*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2015, Intel Corporation.
+ * Copyright (c) 2011, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
peer->ibp_nid = nid;
peer->ibp_error = 0;
peer->ibp_last_alive = 0;
- peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
- peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
+ peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni);
+ peer->ibp_queue_depth = ni->ni_peertxcredits;
atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
kiblnd_debug_tx (kib_tx_t *tx)
{
CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
- "cookie "LPX64" msg %s%s type %x cred %d\n",
+ "cookie %#llx msg %s%s type %x cred %d\n",
tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
tx->tx_status, tx->tx_deadline, tx->tx_cookie,
tx->tx_lntmsg[0] == NULL ? "-" : "!",
continue;
CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
- "incarnation:"LPX64"(%x, "LPX64")\n",
+ "incarnation:%#llx(%x, %#llx)\n",
libcfs_nid2str(peer->ibp_nid),
conn->ibc_version, conn->ibc_incarnation,
version, incarnation);
rx->rx_msgaddr));
KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
- CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
+ CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
i, rx->rx_msg, rx->rx_msgaddr,
(__u64)(page_to_phys(pg) + pg_off));
}
struct ib_mr *
-kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
+kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
int negotiated_nfrags)
{
- __u16 nfrags = (negotiated_nfrags != -1) ?
- negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
+ kib_net_t *net = ni->ni_data;
+ kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables;
+ int mod;
+ __u16 nfrags;
+
+ tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
+ mod = tunables->lnd_map_on_demand;
+ nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
LASSERT(hdev->ibh_mrs != NULL);
- if (*kiblnd_tunables.kib_map_on_demand > 0 &&
- nfrags <= rd->rd_nfrags)
+ if (mod > 0 && nfrags <= rd->rd_nfrags)
return NULL;
return hdev->ibh_mrs;
list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
frd_list) {
list_del(&frd->frd_list);
+#ifndef HAVE_IB_MAP_MR_SG
ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
ib_dereg_mr(frd->frd_mr);
LIBCFS_FREE(frd, sizeof(*frd));
i++;
}
}
-static int kiblnd_fmr_pool_size(int ncpts)
+static int
+kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
+ int ncpts)
{
- int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
+ int size = tunables->lnd_fmr_pool_size / ncpts;
return max(IBLND_FMR_POOL, size);
}
-static int kiblnd_fmr_flush_trigger(int ncpts)
+static int
+kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
+ int ncpts)
{
- int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
+ int size = tunables->lnd_fmr_flush_trigger / ncpts;
return max(IBLND_FMR_POOL_FLUSH, size);
}
.dirty_watermark = fps->fps_flush_trigger,
.flush_function = NULL,
.flush_arg = NULL,
- .cache = !!*kiblnd_tunables.kib_fmr_cache};
+ .cache = !!fps->fps_cache };
int rc = 0;
fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
}
frd->frd_mr = NULL;
+#ifndef HAVE_IB_MAP_MR_SG
frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
LNET_MAX_PAYLOAD/PAGE_SIZE);
if (IS_ERR(frd->frd_frpl)) {
rc = PTR_ERR(frd->frd_frpl);
CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
rc);
+ frd->frd_frpl = NULL;
goto out_middle;
}
+#endif
+#ifdef HAVE_IB_ALLOC_FAST_REG_MR
frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
LNET_MAX_PAYLOAD/PAGE_SIZE);
+#else
+ frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
+ IB_MR_TYPE_MEM_REG,
+ LNET_MAX_PAYLOAD/PAGE_SIZE);
+#endif
if (IS_ERR(frd->frd_mr)) {
rc = PTR_ERR(frd->frd_mr);
CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
+ frd->frd_mr = NULL;
goto out_middle;
}
- frd->frd_valid = true;
+ /* There appears to be a bug in MLX5 code where you must
+ * invalidate the rkey of a new FastReg pool before first
+ * using it. Thus, I am marking the FRD invalid here. */
+ frd->frd_valid = false;
list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
fpo->fast_reg.fpo_pool_size++;
out_middle:
if (frd->frd_mr)
ib_dereg_mr(frd->frd_mr);
+#ifndef HAVE_IB_MAP_MR_SG
if (frd->frd_frpl)
ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
LIBCFS_FREE(frd, sizeof(*frd));
out:
list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
frd_list) {
list_del(&frd->frd_list);
+#ifndef HAVE_IB_MAP_MR_SG
ib_free_fast_reg_page_list(frd->frd_frpl);
+#endif
ib_dereg_mr(frd->frd_mr);
LIBCFS_FREE(frd, sizeof(*frd));
}
kib_fmr_pool_t *fpo;
int rc;
+#ifndef HAVE_IB_DEVICE_ATTRS
dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
if (!dev_attr)
return -ENOMEM;
+#endif
LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
if (!fpo) {
fpo->fpo_hdev = kiblnd_current_hdev(dev);
+#ifdef HAVE_IB_DEVICE_ATTRS
+ dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
+#else
rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
if (rc) {
CERROR("Query device failed for %s: %d\n",
fpo->fpo_hdev->ibh_ibdev->name, rc);
goto out_dev_attr;
}
+#endif
/* Check for FMR or FastReg support */
fpo->fpo_is_fmr = 0;
if (rc)
goto out_fpo;
+#ifndef HAVE_IB_DEVICE_ATTRS
kfree(dev_attr);
+#endif
fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
fpo->fpo_owner = fps;
*pp_fpo = fpo;
LIBCFS_FREE(fpo, sizeof(*fpo));
out_dev_attr:
+#ifndef HAVE_IB_DEVICE_ATTRS
kfree(dev_attr);
+#endif
return rc;
}
}
static int
-kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
- int pool_size, int flush_trigger)
+kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
+ kib_net_t *net,
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables)
{
kib_fmr_pool_t *fpo;
int rc;
fps->fps_net = net;
fps->fps_cpt = cpt;
- fps->fps_pool_size = pool_size;
- fps->fps_flush_trigger = flush_trigger;
+
+ fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
+ fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
+ fps->fps_cache = tunables->lnd_fmr_cache;
+
spin_lock_init(&fps->fps_lock);
INIT_LIST_HEAD(&fps->fps_pool_list);
INIT_LIST_HEAD(&fps->fps_failed_pool_list);
return cfs_time_aftereq(now, fpo->fpo_deadline);
}
+static int
+kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
+{
+ kib_hca_dev_t *hdev;
+ __u64 *pages = tx->tx_pages;
+ int npages;
+ int size;
+ int i;
+
+ hdev = tx->tx_pool->tpo_hdev;
+
+ for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
+ for (size = 0; size < rd->rd_frags[i].rf_nob;
+ size += hdev->ibh_page_size) {
+ pages[npages++] = (rd->rd_frags[i].rf_addr &
+ hdev->ibh_page_mask) + size;
+ }
+ }
+
+ return npages;
+}
+
void
kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
{
}
int
-kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
- __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
+kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
+ __u32 nob, __u64 iov, kib_fmr_t *fmr)
{
kib_fmr_pool_t *fpo;
+ __u64 *pages = tx->tx_pages;
__u64 version;
+ bool is_rx = (rd != tx->tx_rd);
+ bool tx_pages_mapped = 0;
+ int npages = 0;
int rc;
again:
struct ib_pool_fmr *pfmr;
spin_unlock(&fps->fps_lock);
+
+ if (!tx_pages_mapped) {
+ npages = kiblnd_map_tx_pages(tx, rd);
+ tx_pages_mapped = 1;
+ }
+
pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
pages, npages, iov);
if (likely(!IS_ERR(pfmr))) {
rc = PTR_ERR(pfmr);
} else {
if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
- struct ib_send_wr *wr;
struct kib_fast_reg_descriptor *frd;
+#ifdef HAVE_IB_MAP_MR_SG
+ struct ib_reg_wr *wr;
+ int n;
+#else
+ struct ib_rdma_wr *wr;
struct ib_fast_reg_page_list *frpl;
+#endif
struct ib_mr *mr;
frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
list_del(&frd->frd_list);
spin_unlock(&fps->fps_lock);
+#ifndef HAVE_IB_MAP_MR_SG
frpl = frd->frd_frpl;
+#endif
mr = frd->frd_mr;
if (!frd->frd_valid) {
- struct ib_send_wr *inv_wr;
+ struct ib_rdma_wr *inv_wr;
__u32 key = is_rx ? mr->rkey : mr->lkey;
inv_wr = &frd->frd_inv_wr;
memset(inv_wr, 0, sizeof(*inv_wr));
- inv_wr->opcode = IB_WR_LOCAL_INV;
- inv_wr->wr_id = IBLND_WID_MR;
- inv_wr->ex.invalidate_rkey = key;
+
+ inv_wr->wr.opcode = IB_WR_LOCAL_INV;
+ inv_wr->wr.wr_id = IBLND_WID_MR;
+ inv_wr->wr.ex.invalidate_rkey = key;
/* Bump the key */
key = ib_inc_rkey(key);
ib_update_fast_reg_key(mr, key);
}
+#ifdef HAVE_IB_MAP_MR_SG
+#ifdef HAVE_IB_MAP_MR_SG_5ARGS
+ n = ib_map_mr_sg(mr, tx->tx_frags,
+ tx->tx_nfrags, NULL, PAGE_SIZE);
+#else
+ n = ib_map_mr_sg(mr, tx->tx_frags,
+ tx->tx_nfrags, PAGE_SIZE);
+#endif
+ if (unlikely(n != tx->tx_nfrags)) {
+ CERROR("Failed to map mr %d/%d "
+ "elements\n", n, tx->tx_nfrags);
+ return n < 0 ? n : -EINVAL;
+ }
+
+ mr->iova = iov;
+
+ wr = &frd->frd_fastreg_wr;
+ memset(wr, 0, sizeof(*wr));
+
+ wr->wr.opcode = IB_WR_REG_MR;
+ wr->wr.wr_id = IBLND_WID_MR;
+ wr->wr.num_sge = 0;
+ wr->wr.send_flags = 0;
+ wr->mr = mr;
+ wr->key = is_rx ? mr->rkey : mr->lkey;
+ wr->access = (IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_WRITE);
+#else
+ if (!tx_pages_mapped) {
+ npages = kiblnd_map_tx_pages(tx, rd);
+ tx_pages_mapped = 1;
+ }
+
LASSERT(npages <= frpl->max_page_list_len);
memcpy(frpl->page_list, pages,
sizeof(*pages) * npages);
/* Prepare FastReg WR */
wr = &frd->frd_fastreg_wr;
memset(wr, 0, sizeof(*wr));
- wr->opcode = IB_WR_FAST_REG_MR;
- wr->wr_id = IBLND_WID_MR;
- wr->wr.fast_reg.iova_start = iov;
- wr->wr.fast_reg.page_list = frpl;
- wr->wr.fast_reg.page_list_len = npages;
- wr->wr.fast_reg.page_shift = PAGE_SHIFT;
- wr->wr.fast_reg.length = nob;
- wr->wr.fast_reg.rkey = is_rx ? mr->rkey
- : mr->lkey;
- wr->wr.fast_reg.access_flags =
+
+ wr->wr.opcode = IB_WR_FAST_REG_MR;
+ wr->wr.wr_id = IBLND_WID_MR;
+
+ wr->wr.wr.fast_reg.iova_start = iov;
+ wr->wr.wr.fast_reg.page_list = frpl;
+ wr->wr.wr.fast_reg.page_list_len = npages;
+ wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
+ wr->wr.wr.fast_reg.length = nob;
+ wr->wr.wr.fast_reg.rkey =
+ is_rx ? mr->rkey : mr->lkey;
+ wr->wr.wr.fast_reg.access_flags =
(IB_ACCESS_LOCAL_WRITE |
IB_ACCESS_REMOTE_WRITE);
+#endif
fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
fmr->fmr_frd = frd;
sizeof(*tx->tx_pages));
if (tx->tx_frags != NULL)
LIBCFS_FREE(tx->tx_frags,
- IBLND_MAX_RDMA_FRAGS *
- sizeof(*tx->tx_frags));
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_frags));
if (tx->tx_wrq != NULL)
LIBCFS_FREE(tx->tx_wrq,
(1 + IBLND_MAX_RDMA_FRAGS) *
}
LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
- IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
+ (1 + IBLND_MAX_RDMA_FRAGS) *
+ sizeof(*tx->tx_frags));
if (tx->tx_frags == NULL)
break;
- sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
+ sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
(1 + IBLND_MAX_RDMA_FRAGS) *
}
static int
-kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
+kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts)
{
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables;
unsigned long flags;
int cpt;
int rc;
int i;
+ tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
+
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
- if (*kiblnd_tunables.kib_map_on_demand == 0) {
+ if (tunables->lnd_map_on_demand == 0) {
read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
flags);
goto create_tx_pool;
read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- if (*kiblnd_tunables.kib_fmr_pool_size <
- *kiblnd_tunables.kib_ntx / 4) {
+ if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
- *kiblnd_tunables.kib_fmr_pool_size,
+ tunables->lnd_fmr_pool_size,
*kiblnd_tunables.kib_ntx / 4);
rc = -EINVAL;
goto failed;
for (i = 0; i < ncpts; i++) {
cpt = (cpts == NULL) ? i : cpts[i];
- rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
- kiblnd_fmr_pool_size(ncpts),
- kiblnd_fmr_flush_trigger(ncpts));
+ rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
+ net, tunables);
if (rc != 0) {
CERROR("Can't initialize FMR pool for CPT %d: %d\n",
cpt, rc);
static int
kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
{
- struct ib_device_attr *attr;
- int rc;
+#ifndef HAVE_IB_DEVICE_ATTRS
+ struct ib_device_attr *attr;
+ int rc;
+#endif
/* It's safe to assume a HCA can handle a page size
* matching that of the native system */
hdev->ibh_page_size = 1 << PAGE_SHIFT;
hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
+#ifdef HAVE_IB_DEVICE_ATTRS
+ hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
+#else
LIBCFS_ALLOC(attr, sizeof(*attr));
if (attr == NULL) {
CERROR("Out of memory\n");
CERROR("Failed to query IB device: %d\n", rc);
return rc;
}
+#endif
if (hdev->ibh_mr_size == ~0ULL) {
hdev->ibh_mr_shift = 64;
return 0;
}
- CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
+ CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
return -EINVAL;
}
if (dev == NULL)
return NULL;
- memset(dev, 0, sizeof(*dev));
netdev = dev_get_by_name(&init_net, ifname);
if (netdev == NULL) {
dev->ibd_can_failover = 0;
if (net == NULL)
goto failed;
- memset(net, 0, sizeof(*net));
-
do_gettimeofday(&tv);
net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
- ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
- ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
- ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
- ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
+ kiblnd_tunables_setup(ni);
if (ni->ni_interfaces[0] != NULL) {
/* Use the IPoIB interface specified in 'networks=' */
if (rc != 0)
goto failed;
- rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
+ rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
if (rc != 0) {
CERROR("Failed to initialize NI pools: %d\n", rc);
goto failed;
static void __exit ko2iblnd_exit(void)
{
lnet_unregister_lnd(&the_o2iblnd);
- kiblnd_tunables_fini();
}
static int __init ko2iblnd_init(void)