net/rds/ib_ring.c

0001 /*
0002  * Copyright (c) 2006 Oracle.  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
0035 #include "rds.h"
0036 #include "ib.h"
0037
0038 /*
0039  * Locking for IB rings.
0040  * We assume that allocation is always protected by a mutex
0041  * in the caller (this is a valid assumption for the current
0042  * implementation).
0043  *
0044  * Freeing always happens in an interrupt, and hence only
0045  * races with allocations, but not with other free()s.
0046  *
0047  * The interaction between allocation and freeing is that
0048  * the alloc code has to determine the number of free entries.
0049  * To this end, we maintain two counters; an allocation counter
0050  * and a free counter. Both are allowed to run freely, and wrap
0051  * around.
0052  * The number of used entries is always (alloc_ctr - free_ctr) % NR.
0053  *
0054  * The current implementation makes free_ctr atomic. When the
0055  * caller finds an allocation fails, it should set an "alloc fail"
0056  * bit and retry the allocation. The "alloc fail" bit essentially tells
0057  * the CQ completion handlers to wake it up after freeing some
0058  * more entries.
0059  */
0060
0061 /*
0062  * This only happens on shutdown.
0063  */
0064 DECLARE_WAIT_QUEUE_HEAD(rds_ib_ring_empty_wait);
0065
0066 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr)
0067 {
0068     memset(ring, 0, sizeof(*ring));
0069     ring->w_nr = nr;
0070     rdsdebug("ring %p nr %u\n", ring, ring->w_nr);
0071 }
0072
0073 static inline u32 __rds_ib_ring_used(struct rds_ib_work_ring *ring)
0074 {
0075     u32 diff;
0076
0077     /* This assumes that atomic_t has at least as many bits as u32 */
0078     diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr);
0079     BUG_ON(diff > ring->w_nr);
0080
0081     return diff;
0082 }
0083
0084 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr)
0085 {
0086     /* We only ever get called from the connection setup code,
0087      * prior to creating the QP. */
0088     BUG_ON(__rds_ib_ring_used(ring));
0089     ring->w_nr = nr;
0090 }
0091
0092 static int __rds_ib_ring_empty(struct rds_ib_work_ring *ring)
0093 {
0094     return __rds_ib_ring_used(ring) == 0;
0095 }
0096
0097 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos)
0098 {
0099     u32 ret = 0, avail;
0100
0101     avail = ring->w_nr - __rds_ib_ring_used(ring);
0102
0103     rdsdebug("ring %p val %u next %u free %u\n", ring, val,
0104          ring->w_alloc_ptr, avail);
0105
0106     if (val && avail) {
0107         ret = min(val, avail);
0108         *pos = ring->w_alloc_ptr;
0109
0110         ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr;
0111         ring->w_alloc_ctr += ret;
0112     }
0113
0114     return ret;
0115 }
0116
0117 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val)
0118 {
0119     ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr;
0120     atomic_add(val, &ring->w_free_ctr);
0121
0122     if (__rds_ib_ring_empty(ring) &&
0123         waitqueue_active(&rds_ib_ring_empty_wait))
0124         wake_up(&rds_ib_ring_empty_wait);
0125 }
0126
0127 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val)
0128 {
0129     ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr;
0130     ring->w_alloc_ctr -= val;
0131 }
0132
0133 int rds_ib_ring_empty(struct rds_ib_work_ring *ring)
0134 {
0135     return __rds_ib_ring_empty(ring);
0136 }
0137
0138 int rds_ib_ring_low(struct rds_ib_work_ring *ring)
0139 {
0140     return __rds_ib_ring_used(ring) <= (ring->w_nr >> 1);
0141 }
0142
0143 /*
0144  * returns the oldest allocated ring entry.  This will be the next one
0145  * freed.  This can't be called if there are none allocated.
0146  */
0147 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring)
0148 {
0149     return ring->w_free_ptr;
0150 }
0151
0152 /*
0153  * returns the number of completed work requests.
0154  */
0155
0156 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest)
0157 {
0158     u32 ret;
0159
0160     if (oldest <= (unsigned long long)wr_id)
0161         ret = (unsigned long long)wr_id - oldest + 1;
0162     else
0163         ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1;
0164
0165     rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret,
0166          wr_id, oldest);
0167     return ret;
0168 }