X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd_cb.c;h=d9839cef372b48ee182a58449296885ec2340566;hp=92aaf57fc7e4d518b24583fcd47e60362363b7d9;hb=2c800c7ad0102407ea89af4d2a98ba761b957d4a;hpb=31be9f94d72287d7077ffc00d392a71056f61f4d diff --git a/lnet/klnds/gnilnd/gnilnd_cb.c b/lnet/klnds/gnilnd/gnilnd_cb.c index 92aaf57..d9839ce 100644 --- a/lnet/klnds/gnilnd/gnilnd_cb.c +++ b/lnet/klnds/gnilnd/gnilnd_cb.c @@ -490,9 +490,9 @@ kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_ni } int -kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *iov, - lnet_kiov_t *kiov, unsigned int offset, unsigned int nob) - +kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, + struct kvec *iov, lnet_kiov_t *kiov, + unsigned int offset, unsigned int nob) { kgn_msg_t *msg = &tx->tx_msg; int i; @@ -505,8 +505,15 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io if (nob == 0) { tx->tx_buffer = NULL; } else if (kiov != NULL) { + + if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) { + niov = ((nob + offset + kiov->kiov_offset + PAGE_SIZE - 1) / + PAGE_SIZE); + } + LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE, - "bad niov %d\n", niov); + "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n", + niov, msg, kiov, iov, offset, nob); while (offset >= kiov->kiov_len) { offset -= kiov->kiov_len; @@ -599,7 +606,7 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io int kgnilnd_setup_virt_buffer(kgn_tx_t *tx, - unsigned int niov, struct iovec *iov, + unsigned int niov, struct kvec *iov, unsigned int offset, unsigned int nob) { @@ -739,7 +746,7 @@ error: static inline int kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, + struct kvec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob) { int rc; @@ -1158,6 +1165,7 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error) * verified peer notification - the theory is that * a TX error can be communicated in all other cases */ if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED && + error != -GNILND_NOPURG && kgnilnd_check_purgatory_conn(tx->tx_conn)) { kgnilnd_add_purgatory_tx(tx); @@ -1465,7 +1473,6 @@ kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, */ msg->gnm_connstamp = conn->gnc_my_connstamp; msg->gnm_payload_len = immediatenob; - kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex); msg->gnm_seq = atomic_read(&conn->gnc_tx_seq); /* always init here - kgn_checksum is a /sys module tunable @@ -1580,6 +1587,7 @@ kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, timestamp = jiffies; kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex); + kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex); /* delay in jiffies - we are really concerned only with things that * result in a schedule() or really holding this off for long times . * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */ @@ -1624,7 +1632,8 @@ kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob rc = 0; } else { atomic_inc(&conn->gnc_device->gnd_fast_try); - rc = kgnilnd_gl_mutex_trylock(&conn->gnc_device->gnd_cq_mutex); + rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex, + &conn->gnc_smsg_mutex); } if (!rc) { rc = -EAGAIN; @@ -2105,7 +2114,7 @@ kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) int target_is_router = lntmsg->msg_target_is_router; int routing = lntmsg->msg_routing; unsigned int niov = lntmsg->msg_niov; - struct iovec *iov = lntmsg->msg_iov; + struct kvec *iov = lntmsg->msg_iov; lnet_kiov_t *kiov = lntmsg->msg_kiov; unsigned int offset = lntmsg->msg_offset; unsigned int nob = lntmsg->msg_len; @@ -2274,7 +2283,7 @@ kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen) kgn_conn_t *conn = rx->grx_conn; kgn_msg_t *rxmsg = rx->grx_msg; unsigned int niov = lntmsg->msg_niov; - struct iovec *iov = lntmsg->msg_iov; + struct kvec *iov = lntmsg->msg_iov; lnet_kiov_t *kiov = lntmsg->msg_kiov; unsigned int offset = lntmsg->msg_offset; unsigned int nob = lntmsg->msg_len; @@ -2423,7 +2432,7 @@ kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, + struct kvec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen, unsigned int rlen) { kgn_rx_t *rx = private; @@ -3064,7 +3073,7 @@ kgnilnd_reaper(void *arg) next_check_time); mod_timer(&timer, (long) jiffies + timeout); - /* check flag variables before comitting */ + /* check flag variables before committing */ if (!kgnilnd_data.kgn_shutdown && !kgnilnd_data.kgn_quiesce_trigger) { CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n", @@ -3239,6 +3248,11 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev) spin_unlock(&conn->gnc_list_lock); kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex); + if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) { + event_data = 1LL << 48; + rc = 1; + } + if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) { atomic_inc(&dev->gnd_rdma_ntx); atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes); @@ -3293,7 +3307,7 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev) -EFAULT, rcookie, tx->tx_msg.gnm_srcnid); - kgnilnd_tx_done(tx, -EFAULT); + kgnilnd_tx_done(tx, -GNILND_NOPURG); kgnilnd_close_conn(conn, -ECOMM); }