cregit-Linux how code gets into the kernel

Release 4.14 net/rds/ib_recv.c

Directory: net/rds
/*
 * Copyright (c) 2006 Oracle.  All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */
#include <linux/kernel.h>
#include <linux/slab.h>
#include <linux/pci.h>
#include <linux/dma-mapping.h>
#include <rdma/rdma_cm.h>

#include "rds_single_path.h"
#include "rds.h"
#include "ib.h"


static struct kmem_cache *rds_ib_incoming_slab;

static struct kmem_cache *rds_ib_frag_slab;

static atomic_t	rds_ib_allocation = ATOMIC_INIT(0);


void rds_ib_recv_init_ring(struct rds_ib_connection *ic) { struct rds_ib_recv_work *recv; u32 i; for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { struct ib_sge *sge; recv->r_ibinc = NULL; recv->r_frag = NULL; recv->r_wr.next = NULL; recv->r_wr.wr_id = i; recv->r_wr.sg_list = recv->r_sge; recv->r_wr.num_sge = RDS_IB_RECV_SGE; sge = &recv->r_sge[0]; sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); sge->length = sizeof(struct rds_header); sge->lkey = ic->i_pd->local_dma_lkey; sge = &recv->r_sge[1]; sge->addr = 0; sge->length = RDS_FRAG_SIZE; sge->lkey = ic->i_pd->local_dma_lkey; } }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover17397.74%266.67%
Jason Gunthorpe42.26%133.33%
Total177100.00%3100.00%

/* * The entire 'from' list, including the from element itself, is put on * to the tail of the 'to' list. */
static void list_splice_entire_tail(struct list_head *from, struct list_head *to) { struct list_head *from_last = from->prev; list_splice_tail(from_last, to); list_add_tail(from_last, to); }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason2871.79%150.00%
Andy Grover1128.21%150.00%
Total39100.00%2100.00%


static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache) { struct list_head *tmp; tmp = xchg(&cache->xfer, NULL); if (tmp) { if (cache->ready) list_splice_entire_tail(tmp, cache->ready); else cache->ready = tmp; } }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason4580.36%150.00%
Andy Grover1119.64%150.00%
Total56100.00%2100.00%


static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache) { struct rds_ib_cache_head *head; int cpu; cache->percpu = alloc_percpu(struct rds_ib_cache_head); if (!cache->percpu) return -ENOMEM; for_each_possible_cpu(cpu) { head = per_cpu_ptr(cache->percpu, cpu); head->first = NULL; head->count = 0; } cache->xfer = NULL; cache->ready = NULL; return 0; }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason6678.57%133.33%
Andy Grover1821.43%266.67%
Total84100.00%3100.00%


int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic) { int ret; ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs); if (!ret) { ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags); if (ret) free_percpu(ic->i_cache_incs.percpu); } return ret; }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason3867.86%150.00%
Andy Grover1832.14%150.00%
Total56100.00%2100.00%


static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache, struct list_head *caller_list) { struct rds_ib_cache_head *head; int cpu; for_each_possible_cpu(cpu) { head = per_cpu_ptr(cache->percpu, cpu); if (head->first) { list_splice_entire_tail(head->first, caller_list); head->first = NULL; } } if (cache->ready) { list_splice_entire_tail(cache->ready, caller_list); cache->ready = NULL; } }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason7485.06%150.00%
Andy Grover1314.94%150.00%
Total87100.00%2100.00%


void rds_ib_recv_free_caches(struct rds_ib_connection *ic) { struct rds_ib_incoming *inc; struct rds_ib_incoming *inc_tmp; struct rds_page_frag *frag; struct rds_page_frag *frag_tmp; LIST_HEAD(list); rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list); free_percpu(ic->i_cache_incs.percpu); list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) { list_del(&inc->ii_cache_entry); WARN_ON(!list_empty(&inc->ii_frags)); kmem_cache_free(rds_ib_incoming_slab, inc); } rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list); free_percpu(ic->i_cache_frags.percpu); list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) { list_del(&frag->f_cache_entry); WARN_ON(!list_empty(&frag->f_item)); kmem_cache_free(rds_ib_frag_slab, frag); } }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason169100.00%1100.00%
Total169100.00%1100.00%

/* fwd decl */ static void rds_ib_recv_cache_put(struct list_head *new_item, struct rds_ib_refill_cache *cache); static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache); /* Recycle frag and attached recv buffer f_sg */
static void rds_ib_frag_free(struct rds_ib_connection *ic, struct rds_page_frag *frag) { rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg)); rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags); atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason4469.84%150.00%
Santosh Shilimkar1930.16%150.00%
Total63100.00%2100.00%

/* Recycle inc after freeing attached frags */
void rds_ib_inc_free(struct rds_incoming *inc) { struct rds_ib_incoming *ibinc; struct rds_page_frag *frag; struct rds_page_frag *pos; struct rds_ib_connection *ic = inc->i_conn->c_transport_data; ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); /* Free attached frags */ list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { list_del_init(&frag->f_item); rds_ib_frag_free(ic, frag); } BUG_ON(!list_empty(&ibinc->ii_frags)); rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs); }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason110100.00%1100.00%
Total110100.00%1100.00%


static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, struct rds_ib_recv_work *recv) { if (recv->r_ibinc) { rds_inc_put(&recv->r_ibinc->ii_inc); recv->r_ibinc = NULL; } if (recv->r_frag) { ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); rds_ib_frag_free(ic, recv->r_frag); recv->r_frag = NULL; } }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason83100.00%1100.00%
Total83100.00%1100.00%


void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) { u32 i; for (i = 0; i < ic->i_recv_ring.w_nr; i++) rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason43100.00%1100.00%
Total43100.00%1100.00%


static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic, gfp_t slab_mask) { struct rds_ib_incoming *ibinc; struct list_head *cache_item; int avail_allocs; cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs); if (cache_item) { ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry); } else { avail_allocs = atomic_add_unless(&rds_ib_allocation, 1, rds_ib_sysctl_max_recv_allocation); if (!avail_allocs) { rds_ib_stats_inc(s_ib_rx_alloc_limit); return NULL; } ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask); if (!ibinc) { atomic_dec(&rds_ib_allocation); return NULL; } rds_ib_stats_inc(s_ib_rx_total_incs); } INIT_LIST_HEAD(&ibinc->ii_frags); rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr); return ibinc; }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason7954.11%233.33%
Andy Grover6242.47%350.00%
Santosh Shilimkar53.42%116.67%
Total146100.00%6100.00%


static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic, gfp_t slab_mask, gfp_t page_mask) { struct rds_page_frag *frag; struct list_head *cache_item; int ret; cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags); if (cache_item) { frag = container_of(cache_item, struct rds_page_frag, f_cache_entry); atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); } else { frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask); if (!frag) return NULL; sg_init_table(&frag->f_sg, 1); ret = rds_page_remainder_alloc(&frag->f_sg, RDS_FRAG_SIZE, page_mask); if (ret) { kmem_cache_free(rds_ib_frag_slab, frag); return NULL; } rds_ib_stats_inc(s_ib_rx_total_frags); } INIT_LIST_HEAD(&frag->f_item); return frag; }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason13083.87%360.00%
Santosh Shilimkar2415.48%120.00%
Andy Grover10.65%120.00%
Total155100.00%5100.00%


static int rds_ib_recv_refill_one(struct rds_connection *conn, struct rds_ib_recv_work *recv, gfp_t gfp) { struct rds_ib_connection *ic = conn->c_transport_data; struct ib_sge *sge; int ret = -ENOMEM; gfp_t slab_mask = GFP_NOWAIT; gfp_t page_mask = GFP_NOWAIT; if (gfp & __GFP_DIRECT_RECLAIM) { slab_mask = GFP_KERNEL; page_mask = GFP_HIGHUSER; } if (!ic->i_cache_incs.ready) rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); if (!ic->i_cache_frags.ready) rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); /* * ibinc was taken from recv if recv contained the start of a message. * recvs that were continuations will still have this allocated. */ if (!recv->r_ibinc) { recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask); if (!recv->r_ibinc) goto out; } WARN_ON(recv->r_frag); /* leak! */ recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask); if (!recv->r_frag) goto out; ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); WARN_ON(ret != 1); sge = &recv->r_sge[0]; sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); sge->length = sizeof(struct rds_header); sge = &recv->r_sge[1]; sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg); sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg); ret = 0; out: return ret; }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover15753.95%550.00%
Chris Mason11539.52%220.00%
Mike Marciniszyn144.81%110.00%
Santosh Shilimkar41.37%110.00%
Mel Gorman10.34%110.00%
Total291100.00%10100.00%


static int acquire_refill(struct rds_connection *conn) { return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0; }

Contributors

PersonTokensPropCommitsCommitProp
Santosh Shilimkar24100.00%1100.00%
Total24100.00%1100.00%


static void release_refill(struct rds_connection *conn) { clear_bit(RDS_RECV_REFILL, &conn->c_flags); /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a * hot path and finding waiters is very rare. We don't want to walk * the system-wide hashed waitqueue buckets in the fast path only to * almost never find waiters. */ if (waitqueue_active(&conn->c_waitq)) wake_up_all(&conn->c_waitq); }

Contributors

PersonTokensPropCommitsCommitProp
Santosh Shilimkar40100.00%1100.00%
Total40100.00%1100.00%

/* * This tries to allocate and post unused work requests after making sure that * they have all the allocations they need to queue received fragments into * sockets. * * -1 is returned if posting fails due to temporary resource exhaustion. */
void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp) { struct rds_ib_connection *ic = conn->c_transport_data; struct rds_ib_recv_work *recv; struct ib_recv_wr *failed_wr; unsigned int posted = 0; int ret = 0; bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM); u32 pos; /* the goal here is to just make sure that someone, somewhere * is posting buffers. If we can't get the refill lock, * let them do their thing */ if (!acquire_refill(conn)) return; while ((prefill || rds_conn_up(conn)) && rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { if (pos >= ic->i_recv_ring.w_nr) { printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", pos); break; } recv = &ic->i_recvs[pos]; ret = rds_ib_recv_refill_one(conn, recv, gfp); if (ret) { break; } rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv, recv->r_ibinc, sg_page(&recv->r_frag->f_sg), (long) ib_sg_dma_address( ic->i_cm_id->device, &recv->r_frag->f_sg)); /* XXX when can this fail? */ ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); if (ret) { rds_ib_conn_error(conn, "recv post on " "%pI4 returned %d, disconnecting and " "reconnecting\n", &conn->c_faddr, ret); break; } posted++; } /* We're doing flow control - update the window. */ if (ic->i_flowctl && posted) rds_ib_advertise_credits(conn, posted); if (ret) rds_ib_ring_unalloc(&ic->i_recv_ring, 1); release_refill(conn); /* if we're called from the softirq handler, we'll be GFP_NOWAIT. * in this case the ring being low is going to lead to more interrupts * and we can safely let the softirq code take care of it unless the * ring is completely empty. * * if we're called from krdsd, we'll be GFP_KERNEL. In this case * we might have raced with the softirq code while we had the refill * lock held. Use rds_ib_ring_low() instead of ring_empty to decide * if we should requeue. */ if (rds_conn_up(conn) && ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) || rds_ib_ring_empty(&ic->i_recv_ring))) { queue_delayed_work(rds_wq, &conn->c_recv_w, 1); } }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover20867.10%333.33%
Santosh Shilimkar6821.94%111.11%
Håkon Bugge206.45%111.11%
Mike Marciniszyn72.26%111.11%
David S. Miller51.61%111.11%
Mel Gorman10.32%111.11%
Chris Mason10.32%111.11%
Total310100.00%9100.00%

/* * We want to recycle several types of recv allocations, like incs and frags. * To use this, the *_free() function passes in the ptr to a list_head within * the recyclee, as well as the cache to put it on. * * First, we put the memory on a percpu list. When this reaches a certain size, * We move it to an intermediate non-percpu list in a lockless manner, with some * xchg/compxchg wizardry. * * N.B. Instead of a list_head as the anchor, we use a single pointer, which can * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and * list_empty() will return true with one element is actually present. */
static void rds_ib_recv_cache_put(struct list_head *new_item, struct rds_ib_refill_cache *cache) { unsigned long flags; struct list_head *old, *chpfirst; local_irq_save(flags); chpfirst = __this_cpu_read(cache->percpu->first); if (!chpfirst) INIT_LIST_HEAD(new_item); else /* put on front */ list_add_tail(new_item, chpfirst); __this_cpu_write(cache->percpu->first, new_item); __this_cpu_inc(cache->percpu->count); if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT) goto end; /* * Return our per-cpu first list to the cache's xfer by atomically * grabbing the current xfer list, appending it to our per-cpu list, * and then atomically returning that entire list back to the * cache's xfer list as long as it's still empty. */ do { old = xchg(&cache->xfer, NULL); if (old) list_splice_entire_tail(old, chpfirst); old = cmpxchg(&cache->xfer, NULL, chpfirst); } while (old); __this_cpu_write(cache->percpu->first, NULL); __this_cpu_write(cache->percpu->count, 0); end: local_irq_restore(flags); }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason9252.87%120.00%
Shan Wei3821.84%120.00%
Andy Grover3318.97%240.00%
Gerald Schaefer116.32%120.00%
Total174100.00%5100.00%


static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache) { struct list_head *head = cache->ready; if (head) { if (!list_empty(head)) { cache->ready = head->next; list_del_init(head); } else cache->ready = NULL; } return head; }

Contributors

PersonTokensPropCommitsCommitProp
Chris Mason4370.49%150.00%
Andy Grover1829.51%150.00%
Total61100.00%2100.00%


int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) { struct rds_ib_incoming *ibinc; struct rds_page_frag *frag; unsigned long to_copy; unsigned long frag_off = 0; int copied = 0; int ret; u32 len; ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); len = be32_to_cpu(inc->i_hdr.h_len); while (iov_iter_count(to) && copied < len) { if (frag_off == RDS_FRAG_SIZE) { frag = list_entry(frag->f_item.next, struct rds_page_frag, f_item); frag_off = 0; } to_copy = min_t(unsigned long, iov_iter_count(to), RDS_FRAG_SIZE - frag_off); to_copy = min_t(unsigned long, to_copy, len - copied); /* XXX needs + offset for multiple recvs per page */ rds_stats_add(s_copy_to_user, to_copy); ret = copy_page_to_iter(sg_page(&frag->f_sg), frag->f_sg.offset + frag_off, to_copy, to); if (ret != to_copy) return -EFAULT; frag_off += to_copy; copied += to_copy; } return copied; }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover18189.60%266.67%
Al Viro2110.40%133.33%
Total202100.00%3100.00%

/* ic starts out kzalloc()ed */
void rds_ib_recv_init_ack(struct rds_ib_connection *ic) { struct ib_send_wr *wr = &ic->i_ack_wr; struct ib_sge *sge = &ic->i_ack_sge; sge->addr = ic->i_ack_dma; sge->length = sizeof(struct rds_header); sge->lkey = ic->i_pd->local_dma_lkey; wr->sg_list = sge; wr->num_sge = 1; wr->opcode = IB_WR_SEND; wr->wr_id = RDS_IB_ACK_WR_ID; wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover8897.78%150.00%
Jason Gunthorpe22.22%150.00%
Total90100.00%2100.00%

/* * You'd think that with reliable IB connections you wouldn't need to ack * messages that have been received. The problem is that IB hardware generates * an ack message before it has DMAed the message into memory. This creates a * potential message loss if the HCA is disabled for any reason between when it * sends the ack and before the message is DMAed and processed. This is only a * potential issue if another HCA is available for fail-over. * * When the remote host receives our ack they'll free the sent message from * their send queue. To decrease the latency of this we always send an ack * immediately after we've received messages. * * For simplicity, we only have one ack in flight at a time. This puts * pressure on senders to have deep enough send queues to absorb the latency of * a single ack frame being in flight. This might not be good enough. * * This is implemented by have a long-lived send_wr and sge which point to a * statically allocated ack frame. This ack wr does not fall under the ring * accounting that the tx and rx wrs do. The QP attribute specifically makes * room for it beyond the ring size. Send completion notices its special * wr_id and avoids working with the ring in that case. */ #ifndef KERNEL_HAS_ATOMIC64
void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) { unsigned long flags; spin_lock_irqsave(&ic->i_ack_lock, flags); ic->i_ack_next = seq; if (ack_required) set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); spin_unlock_irqrestore(&ic->i_ack_lock, flags); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover60100.00%1100.00%
Total60100.00%1100.00%


static u64 rds_ib_get_ack(struct rds_ib_connection *ic) { unsigned long flags; u64 seq; clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); spin_lock_irqsave(&ic->i_ack_lock, flags); seq = ic->i_ack_next; spin_unlock_irqrestore(&ic->i_ack_lock, flags); return seq; }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover57100.00%1100.00%
Total57100.00%1100.00%

#else
void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) { atomic64_set(&ic->i_ack_next, seq); if (ack_required) { smp_mb__before_atomic(); set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); } }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover4497.78%266.67%
Peter Zijlstra12.22%133.33%
Total45100.00%3100.00%


static u64 rds_ib_get_ack(struct rds_ib_connection *ic) { clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); smp_mb__after_atomic(); return atomic64_read(&ic->i_ack_next); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover3296.97%266.67%
Peter Zijlstra13.03%133.33%
Total33100.00%3100.00%

#endif
static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) { struct rds_header *hdr = ic->i_ack; struct ib_send_wr *failed_wr; u64 seq; int ret; seq = rds_ib_get_ack(ic); rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); rds_message_populate_header(hdr, 0, 0, 0); hdr->h_ack = cpu_to_be64(seq); hdr->h_credit = adv_credits; rds_message_make_checksum(hdr); ic->i_ack_queued = jiffies; ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); if (unlikely(ret)) { /* Failed to send. Release the WR, and * force another ACK. */ clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); rds_ib_stats_inc(s_ib_ack_send_failure); rds_ib_conn_error(ic->conn, "sending ack failed\n"); } else rds_ib_stats_inc(s_ib_ack_sent); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover162100.00%2100.00%
Total162100.00%2100.00%

/* * There are 3 ways of getting acknowledgements to the peer: * 1. We call rds_ib_attempt_ack from the recv completion handler * to send an ACK-only frame. * However, there can be only one such frame in the send queue * at any time, so we may have to postpone it. * 2. When another (data) packet is transmitted while there's * an ACK in the queue, we piggyback the ACK sequence number * on the data packet. * 3. If the ACK WR is done sending, we get called from the * send queue completion handler, and check whether there's * another ACK pending (postponed because the WR was on the * queue). If so, we transmit it. * * We maintain 2 variables: * - i_ack_flags, which keeps track of whether the ACK WR * is currently in the send queue or not (IB_ACK_IN_FLIGHT) * - i_ack_next, which is the last sequence number we received * * Potentially, send queue and receive queue handlers can run concurrently. * It would be nice to not have to use a spinlock to synchronize things, * but the one problem that rules this out is that 64bit updates are * not atomic on all platforms. Things would be a lot simpler if * we had atomic64 or maybe cmpxchg64 everywhere. * * Reconnecting complicates this picture just slightly. When we * reconnect, we may be seeing duplicate packets. The peer * is retransmitting them, because it hasn't seen an ACK for * them. It is important that we ACK these. * * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with * this flag set *MUST* be acknowledged immediately. */ /* * When we get here, we're called from the recv queue handler. * Check whether we ought to transmit an ACK. */
void rds_ib_attempt_ack(struct rds_ib_connection *ic) { unsigned int adv_credits; if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) return; if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { rds_ib_stats_inc(s_ib_ack_send_delayed); return; } /* Can we get a send credit? */ if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { rds_ib_stats_inc(s_ib_tx_throttle); clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); return; } clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); rds_ib_send_ack(ic, adv_credits); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover9998.02%150.00%
Steve Wise21.98%150.00%
Total101100.00%2100.00%

/* * We get here from the send completion handler, when the * adapter tells us the ACK frame was sent. */
void rds_ib_ack_send_complete(struct rds_ib_connection *ic) { clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); rds_ib_attempt_ack(ic); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover25100.00%1100.00%
Total25100.00%1100.00%

/* * This is called by the regular xmit code when it wants to piggyback * an ACK on an outgoing frame. */
u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) { if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) rds_ib_stats_inc(s_ib_ack_send_piggybacked); return rds_ib_get_ack(ic); }

Contributors

PersonTokensPropCommitsCommitProp
Andy Grover33100.00%1100.00%
Total33100.00%1100.00%

/* * It's kind of lame that we're copying from the posted receive pages into * long-lived bitmaps. We could have posted the bitmaps and rdma written into * them. But receiving new congestion bitmaps should be a *rare* event, so * hopefully we won't need to invest that complexity in making it more * efficient. By copying we can share a simpler core with TCP which has to * copy. */
static void rds_ib_cong_recv(struct rds_connection *conn, struct rds_ib_incoming *ibinc) { struct rds_cong_map