linux/drivers/infiniband/hw/mthca/mthca_srq.c
Arthur Kepner 1f5c23e2c1 IB/mthca: Use mmiowb after doorbell ring
We discovered a problem when running IPoIB applications on multiple
CPUs on an Altix system. Many messages such as:

ib_mthca 0002:01:00.0: SQ 000014 full (19941644 head, 19941707 tail, 64 max, 0 nreq)

appear in syslog, and the driver wedges up.

Apparently this is because writes to the doorbells from different CPUs
reach the device out of order. The following patch adds mmiowb() calls
after doorbell rings to ensure the doorbell writes are ordered.

Signed-off-by: Arthur Kepner <akepner@sgi.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
2006-10-16 20:22:35 -07:00

750 lines
18 KiB
C

/*
* Copyright (c) 2005 Cisco Systems. 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.
*
* $Id: mthca_srq.c 3047 2005-08-10 03:59:35Z roland $
*/
#include <linux/slab.h>
#include <linux/string.h>
#include <asm/io.h>
#include "mthca_dev.h"
#include "mthca_cmd.h"
#include "mthca_memfree.h"
#include "mthca_wqe.h"
enum {
MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
};
struct mthca_tavor_srq_context {
__be64 wqe_base_ds; /* low 6 bits is descriptor size */
__be32 state_pd;
__be32 lkey;
__be32 uar;
__be16 limit_watermark;
__be16 wqe_cnt;
u32 reserved[2];
};
struct mthca_arbel_srq_context {
__be32 state_logsize_srqn;
__be32 lkey;
__be32 db_index;
__be32 logstride_usrpage;
__be64 wqe_base;
__be32 eq_pd;
__be16 limit_watermark;
__be16 wqe_cnt;
u16 reserved1;
__be16 wqe_counter;
u32 reserved2[3];
};
static void *get_wqe(struct mthca_srq *srq, int n)
{
if (srq->is_direct)
return srq->queue.direct.buf + (n << srq->wqe_shift);
else
return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
((n << srq->wqe_shift) & (PAGE_SIZE - 1));
}
/*
* Return a pointer to the location within a WQE that we're using as a
* link when the WQE is in the free list. We use the imm field
* because in the Tavor case, posting a WQE may overwrite the next
* segment of the previous WQE, but a receive WQE will never touch the
* imm field. This avoids corrupting our free list if the previous
* WQE has already completed and been put on the free list when we
* post the next WQE.
*/
static inline int *wqe_to_link(void *wqe)
{
return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
}
static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
struct mthca_pd *pd,
struct mthca_srq *srq,
struct mthca_tavor_srq_context *context)
{
memset(context, 0, sizeof *context);
context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
context->state_pd = cpu_to_be32(pd->pd_num);
context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
if (pd->ibpd.uobject)
context->uar =
cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
else
context->uar = cpu_to_be32(dev->driver_uar.index);
}
static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
struct mthca_pd *pd,
struct mthca_srq *srq,
struct mthca_arbel_srq_context *context)
{
int logsize;
memset(context, 0, sizeof *context);
logsize = long_log2(srq->max) + srq->wqe_shift;
context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
context->db_index = cpu_to_be32(srq->db_index);
context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
if (pd->ibpd.uobject)
context->logstride_usrpage |=
cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
else
context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
}
static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
{
mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
srq->is_direct, &srq->mr);
kfree(srq->wrid);
}
static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
struct mthca_srq *srq)
{
struct mthca_data_seg *scatter;
void *wqe;
int err;
int i;
if (pd->ibpd.uobject)
return 0;
srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
if (!srq->wrid)
return -ENOMEM;
err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
MTHCA_MAX_DIRECT_SRQ_SIZE,
&srq->queue, &srq->is_direct, pd, 1, &srq->mr);
if (err) {
kfree(srq->wrid);
return err;
}
/*
* Now initialize the SRQ buffer so that all of the WQEs are
* linked into the list of free WQEs. In addition, set the
* scatter list L_Keys to the sentry value of 0x100.
*/
for (i = 0; i < srq->max; ++i) {
wqe = get_wqe(srq, i);
*wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
for (scatter = wqe + sizeof (struct mthca_next_seg);
(void *) scatter < wqe + (1 << srq->wqe_shift);
++scatter)
scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
}
srq->last = get_wqe(srq, srq->max - 1);
return 0;
}
int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
struct ib_srq_attr *attr, struct mthca_srq *srq)
{
struct mthca_mailbox *mailbox;
u8 status;
int ds;
int err;
/* Sanity check SRQ size before proceeding */
if (attr->max_wr > dev->limits.max_srq_wqes ||
attr->max_sge > dev->limits.max_srq_sge)
return -EINVAL;
srq->max = attr->max_wr;
srq->max_gs = attr->max_sge;
srq->counter = 0;
if (mthca_is_memfree(dev))
srq->max = roundup_pow_of_two(srq->max + 1);
else
srq->max = srq->max + 1;
ds = max(64UL,
roundup_pow_of_two(sizeof (struct mthca_next_seg) +
srq->max_gs * sizeof (struct mthca_data_seg)));
if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
return -EINVAL;
srq->wqe_shift = long_log2(ds);
srq->srqn = mthca_alloc(&dev->srq_table.alloc);
if (srq->srqn == -1)
return -ENOMEM;
if (mthca_is_memfree(dev)) {
err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
if (err)
goto err_out;
if (!pd->ibpd.uobject) {
srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
srq->srqn, &srq->db);
if (srq->db_index < 0) {
err = -ENOMEM;
goto err_out_icm;
}
}
}
mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
if (IS_ERR(mailbox)) {
err = PTR_ERR(mailbox);
goto err_out_db;
}
err = mthca_alloc_srq_buf(dev, pd, srq);
if (err)
goto err_out_mailbox;
spin_lock_init(&srq->lock);
srq->refcount = 1;
init_waitqueue_head(&srq->wait);
mutex_init(&srq->mutex);
if (mthca_is_memfree(dev))
mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
else
mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
if (err) {
mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
goto err_out_free_buf;
}
if (status) {
mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
status);
err = -EINVAL;
goto err_out_free_buf;
}
spin_lock_irq(&dev->srq_table.lock);
if (mthca_array_set(&dev->srq_table.srq,
srq->srqn & (dev->limits.num_srqs - 1),
srq)) {
spin_unlock_irq(&dev->srq_table.lock);
goto err_out_free_srq;
}
spin_unlock_irq(&dev->srq_table.lock);
mthca_free_mailbox(dev, mailbox);
srq->first_free = 0;
srq->last_free = srq->max - 1;
attr->max_wr = srq->max - 1;
attr->max_sge = srq->max_gs;
return 0;
err_out_free_srq:
err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
if (err)
mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
else if (status)
mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
err_out_free_buf:
if (!pd->ibpd.uobject)
mthca_free_srq_buf(dev, srq);
err_out_mailbox:
mthca_free_mailbox(dev, mailbox);
err_out_db:
if (!pd->ibpd.uobject && mthca_is_memfree(dev))
mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
err_out_icm:
mthca_table_put(dev, dev->srq_table.table, srq->srqn);
err_out:
mthca_free(&dev->srq_table.alloc, srq->srqn);
return err;
}
static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
{
int c;
spin_lock_irq(&dev->srq_table.lock);
c = srq->refcount;
spin_unlock_irq(&dev->srq_table.lock);
return c;
}
void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
{
struct mthca_mailbox *mailbox;
int err;
u8 status;
mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
if (IS_ERR(mailbox)) {
mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
return;
}
err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
if (err)
mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
else if (status)
mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
spin_lock_irq(&dev->srq_table.lock);
mthca_array_clear(&dev->srq_table.srq,
srq->srqn & (dev->limits.num_srqs - 1));
--srq->refcount;
spin_unlock_irq(&dev->srq_table.lock);
wait_event(srq->wait, !get_srq_refcount(dev, srq));
if (!srq->ibsrq.uobject) {
mthca_free_srq_buf(dev, srq);
if (mthca_is_memfree(dev))
mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
}
mthca_table_put(dev, dev->srq_table.table, srq->srqn);
mthca_free(&dev->srq_table.alloc, srq->srqn);
mthca_free_mailbox(dev, mailbox);
}
int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
{
struct mthca_dev *dev = to_mdev(ibsrq->device);
struct mthca_srq *srq = to_msrq(ibsrq);
int ret;
u8 status;
/* We don't support resizing SRQs (yet?) */
if (attr_mask & IB_SRQ_MAX_WR)
return -EINVAL;
if (attr_mask & IB_SRQ_LIMIT) {
u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
if (attr->srq_limit > max_wr)
return -EINVAL;
mutex_lock(&srq->mutex);
ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
mutex_unlock(&srq->mutex);
if (ret)
return ret;
if (status)
return -EINVAL;
}
return 0;
}
int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
{
struct mthca_dev *dev = to_mdev(ibsrq->device);
struct mthca_srq *srq = to_msrq(ibsrq);
struct mthca_mailbox *mailbox;
struct mthca_arbel_srq_context *arbel_ctx;
struct mthca_tavor_srq_context *tavor_ctx;
u8 status;
int err;
mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
if (IS_ERR(mailbox))
return PTR_ERR(mailbox);
err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox, &status);
if (err)
goto out;
if (mthca_is_memfree(dev)) {
arbel_ctx = mailbox->buf;
srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
} else {
tavor_ctx = mailbox->buf;
srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
}
srq_attr->max_wr = srq->max - 1;
srq_attr->max_sge = srq->max_gs;
out:
mthca_free_mailbox(dev, mailbox);
return err;
}
void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
enum ib_event_type event_type)
{
struct mthca_srq *srq;
struct ib_event event;
spin_lock(&dev->srq_table.lock);
srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
if (srq)
++srq->refcount;
spin_unlock(&dev->srq_table.lock);
if (!srq) {
mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
return;
}
if (!srq->ibsrq.event_handler)
goto out;
event.device = &dev->ib_dev;
event.event = event_type;
event.element.srq = &srq->ibsrq;
srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
out:
spin_lock(&dev->srq_table.lock);
if (!--srq->refcount)
wake_up(&srq->wait);
spin_unlock(&dev->srq_table.lock);
}
/*
* This function must be called with IRQs disabled.
*/
void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
{
int ind;
ind = wqe_addr >> srq->wqe_shift;
spin_lock(&srq->lock);
if (likely(srq->first_free >= 0))
*wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
else
srq->first_free = ind;
*wqe_to_link(get_wqe(srq, ind)) = -1;
srq->last_free = ind;
spin_unlock(&srq->lock);
}
int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct mthca_dev *dev = to_mdev(ibsrq->device);
struct mthca_srq *srq = to_msrq(ibsrq);
__be32 doorbell[2];
unsigned long flags;
int err = 0;
int first_ind;
int ind;
int next_ind;
int nreq;
int i;
void *wqe;
void *prev_wqe;
spin_lock_irqsave(&srq->lock, flags);
first_ind = srq->first_free;
for (nreq = 0; wr; wr = wr->next) {
ind = srq->first_free;
if (ind < 0) {
mthca_err(dev, "SRQ %06x full\n", srq->srqn);
err = -ENOMEM;
*bad_wr = wr;
break;
}
wqe = get_wqe(srq, ind);
next_ind = *wqe_to_link(wqe);
if (next_ind < 0) {
mthca_err(dev, "SRQ %06x full\n", srq->srqn);
err = -ENOMEM;
*bad_wr = wr;
break;
}
prev_wqe = srq->last;
srq->last = wqe;
((struct mthca_next_seg *) wqe)->nda_op = 0;
((struct mthca_next_seg *) wqe)->ee_nds = 0;
/* flags field will always remain 0 */
wqe += sizeof (struct mthca_next_seg);
if (unlikely(wr->num_sge > srq->max_gs)) {
err = -EINVAL;
*bad_wr = wr;
srq->last = prev_wqe;
break;
}
for (i = 0; i < wr->num_sge; ++i) {
((struct mthca_data_seg *) wqe)->byte_count =
cpu_to_be32(wr->sg_list[i].length);
((struct mthca_data_seg *) wqe)->lkey =
cpu_to_be32(wr->sg_list[i].lkey);
((struct mthca_data_seg *) wqe)->addr =
cpu_to_be64(wr->sg_list[i].addr);
wqe += sizeof (struct mthca_data_seg);
}
if (i < srq->max_gs) {
((struct mthca_data_seg *) wqe)->byte_count = 0;
((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
((struct mthca_data_seg *) wqe)->addr = 0;
}
((struct mthca_next_seg *) prev_wqe)->nda_op =
cpu_to_be32((ind << srq->wqe_shift) | 1);
wmb();
((struct mthca_next_seg *) prev_wqe)->ee_nds =
cpu_to_be32(MTHCA_NEXT_DBD);
srq->wrid[ind] = wr->wr_id;
srq->first_free = next_ind;
++nreq;
if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
nreq = 0;
doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
doorbell[1] = cpu_to_be32(srq->srqn << 8);
/*
* Make sure that descriptors are written
* before doorbell is rung.
*/
wmb();
mthca_write64(doorbell,
dev->kar + MTHCA_RECEIVE_DOORBELL,
MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
first_ind = srq->first_free;
}
}
if (likely(nreq)) {
doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
/*
* Make sure that descriptors are written before
* doorbell is rung.
*/
wmb();
mthca_write64(doorbell,
dev->kar + MTHCA_RECEIVE_DOORBELL,
MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
}
/*
* Make sure doorbells don't leak out of SRQ spinlock and
* reach the HCA out of order:
*/
mmiowb();
spin_unlock_irqrestore(&srq->lock, flags);
return err;
}
int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct mthca_dev *dev = to_mdev(ibsrq->device);
struct mthca_srq *srq = to_msrq(ibsrq);
unsigned long flags;
int err = 0;
int ind;
int next_ind;
int nreq;
int i;
void *wqe;
spin_lock_irqsave(&srq->lock, flags);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
ind = srq->first_free;
if (ind < 0) {
mthca_err(dev, "SRQ %06x full\n", srq->srqn);
err = -ENOMEM;
*bad_wr = wr;
break;
}
wqe = get_wqe(srq, ind);
next_ind = *wqe_to_link(wqe);
if (next_ind < 0) {
mthca_err(dev, "SRQ %06x full\n", srq->srqn);
err = -ENOMEM;
*bad_wr = wr;
break;
}
((struct mthca_next_seg *) wqe)->nda_op =
cpu_to_be32((next_ind << srq->wqe_shift) | 1);
((struct mthca_next_seg *) wqe)->ee_nds = 0;
/* flags field will always remain 0 */
wqe += sizeof (struct mthca_next_seg);
if (unlikely(wr->num_sge > srq->max_gs)) {
err = -EINVAL;
*bad_wr = wr;
break;
}
for (i = 0; i < wr->num_sge; ++i) {
((struct mthca_data_seg *) wqe)->byte_count =
cpu_to_be32(wr->sg_list[i].length);
((struct mthca_data_seg *) wqe)->lkey =
cpu_to_be32(wr->sg_list[i].lkey);
((struct mthca_data_seg *) wqe)->addr =
cpu_to_be64(wr->sg_list[i].addr);
wqe += sizeof (struct mthca_data_seg);
}
if (i < srq->max_gs) {
((struct mthca_data_seg *) wqe)->byte_count = 0;
((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
((struct mthca_data_seg *) wqe)->addr = 0;
}
srq->wrid[ind] = wr->wr_id;
srq->first_free = next_ind;
}
if (likely(nreq)) {
srq->counter += nreq;
/*
* Make sure that descriptors are written before
* we write doorbell record.
*/
wmb();
*srq->db = cpu_to_be32(srq->counter);
}
spin_unlock_irqrestore(&srq->lock, flags);
return err;
}
int mthca_max_srq_sge(struct mthca_dev *dev)
{
if (mthca_is_memfree(dev))
return dev->limits.max_sg;
/*
* SRQ allocations are based on powers of 2 for Tavor,
* (although they only need to be multiples of 16 bytes).
*
* Therefore, we need to base the max number of sg entries on
* the largest power of 2 descriptor size that is <= to the
* actual max WQE descriptor size, rather than return the
* max_sg value given by the firmware (which is based on WQE
* sizes as multiples of 16, not powers of 2).
*
* If SRQ implementation is changed for Tavor to be based on
* multiples of 16, the calculation below can be deleted and
* the FW max_sg value returned.
*/
return min_t(int, dev->limits.max_sg,
((1 << (fls(dev->limits.max_desc_sz) - 1)) -
sizeof (struct mthca_next_seg)) /
sizeof (struct mthca_data_seg));
}
int __devinit mthca_init_srq_table(struct mthca_dev *dev)
{
int err;
if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
return 0;
spin_lock_init(&dev->srq_table.lock);
err = mthca_alloc_init(&dev->srq_table.alloc,
dev->limits.num_srqs,
dev->limits.num_srqs - 1,
dev->limits.reserved_srqs);
if (err)
return err;
err = mthca_array_init(&dev->srq_table.srq,
dev->limits.num_srqs);
if (err)
mthca_alloc_cleanup(&dev->srq_table.alloc);
return err;
}
void mthca_cleanup_srq_table(struct mthca_dev *dev)
{
if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
return;
mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
mthca_alloc_cleanup(&dev->srq_table.alloc);
}