linux/net/smc/smc_ib.c
Ursula Braun 10428dd835 net/smc: synchronize buffer usage with device
Usage of send buffer "sndbuf" is synced
(a) before filling sndbuf for cpu access
(b) after filling sndbuf for device access

Usage of receive buffer "RMB" is synced
(a) before reading RMB content for cpu access
(b) after reading RMB content for device access

Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2017-07-29 11:22:58 -07:00

530 lines
14 KiB
C

/*
* Shared Memory Communications over RDMA (SMC-R) and RoCE
*
* IB infrastructure:
* Establish SMC-R as an Infiniband Client to be notified about added and
* removed IB devices of type RDMA.
* Determine device and port characteristics for these IB devices.
*
* Copyright IBM Corp. 2016
*
* Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
*/
#include <linux/random.h>
#include <linux/workqueue.h>
#include <linux/scatterlist.h>
#include <rdma/ib_verbs.h>
#include "smc_pnet.h"
#include "smc_ib.h"
#include "smc_core.h"
#include "smc_wr.h"
#include "smc.h"
#define SMC_QP_MIN_RNR_TIMER 5
#define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */
#define SMC_QP_RETRY_CNT 7 /* 7: infinite */
#define SMC_QP_RNR_RETRY 7 /* 7: infinite */
struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */
.lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
.list = LIST_HEAD_INIT(smc_ib_devices.list),
};
#define SMC_LOCAL_SYSTEMID_RESET "%%%%%%%"
u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
* identifier
*/
static int smc_ib_modify_qp_init(struct smc_link *lnk)
{
struct ib_qp_attr qp_attr;
memset(&qp_attr, 0, sizeof(qp_attr));
qp_attr.qp_state = IB_QPS_INIT;
qp_attr.pkey_index = 0;
qp_attr.port_num = lnk->ibport;
qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
| IB_ACCESS_REMOTE_WRITE;
return ib_modify_qp(lnk->roce_qp, &qp_attr,
IB_QP_STATE | IB_QP_PKEY_INDEX |
IB_QP_ACCESS_FLAGS | IB_QP_PORT);
}
static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
{
enum ib_qp_attr_mask qp_attr_mask =
IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
struct ib_qp_attr qp_attr;
memset(&qp_attr, 0, sizeof(qp_attr));
qp_attr.qp_state = IB_QPS_RTR;
qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0);
rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
sizeof(lnk->peer_mac));
qp_attr.dest_qp_num = lnk->peer_qpn;
qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
* requests
*/
qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
}
int smc_ib_modify_qp_rts(struct smc_link *lnk)
{
struct ib_qp_attr qp_attr;
memset(&qp_attr, 0, sizeof(qp_attr));
qp_attr.qp_state = IB_QPS_RTS;
qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */
qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */
qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */
qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */
qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and
* atomic ops allowed
*/
return ib_modify_qp(lnk->roce_qp, &qp_attr,
IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
IB_QP_MAX_QP_RD_ATOMIC);
}
int smc_ib_modify_qp_reset(struct smc_link *lnk)
{
struct ib_qp_attr qp_attr;
memset(&qp_attr, 0, sizeof(qp_attr));
qp_attr.qp_state = IB_QPS_RESET;
return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
}
int smc_ib_ready_link(struct smc_link *lnk)
{
struct smc_link_group *lgr =
container_of(lnk, struct smc_link_group, lnk[0]);
int rc = 0;
rc = smc_ib_modify_qp_init(lnk);
if (rc)
goto out;
rc = smc_ib_modify_qp_rtr(lnk);
if (rc)
goto out;
smc_wr_remember_qp_attr(lnk);
rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
IB_CQ_SOLICITED_MASK);
if (rc)
goto out;
rc = smc_wr_rx_post_init(lnk);
if (rc)
goto out;
smc_wr_remember_qp_attr(lnk);
if (lgr->role == SMC_SERV) {
rc = smc_ib_modify_qp_rts(lnk);
if (rc)
goto out;
smc_wr_remember_qp_attr(lnk);
}
out:
return rc;
}
/* process context wrapper for might_sleep smc_ib_remember_port_attr */
static void smc_ib_port_event_work(struct work_struct *work)
{
struct smc_ib_device *smcibdev = container_of(
work, struct smc_ib_device, port_event_work);
u8 port_idx;
for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
smc_ib_remember_port_attr(smcibdev, port_idx + 1);
clear_bit(port_idx, &smcibdev->port_event_mask);
}
}
/* can be called in IRQ context */
static void smc_ib_global_event_handler(struct ib_event_handler *handler,
struct ib_event *ibevent)
{
struct smc_ib_device *smcibdev;
u8 port_idx;
smcibdev = container_of(handler, struct smc_ib_device, event_handler);
switch (ibevent->event) {
case IB_EVENT_PORT_ERR:
port_idx = ibevent->element.port_num - 1;
set_bit(port_idx, &smcibdev->port_event_mask);
schedule_work(&smcibdev->port_event_work);
/* fall through */
case IB_EVENT_DEVICE_FATAL:
/* tbd in follow-on patch:
* abnormal close of corresponding connections
*/
break;
case IB_EVENT_PORT_ACTIVE:
port_idx = ibevent->element.port_num - 1;
set_bit(port_idx, &smcibdev->port_event_mask);
schedule_work(&smcibdev->port_event_work);
break;
default:
break;
}
}
void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
{
ib_dealloc_pd(lnk->roce_pd);
lnk->roce_pd = NULL;
}
int smc_ib_create_protection_domain(struct smc_link *lnk)
{
int rc;
lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
if (IS_ERR(lnk->roce_pd))
lnk->roce_pd = NULL;
return rc;
}
static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
{
switch (ibevent->event) {
case IB_EVENT_DEVICE_FATAL:
case IB_EVENT_GID_CHANGE:
case IB_EVENT_PORT_ERR:
case IB_EVENT_QP_ACCESS_ERR:
/* tbd in follow-on patch:
* abnormal close of corresponding connections
*/
break;
default:
break;
}
}
void smc_ib_destroy_queue_pair(struct smc_link *lnk)
{
ib_destroy_qp(lnk->roce_qp);
lnk->roce_qp = NULL;
}
/* create a queue pair within the protection domain for a link */
int smc_ib_create_queue_pair(struct smc_link *lnk)
{
struct ib_qp_init_attr qp_attr = {
.event_handler = smc_ib_qp_event_handler,
.qp_context = lnk,
.send_cq = lnk->smcibdev->roce_cq_send,
.recv_cq = lnk->smcibdev->roce_cq_recv,
.srq = NULL,
.cap = {
/* include unsolicited rdma_writes as well,
* there are max. 2 RDMA_WRITE per 1 WR_SEND
*/
.max_send_wr = SMC_WR_BUF_CNT * 3,
.max_recv_wr = SMC_WR_BUF_CNT * 3,
.max_send_sge = SMC_IB_MAX_SEND_SGE,
.max_recv_sge = 1,
},
.sq_sig_type = IB_SIGNAL_REQ_WR,
.qp_type = IB_QPT_RC,
};
int rc;
lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
if (IS_ERR(lnk->roce_qp))
lnk->roce_qp = NULL;
else
smc_wr_remember_qp_attr(lnk);
return rc;
}
void smc_ib_put_memory_region(struct ib_mr *mr)
{
ib_dereg_mr(mr);
}
static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot)
{
unsigned int offset = 0;
int sg_num;
/* map the largest prefix of a dma mapped SG list */
sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK],
buf_slot->sgt[SMC_SINGLE_LINK].sgl,
buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
&offset, PAGE_SIZE);
return sg_num;
}
/* Allocate a memory region and map the dma mapped SG list of buf_slot */
int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
struct smc_buf_desc *buf_slot)
{
if (buf_slot->mr_rx[SMC_SINGLE_LINK])
return 0; /* already done */
buf_slot->mr_rx[SMC_SINGLE_LINK] =
ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) {
int rc;
rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]);
buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL;
return rc;
}
if (smc_ib_map_mr_sg(buf_slot) != 1)
return -EINVAL;
return 0;
}
/* synchronize buffer usage for cpu access */
void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev,
struct smc_buf_desc *buf_slot,
enum dma_data_direction data_direction)
{
struct scatterlist *sg;
unsigned int i;
/* for now there is just one DMA address */
for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
if (!sg_dma_len(sg))
break;
ib_dma_sync_single_for_cpu(smcibdev->ibdev,
sg_dma_address(sg),
sg_dma_len(sg),
data_direction);
}
}
/* synchronize buffer usage for device access */
void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev,
struct smc_buf_desc *buf_slot,
enum dma_data_direction data_direction)
{
struct scatterlist *sg;
unsigned int i;
/* for now there is just one DMA address */
for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
if (!sg_dma_len(sg))
break;
ib_dma_sync_single_for_device(smcibdev->ibdev,
sg_dma_address(sg),
sg_dma_len(sg),
data_direction);
}
}
/* Map a new TX or RX buffer SG-table to DMA */
int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev,
struct smc_buf_desc *buf_slot,
enum dma_data_direction data_direction)
{
int mapped_nents;
mapped_nents = ib_dma_map_sg(smcibdev->ibdev,
buf_slot->sgt[SMC_SINGLE_LINK].sgl,
buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
data_direction);
if (!mapped_nents)
return -ENOMEM;
return mapped_nents;
}
void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev,
struct smc_buf_desc *buf_slot,
enum dma_data_direction data_direction)
{
if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address)
return; /* already unmapped */
ib_dma_unmap_sg(smcibdev->ibdev,
buf_slot->sgt[SMC_SINGLE_LINK].sgl,
buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
data_direction);
buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0;
}
static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport)
{
struct net_device *ndev;
int rc;
rc = ib_query_gid(smcibdev->ibdev, ibport, 0,
&smcibdev->gid[ibport - 1], NULL);
/* the SMC protocol requires specification of the roce MAC address;
* if net_device cannot be determined, it can be derived from gid 0
*/
ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport);
if (ndev) {
memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN);
} else if (!rc) {
memcpy(&smcibdev->mac[ibport - 1][0],
&smcibdev->gid[ibport - 1].raw[8], 3);
memcpy(&smcibdev->mac[ibport - 1][3],
&smcibdev->gid[ibport - 1].raw[13], 3);
smcibdev->mac[ibport - 1][0] &= ~0x02;
}
return rc;
}
/* Create an identifier unique for this instance of SMC-R.
* The MAC-address of the first active registered IB device
* plus a random 2-byte number is used to create this identifier.
* This name is delivered to the peer during connection initialization.
*/
static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
u8 ibport)
{
memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
sizeof(smcibdev->mac[ibport - 1]));
get_random_bytes(&local_systemid[0], 2);
}
bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
{
return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
}
int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
{
int rc;
memset(&smcibdev->pattr[ibport - 1], 0,
sizeof(smcibdev->pattr[ibport - 1]));
rc = ib_query_port(smcibdev->ibdev, ibport,
&smcibdev->pattr[ibport - 1]);
if (rc)
goto out;
rc = smc_ib_fill_gid_and_mac(smcibdev, ibport);
if (rc)
goto out;
if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
sizeof(local_systemid)) &&
smc_ib_port_active(smcibdev, ibport))
/* create unique system identifier */
smc_ib_define_local_systemid(smcibdev, ibport);
out:
return rc;
}
long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
{
struct ib_cq_init_attr cqattr = {
.cqe = SMC_WR_MAX_CQE, .comp_vector = 0 };
long rc;
smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
smc_wr_tx_cq_handler, NULL,
smcibdev, &cqattr);
rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
if (IS_ERR(smcibdev->roce_cq_send)) {
smcibdev->roce_cq_send = NULL;
return rc;
}
smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
smc_wr_rx_cq_handler, NULL,
smcibdev, &cqattr);
rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
if (IS_ERR(smcibdev->roce_cq_recv)) {
smcibdev->roce_cq_recv = NULL;
goto err;
}
INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
smc_ib_global_event_handler);
ib_register_event_handler(&smcibdev->event_handler);
smc_wr_add_dev(smcibdev);
smcibdev->initialized = 1;
return rc;
err:
ib_destroy_cq(smcibdev->roce_cq_send);
return rc;
}
static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
{
if (!smcibdev->initialized)
return;
smc_wr_remove_dev(smcibdev);
ib_unregister_event_handler(&smcibdev->event_handler);
ib_destroy_cq(smcibdev->roce_cq_recv);
ib_destroy_cq(smcibdev->roce_cq_send);
}
static struct ib_client smc_ib_client;
/* callback function for ib_register_client() */
static void smc_ib_add_dev(struct ib_device *ibdev)
{
struct smc_ib_device *smcibdev;
if (ibdev->node_type != RDMA_NODE_IB_CA)
return;
smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
if (!smcibdev)
return;
smcibdev->ibdev = ibdev;
INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
spin_lock(&smc_ib_devices.lock);
list_add_tail(&smcibdev->list, &smc_ib_devices.list);
spin_unlock(&smc_ib_devices.lock);
ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
}
/* callback function for ib_register_client() */
static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
{
struct smc_ib_device *smcibdev;
smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
ib_set_client_data(ibdev, &smc_ib_client, NULL);
spin_lock(&smc_ib_devices.lock);
list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
spin_unlock(&smc_ib_devices.lock);
smc_pnet_remove_by_ibdev(smcibdev);
smc_ib_cleanup_per_ibdev(smcibdev);
kfree(smcibdev);
}
static struct ib_client smc_ib_client = {
.name = "smc_ib",
.add = smc_ib_add_dev,
.remove = smc_ib_remove_dev,
};
int __init smc_ib_register_client(void)
{
return ib_register_client(&smc_ib_client);
}
void smc_ib_unregister_client(void)
{
ib_unregister_client(&smc_ib_client);
}