Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2006 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 <linux/kernel.h> |
| 34 | #include <linux/pci.h> |
| 35 | #include <linux/dma-mapping.h> |
| 36 | #include <rdma/rdma_cm.h> |
| 37 | |
| 38 | #include "rds.h" |
| 39 | #include "ib.h" |
| 40 | |
| 41 | static struct kmem_cache *rds_ib_incoming_slab; |
| 42 | static struct kmem_cache *rds_ib_frag_slab; |
| 43 | static atomic_t rds_ib_allocation = ATOMIC_INIT(0); |
| 44 | |
| 45 | static void rds_ib_frag_drop_page(struct rds_page_frag *frag) |
| 46 | { |
| 47 | rdsdebug("frag %p page %p\n", frag, frag->f_page); |
| 48 | __free_page(frag->f_page); |
| 49 | frag->f_page = NULL; |
| 50 | } |
| 51 | |
| 52 | static void rds_ib_frag_free(struct rds_page_frag *frag) |
| 53 | { |
| 54 | rdsdebug("frag %p page %p\n", frag, frag->f_page); |
| 55 | BUG_ON(frag->f_page != NULL); |
| 56 | kmem_cache_free(rds_ib_frag_slab, frag); |
| 57 | } |
| 58 | |
| 59 | /* |
| 60 | * We map a page at a time. Its fragments are posted in order. This |
| 61 | * is called in fragment order as the fragments get send completion events. |
| 62 | * Only the last frag in the page performs the unmapping. |
| 63 | * |
| 64 | * It's OK for ring cleanup to call this in whatever order it likes because |
| 65 | * DMA is not in flight and so we can unmap while other ring entries still |
| 66 | * hold page references in their frags. |
| 67 | */ |
| 68 | static void rds_ib_recv_unmap_page(struct rds_ib_connection *ic, |
| 69 | struct rds_ib_recv_work *recv) |
| 70 | { |
| 71 | struct rds_page_frag *frag = recv->r_frag; |
| 72 | |
| 73 | rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page); |
| 74 | if (frag->f_mapped) |
| 75 | ib_dma_unmap_page(ic->i_cm_id->device, |
| 76 | frag->f_mapped, |
| 77 | RDS_FRAG_SIZE, DMA_FROM_DEVICE); |
| 78 | frag->f_mapped = 0; |
| 79 | } |
| 80 | |
| 81 | void rds_ib_recv_init_ring(struct rds_ib_connection *ic) |
| 82 | { |
| 83 | struct rds_ib_recv_work *recv; |
| 84 | u32 i; |
| 85 | |
| 86 | for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { |
| 87 | struct ib_sge *sge; |
| 88 | |
| 89 | recv->r_ibinc = NULL; |
| 90 | recv->r_frag = NULL; |
| 91 | |
| 92 | recv->r_wr.next = NULL; |
| 93 | recv->r_wr.wr_id = i; |
| 94 | recv->r_wr.sg_list = recv->r_sge; |
| 95 | recv->r_wr.num_sge = RDS_IB_RECV_SGE; |
| 96 | |
| 97 | sge = rds_ib_data_sge(ic, recv->r_sge); |
| 98 | sge->addr = 0; |
| 99 | sge->length = RDS_FRAG_SIZE; |
| 100 | sge->lkey = ic->i_mr->lkey; |
| 101 | |
| 102 | sge = rds_ib_header_sge(ic, recv->r_sge); |
| 103 | sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); |
| 104 | sge->length = sizeof(struct rds_header); |
| 105 | sge->lkey = ic->i_mr->lkey; |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, |
| 110 | struct rds_ib_recv_work *recv) |
| 111 | { |
| 112 | if (recv->r_ibinc) { |
| 113 | rds_inc_put(&recv->r_ibinc->ii_inc); |
| 114 | recv->r_ibinc = NULL; |
| 115 | } |
| 116 | if (recv->r_frag) { |
| 117 | rds_ib_recv_unmap_page(ic, recv); |
| 118 | if (recv->r_frag->f_page) |
| 119 | rds_ib_frag_drop_page(recv->r_frag); |
| 120 | rds_ib_frag_free(recv->r_frag); |
| 121 | recv->r_frag = NULL; |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) |
| 126 | { |
| 127 | u32 i; |
| 128 | |
| 129 | for (i = 0; i < ic->i_recv_ring.w_nr; i++) |
| 130 | rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); |
| 131 | |
| 132 | if (ic->i_frag.f_page) |
| 133 | rds_ib_frag_drop_page(&ic->i_frag); |
| 134 | } |
| 135 | |
| 136 | static int rds_ib_recv_refill_one(struct rds_connection *conn, |
| 137 | struct rds_ib_recv_work *recv, |
| 138 | gfp_t kptr_gfp, gfp_t page_gfp) |
| 139 | { |
| 140 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 141 | dma_addr_t dma_addr; |
| 142 | struct ib_sge *sge; |
| 143 | int ret = -ENOMEM; |
| 144 | |
| 145 | if (recv->r_ibinc == NULL) { |
| 146 | if (atomic_read(&rds_ib_allocation) >= rds_ib_sysctl_max_recv_allocation) { |
| 147 | rds_ib_stats_inc(s_ib_rx_alloc_limit); |
| 148 | goto out; |
| 149 | } |
| 150 | recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab, |
| 151 | kptr_gfp); |
| 152 | if (recv->r_ibinc == NULL) |
| 153 | goto out; |
| 154 | atomic_inc(&rds_ib_allocation); |
| 155 | INIT_LIST_HEAD(&recv->r_ibinc->ii_frags); |
| 156 | rds_inc_init(&recv->r_ibinc->ii_inc, conn, conn->c_faddr); |
| 157 | } |
| 158 | |
| 159 | if (recv->r_frag == NULL) { |
| 160 | recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp); |
| 161 | if (recv->r_frag == NULL) |
| 162 | goto out; |
| 163 | INIT_LIST_HEAD(&recv->r_frag->f_item); |
| 164 | recv->r_frag->f_page = NULL; |
| 165 | } |
| 166 | |
| 167 | if (ic->i_frag.f_page == NULL) { |
| 168 | ic->i_frag.f_page = alloc_page(page_gfp); |
| 169 | if (ic->i_frag.f_page == NULL) |
| 170 | goto out; |
| 171 | ic->i_frag.f_offset = 0; |
| 172 | } |
| 173 | |
| 174 | dma_addr = ib_dma_map_page(ic->i_cm_id->device, |
| 175 | ic->i_frag.f_page, |
| 176 | ic->i_frag.f_offset, |
| 177 | RDS_FRAG_SIZE, |
| 178 | DMA_FROM_DEVICE); |
| 179 | if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr)) |
| 180 | goto out; |
| 181 | |
| 182 | /* |
| 183 | * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap() |
| 184 | * must be called on this recv. This happens as completions hit |
| 185 | * in order or on connection shutdown. |
| 186 | */ |
| 187 | recv->r_frag->f_page = ic->i_frag.f_page; |
| 188 | recv->r_frag->f_offset = ic->i_frag.f_offset; |
| 189 | recv->r_frag->f_mapped = dma_addr; |
| 190 | |
| 191 | sge = rds_ib_data_sge(ic, recv->r_sge); |
| 192 | sge->addr = dma_addr; |
| 193 | sge->length = RDS_FRAG_SIZE; |
| 194 | |
| 195 | sge = rds_ib_header_sge(ic, recv->r_sge); |
| 196 | sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); |
| 197 | sge->length = sizeof(struct rds_header); |
| 198 | |
| 199 | get_page(recv->r_frag->f_page); |
| 200 | |
| 201 | if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) { |
| 202 | ic->i_frag.f_offset += RDS_FRAG_SIZE; |
| 203 | } else { |
| 204 | put_page(ic->i_frag.f_page); |
| 205 | ic->i_frag.f_page = NULL; |
| 206 | ic->i_frag.f_offset = 0; |
| 207 | } |
| 208 | |
| 209 | ret = 0; |
| 210 | out: |
| 211 | return ret; |
| 212 | } |
| 213 | |
| 214 | /* |
| 215 | * This tries to allocate and post unused work requests after making sure that |
| 216 | * they have all the allocations they need to queue received fragments into |
| 217 | * sockets. The i_recv_mutex is held here so that ring_alloc and _unalloc |
| 218 | * pairs don't go unmatched. |
| 219 | * |
| 220 | * -1 is returned if posting fails due to temporary resource exhaustion. |
| 221 | */ |
| 222 | int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, |
| 223 | gfp_t page_gfp, int prefill) |
| 224 | { |
| 225 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 226 | struct rds_ib_recv_work *recv; |
| 227 | struct ib_recv_wr *failed_wr; |
| 228 | unsigned int posted = 0; |
| 229 | int ret = 0; |
| 230 | u32 pos; |
| 231 | |
| 232 | while ((prefill || rds_conn_up(conn)) |
| 233 | && rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { |
| 234 | if (pos >= ic->i_recv_ring.w_nr) { |
| 235 | printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", |
| 236 | pos); |
| 237 | ret = -EINVAL; |
| 238 | break; |
| 239 | } |
| 240 | |
| 241 | recv = &ic->i_recvs[pos]; |
| 242 | ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp); |
| 243 | if (ret) { |
| 244 | ret = -1; |
| 245 | break; |
| 246 | } |
| 247 | |
| 248 | /* XXX when can this fail? */ |
| 249 | ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); |
| 250 | rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv, |
| 251 | recv->r_ibinc, recv->r_frag->f_page, |
| 252 | (long) recv->r_frag->f_mapped, ret); |
| 253 | if (ret) { |
| 254 | rds_ib_conn_error(conn, "recv post on " |
| 255 | "%pI4 returned %d, disconnecting and " |
| 256 | "reconnecting\n", &conn->c_faddr, |
| 257 | ret); |
| 258 | ret = -1; |
| 259 | break; |
| 260 | } |
| 261 | |
| 262 | posted++; |
| 263 | } |
| 264 | |
| 265 | /* We're doing flow control - update the window. */ |
| 266 | if (ic->i_flowctl && posted) |
| 267 | rds_ib_advertise_credits(conn, posted); |
| 268 | |
| 269 | if (ret) |
| 270 | rds_ib_ring_unalloc(&ic->i_recv_ring, 1); |
| 271 | return ret; |
| 272 | } |
| 273 | |
| 274 | void rds_ib_inc_purge(struct rds_incoming *inc) |
| 275 | { |
| 276 | struct rds_ib_incoming *ibinc; |
| 277 | struct rds_page_frag *frag; |
| 278 | struct rds_page_frag *pos; |
| 279 | |
| 280 | ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); |
| 281 | rdsdebug("purging ibinc %p inc %p\n", ibinc, inc); |
| 282 | |
| 283 | list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { |
| 284 | list_del_init(&frag->f_item); |
| 285 | rds_ib_frag_drop_page(frag); |
| 286 | rds_ib_frag_free(frag); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | void rds_ib_inc_free(struct rds_incoming *inc) |
| 291 | { |
| 292 | struct rds_ib_incoming *ibinc; |
| 293 | |
| 294 | ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); |
| 295 | |
| 296 | rds_ib_inc_purge(inc); |
| 297 | rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); |
| 298 | BUG_ON(!list_empty(&ibinc->ii_frags)); |
| 299 | kmem_cache_free(rds_ib_incoming_slab, ibinc); |
| 300 | atomic_dec(&rds_ib_allocation); |
| 301 | BUG_ON(atomic_read(&rds_ib_allocation) < 0); |
| 302 | } |
| 303 | |
| 304 | int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov, |
| 305 | size_t size) |
| 306 | { |
| 307 | struct rds_ib_incoming *ibinc; |
| 308 | struct rds_page_frag *frag; |
| 309 | struct iovec *iov = first_iov; |
| 310 | unsigned long to_copy; |
| 311 | unsigned long frag_off = 0; |
| 312 | unsigned long iov_off = 0; |
| 313 | int copied = 0; |
| 314 | int ret; |
| 315 | u32 len; |
| 316 | |
| 317 | ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); |
| 318 | frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); |
| 319 | len = be32_to_cpu(inc->i_hdr.h_len); |
| 320 | |
| 321 | while (copied < size && copied < len) { |
| 322 | if (frag_off == RDS_FRAG_SIZE) { |
| 323 | frag = list_entry(frag->f_item.next, |
| 324 | struct rds_page_frag, f_item); |
| 325 | frag_off = 0; |
| 326 | } |
| 327 | while (iov_off == iov->iov_len) { |
| 328 | iov_off = 0; |
| 329 | iov++; |
| 330 | } |
| 331 | |
| 332 | to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off); |
| 333 | to_copy = min_t(size_t, to_copy, size - copied); |
| 334 | to_copy = min_t(unsigned long, to_copy, len - copied); |
| 335 | |
| 336 | rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag " |
| 337 | "[%p, %lu] + %lu\n", |
| 338 | to_copy, iov->iov_base, iov->iov_len, iov_off, |
| 339 | frag->f_page, frag->f_offset, frag_off); |
| 340 | |
| 341 | /* XXX needs + offset for multiple recvs per page */ |
| 342 | ret = rds_page_copy_to_user(frag->f_page, |
| 343 | frag->f_offset + frag_off, |
| 344 | iov->iov_base + iov_off, |
| 345 | to_copy); |
| 346 | if (ret) { |
| 347 | copied = ret; |
| 348 | break; |
| 349 | } |
| 350 | |
| 351 | iov_off += to_copy; |
| 352 | frag_off += to_copy; |
| 353 | copied += to_copy; |
| 354 | } |
| 355 | |
| 356 | return copied; |
| 357 | } |
| 358 | |
| 359 | /* ic starts out kzalloc()ed */ |
| 360 | void rds_ib_recv_init_ack(struct rds_ib_connection *ic) |
| 361 | { |
| 362 | struct ib_send_wr *wr = &ic->i_ack_wr; |
| 363 | struct ib_sge *sge = &ic->i_ack_sge; |
| 364 | |
| 365 | sge->addr = ic->i_ack_dma; |
| 366 | sge->length = sizeof(struct rds_header); |
| 367 | sge->lkey = ic->i_mr->lkey; |
| 368 | |
| 369 | wr->sg_list = sge; |
| 370 | wr->num_sge = 1; |
| 371 | wr->opcode = IB_WR_SEND; |
| 372 | wr->wr_id = RDS_IB_ACK_WR_ID; |
| 373 | wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; |
| 374 | } |
| 375 | |
| 376 | /* |
| 377 | * You'd think that with reliable IB connections you wouldn't need to ack |
| 378 | * messages that have been received. The problem is that IB hardware generates |
| 379 | * an ack message before it has DMAed the message into memory. This creates a |
| 380 | * potential message loss if the HCA is disabled for any reason between when it |
| 381 | * sends the ack and before the message is DMAed and processed. This is only a |
| 382 | * potential issue if another HCA is available for fail-over. |
| 383 | * |
| 384 | * When the remote host receives our ack they'll free the sent message from |
| 385 | * their send queue. To decrease the latency of this we always send an ack |
| 386 | * immediately after we've received messages. |
| 387 | * |
| 388 | * For simplicity, we only have one ack in flight at a time. This puts |
| 389 | * pressure on senders to have deep enough send queues to absorb the latency of |
| 390 | * a single ack frame being in flight. This might not be good enough. |
| 391 | * |
| 392 | * This is implemented by have a long-lived send_wr and sge which point to a |
| 393 | * statically allocated ack frame. This ack wr does not fall under the ring |
| 394 | * accounting that the tx and rx wrs do. The QP attribute specifically makes |
| 395 | * room for it beyond the ring size. Send completion notices its special |
| 396 | * wr_id and avoids working with the ring in that case. |
| 397 | */ |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame^] | 398 | #ifndef KERNEL_HAS_ATOMIC64 |
Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 399 | static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, |
| 400 | int ack_required) |
| 401 | { |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame^] | 402 | unsigned long flags; |
| 403 | |
| 404 | spin_lock_irqsave(&ic->i_ack_lock, flags); |
| 405 | ic->i_ack_next = seq; |
| 406 | if (ack_required) |
| 407 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 408 | spin_unlock_irqrestore(&ic->i_ack_lock, flags); |
| 409 | } |
| 410 | |
| 411 | static u64 rds_ib_get_ack(struct rds_ib_connection *ic) |
| 412 | { |
| 413 | unsigned long flags; |
| 414 | u64 seq; |
| 415 | |
| 416 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 417 | |
| 418 | spin_lock_irqsave(&ic->i_ack_lock, flags); |
| 419 | seq = ic->i_ack_next; |
| 420 | spin_unlock_irqrestore(&ic->i_ack_lock, flags); |
| 421 | |
| 422 | return seq; |
| 423 | } |
| 424 | #else |
| 425 | static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, |
| 426 | int ack_required) |
| 427 | { |
| 428 | atomic64_set(&ic->i_ack_next, seq); |
Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 429 | if (ack_required) { |
| 430 | smp_mb__before_clear_bit(); |
| 431 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | static u64 rds_ib_get_ack(struct rds_ib_connection *ic) |
| 436 | { |
| 437 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 438 | smp_mb__after_clear_bit(); |
| 439 | |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame^] | 440 | return atomic64_read(&ic->i_ack_next); |
Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 441 | } |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame^] | 442 | #endif |
| 443 | |
Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 444 | |
| 445 | static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) |
| 446 | { |
| 447 | struct rds_header *hdr = ic->i_ack; |
| 448 | struct ib_send_wr *failed_wr; |
| 449 | u64 seq; |
| 450 | int ret; |
| 451 | |
| 452 | seq = rds_ib_get_ack(ic); |
| 453 | |
| 454 | rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); |
| 455 | rds_message_populate_header(hdr, 0, 0, 0); |
| 456 | hdr->h_ack = cpu_to_be64(seq); |
| 457 | hdr->h_credit = adv_credits; |
| 458 | rds_message_make_checksum(hdr); |
| 459 | ic->i_ack_queued = jiffies; |
| 460 | |
| 461 | ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); |
| 462 | if (unlikely(ret)) { |
| 463 | /* Failed to send. Release the WR, and |
| 464 | * force another ACK. |
| 465 | */ |
| 466 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); |
| 467 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 468 | |
| 469 | rds_ib_stats_inc(s_ib_ack_send_failure); |
| 470 | /* Need to finesse this later. */ |
| 471 | BUG(); |
| 472 | } else |
| 473 | rds_ib_stats_inc(s_ib_ack_sent); |
| 474 | } |
| 475 | |
| 476 | /* |
| 477 | * There are 3 ways of getting acknowledgements to the peer: |
| 478 | * 1. We call rds_ib_attempt_ack from the recv completion handler |
| 479 | * to send an ACK-only frame. |
| 480 | * However, there can be only one such frame in the send queue |
| 481 | * at any time, so we may have to postpone it. |
| 482 | * 2. When another (data) packet is transmitted while there's |
| 483 | * an ACK in the queue, we piggyback the ACK sequence number |
| 484 | * on the data packet. |
| 485 | * 3. If the ACK WR is done sending, we get called from the |
| 486 | * send queue completion handler, and check whether there's |
| 487 | * another ACK pending (postponed because the WR was on the |
| 488 | * queue). If so, we transmit it. |
| 489 | * |
| 490 | * We maintain 2 variables: |
| 491 | * - i_ack_flags, which keeps track of whether the ACK WR |
| 492 | * is currently in the send queue or not (IB_ACK_IN_FLIGHT) |
| 493 | * - i_ack_next, which is the last sequence number we received |
| 494 | * |
| 495 | * Potentially, send queue and receive queue handlers can run concurrently. |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame^] | 496 | * It would be nice to not have to use a spinlock to synchronize things, |
| 497 | * but the one problem that rules this out is that 64bit updates are |
| 498 | * not atomic on all platforms. Things would be a lot simpler if |
| 499 | * we had atomic64 or maybe cmpxchg64 everywhere. |
Andy Grover | 1e23b3e | 2009-02-24 15:30:34 +0000 | [diff] [blame] | 500 | * |
| 501 | * Reconnecting complicates this picture just slightly. When we |
| 502 | * reconnect, we may be seeing duplicate packets. The peer |
| 503 | * is retransmitting them, because it hasn't seen an ACK for |
| 504 | * them. It is important that we ACK these. |
| 505 | * |
| 506 | * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with |
| 507 | * this flag set *MUST* be acknowledged immediately. |
| 508 | */ |
| 509 | |
| 510 | /* |
| 511 | * When we get here, we're called from the recv queue handler. |
| 512 | * Check whether we ought to transmit an ACK. |
| 513 | */ |
| 514 | void rds_ib_attempt_ack(struct rds_ib_connection *ic) |
| 515 | { |
| 516 | unsigned int adv_credits; |
| 517 | |
| 518 | if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) |
| 519 | return; |
| 520 | |
| 521 | if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { |
| 522 | rds_ib_stats_inc(s_ib_ack_send_delayed); |
| 523 | return; |
| 524 | } |
| 525 | |
| 526 | /* Can we get a send credit? */ |
| 527 | if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0)) { |
| 528 | rds_ib_stats_inc(s_ib_tx_throttle); |
| 529 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); |
| 530 | return; |
| 531 | } |
| 532 | |
| 533 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
| 534 | rds_ib_send_ack(ic, adv_credits); |
| 535 | } |
| 536 | |
| 537 | /* |
| 538 | * We get here from the send completion handler, when the |
| 539 | * adapter tells us the ACK frame was sent. |
| 540 | */ |
| 541 | void rds_ib_ack_send_complete(struct rds_ib_connection *ic) |
| 542 | { |
| 543 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); |
| 544 | rds_ib_attempt_ack(ic); |
| 545 | } |
| 546 | |
| 547 | /* |
| 548 | * This is called by the regular xmit code when it wants to piggyback |
| 549 | * an ACK on an outgoing frame. |
| 550 | */ |
| 551 | u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) |
| 552 | { |
| 553 | if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) |
| 554 | rds_ib_stats_inc(s_ib_ack_send_piggybacked); |
| 555 | return rds_ib_get_ack(ic); |
| 556 | } |
| 557 | |
| 558 | /* |
| 559 | * It's kind of lame that we're copying from the posted receive pages into |
| 560 | * long-lived bitmaps. We could have posted the bitmaps and rdma written into |
| 561 | * them. But receiving new congestion bitmaps should be a *rare* event, so |
| 562 | * hopefully we won't need to invest that complexity in making it more |
| 563 | * efficient. By copying we can share a simpler core with TCP which has to |
| 564 | * copy. |
| 565 | */ |
| 566 | static void rds_ib_cong_recv(struct rds_connection *conn, |
| 567 | struct rds_ib_incoming *ibinc) |
| 568 | { |
| 569 | struct rds_cong_map *map; |
| 570 | unsigned int map_off; |
| 571 | unsigned int map_page; |
| 572 | struct rds_page_frag *frag; |
| 573 | unsigned long frag_off; |
| 574 | unsigned long to_copy; |
| 575 | unsigned long copied; |
| 576 | uint64_t uncongested = 0; |
| 577 | void *addr; |
| 578 | |
| 579 | /* catch completely corrupt packets */ |
| 580 | if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) |
| 581 | return; |
| 582 | |
| 583 | map = conn->c_fcong; |
| 584 | map_page = 0; |
| 585 | map_off = 0; |
| 586 | |
| 587 | frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); |
| 588 | frag_off = 0; |
| 589 | |
| 590 | copied = 0; |
| 591 | |
| 592 | while (copied < RDS_CONG_MAP_BYTES) { |
| 593 | uint64_t *src, *dst; |
| 594 | unsigned int k; |
| 595 | |
| 596 | to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); |
| 597 | BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ |
| 598 | |
| 599 | addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0); |
| 600 | |
| 601 | src = addr + frag_off; |
| 602 | dst = (void *)map->m_page_addrs[map_page] + map_off; |
| 603 | for (k = 0; k < to_copy; k += 8) { |
| 604 | /* Record ports that became uncongested, ie |
| 605 | * bits that changed from 0 to 1. */ |
| 606 | uncongested |= ~(*src) & *dst; |
| 607 | *dst++ = *src++; |
| 608 | } |
| 609 | kunmap_atomic(addr, KM_SOFTIRQ0); |
| 610 | |
| 611 | copied += to_copy; |
| 612 | |
| 613 | map_off += to_copy; |
| 614 | if (map_off == PAGE_SIZE) { |
| 615 | map_off = 0; |
| 616 | map_page++; |
| 617 | } |
| 618 | |
| 619 | frag_off += to_copy; |
| 620 | if (frag_off == RDS_FRAG_SIZE) { |
| 621 | frag = list_entry(frag->f_item.next, |
| 622 | struct rds_page_frag, f_item); |
| 623 | frag_off = 0; |
| 624 | } |
| 625 | } |
| 626 | |
| 627 | /* the congestion map is in little endian order */ |
| 628 | uncongested = le64_to_cpu(uncongested); |
| 629 | |
| 630 | rds_cong_map_updated(map, uncongested); |
| 631 | } |
| 632 | |
| 633 | /* |
| 634 | * Rings are posted with all the allocations they'll need to queue the |
| 635 | * incoming message to the receiving socket so this can't fail. |
| 636 | * All fragments start with a header, so we can make sure we're not receiving |
| 637 | * garbage, and we can tell a small 8 byte fragment from an ACK frame. |
| 638 | */ |
| 639 | struct rds_ib_ack_state { |
| 640 | u64 ack_next; |
| 641 | u64 ack_recv; |
| 642 | unsigned int ack_required:1; |
| 643 | unsigned int ack_next_valid:1; |
| 644 | unsigned int ack_recv_valid:1; |
| 645 | }; |
| 646 | |
| 647 | static void rds_ib_process_recv(struct rds_connection *conn, |
| 648 | struct rds_ib_recv_work *recv, u32 byte_len, |
| 649 | struct rds_ib_ack_state *state) |
| 650 | { |
| 651 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 652 | struct rds_ib_incoming *ibinc = ic->i_ibinc; |
| 653 | struct rds_header *ihdr, *hdr; |
| 654 | |
| 655 | /* XXX shut down the connection if port 0,0 are seen? */ |
| 656 | |
| 657 | rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, |
| 658 | byte_len); |
| 659 | |
| 660 | if (byte_len < sizeof(struct rds_header)) { |
| 661 | rds_ib_conn_error(conn, "incoming message " |
| 662 | "from %pI4 didn't inclue a " |
| 663 | "header, disconnecting and " |
| 664 | "reconnecting\n", |
| 665 | &conn->c_faddr); |
| 666 | return; |
| 667 | } |
| 668 | byte_len -= sizeof(struct rds_header); |
| 669 | |
| 670 | ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; |
| 671 | |
| 672 | /* Validate the checksum. */ |
| 673 | if (!rds_message_verify_checksum(ihdr)) { |
| 674 | rds_ib_conn_error(conn, "incoming message " |
| 675 | "from %pI4 has corrupted header - " |
| 676 | "forcing a reconnect\n", |
| 677 | &conn->c_faddr); |
| 678 | rds_stats_inc(s_recv_drop_bad_checksum); |
| 679 | return; |
| 680 | } |
| 681 | |
| 682 | /* Process the ACK sequence which comes with every packet */ |
| 683 | state->ack_recv = be64_to_cpu(ihdr->h_ack); |
| 684 | state->ack_recv_valid = 1; |
| 685 | |
| 686 | /* Process the credits update if there was one */ |
| 687 | if (ihdr->h_credit) |
| 688 | rds_ib_send_add_credits(conn, ihdr->h_credit); |
| 689 | |
| 690 | if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && byte_len == 0) { |
| 691 | /* This is an ACK-only packet. The fact that it gets |
| 692 | * special treatment here is that historically, ACKs |
| 693 | * were rather special beasts. |
| 694 | */ |
| 695 | rds_ib_stats_inc(s_ib_ack_received); |
| 696 | |
| 697 | /* |
| 698 | * Usually the frags make their way on to incs and are then freed as |
| 699 | * the inc is freed. We don't go that route, so we have to drop the |
| 700 | * page ref ourselves. We can't just leave the page on the recv |
| 701 | * because that confuses the dma mapping of pages and each recv's use |
| 702 | * of a partial page. We can leave the frag, though, it will be |
| 703 | * reused. |
| 704 | * |
| 705 | * FIXME: Fold this into the code path below. |
| 706 | */ |
| 707 | rds_ib_frag_drop_page(recv->r_frag); |
| 708 | return; |
| 709 | } |
| 710 | |
| 711 | /* |
| 712 | * If we don't already have an inc on the connection then this |
| 713 | * fragment has a header and starts a message.. copy its header |
| 714 | * into the inc and save the inc so we can hang upcoming fragments |
| 715 | * off its list. |
| 716 | */ |
| 717 | if (ibinc == NULL) { |
| 718 | ibinc = recv->r_ibinc; |
| 719 | recv->r_ibinc = NULL; |
| 720 | ic->i_ibinc = ibinc; |
| 721 | |
| 722 | hdr = &ibinc->ii_inc.i_hdr; |
| 723 | memcpy(hdr, ihdr, sizeof(*hdr)); |
| 724 | ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); |
| 725 | |
| 726 | rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, |
| 727 | ic->i_recv_data_rem, hdr->h_flags); |
| 728 | } else { |
| 729 | hdr = &ibinc->ii_inc.i_hdr; |
| 730 | /* We can't just use memcmp here; fragments of a |
| 731 | * single message may carry different ACKs */ |
| 732 | if (hdr->h_sequence != ihdr->h_sequence |
| 733 | || hdr->h_len != ihdr->h_len |
| 734 | || hdr->h_sport != ihdr->h_sport |
| 735 | || hdr->h_dport != ihdr->h_dport) { |
| 736 | rds_ib_conn_error(conn, |
| 737 | "fragment header mismatch; forcing reconnect\n"); |
| 738 | return; |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); |
| 743 | recv->r_frag = NULL; |
| 744 | |
| 745 | if (ic->i_recv_data_rem > RDS_FRAG_SIZE) |
| 746 | ic->i_recv_data_rem -= RDS_FRAG_SIZE; |
| 747 | else { |
| 748 | ic->i_recv_data_rem = 0; |
| 749 | ic->i_ibinc = NULL; |
| 750 | |
| 751 | if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) |
| 752 | rds_ib_cong_recv(conn, ibinc); |
| 753 | else { |
| 754 | rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr, |
| 755 | &ibinc->ii_inc, GFP_ATOMIC, |
| 756 | KM_SOFTIRQ0); |
| 757 | state->ack_next = be64_to_cpu(hdr->h_sequence); |
| 758 | state->ack_next_valid = 1; |
| 759 | } |
| 760 | |
| 761 | /* Evaluate the ACK_REQUIRED flag *after* we received |
| 762 | * the complete frame, and after bumping the next_rx |
| 763 | * sequence. */ |
| 764 | if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { |
| 765 | rds_stats_inc(s_recv_ack_required); |
| 766 | state->ack_required = 1; |
| 767 | } |
| 768 | |
| 769 | rds_inc_put(&ibinc->ii_inc); |
| 770 | } |
| 771 | } |
| 772 | |
| 773 | /* |
| 774 | * Plucking the oldest entry from the ring can be done concurrently with |
| 775 | * the thread refilling the ring. Each ring operation is protected by |
| 776 | * spinlocks and the transient state of refilling doesn't change the |
| 777 | * recording of which entry is oldest. |
| 778 | * |
| 779 | * This relies on IB only calling one cq comp_handler for each cq so that |
| 780 | * there will only be one caller of rds_recv_incoming() per RDS connection. |
| 781 | */ |
| 782 | void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context) |
| 783 | { |
| 784 | struct rds_connection *conn = context; |
| 785 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 786 | struct ib_wc wc; |
| 787 | struct rds_ib_ack_state state = { 0, }; |
| 788 | struct rds_ib_recv_work *recv; |
| 789 | |
| 790 | rdsdebug("conn %p cq %p\n", conn, cq); |
| 791 | |
| 792 | rds_ib_stats_inc(s_ib_rx_cq_call); |
| 793 | |
| 794 | ib_req_notify_cq(cq, IB_CQ_SOLICITED); |
| 795 | |
| 796 | while (ib_poll_cq(cq, 1, &wc) > 0) { |
| 797 | rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", |
| 798 | (unsigned long long)wc.wr_id, wc.status, wc.byte_len, |
| 799 | be32_to_cpu(wc.ex.imm_data)); |
| 800 | rds_ib_stats_inc(s_ib_rx_cq_event); |
| 801 | |
| 802 | recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; |
| 803 | |
| 804 | rds_ib_recv_unmap_page(ic, recv); |
| 805 | |
| 806 | /* |
| 807 | * Also process recvs in connecting state because it is possible |
| 808 | * to get a recv completion _before_ the rdmacm ESTABLISHED |
| 809 | * event is processed. |
| 810 | */ |
| 811 | if (rds_conn_up(conn) || rds_conn_connecting(conn)) { |
| 812 | /* We expect errors as the qp is drained during shutdown */ |
| 813 | if (wc.status == IB_WC_SUCCESS) { |
| 814 | rds_ib_process_recv(conn, recv, wc.byte_len, &state); |
| 815 | } else { |
| 816 | rds_ib_conn_error(conn, "recv completion on " |
| 817 | "%pI4 had status %u, disconnecting and " |
| 818 | "reconnecting\n", &conn->c_faddr, |
| 819 | wc.status); |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | rds_ib_ring_free(&ic->i_recv_ring, 1); |
| 824 | } |
| 825 | |
| 826 | if (state.ack_next_valid) |
| 827 | rds_ib_set_ack(ic, state.ack_next, state.ack_required); |
| 828 | if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { |
| 829 | rds_send_drop_acked(conn, state.ack_recv, NULL); |
| 830 | ic->i_ack_recv = state.ack_recv; |
| 831 | } |
| 832 | if (rds_conn_up(conn)) |
| 833 | rds_ib_attempt_ack(ic); |
| 834 | |
| 835 | /* If we ever end up with a really empty receive ring, we're |
| 836 | * in deep trouble, as the sender will definitely see RNR |
| 837 | * timeouts. */ |
| 838 | if (rds_ib_ring_empty(&ic->i_recv_ring)) |
| 839 | rds_ib_stats_inc(s_ib_rx_ring_empty); |
| 840 | |
| 841 | /* |
| 842 | * If the ring is running low, then schedule the thread to refill. |
| 843 | */ |
| 844 | if (rds_ib_ring_low(&ic->i_recv_ring)) |
| 845 | queue_delayed_work(rds_wq, &conn->c_recv_w, 0); |
| 846 | } |
| 847 | |
| 848 | int rds_ib_recv(struct rds_connection *conn) |
| 849 | { |
| 850 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 851 | int ret = 0; |
| 852 | |
| 853 | rdsdebug("conn %p\n", conn); |
| 854 | |
| 855 | /* |
| 856 | * If we get a temporary posting failure in this context then |
| 857 | * we're really low and we want the caller to back off for a bit. |
| 858 | */ |
| 859 | mutex_lock(&ic->i_recv_mutex); |
| 860 | if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0)) |
| 861 | ret = -ENOMEM; |
| 862 | else |
| 863 | rds_ib_stats_inc(s_ib_rx_refill_from_thread); |
| 864 | mutex_unlock(&ic->i_recv_mutex); |
| 865 | |
| 866 | if (rds_conn_up(conn)) |
| 867 | rds_ib_attempt_ack(ic); |
| 868 | |
| 869 | return ret; |
| 870 | } |
| 871 | |
| 872 | int __init rds_ib_recv_init(void) |
| 873 | { |
| 874 | struct sysinfo si; |
| 875 | int ret = -ENOMEM; |
| 876 | |
| 877 | /* Default to 30% of all available RAM for recv memory */ |
| 878 | si_meminfo(&si); |
| 879 | rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; |
| 880 | |
| 881 | rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", |
| 882 | sizeof(struct rds_ib_incoming), |
| 883 | 0, 0, NULL); |
| 884 | if (rds_ib_incoming_slab == NULL) |
| 885 | goto out; |
| 886 | |
| 887 | rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", |
| 888 | sizeof(struct rds_page_frag), |
| 889 | 0, 0, NULL); |
| 890 | if (rds_ib_frag_slab == NULL) |
| 891 | kmem_cache_destroy(rds_ib_incoming_slab); |
| 892 | else |
| 893 | ret = 0; |
| 894 | out: |
| 895 | return ret; |
| 896 | } |
| 897 | |
| 898 | void rds_ib_recv_exit(void) |
| 899 | { |
| 900 | kmem_cache_destroy(rds_ib_incoming_slab); |
| 901 | kmem_cache_destroy(rds_ib_frag_slab); |
| 902 | } |