ret = -ENOMEM; |
| 61 | goto bail_srq; |
| 62 | } |
| 63 | |
| 64 | /* |
| 65 | * Return the address of the RWQ as the offset to mmap. |
| 66 | * See rvt_mmap() for details. |
| 67 | */ |
| 68 | if (udata && udata->outlen >= sizeof(__u64)) { |
| 69 | u32 s = sizeof(struct rvt_rwq) + srq->rq.size * sz; |
| 70 | |
| 71 | srq->ip = rvt_create_mmap_info(rdi: dev, size: s, udata, obj: srq->rq.wq); |
| 72 | if (IS_ERR(ptr: srq->ip)) { |
| 73 | ret = PTR_ERR(ptr: srq->ip); |
| 74 | goto bail_wq; |
| 75 | } |
| 76 | |
| 77 | ret = ib_copy_to_udata(udata, src: &srq->ip->offset, |
| 78 | len: sizeof(srq->ip->offset)); |
| 79 | if (ret) |
| 80 | goto bail_ip; |
| 81 | } |
| 82 | |
| 83 | /* |
| 84 | * ib_create_srq() will initialize srq->ibsrq. |
| 85 | */ |
| 86 | spin_lock_init(&srq->rq.lock); |
| 87 | srq->limit = srq_init_attr->attr.srq_limit; |
| 88 | |
| 89 | spin_lock(lock: &dev->n_srqs_lock); |
| 90 | if (dev->n_srqs_allocated == dev->dparms.props.max_srq) { |
| 91 | spin_unlock(lock: &dev->n_srqs_lock); |
| 92 | ret = -ENOMEM; |
| 93 | goto bail_ip; |
| 94 | } |
| 95 | |
| 96 | dev->n_srqs_allocated++; |
| 97 | spin_unlock(lock: &dev->n_srqs_lock); |
| 98 | |
| 99 | if (srq->ip) { |
| 100 | spin_lock_irq(lock: &dev->pending_lock); |
| 101 | list_add(new: &srq->ip->pending_mmaps, head: &dev->pending_mmaps); |
| 102 | spin_unlock_irq(lock: &dev->pending_lock); |
| 103 | } |
| 104 | |
| 105 | return 0; |
| 106 | |
| 107 | bail_ip: |
| 108 | kfree(objp: srq->ip); |
| 109 | bail_wq: |
| 110 | rvt_free_rq(rq: &srq->rq); |
| 111 | bail_srq: |
| 112 | return ret; |
| 113 | } |
| 114 | |
| 115 | /** |
| 116 | * rvt_modify_srq - modify a shared receive queue |
| 117 | * @ibsrq: the SRQ to modify |
| 118 | * @attr: the new attributes of the SRQ |
| 119 | * @attr_mask: indicates which attributes to modify |
| 120 | * @udata: user data for libibverbs.so |
| 121 | * |
| 122 | * Return: 0 on success |
| 123 | */ |
| 124 | int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, |
| 125 | enum ib_srq_attr_mask attr_mask, |
| 126 | struct ib_udata *udata) |
| 127 | { |
| 128 | struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); |
| 129 | struct rvt_dev_info *dev = ib_to_rvt(ibdev: ibsrq->device); |
| 130 | struct rvt_rq tmp_rq = {}; |
| 131 | int ret = 0; |
| 132 | |
| 133 | if (attr_mask & IB_SRQ_MAX_WR) { |
| 134 | struct rvt_krwq *okwq = NULL; |
| 135 | struct rvt_rwq *owq = NULL; |
| 136 | struct rvt_rwqe *p; |
| 137 | u32 sz, size, n, head, tail; |
| 138 | |
| 139 | /* Check that the requested sizes are below the limits. */ |
| 140 | if ((attr->max_wr > dev->dparms.props.max_srq_wr) || |
| 141 | ((attr_mask & IB_SRQ_LIMIT) ? |
| 142 | attr->srq_limit : srq->limit) > attr->max_wr) |
| 143 | return -EINVAL; |
| 144 | sz = sizeof(struct rvt_rwqe) + |
| 145 | srq->rq.max_sge * sizeof(struct ib_sge); |
| 146 | size = attr->max_wr + 1; |
| 147 | if (rvt_alloc_rq(rq: &tmp_rq, size: size * sz, node: dev->dparms.node, |
| 148 | udata)) |
| 149 | return -ENOMEM; |
| 150 | /* Check that we can write the offset to mmap. */ |
| 151 | if (udata && udata->inlen >= sizeof(__u64)) { |
| 152 | __u64 offset_addr; |
| 153 | __u64 offset = 0; |
| 154 | |
| 155 | ret = ib_copy_from_udata(dest: &offset_addr, udata, |
| 156 | len: sizeof(offset_addr)); |
| 157 | if (ret) |
| 158 | goto bail_free; |
| 159 | udata->outbuf = (void __user *) |
| 160 | (unsigned long)offset_addr; |
| 161 | ret = ib_copy_to_udata(udata, src: &offset, |
| 162 | len: sizeof(offset)); |
| 163 | if (ret) |
| 164 | goto bail_free; |
| 165 | } |
| 166 | |
| 167 | spin_lock_irq(lock: &srq->rq.kwq->c_lock); |
| 168 | /* |
| 169 | * validate head and tail pointer values and compute |
| 170 | * the number of remaining WQEs. |
| 171 | */ |
| 172 | if (udata) { |
| 173 | owq = srq->rq.wq; |
| 174 | head = RDMA_READ_UAPI_ATOMIC(owq->head); |
| 175 | tail = RDMA_READ_UAPI_ATOMIC(owq->tail); |
| 176 | } else { |
| 177 | okwq = srq->rq.kwq; |
| 178 | head = okwq->head; |
| 179 | tail = okwq->tail; |
| 180 | } |
| 181 | if (head >= srq->rq.size || tail >= srq->rq.size) { |
| 182 | ret = -EINVAL; |
| 183 | goto bail_unlock; |
| 184 | } |
| 185 | n = head; |
| 186 | if (n < tail) |
| 187 | n += srq->rq.size - tail; |
| 188 | else |
| 189 | n -= tail; |
| 190 | if (size <= n) { |
| 191 | ret = -EINVAL; |
| 192 | goto bail_unlock; |
| 193 | } |
| 194 | n = 0; |
| 195 | p = tmp_rq.kwq->curr_wq; |
| 196 | while (tail != head) { |
| 197 | struct rvt_rwqe *wqe; |
| 198 | int i; |
| 199 | |
| 200 | wqe = rvt_get_rwqe_ptr(rq: &srq->rq, n: tail); |
| 201 | p->wr_id = wqe->wr_id; |
| 202 | p->num_sge = wqe->num_sge; |
| 203 | for (i = 0; i < wqe->num_sge; i++) |
| 204 | p->sg_list[i] = wqe->sg_list[i]; |
| 205 | n++; |
| 206 | p = (struct rvt_rwqe *)((char *)p + sz); |
| 207 | if (++tail >= srq->rq.size) |
| 208 | tail = 0; |
| 209 | } |
| 210 | srq->rq.kwq = tmp_rq.kwq; |
| 211 | if (udata) { |
| 212 | srq->rq.wq = tmp_rq.wq; |
| 213 | RDMA_WRITE_UAPI_ATOMIC(tmp_rq.wq->head, n); |
| 214 | RDMA_WRITE_UAPI_ATOMIC(tmp_rq.wq->tail, 0); |
| 215 | } else { |
| 216 | tmp_rq.kwq->head = n; |
| 217 | tmp_rq.kwq->tail = 0; |
| 218 | } |
| 219 | srq->rq.size = size; |
| 220 | if (attr_mask & IB_SRQ_LIMIT) |
| 221 | srq->limit = attr->srq_limit; |
| 222 | spin_unlock_irq(lock: &srq->rq.kwq->c_lock); |
| 223 | |
| 224 | vfree(addr: owq); |
| 225 | kvfree(addr: okwq); |
| 226 | |
| 227 | if (srq->ip) { |
| 228 | struct rvt_mmap_info *ip = srq->ip; |
| 229 | struct rvt_dev_info *dev = ib_to_rvt(ibdev: srq->ibsrq.device); |
| 230 | u32 s = sizeof(struct rvt_rwq) + size * sz; |
| 231 | |
| 232 | rvt_update_mmap_info(rdi: dev, ip, size: s, obj: tmp_rq.wq); |
| 233 | |
| 234 | /* |
| 235 | * Return the offset to mmap. |
| 236 | * See rvt_mmap() for details. |
| 237 | */ |
| 238 | if (udata && udata->inlen >= sizeof(__u64)) { |
| 239 | ret = ib_copy_to_udata(udata, src: &ip->offset, |
| 240 | len: sizeof(ip->offset)); |
| 241 | if (ret) |
| 242 | return ret; |
| 243 | } |
| 244 | |
| 245 | /* |
| 246 | * Put user mapping info onto the pending list |
| 247 | * unless it already is on the list. |
| 248 | */ |
| 249 | spin_lock_irq(lock: &dev->pending_lock); |
| 250 | if (list_empty(head: &ip->pending_mmaps)) |
| 251 | list_add(new: &ip->pending_mmaps, |
| 252 | head: &dev->pending_mmaps); |
| 253 | spin_unlock_irq(lock: &dev->pending_lock); |
| 254 | } |
| 255 | } else if (attr_mask & IB_SRQ_LIMIT) { |
| 256 | spin_lock_irq(lock: &srq->rq.kwq->c_lock); |
| 257 | if (attr->srq_limit >= srq->rq.size) |
| 258 | ret = -EINVAL; |
| 259 | else |
| 260 | srq->limit = attr->srq_limit; |
| 261 | spin_unlock_irq(lock: &srq->rq.kwq->c_lock); |
| 262 | } |
| 263 | return ret; |
| 264 | |
| 265 | bail_unlock: |
| 266 | spin_unlock_irq(lock: &srq->rq.kwq->c_lock); |
| 267 | bail_free: |
| 268 | rvt_free_rq(rq: &tmp_rq); |
| 269 | return ret; |
| 270 | } |
| 271 | |
| 272 | /** |
| 273 | * rvt_query_srq - query srq data |
| 274 | * @ibsrq: srq to query |
| 275 | * @attr: return info in attr |
| 276 | * |
| 277 | * Return: always 0 |
| 278 | */ |
| 279 | int rvt_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr) |
| 280 | { |
| 281 | struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); |
| 282 | |
| 283 | attr->max_wr = srq->rq.size - 1; |
| 284 | attr->max_sge = srq->rq.max_sge; |
| 285 | attr->srq_limit = srq->limit; |
| 286 | return 0; |
| 287 | } |
| 288 | |
| 289 | /** |
| 290 | * rvt_destroy_srq - destory an srq |
| 291 | * @ibsrq: srq object to destroy |
| 292 | * @udata: user data for libibverbs.so |
| 293 | */ |
| 294 | int rvt_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata) |
| 295 | { |
| 296 | struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); |
| 297 | struct rvt_dev_info *dev = ib_to_rvt(ibdev: ibsrq->device); |
| 298 | |
| 299 | spin_lock(lock: &dev->n_srqs_lock); |
| 300 | dev->n_srqs_allocated--; |
| 301 | spin_unlock(lock: &dev->n_srqs_lock); |
| 302 | if (srq->ip) |
| 303 | kref_put(kref: &srq->ip->ref, release: rvt_release_mmap_info); |
| 304 | kvfree(addr: srq->rq.kwq); |
| 305 | return 0; |
| 306 | } |
| 307 | |