linux/net/rds/ib_frmr.c
Santosh Shilimkar 5601245931 RDS: IB: split the mr registration and invalidation path
MR invalidation in RDS is done in background thread and not in
data path like registration. So break the dependency between them
which helps to remove the performance bottleneck.

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2017-01-02 14:02:47 -08:00

378 lines
9.9 KiB
C

/*
* Copyright (c) 2016 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 "ib_mr.h"
static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
int npages)
{
struct rds_ib_mr_pool *pool;
struct rds_ib_mr *ibmr = NULL;
struct rds_ib_frmr *frmr;
int err = 0;
if (npages <= RDS_MR_8K_MSG_SIZE)
pool = rds_ibdev->mr_8k_pool;
else
pool = rds_ibdev->mr_1m_pool;
ibmr = rds_ib_try_reuse_ibmr(pool);
if (ibmr)
return ibmr;
ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
rdsibdev_to_node(rds_ibdev));
if (!ibmr) {
err = -ENOMEM;
goto out_no_cigar;
}
frmr = &ibmr->u.frmr;
frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG,
pool->fmr_attr.max_pages);
if (IS_ERR(frmr->mr)) {
pr_warn("RDS/IB: %s failed to allocate MR", __func__);
goto out_no_cigar;
}
ibmr->pool = pool;
if (pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
if (atomic_read(&pool->item_count) > pool->max_items_soft)
pool->max_items_soft = pool->max_items;
frmr->fr_state = FRMR_IS_FREE;
return ibmr;
out_no_cigar:
kfree(ibmr);
atomic_dec(&pool->item_count);
return ERR_PTR(err);
}
static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop)
{
struct rds_ib_mr_pool *pool = ibmr->pool;
if (drop)
llist_add(&ibmr->llnode, &pool->drop_list);
else
llist_add(&ibmr->llnode, &pool->free_list);
atomic_add(ibmr->sg_len, &pool->free_pinned);
atomic_inc(&pool->dirty_count);
/* If we've pinned too many pages, request a flush */
if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
atomic_read(&pool->dirty_count) >= pool->max_items / 5)
queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
}
static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
{
struct rds_ib_frmr *frmr = &ibmr->u.frmr;
struct ib_send_wr *failed_wr;
struct ib_reg_wr reg_wr;
int ret, off = 0;
while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
atomic_inc(&ibmr->ic->i_fastreg_wrs);
cpu_relax();
}
ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_len,
&off, PAGE_SIZE);
if (unlikely(ret != ibmr->sg_len))
return ret < 0 ? ret : -EINVAL;
/* Perform a WR for the fast_reg_mr. Each individual page
* in the sg list is added to the fast reg page list and placed
* inside the fast_reg_mr WR. The key used is a rolling 8bit
* counter, which should guarantee uniqueness.
*/
ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
frmr->fr_state = FRMR_IS_INUSE;
memset(&reg_wr, 0, sizeof(reg_wr));
reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
reg_wr.wr.opcode = IB_WR_REG_MR;
reg_wr.wr.num_sge = 0;
reg_wr.mr = frmr->mr;
reg_wr.key = frmr->mr->rkey;
reg_wr.access = IB_ACCESS_LOCAL_WRITE |
IB_ACCESS_REMOTE_READ |
IB_ACCESS_REMOTE_WRITE;
reg_wr.wr.send_flags = IB_SEND_SIGNALED;
failed_wr = &reg_wr.wr;
ret = ib_post_send(ibmr->ic->i_cm_id->qp, &reg_wr.wr, &failed_wr);
WARN_ON(failed_wr != &reg_wr.wr);
if (unlikely(ret)) {
/* Failure here can be because of -ENOMEM as well */
frmr->fr_state = FRMR_IS_STALE;
atomic_inc(&ibmr->ic->i_fastreg_wrs);
if (printk_ratelimit())
pr_warn("RDS/IB: %s returned error(%d)\n",
__func__, ret);
}
return ret;
}
static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev,
struct rds_ib_mr_pool *pool,
struct rds_ib_mr *ibmr,
struct scatterlist *sg, unsigned int sg_len)
{
struct ib_device *dev = rds_ibdev->dev;
struct rds_ib_frmr *frmr = &ibmr->u.frmr;
int i;
u32 len;
int ret = 0;
/* We want to teardown old ibmr values here and fill it up with
* new sg values
*/
rds_ib_teardown_mr(ibmr);
ibmr->sg = sg;
ibmr->sg_len = sg_len;
ibmr->sg_dma_len = 0;
frmr->sg_byte_len = 0;
WARN_ON(ibmr->sg_dma_len);
ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len,
DMA_BIDIRECTIONAL);
if (unlikely(!ibmr->sg_dma_len)) {
pr_warn("RDS/IB: %s failed!\n", __func__);
return -EBUSY;
}
frmr->sg_byte_len = 0;
frmr->dma_npages = 0;
len = 0;
ret = -EINVAL;
for (i = 0; i < ibmr->sg_dma_len; ++i) {
unsigned int dma_len = ib_sg_dma_len(dev, &ibmr->sg[i]);
u64 dma_addr = ib_sg_dma_address(dev, &ibmr->sg[i]);
frmr->sg_byte_len += dma_len;
if (dma_addr & ~PAGE_MASK) {
if (i > 0)
goto out_unmap;
else
++frmr->dma_npages;
}
if ((dma_addr + dma_len) & ~PAGE_MASK) {
if (i < ibmr->sg_dma_len - 1)
goto out_unmap;
else
++frmr->dma_npages;
}
len += dma_len;
}
frmr->dma_npages += len >> PAGE_SHIFT;
if (frmr->dma_npages > ibmr->pool->fmr_attr.max_pages) {
ret = -EMSGSIZE;
goto out_unmap;
}
ret = rds_ib_post_reg_frmr(ibmr);
if (ret)
goto out_unmap;
if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
return ret;
out_unmap:
ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len,
DMA_BIDIRECTIONAL);
ibmr->sg_dma_len = 0;
return ret;
}
static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
{
struct ib_send_wr *s_wr, *failed_wr;
struct rds_ib_frmr *frmr = &ibmr->u.frmr;
struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id;
int ret = -EINVAL;
if (!i_cm_id || !i_cm_id->qp || !frmr->mr)
goto out;
if (frmr->fr_state != FRMR_IS_INUSE)
goto out;
while (atomic_dec_return(&ibmr->ic->i_fastunreg_wrs) <= 0) {
atomic_inc(&ibmr->ic->i_fastunreg_wrs);
cpu_relax();
}
frmr->fr_inv = true;
s_wr = &frmr->fr_wr;
memset(s_wr, 0, sizeof(*s_wr));
s_wr->wr_id = (unsigned long)(void *)ibmr;
s_wr->opcode = IB_WR_LOCAL_INV;
s_wr->ex.invalidate_rkey = frmr->mr->rkey;
s_wr->send_flags = IB_SEND_SIGNALED;
failed_wr = s_wr;
ret = ib_post_send(i_cm_id->qp, s_wr, &failed_wr);
WARN_ON(failed_wr != s_wr);
if (unlikely(ret)) {
frmr->fr_state = FRMR_IS_STALE;
frmr->fr_inv = false;
atomic_inc(&ibmr->ic->i_fastunreg_wrs);
pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
goto out;
}
out:
return ret;
}
void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
{
struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id;
struct rds_ib_frmr *frmr = &ibmr->u.frmr;
if (wc->status != IB_WC_SUCCESS) {
frmr->fr_state = FRMR_IS_STALE;
if (rds_conn_up(ic->conn))
rds_ib_conn_error(ic->conn,
"frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
&ic->conn->c_laddr,
&ic->conn->c_faddr,
wc->status,
ib_wc_status_msg(wc->status),
wc->vendor_err);
}
if (frmr->fr_inv) {
frmr->fr_state = FRMR_IS_FREE;
frmr->fr_inv = false;
atomic_inc(&ic->i_fastreg_wrs);
} else {
atomic_inc(&ic->i_fastunreg_wrs);
}
}
void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
unsigned long *unpinned, unsigned int goal)
{
struct rds_ib_mr *ibmr, *next;
struct rds_ib_frmr *frmr;
int ret = 0;
unsigned int freed = *nfreed;
/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
list_for_each_entry(ibmr, list, unmap_list) {
if (ibmr->sg_dma_len)
ret |= rds_ib_post_inv(ibmr);
}
if (ret)
pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
/* Now we can destroy the DMA mapping and unpin any pages */
list_for_each_entry_safe(ibmr, next, list, unmap_list) {
*unpinned += ibmr->sg_len;
frmr = &ibmr->u.frmr;
__rds_ib_teardown_mr(ibmr);
if (freed < goal || frmr->fr_state == FRMR_IS_STALE) {
/* Don't de-allocate if the MR is not free yet */
if (frmr->fr_state == FRMR_IS_INUSE)
continue;
if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
list_del(&ibmr->unmap_list);
if (frmr->mr)
ib_dereg_mr(frmr->mr);
kfree(ibmr);
freed++;
}
}
*nfreed = freed;
}
struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
struct rds_ib_connection *ic,
struct scatterlist *sg,
unsigned long nents, u32 *key)
{
struct rds_ib_mr *ibmr = NULL;
struct rds_ib_frmr *frmr;
int ret;
do {
if (ibmr)
rds_ib_free_frmr(ibmr, true);
ibmr = rds_ib_alloc_frmr(rds_ibdev, nents);
if (IS_ERR(ibmr))
return ibmr;
frmr = &ibmr->u.frmr;
} while (frmr->fr_state != FRMR_IS_FREE);
ibmr->ic = ic;
ibmr->device = rds_ibdev;
ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents);
if (ret == 0) {
*key = frmr->mr->rkey;
} else {
rds_ib_free_frmr(ibmr, false);
ibmr = ERR_PTR(ret);
}
return ibmr;
}
void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr)
{
struct rds_ib_mr_pool *pool = ibmr->pool;
struct rds_ib_frmr *frmr = &ibmr->u.frmr;
if (frmr->fr_state == FRMR_IS_STALE)
llist_add(&ibmr->llnode, &pool->drop_list);
else
llist_add(&ibmr->llnode, &pool->free_list);
}