Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
ib_ring.c
Go to the documentation of this file.
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 
35 #include "rds.h"
36 #include "ib.h"
37 
38 /*
39  * Locking for IB rings.
40  * We assume that allocation is always protected by a mutex
41  * in the caller (this is a valid assumption for the current
42  * implementation).
43  *
44  * Freeing always happens in an interrupt, and hence only
45  * races with allocations, but not with other free()s.
46  *
47  * The interaction between allocation and freeing is that
48  * the alloc code has to determine the number of free entries.
49  * To this end, we maintain two counters; an allocation counter
50  * and a free counter. Both are allowed to run freely, and wrap
51  * around.
52  * The number of used entries is always (alloc_ctr - free_ctr) % NR.
53  *
54  * The current implementation makes free_ctr atomic. When the
55  * caller finds an allocation fails, it should set an "alloc fail"
56  * bit and retry the allocation. The "alloc fail" bit essentially tells
57  * the CQ completion handlers to wake it up after freeing some
58  * more entries.
59  */
60 
61 /*
62  * This only happens on shutdown.
63  */
64 DECLARE_WAIT_QUEUE_HEAD(rds_ib_ring_empty_wait);
65 
67 {
68  memset(ring, 0, sizeof(*ring));
69  ring->w_nr = nr;
70  rdsdebug("ring %p nr %u\n", ring, ring->w_nr);
71 }
72 
73 static inline u32 __rds_ib_ring_used(struct rds_ib_work_ring *ring)
74 {
75  u32 diff;
76 
77  /* This assumes that atomic_t has at least as many bits as u32 */
78  diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr);
79  BUG_ON(diff > ring->w_nr);
80 
81  return diff;
82 }
83 
85 {
86  /* We only ever get called from the connection setup code,
87  * prior to creating the QP. */
88  BUG_ON(__rds_ib_ring_used(ring));
89  ring->w_nr = nr;
90 }
91 
92 static int __rds_ib_ring_empty(struct rds_ib_work_ring *ring)
93 {
94  return __rds_ib_ring_used(ring) == 0;
95 }
96 
98 {
99  u32 ret = 0, avail;
100 
101  avail = ring->w_nr - __rds_ib_ring_used(ring);
102 
103  rdsdebug("ring %p val %u next %u free %u\n", ring, val,
104  ring->w_alloc_ptr, avail);
105 
106  if (val && avail) {
107  ret = min(val, avail);
108  *pos = ring->w_alloc_ptr;
109 
110  ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr;
111  ring->w_alloc_ctr += ret;
112  }
113 
114  return ret;
115 }
116 
118 {
119  ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr;
120  atomic_add(val, &ring->w_free_ctr);
121 
122  if (__rds_ib_ring_empty(ring) &&
123  waitqueue_active(&rds_ib_ring_empty_wait))
125 }
126 
128 {
129  ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr;
130  ring->w_alloc_ctr -= val;
131 }
132 
134 {
135  return __rds_ib_ring_empty(ring);
136 }
137 
139 {
140  return __rds_ib_ring_used(ring) <= (ring->w_nr >> 1);
141 }
142 
143 /*
144  * returns the oldest alloced ring entry. This will be the next one
145  * freed. This can't be called if there are none allocated.
146  */
148 {
149  return ring->w_free_ptr;
150 }
151 
152 /*
153  * returns the number of completed work requests.
154  */
155 
156 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest)
157 {
158  u32 ret;
159 
160  if (oldest <= (unsigned long long)wr_id)
161  ret = (unsigned long long)wr_id - oldest + 1;
162  else
163  ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1;
164 
165  rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret,
166  wr_id, oldest);
167  return ret;
168 }