1 /* 2 * Copyright (c) 2016 Oracle. 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 33 #include "ib_mr.h" 34 35 static inline void 36 rds_transition_frwr_state(struct rds_ib_mr *ibmr, 37 enum rds_ib_fr_state old_state, 38 enum rds_ib_fr_state new_state) 39 { 40 if (cmpxchg(&ibmr->u.frmr.fr_state, 41 old_state, new_state) == old_state && 42 old_state == FRMR_IS_INUSE) { 43 /* enforce order of ibmr->u.frmr.fr_state update 44 * before decrementing i_fastreg_inuse_count 45 */ 46 smp_mb__before_atomic(); 47 atomic_dec(&ibmr->ic->i_fastreg_inuse_count); 48 if (waitqueue_active(&rds_ib_ring_empty_wait)) 49 wake_up(&rds_ib_ring_empty_wait); 50 } 51 } 52 53 static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev, 54 int npages) 55 { 56 struct rds_ib_mr_pool *pool; 57 struct rds_ib_mr *ibmr = NULL; 58 struct rds_ib_frmr *frmr; 59 int err = 0; 60 61 if (npages <= RDS_MR_8K_MSG_SIZE) 62 pool = rds_ibdev->mr_8k_pool; 63 else 64 pool = rds_ibdev->mr_1m_pool; 65 66 ibmr = rds_ib_try_reuse_ibmr(pool); 67 if (ibmr) 68 return ibmr; 69 70 ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, 71 rdsibdev_to_node(rds_ibdev)); 72 if (!ibmr) { 73 err = -ENOMEM; 74 goto out_no_cigar; 75 } 76 77 frmr = &ibmr->u.frmr; 78 frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG, 79 pool->max_pages); 80 if (IS_ERR(frmr->mr)) { 81 pr_warn("RDS/IB: %s failed to allocate MR", __func__); 82 err = PTR_ERR(frmr->mr); 83 goto out_no_cigar; 84 } 85 86 ibmr->pool = pool; 87 if (pool->pool_type == RDS_IB_MR_8K_POOL) 88 rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc); 89 else 90 rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc); 91 92 if (atomic_read(&pool->item_count) > pool->max_items_soft) 93 pool->max_items_soft = pool->max_items; 94 95 frmr->fr_state = FRMR_IS_FREE; 96 init_waitqueue_head(&frmr->fr_inv_done); 97 init_waitqueue_head(&frmr->fr_reg_done); 98 return ibmr; 99 100 out_no_cigar: 101 kfree(ibmr); 102 atomic_dec(&pool->item_count); 103 return ERR_PTR(err); 104 } 105 106 static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop) 107 { 108 struct rds_ib_mr_pool *pool = ibmr->pool; 109 110 if (drop) 111 llist_add(&ibmr->llnode, &pool->drop_list); 112 else 113 llist_add(&ibmr->llnode, &pool->free_list); 114 atomic_add(ibmr->sg_len, &pool->free_pinned); 115 atomic_inc(&pool->dirty_count); 116 117 /* If we've pinned too many pages, request a flush */ 118 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || 119 atomic_read(&pool->dirty_count) >= pool->max_items / 5) 120 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); 121 } 122 123 static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr) 124 { 125 struct rds_ib_frmr *frmr = &ibmr->u.frmr; 126 struct ib_reg_wr reg_wr; 127 int ret, off = 0; 128 129 while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) { 130 atomic_inc(&ibmr->ic->i_fastreg_wrs); 131 cpu_relax(); 132 } 133 134 ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_dma_len, 135 &off, PAGE_SIZE); 136 if (unlikely(ret != ibmr->sg_dma_len)) { 137 ret = ret < 0 ? ret : -EINVAL; 138 goto out_inc; 139 } 140 141 if (cmpxchg(&frmr->fr_state, FRMR_IS_FREE, FRMR_IS_INUSE) != FRMR_IS_FREE) { 142 ret = -EBUSY; 143 goto out_inc; 144 } 145 146 atomic_inc(&ibmr->ic->i_fastreg_inuse_count); 147 148 /* Perform a WR for the fast_reg_mr. Each individual page 149 * in the sg list is added to the fast reg page list and placed 150 * inside the fast_reg_mr WR. The key used is a rolling 8bit 151 * counter, which should guarantee uniqueness. 152 */ 153 ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++); 154 frmr->fr_reg = true; 155 156 memset(®_wr, 0, sizeof(reg_wr)); 157 reg_wr.wr.wr_id = (unsigned long)(void *)ibmr; 158 reg_wr.wr.opcode = IB_WR_REG_MR; 159 reg_wr.wr.num_sge = 0; 160 reg_wr.mr = frmr->mr; 161 reg_wr.key = frmr->mr->rkey; 162 reg_wr.access = IB_ACCESS_LOCAL_WRITE | 163 IB_ACCESS_REMOTE_READ | 164 IB_ACCESS_REMOTE_WRITE; 165 reg_wr.wr.send_flags = IB_SEND_SIGNALED; 166 167 ret = ib_post_send(ibmr->ic->i_cm_id->qp, ®_wr.wr, NULL); 168 if (unlikely(ret)) { 169 /* Failure here can be because of -ENOMEM as well */ 170 rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE); 171 172 if (printk_ratelimit()) 173 pr_warn("RDS/IB: %s returned error(%d)\n", 174 __func__, ret); 175 goto out_inc; 176 } 177 178 /* Wait for the registration to complete in order to prevent an invalid 179 * access error resulting from a race between the memory region already 180 * being accessed while registration is still pending. 181 */ 182 wait_event(frmr->fr_reg_done, !frmr->fr_reg); 183 184 return ret; 185 186 out_inc: 187 atomic_inc(&ibmr->ic->i_fastreg_wrs); 188 return ret; 189 } 190 191 static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev, 192 struct rds_ib_mr_pool *pool, 193 struct rds_ib_mr *ibmr, 194 struct scatterlist *sg, unsigned int sg_len) 195 { 196 struct ib_device *dev = rds_ibdev->dev; 197 struct rds_ib_frmr *frmr = &ibmr->u.frmr; 198 int i; 199 u32 len; 200 int ret = 0; 201 202 /* We want to teardown old ibmr values here and fill it up with 203 * new sg values 204 */ 205 rds_ib_teardown_mr(ibmr); 206 207 frmr->sg_byte_len = 0; 208 ibmr->sg_dma_len = ib_dma_map_sg(dev, sg, sg_len, 209 DMA_BIDIRECTIONAL); 210 if (unlikely(!ibmr->sg_dma_len)) { 211 pr_warn("RDS/IB: %s failed!\n", __func__); 212 return -EBUSY; 213 } 214 215 ibmr->sg = sg; 216 ibmr->sg_len = sg_len; 217 frmr->dma_npages = 0; 218 len = 0; 219 220 ret = -EINVAL; 221 for (i = 0; i < ibmr->sg_dma_len; ++i) { 222 unsigned int dma_len = sg_dma_len(&ibmr->sg[i]); 223 u64 dma_addr = sg_dma_address(&ibmr->sg[i]); 224 225 frmr->sg_byte_len += dma_len; 226 if (dma_addr & ~PAGE_MASK) { 227 if (i > 0) 228 goto out_unmap; 229 else 230 ++frmr->dma_npages; 231 } 232 233 if ((dma_addr + dma_len) & ~PAGE_MASK) { 234 if (i < ibmr->sg_dma_len - 1) 235 goto out_unmap; 236 else 237 ++frmr->dma_npages; 238 } 239 240 len += dma_len; 241 } 242 frmr->dma_npages += len >> PAGE_SHIFT; 243 244 if (frmr->dma_npages > ibmr->pool->max_pages) { 245 ret = -EMSGSIZE; 246 goto out_unmap; 247 } 248 249 ret = rds_ib_post_reg_frmr(ibmr); 250 if (ret) 251 goto out_unmap; 252 253 if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) 254 rds_ib_stats_inc(s_ib_rdma_mr_8k_used); 255 else 256 rds_ib_stats_inc(s_ib_rdma_mr_1m_used); 257 258 return ret; 259 260 out_unmap: 261 ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len, 262 DMA_BIDIRECTIONAL); 263 ibmr->sg_dma_len = 0; 264 ibmr->sg = NULL; 265 ibmr->sg_len = 0; 266 return ret; 267 } 268 269 static int rds_ib_post_inv(struct rds_ib_mr *ibmr) 270 { 271 struct ib_send_wr *s_wr; 272 struct rds_ib_frmr *frmr = &ibmr->u.frmr; 273 struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id; 274 int ret = -EINVAL; 275 276 if (!i_cm_id || !i_cm_id->qp || !frmr->mr) 277 goto out; 278 279 if (frmr->fr_state != FRMR_IS_INUSE) 280 goto out; 281 282 while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) { 283 atomic_inc(&ibmr->ic->i_fastreg_wrs); 284 cpu_relax(); 285 } 286 287 frmr->fr_inv = true; 288 s_wr = &frmr->fr_wr; 289 290 memset(s_wr, 0, sizeof(*s_wr)); 291 s_wr->wr_id = (unsigned long)(void *)ibmr; 292 s_wr->opcode = IB_WR_LOCAL_INV; 293 s_wr->ex.invalidate_rkey = frmr->mr->rkey; 294 s_wr->send_flags = IB_SEND_SIGNALED; 295 296 ret = ib_post_send(i_cm_id->qp, s_wr, NULL); 297 if (unlikely(ret)) { 298 rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE); 299 frmr->fr_inv = false; 300 /* enforce order of frmr->fr_inv update 301 * before incrementing i_fastreg_wrs 302 */ 303 smp_mb__before_atomic(); 304 atomic_inc(&ibmr->ic->i_fastreg_wrs); 305 pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret); 306 goto out; 307 } 308 309 /* Wait for the FRMR_IS_FREE (or FRMR_IS_STALE) transition in order to 310 * 1) avoid a silly bouncing between "clean_list" and "drop_list" 311 * triggered by function "rds_ib_reg_frmr" as it is releases frmr 312 * regions whose state is not "FRMR_IS_FREE" right away. 313 * 2) prevents an invalid access error in a race 314 * from a pending "IB_WR_LOCAL_INV" operation 315 * with a teardown ("dma_unmap_sg", "put_page") 316 * and de-registration ("ib_dereg_mr") of the corresponding 317 * memory region. 318 */ 319 wait_event(frmr->fr_inv_done, frmr->fr_state != FRMR_IS_INUSE); 320 321 out: 322 return ret; 323 } 324 325 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc) 326 { 327 struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id; 328 struct rds_ib_frmr *frmr = &ibmr->u.frmr; 329 330 if (wc->status != IB_WC_SUCCESS) { 331 rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE); 332 if (rds_conn_up(ic->conn)) 333 rds_ib_conn_error(ic->conn, 334 "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n", 335 &ic->conn->c_laddr, 336 &ic->conn->c_faddr, 337 wc->status, 338 ib_wc_status_msg(wc->status), 339 wc->vendor_err); 340 } 341 342 if (frmr->fr_inv) { 343 rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_FREE); 344 frmr->fr_inv = false; 345 wake_up(&frmr->fr_inv_done); 346 } 347 348 if (frmr->fr_reg) { 349 frmr->fr_reg = false; 350 wake_up(&frmr->fr_reg_done); 351 } 352 353 /* enforce order of frmr->{fr_reg,fr_inv} update 354 * before incrementing i_fastreg_wrs 355 */ 356 smp_mb__before_atomic(); 357 atomic_inc(&ic->i_fastreg_wrs); 358 } 359 360 void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed, 361 unsigned long *unpinned, unsigned int goal) 362 { 363 struct rds_ib_mr *ibmr, *next; 364 struct rds_ib_frmr *frmr; 365 int ret = 0, ret2; 366 unsigned int freed = *nfreed; 367 368 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */ 369 list_for_each_entry(ibmr, list, unmap_list) { 370 if (ibmr->sg_dma_len) { 371 ret2 = rds_ib_post_inv(ibmr); 372 if (ret2 && !ret) 373 ret = ret2; 374 } 375 } 376 377 if (ret) 378 pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret); 379 380 /* Now we can destroy the DMA mapping and unpin any pages */ 381 list_for_each_entry_safe(ibmr, next, list, unmap_list) { 382 *unpinned += ibmr->sg_len; 383 frmr = &ibmr->u.frmr; 384 __rds_ib_teardown_mr(ibmr); 385 if (freed < goal || frmr->fr_state == FRMR_IS_STALE) { 386 /* Don't de-allocate if the MR is not free yet */ 387 if (frmr->fr_state == FRMR_IS_INUSE) 388 continue; 389 390 if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) 391 rds_ib_stats_inc(s_ib_rdma_mr_8k_free); 392 else 393 rds_ib_stats_inc(s_ib_rdma_mr_1m_free); 394 list_del(&ibmr->unmap_list); 395 if (frmr->mr) 396 ib_dereg_mr(frmr->mr); 397 kfree(ibmr); 398 freed++; 399 } 400 } 401 *nfreed = freed; 402 } 403 404 struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev, 405 struct rds_ib_connection *ic, 406 struct scatterlist *sg, 407 unsigned long nents, u32 *key) 408 { 409 struct rds_ib_mr *ibmr = NULL; 410 struct rds_ib_frmr *frmr; 411 int ret; 412 413 if (!ic) { 414 /* TODO: Add FRWR support for RDS_GET_MR using proxy qp*/ 415 return ERR_PTR(-EOPNOTSUPP); 416 } 417 418 do { 419 if (ibmr) 420 rds_ib_free_frmr(ibmr, true); 421 ibmr = rds_ib_alloc_frmr(rds_ibdev, nents); 422 if (IS_ERR(ibmr)) 423 return ibmr; 424 frmr = &ibmr->u.frmr; 425 } while (frmr->fr_state != FRMR_IS_FREE); 426 427 ibmr->ic = ic; 428 ibmr->device = rds_ibdev; 429 ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents); 430 if (ret == 0) { 431 *key = frmr->mr->rkey; 432 } else { 433 rds_ib_free_frmr(ibmr, false); 434 ibmr = ERR_PTR(ret); 435 } 436 437 return ibmr; 438 } 439 440 void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr) 441 { 442 struct rds_ib_mr_pool *pool = ibmr->pool; 443 struct rds_ib_frmr *frmr = &ibmr->u.frmr; 444 445 if (frmr->fr_state == FRMR_IS_STALE) 446 llist_add(&ibmr->llnode, &pool->drop_list); 447 else 448 llist_add(&ibmr->llnode, &pool->free_list); 449 } 450