Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/slab.h>
0035 #include <linux/pci.h>
0036 #include <linux/dma-mapping.h>
0037 #include <rdma/rdma_cm.h>
0038 
0039 #include "rds_single_path.h"
0040 #include "rds.h"
0041 #include "ib.h"
0042 
0043 static struct kmem_cache *rds_ib_incoming_slab;
0044 static struct kmem_cache *rds_ib_frag_slab;
0045 static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
0046 
0047 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
0048 {
0049     struct rds_ib_recv_work *recv;
0050     u32 i;
0051 
0052     for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
0053         struct ib_sge *sge;
0054 
0055         recv->r_ibinc = NULL;
0056         recv->r_frag = NULL;
0057 
0058         recv->r_wr.next = NULL;
0059         recv->r_wr.wr_id = i;
0060         recv->r_wr.sg_list = recv->r_sge;
0061         recv->r_wr.num_sge = RDS_IB_RECV_SGE;
0062 
0063         sge = &recv->r_sge[0];
0064         sge->addr = ic->i_recv_hdrs_dma[i];
0065         sge->length = sizeof(struct rds_header);
0066         sge->lkey = ic->i_pd->local_dma_lkey;
0067 
0068         sge = &recv->r_sge[1];
0069         sge->addr = 0;
0070         sge->length = RDS_FRAG_SIZE;
0071         sge->lkey = ic->i_pd->local_dma_lkey;
0072     }
0073 }
0074 
0075 /*
0076  * The entire 'from' list, including the from element itself, is put on
0077  * to the tail of the 'to' list.
0078  */
0079 static void list_splice_entire_tail(struct list_head *from,
0080                     struct list_head *to)
0081 {
0082     struct list_head *from_last = from->prev;
0083 
0084     list_splice_tail(from_last, to);
0085     list_add_tail(from_last, to);
0086 }
0087 
0088 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
0089 {
0090     struct list_head *tmp;
0091 
0092     tmp = xchg(&cache->xfer, NULL);
0093     if (tmp) {
0094         if (cache->ready)
0095             list_splice_entire_tail(tmp, cache->ready);
0096         else
0097             cache->ready = tmp;
0098     }
0099 }
0100 
0101 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
0102 {
0103     struct rds_ib_cache_head *head;
0104     int cpu;
0105 
0106     cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
0107     if (!cache->percpu)
0108            return -ENOMEM;
0109 
0110     for_each_possible_cpu(cpu) {
0111         head = per_cpu_ptr(cache->percpu, cpu);
0112         head->first = NULL;
0113         head->count = 0;
0114     }
0115     cache->xfer = NULL;
0116     cache->ready = NULL;
0117 
0118     return 0;
0119 }
0120 
0121 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
0122 {
0123     int ret;
0124 
0125     ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
0126     if (!ret) {
0127         ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
0128         if (ret)
0129             free_percpu(ic->i_cache_incs.percpu);
0130     }
0131 
0132     return ret;
0133 }
0134 
0135 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
0136                       struct list_head *caller_list)
0137 {
0138     struct rds_ib_cache_head *head;
0139     int cpu;
0140 
0141     for_each_possible_cpu(cpu) {
0142         head = per_cpu_ptr(cache->percpu, cpu);
0143         if (head->first) {
0144             list_splice_entire_tail(head->first, caller_list);
0145             head->first = NULL;
0146         }
0147     }
0148 
0149     if (cache->ready) {
0150         list_splice_entire_tail(cache->ready, caller_list);
0151         cache->ready = NULL;
0152     }
0153 }
0154 
0155 void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
0156 {
0157     struct rds_ib_incoming *inc;
0158     struct rds_ib_incoming *inc_tmp;
0159     struct rds_page_frag *frag;
0160     struct rds_page_frag *frag_tmp;
0161     LIST_HEAD(list);
0162 
0163     rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
0164     rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
0165     free_percpu(ic->i_cache_incs.percpu);
0166 
0167     list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
0168         list_del(&inc->ii_cache_entry);
0169         WARN_ON(!list_empty(&inc->ii_frags));
0170         kmem_cache_free(rds_ib_incoming_slab, inc);
0171         atomic_dec(&rds_ib_allocation);
0172     }
0173 
0174     rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
0175     rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
0176     free_percpu(ic->i_cache_frags.percpu);
0177 
0178     list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
0179         list_del(&frag->f_cache_entry);
0180         WARN_ON(!list_empty(&frag->f_item));
0181         kmem_cache_free(rds_ib_frag_slab, frag);
0182     }
0183 }
0184 
0185 /* fwd decl */
0186 static void rds_ib_recv_cache_put(struct list_head *new_item,
0187                   struct rds_ib_refill_cache *cache);
0188 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
0189 
0190 
0191 /* Recycle frag and attached recv buffer f_sg */
0192 static void rds_ib_frag_free(struct rds_ib_connection *ic,
0193                  struct rds_page_frag *frag)
0194 {
0195     rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
0196 
0197     rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
0198     atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
0199     rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
0200 }
0201 
0202 /* Recycle inc after freeing attached frags */
0203 void rds_ib_inc_free(struct rds_incoming *inc)
0204 {
0205     struct rds_ib_incoming *ibinc;
0206     struct rds_page_frag *frag;
0207     struct rds_page_frag *pos;
0208     struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
0209 
0210     ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
0211 
0212     /* Free attached frags */
0213     list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
0214         list_del_init(&frag->f_item);
0215         rds_ib_frag_free(ic, frag);
0216     }
0217     BUG_ON(!list_empty(&ibinc->ii_frags));
0218 
0219     rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
0220     rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
0221 }
0222 
0223 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
0224                   struct rds_ib_recv_work *recv)
0225 {
0226     if (recv->r_ibinc) {
0227         rds_inc_put(&recv->r_ibinc->ii_inc);
0228         recv->r_ibinc = NULL;
0229     }
0230     if (recv->r_frag) {
0231         ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
0232         rds_ib_frag_free(ic, recv->r_frag);
0233         recv->r_frag = NULL;
0234     }
0235 }
0236 
0237 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
0238 {
0239     u32 i;
0240 
0241     for (i = 0; i < ic->i_recv_ring.w_nr; i++)
0242         rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
0243 }
0244 
0245 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
0246                              gfp_t slab_mask)
0247 {
0248     struct rds_ib_incoming *ibinc;
0249     struct list_head *cache_item;
0250     int avail_allocs;
0251 
0252     cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
0253     if (cache_item) {
0254         ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
0255     } else {
0256         avail_allocs = atomic_add_unless(&rds_ib_allocation,
0257                          1, rds_ib_sysctl_max_recv_allocation);
0258         if (!avail_allocs) {
0259             rds_ib_stats_inc(s_ib_rx_alloc_limit);
0260             return NULL;
0261         }
0262         ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
0263         if (!ibinc) {
0264             atomic_dec(&rds_ib_allocation);
0265             return NULL;
0266         }
0267         rds_ib_stats_inc(s_ib_rx_total_incs);
0268     }
0269     INIT_LIST_HEAD(&ibinc->ii_frags);
0270     rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
0271 
0272     return ibinc;
0273 }
0274 
0275 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
0276                             gfp_t slab_mask, gfp_t page_mask)
0277 {
0278     struct rds_page_frag *frag;
0279     struct list_head *cache_item;
0280     int ret;
0281 
0282     cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
0283     if (cache_item) {
0284         frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
0285         atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
0286         rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
0287     } else {
0288         frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
0289         if (!frag)
0290             return NULL;
0291 
0292         sg_init_table(&frag->f_sg, 1);
0293         ret = rds_page_remainder_alloc(&frag->f_sg,
0294                            RDS_FRAG_SIZE, page_mask);
0295         if (ret) {
0296             kmem_cache_free(rds_ib_frag_slab, frag);
0297             return NULL;
0298         }
0299         rds_ib_stats_inc(s_ib_rx_total_frags);
0300     }
0301 
0302     INIT_LIST_HEAD(&frag->f_item);
0303 
0304     return frag;
0305 }
0306 
0307 static int rds_ib_recv_refill_one(struct rds_connection *conn,
0308                   struct rds_ib_recv_work *recv, gfp_t gfp)
0309 {
0310     struct rds_ib_connection *ic = conn->c_transport_data;
0311     struct ib_sge *sge;
0312     int ret = -ENOMEM;
0313     gfp_t slab_mask = gfp;
0314     gfp_t page_mask = gfp;
0315 
0316     if (gfp & __GFP_DIRECT_RECLAIM) {
0317         slab_mask = GFP_KERNEL;
0318         page_mask = GFP_HIGHUSER;
0319     }
0320 
0321     if (!ic->i_cache_incs.ready)
0322         rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
0323     if (!ic->i_cache_frags.ready)
0324         rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
0325 
0326     /*
0327      * ibinc was taken from recv if recv contained the start of a message.
0328      * recvs that were continuations will still have this allocated.
0329      */
0330     if (!recv->r_ibinc) {
0331         recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
0332         if (!recv->r_ibinc)
0333             goto out;
0334     }
0335 
0336     WARN_ON(recv->r_frag); /* leak! */
0337     recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
0338     if (!recv->r_frag)
0339         goto out;
0340 
0341     ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
0342                 1, DMA_FROM_DEVICE);
0343     WARN_ON(ret != 1);
0344 
0345     sge = &recv->r_sge[0];
0346     sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
0347     sge->length = sizeof(struct rds_header);
0348 
0349     sge = &recv->r_sge[1];
0350     sge->addr = sg_dma_address(&recv->r_frag->f_sg);
0351     sge->length = sg_dma_len(&recv->r_frag->f_sg);
0352 
0353     ret = 0;
0354 out:
0355     return ret;
0356 }
0357 
0358 static int acquire_refill(struct rds_connection *conn)
0359 {
0360     return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
0361 }
0362 
0363 static void release_refill(struct rds_connection *conn)
0364 {
0365     clear_bit(RDS_RECV_REFILL, &conn->c_flags);
0366     smp_mb__after_atomic();
0367 
0368     /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
0369      * hot path and finding waiters is very rare.  We don't want to walk
0370      * the system-wide hashed waitqueue buckets in the fast path only to
0371      * almost never find waiters.
0372      */
0373     if (waitqueue_active(&conn->c_waitq))
0374         wake_up_all(&conn->c_waitq);
0375 }
0376 
0377 /*
0378  * This tries to allocate and post unused work requests after making sure that
0379  * they have all the allocations they need to queue received fragments into
0380  * sockets.
0381  */
0382 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
0383 {
0384     struct rds_ib_connection *ic = conn->c_transport_data;
0385     struct rds_ib_recv_work *recv;
0386     unsigned int posted = 0;
0387     int ret = 0;
0388     bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
0389     bool must_wake = false;
0390     u32 pos;
0391 
0392     /* the goal here is to just make sure that someone, somewhere
0393      * is posting buffers.  If we can't get the refill lock,
0394      * let them do their thing
0395      */
0396     if (!acquire_refill(conn))
0397         return;
0398 
0399     while ((prefill || rds_conn_up(conn)) &&
0400            rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
0401         if (pos >= ic->i_recv_ring.w_nr) {
0402             printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
0403                     pos);
0404             break;
0405         }
0406 
0407         recv = &ic->i_recvs[pos];
0408         ret = rds_ib_recv_refill_one(conn, recv, gfp);
0409         if (ret) {
0410             must_wake = true;
0411             break;
0412         }
0413 
0414         rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
0415              recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
0416              (long)sg_dma_address(&recv->r_frag->f_sg));
0417 
0418         /* XXX when can this fail? */
0419         ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
0420         if (ret) {
0421             rds_ib_conn_error(conn, "recv post on "
0422                    "%pI6c returned %d, disconnecting and "
0423                    "reconnecting\n", &conn->c_faddr,
0424                    ret);
0425             break;
0426         }
0427 
0428         posted++;
0429 
0430         if ((posted > 128 && need_resched()) || posted > 8192) {
0431             must_wake = true;
0432             break;
0433         }
0434     }
0435 
0436     /* We're doing flow control - update the window. */
0437     if (ic->i_flowctl && posted)
0438         rds_ib_advertise_credits(conn, posted);
0439 
0440     if (ret)
0441         rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
0442 
0443     release_refill(conn);
0444 
0445     /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
0446      * in this case the ring being low is going to lead to more interrupts
0447      * and we can safely let the softirq code take care of it unless the
0448      * ring is completely empty.
0449      *
0450      * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
0451      * we might have raced with the softirq code while we had the refill
0452      * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
0453      * if we should requeue.
0454      */
0455     if (rds_conn_up(conn) &&
0456         (must_wake ||
0457         (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
0458         rds_ib_ring_empty(&ic->i_recv_ring))) {
0459         queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
0460     }
0461     if (can_wait)
0462         cond_resched();
0463 }
0464 
0465 /*
0466  * We want to recycle several types of recv allocations, like incs and frags.
0467  * To use this, the *_free() function passes in the ptr to a list_head within
0468  * the recyclee, as well as the cache to put it on.
0469  *
0470  * First, we put the memory on a percpu list. When this reaches a certain size,
0471  * We move it to an intermediate non-percpu list in a lockless manner, with some
0472  * xchg/compxchg wizardry.
0473  *
0474  * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
0475  * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
0476  * list_empty() will return true with one element is actually present.
0477  */
0478 static void rds_ib_recv_cache_put(struct list_head *new_item,
0479                  struct rds_ib_refill_cache *cache)
0480 {
0481     unsigned long flags;
0482     struct list_head *old, *chpfirst;
0483 
0484     local_irq_save(flags);
0485 
0486     chpfirst = __this_cpu_read(cache->percpu->first);
0487     if (!chpfirst)
0488         INIT_LIST_HEAD(new_item);
0489     else /* put on front */
0490         list_add_tail(new_item, chpfirst);
0491 
0492     __this_cpu_write(cache->percpu->first, new_item);
0493     __this_cpu_inc(cache->percpu->count);
0494 
0495     if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
0496         goto end;
0497 
0498     /*
0499      * Return our per-cpu first list to the cache's xfer by atomically
0500      * grabbing the current xfer list, appending it to our per-cpu list,
0501      * and then atomically returning that entire list back to the
0502      * cache's xfer list as long as it's still empty.
0503      */
0504     do {
0505         old = xchg(&cache->xfer, NULL);
0506         if (old)
0507             list_splice_entire_tail(old, chpfirst);
0508         old = cmpxchg(&cache->xfer, NULL, chpfirst);
0509     } while (old);
0510 
0511 
0512     __this_cpu_write(cache->percpu->first, NULL);
0513     __this_cpu_write(cache->percpu->count, 0);
0514 end:
0515     local_irq_restore(flags);
0516 }
0517 
0518 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
0519 {
0520     struct list_head *head = cache->ready;
0521 
0522     if (head) {
0523         if (!list_empty(head)) {
0524             cache->ready = head->next;
0525             list_del_init(head);
0526         } else
0527             cache->ready = NULL;
0528     }
0529 
0530     return head;
0531 }
0532 
0533 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
0534 {
0535     struct rds_ib_incoming *ibinc;
0536     struct rds_page_frag *frag;
0537     unsigned long to_copy;
0538     unsigned long frag_off = 0;
0539     int copied = 0;
0540     int ret;
0541     u32 len;
0542 
0543     ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
0544     frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
0545     len = be32_to_cpu(inc->i_hdr.h_len);
0546 
0547     while (iov_iter_count(to) && copied < len) {
0548         if (frag_off == RDS_FRAG_SIZE) {
0549             frag = list_entry(frag->f_item.next,
0550                       struct rds_page_frag, f_item);
0551             frag_off = 0;
0552         }
0553         to_copy = min_t(unsigned long, iov_iter_count(to),
0554                 RDS_FRAG_SIZE - frag_off);
0555         to_copy = min_t(unsigned long, to_copy, len - copied);
0556 
0557         /* XXX needs + offset for multiple recvs per page */
0558         rds_stats_add(s_copy_to_user, to_copy);
0559         ret = copy_page_to_iter(sg_page(&frag->f_sg),
0560                     frag->f_sg.offset + frag_off,
0561                     to_copy,
0562                     to);
0563         if (ret != to_copy)
0564             return -EFAULT;
0565 
0566         frag_off += to_copy;
0567         copied += to_copy;
0568     }
0569 
0570     return copied;
0571 }
0572 
0573 /* ic starts out kzalloc()ed */
0574 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
0575 {
0576     struct ib_send_wr *wr = &ic->i_ack_wr;
0577     struct ib_sge *sge = &ic->i_ack_sge;
0578 
0579     sge->addr = ic->i_ack_dma;
0580     sge->length = sizeof(struct rds_header);
0581     sge->lkey = ic->i_pd->local_dma_lkey;
0582 
0583     wr->sg_list = sge;
0584     wr->num_sge = 1;
0585     wr->opcode = IB_WR_SEND;
0586     wr->wr_id = RDS_IB_ACK_WR_ID;
0587     wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
0588 }
0589 
0590 /*
0591  * You'd think that with reliable IB connections you wouldn't need to ack
0592  * messages that have been received.  The problem is that IB hardware generates
0593  * an ack message before it has DMAed the message into memory.  This creates a
0594  * potential message loss if the HCA is disabled for any reason between when it
0595  * sends the ack and before the message is DMAed and processed.  This is only a
0596  * potential issue if another HCA is available for fail-over.
0597  *
0598  * When the remote host receives our ack they'll free the sent message from
0599  * their send queue.  To decrease the latency of this we always send an ack
0600  * immediately after we've received messages.
0601  *
0602  * For simplicity, we only have one ack in flight at a time.  This puts
0603  * pressure on senders to have deep enough send queues to absorb the latency of
0604  * a single ack frame being in flight.  This might not be good enough.
0605  *
0606  * This is implemented by have a long-lived send_wr and sge which point to a
0607  * statically allocated ack frame.  This ack wr does not fall under the ring
0608  * accounting that the tx and rx wrs do.  The QP attribute specifically makes
0609  * room for it beyond the ring size.  Send completion notices its special
0610  * wr_id and avoids working with the ring in that case.
0611  */
0612 #ifndef KERNEL_HAS_ATOMIC64
0613 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
0614 {
0615     unsigned long flags;
0616 
0617     spin_lock_irqsave(&ic->i_ack_lock, flags);
0618     ic->i_ack_next = seq;
0619     if (ack_required)
0620         set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0621     spin_unlock_irqrestore(&ic->i_ack_lock, flags);
0622 }
0623 
0624 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
0625 {
0626     unsigned long flags;
0627     u64 seq;
0628 
0629     clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0630 
0631     spin_lock_irqsave(&ic->i_ack_lock, flags);
0632     seq = ic->i_ack_next;
0633     spin_unlock_irqrestore(&ic->i_ack_lock, flags);
0634 
0635     return seq;
0636 }
0637 #else
0638 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
0639 {
0640     atomic64_set(&ic->i_ack_next, seq);
0641     if (ack_required) {
0642         smp_mb__before_atomic();
0643         set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0644     }
0645 }
0646 
0647 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
0648 {
0649     clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0650     smp_mb__after_atomic();
0651 
0652     return atomic64_read(&ic->i_ack_next);
0653 }
0654 #endif
0655 
0656 
0657 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
0658 {
0659     struct rds_header *hdr = ic->i_ack;
0660     u64 seq;
0661     int ret;
0662 
0663     seq = rds_ib_get_ack(ic);
0664 
0665     rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
0666 
0667     ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
0668                    sizeof(*hdr), DMA_TO_DEVICE);
0669     rds_message_populate_header(hdr, 0, 0, 0);
0670     hdr->h_ack = cpu_to_be64(seq);
0671     hdr->h_credit = adv_credits;
0672     rds_message_make_checksum(hdr);
0673     ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
0674                       sizeof(*hdr), DMA_TO_DEVICE);
0675 
0676     ic->i_ack_queued = jiffies;
0677 
0678     ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
0679     if (unlikely(ret)) {
0680         /* Failed to send. Release the WR, and
0681          * force another ACK.
0682          */
0683         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
0684         set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0685 
0686         rds_ib_stats_inc(s_ib_ack_send_failure);
0687 
0688         rds_ib_conn_error(ic->conn, "sending ack failed\n");
0689     } else
0690         rds_ib_stats_inc(s_ib_ack_sent);
0691 }
0692 
0693 /*
0694  * There are 3 ways of getting acknowledgements to the peer:
0695  *  1.  We call rds_ib_attempt_ack from the recv completion handler
0696  *  to send an ACK-only frame.
0697  *  However, there can be only one such frame in the send queue
0698  *  at any time, so we may have to postpone it.
0699  *  2.  When another (data) packet is transmitted while there's
0700  *  an ACK in the queue, we piggyback the ACK sequence number
0701  *  on the data packet.
0702  *  3.  If the ACK WR is done sending, we get called from the
0703  *  send queue completion handler, and check whether there's
0704  *  another ACK pending (postponed because the WR was on the
0705  *  queue). If so, we transmit it.
0706  *
0707  * We maintain 2 variables:
0708  *  -   i_ack_flags, which keeps track of whether the ACK WR
0709  *  is currently in the send queue or not (IB_ACK_IN_FLIGHT)
0710  *  -   i_ack_next, which is the last sequence number we received
0711  *
0712  * Potentially, send queue and receive queue handlers can run concurrently.
0713  * It would be nice to not have to use a spinlock to synchronize things,
0714  * but the one problem that rules this out is that 64bit updates are
0715  * not atomic on all platforms. Things would be a lot simpler if
0716  * we had atomic64 or maybe cmpxchg64 everywhere.
0717  *
0718  * Reconnecting complicates this picture just slightly. When we
0719  * reconnect, we may be seeing duplicate packets. The peer
0720  * is retransmitting them, because it hasn't seen an ACK for
0721  * them. It is important that we ACK these.
0722  *
0723  * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
0724  * this flag set *MUST* be acknowledged immediately.
0725  */
0726 
0727 /*
0728  * When we get here, we're called from the recv queue handler.
0729  * Check whether we ought to transmit an ACK.
0730  */
0731 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
0732 {
0733     unsigned int adv_credits;
0734 
0735     if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
0736         return;
0737 
0738     if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
0739         rds_ib_stats_inc(s_ib_ack_send_delayed);
0740         return;
0741     }
0742 
0743     /* Can we get a send credit? */
0744     if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
0745         rds_ib_stats_inc(s_ib_tx_throttle);
0746         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
0747         return;
0748     }
0749 
0750     clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
0751     rds_ib_send_ack(ic, adv_credits);
0752 }
0753 
0754 /*
0755  * We get here from the send completion handler, when the
0756  * adapter tells us the ACK frame was sent.
0757  */
0758 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
0759 {
0760     clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
0761     rds_ib_attempt_ack(ic);
0762 }
0763 
0764 /*
0765  * This is called by the regular xmit code when it wants to piggyback
0766  * an ACK on an outgoing frame.
0767  */
0768 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
0769 {
0770     if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
0771         rds_ib_stats_inc(s_ib_ack_send_piggybacked);
0772     return rds_ib_get_ack(ic);
0773 }
0774 
0775 /*
0776  * It's kind of lame that we're copying from the posted receive pages into
0777  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
0778  * them.  But receiving new congestion bitmaps should be a *rare* event, so
0779  * hopefully we won't need to invest that complexity in making it more
0780  * efficient.  By copying we can share a simpler core with TCP which has to
0781  * copy.
0782  */
0783 static void rds_ib_cong_recv(struct rds_connection *conn,
0784                   struct rds_ib_incoming *ibinc)
0785 {
0786     struct rds_cong_map *map;
0787     unsigned int map_off;
0788     unsigned int map_page;
0789     struct rds_page_frag *frag;
0790     unsigned long frag_off;
0791     unsigned long to_copy;
0792     unsigned long copied;
0793     __le64 uncongested = 0;
0794     void *addr;
0795 
0796     /* catch completely corrupt packets */
0797     if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
0798         return;
0799 
0800     map = conn->c_fcong;
0801     map_page = 0;
0802     map_off = 0;
0803 
0804     frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
0805     frag_off = 0;
0806 
0807     copied = 0;
0808 
0809     while (copied < RDS_CONG_MAP_BYTES) {
0810         __le64 *src, *dst;
0811         unsigned int k;
0812 
0813         to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
0814         BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
0815 
0816         addr = kmap_atomic(sg_page(&frag->f_sg));
0817 
0818         src = addr + frag->f_sg.offset + frag_off;
0819         dst = (void *)map->m_page_addrs[map_page] + map_off;
0820         for (k = 0; k < to_copy; k += 8) {
0821             /* Record ports that became uncongested, ie
0822              * bits that changed from 0 to 1. */
0823             uncongested |= ~(*src) & *dst;
0824             *dst++ = *src++;
0825         }
0826         kunmap_atomic(addr);
0827 
0828         copied += to_copy;
0829 
0830         map_off += to_copy;
0831         if (map_off == PAGE_SIZE) {
0832             map_off = 0;
0833             map_page++;
0834         }
0835 
0836         frag_off += to_copy;
0837         if (frag_off == RDS_FRAG_SIZE) {
0838             frag = list_entry(frag->f_item.next,
0839                       struct rds_page_frag, f_item);
0840             frag_off = 0;
0841         }
0842     }
0843 
0844     /* the congestion map is in little endian order */
0845     rds_cong_map_updated(map, le64_to_cpu(uncongested));
0846 }
0847 
0848 static void rds_ib_process_recv(struct rds_connection *conn,
0849                 struct rds_ib_recv_work *recv, u32 data_len,
0850                 struct rds_ib_ack_state *state)
0851 {
0852     struct rds_ib_connection *ic = conn->c_transport_data;
0853     struct rds_ib_incoming *ibinc = ic->i_ibinc;
0854     struct rds_header *ihdr, *hdr;
0855     dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
0856 
0857     /* XXX shut down the connection if port 0,0 are seen? */
0858 
0859     rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
0860          data_len);
0861 
0862     if (data_len < sizeof(struct rds_header)) {
0863         rds_ib_conn_error(conn, "incoming message "
0864                "from %pI6c didn't include a "
0865                "header, disconnecting and "
0866                "reconnecting\n",
0867                &conn->c_faddr);
0868         return;
0869     }
0870     data_len -= sizeof(struct rds_header);
0871 
0872     ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
0873 
0874     ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
0875                    sizeof(*ihdr), DMA_FROM_DEVICE);
0876     /* Validate the checksum. */
0877     if (!rds_message_verify_checksum(ihdr)) {
0878         rds_ib_conn_error(conn, "incoming message "
0879                "from %pI6c has corrupted header - "
0880                "forcing a reconnect\n",
0881                &conn->c_faddr);
0882         rds_stats_inc(s_recv_drop_bad_checksum);
0883         goto done;
0884     }
0885 
0886     /* Process the ACK sequence which comes with every packet */
0887     state->ack_recv = be64_to_cpu(ihdr->h_ack);
0888     state->ack_recv_valid = 1;
0889 
0890     /* Process the credits update if there was one */
0891     if (ihdr->h_credit)
0892         rds_ib_send_add_credits(conn, ihdr->h_credit);
0893 
0894     if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
0895         /* This is an ACK-only packet. The fact that it gets
0896          * special treatment here is that historically, ACKs
0897          * were rather special beasts.
0898          */
0899         rds_ib_stats_inc(s_ib_ack_received);
0900 
0901         /*
0902          * Usually the frags make their way on to incs and are then freed as
0903          * the inc is freed.  We don't go that route, so we have to drop the
0904          * page ref ourselves.  We can't just leave the page on the recv
0905          * because that confuses the dma mapping of pages and each recv's use
0906          * of a partial page.
0907          *
0908          * FIXME: Fold this into the code path below.
0909          */
0910         rds_ib_frag_free(ic, recv->r_frag);
0911         recv->r_frag = NULL;
0912         goto done;
0913     }
0914 
0915     /*
0916      * If we don't already have an inc on the connection then this
0917      * fragment has a header and starts a message.. copy its header
0918      * into the inc and save the inc so we can hang upcoming fragments
0919      * off its list.
0920      */
0921     if (!ibinc) {
0922         ibinc = recv->r_ibinc;
0923         recv->r_ibinc = NULL;
0924         ic->i_ibinc = ibinc;
0925 
0926         hdr = &ibinc->ii_inc.i_hdr;
0927         ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
0928                 local_clock();
0929         memcpy(hdr, ihdr, sizeof(*hdr));
0930         ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
0931         ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
0932                 local_clock();
0933 
0934         rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
0935              ic->i_recv_data_rem, hdr->h_flags);
0936     } else {
0937         hdr = &ibinc->ii_inc.i_hdr;
0938         /* We can't just use memcmp here; fragments of a
0939          * single message may carry different ACKs */
0940         if (hdr->h_sequence != ihdr->h_sequence ||
0941             hdr->h_len != ihdr->h_len ||
0942             hdr->h_sport != ihdr->h_sport ||
0943             hdr->h_dport != ihdr->h_dport) {
0944             rds_ib_conn_error(conn,
0945                 "fragment header mismatch; forcing reconnect\n");
0946             goto done;
0947         }
0948     }
0949 
0950     list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
0951     recv->r_frag = NULL;
0952 
0953     if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
0954         ic->i_recv_data_rem -= RDS_FRAG_SIZE;
0955     else {
0956         ic->i_recv_data_rem = 0;
0957         ic->i_ibinc = NULL;
0958 
0959         if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
0960             rds_ib_cong_recv(conn, ibinc);
0961         } else {
0962             rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
0963                       &ibinc->ii_inc, GFP_ATOMIC);
0964             state->ack_next = be64_to_cpu(hdr->h_sequence);
0965             state->ack_next_valid = 1;
0966         }
0967 
0968         /* Evaluate the ACK_REQUIRED flag *after* we received
0969          * the complete frame, and after bumping the next_rx
0970          * sequence. */
0971         if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
0972             rds_stats_inc(s_recv_ack_required);
0973             state->ack_required = 1;
0974         }
0975 
0976         rds_inc_put(&ibinc->ii_inc);
0977     }
0978 done:
0979     ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
0980                       sizeof(*ihdr), DMA_FROM_DEVICE);
0981 }
0982 
0983 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
0984                  struct ib_wc *wc,
0985                  struct rds_ib_ack_state *state)
0986 {
0987     struct rds_connection *conn = ic->conn;
0988     struct rds_ib_recv_work *recv;
0989 
0990     rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
0991          (unsigned long long)wc->wr_id, wc->status,
0992          ib_wc_status_msg(wc->status), wc->byte_len,
0993          be32_to_cpu(wc->ex.imm_data));
0994 
0995     rds_ib_stats_inc(s_ib_rx_cq_event);
0996     recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
0997     ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
0998             DMA_FROM_DEVICE);
0999 
1000     /* Also process recvs in connecting state because it is possible
1001      * to get a recv completion _before_ the rdmacm ESTABLISHED
1002      * event is processed.
1003      */
1004     if (wc->status == IB_WC_SUCCESS) {
1005         rds_ib_process_recv(conn, recv, wc->byte_len, state);
1006     } else {
1007         /* We expect errors as the qp is drained during shutdown */
1008         if (rds_conn_up(conn) || rds_conn_connecting(conn))
1009             rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
1010                       &conn->c_laddr, &conn->c_faddr,
1011                       conn->c_tos, wc->status,
1012                       ib_wc_status_msg(wc->status),
1013                       wc->vendor_err);
1014     }
1015 
1016     /* rds_ib_process_recv() doesn't always consume the frag, and
1017      * we might not have called it at all if the wc didn't indicate
1018      * success. We already unmapped the frag's pages, though, and
1019      * the following rds_ib_ring_free() call tells the refill path
1020      * that it will not find an allocated frag here. Make sure we
1021      * keep that promise by freeing a frag that's still on the ring.
1022      */
1023     if (recv->r_frag) {
1024         rds_ib_frag_free(ic, recv->r_frag);
1025         recv->r_frag = NULL;
1026     }
1027     rds_ib_ring_free(&ic->i_recv_ring, 1);
1028 
1029     /* If we ever end up with a really empty receive ring, we're
1030      * in deep trouble, as the sender will definitely see RNR
1031      * timeouts. */
1032     if (rds_ib_ring_empty(&ic->i_recv_ring))
1033         rds_ib_stats_inc(s_ib_rx_ring_empty);
1034 
1035     if (rds_ib_ring_low(&ic->i_recv_ring)) {
1036         rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
1037         rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1038     }
1039 }
1040 
1041 int rds_ib_recv_path(struct rds_conn_path *cp)
1042 {
1043     struct rds_connection *conn = cp->cp_conn;
1044     struct rds_ib_connection *ic = conn->c_transport_data;
1045 
1046     rdsdebug("conn %p\n", conn);
1047     if (rds_conn_up(conn)) {
1048         rds_ib_attempt_ack(ic);
1049         rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1050         rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1051     }
1052 
1053     return 0;
1054 }
1055 
1056 int rds_ib_recv_init(void)
1057 {
1058     struct sysinfo si;
1059     int ret = -ENOMEM;
1060 
1061     /* Default to 30% of all available RAM for recv memory */
1062     si_meminfo(&si);
1063     rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1064 
1065     rds_ib_incoming_slab =
1066         kmem_cache_create_usercopy("rds_ib_incoming",
1067                        sizeof(struct rds_ib_incoming),
1068                        0, SLAB_HWCACHE_ALIGN,
1069                        offsetof(struct rds_ib_incoming,
1070                             ii_inc.i_usercopy),
1071                        sizeof(struct rds_inc_usercopy),
1072                        NULL);
1073     if (!rds_ib_incoming_slab)
1074         goto out;
1075 
1076     rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1077                     sizeof(struct rds_page_frag),
1078                     0, SLAB_HWCACHE_ALIGN, NULL);
1079     if (!rds_ib_frag_slab) {
1080         kmem_cache_destroy(rds_ib_incoming_slab);
1081         rds_ib_incoming_slab = NULL;
1082     } else
1083         ret = 0;
1084 out:
1085     return ret;
1086 }
1087 
1088 void rds_ib_recv_exit(void)
1089 {
1090     WARN_ON(atomic_read(&rds_ib_allocation));
1091 
1092     kmem_cache_destroy(rds_ib_incoming_slab);
1093     kmem_cache_destroy(rds_ib_frag_slab);
1094 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!