InfiniBand fixes
Linus, please pull from
master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband.git
This tree is also available from kernel.org mirrors at:
rsync://rsync.kernel.org/pub/scm/linux/kernel/git/roland/infiniband.git
This will pull the following changes (patches also sent as replies to this email):
Hal Rosenstock:
IPoIB: Fix SA client retransmission strategy
IB: Fix data length for RMPP SA sends
Michael S. Tsirkin:
IPoIB: fix module removal race
IB/mthca: Fix device removal memory leak
Roland Dreier:
IB/mthca: assign ACK timeout field correctly IB/mthca: fix posting of first work request IB/mthca: Initialize eq->nent before we use it IB/mthca: Fix posting work requests to shared receive queues IB/mthca: Don't try to set srq->last for userspace SRQs
IPoIB: Don't flush workqueue from within workqueue
drivers/infiniband/core/user_mad.c | 5 +- drivers/infiniband/hw/mthca/mthca_eq.c | 16 ++------ drivers/infiniband/hw/mthca/mthca_qp.c | 51 +++++++++++------------- drivers/infiniband/hw/mthca/mthca_srq.c | 25 +++++------- drivers/infiniband/ulp/ipoib/ipoib.h | 2 - drivers/infiniband/ulp/ipoib/ipoib_ib.c | 4 +- drivers/infiniband/ulp/ipoib/ipoib_main.c | 2 + drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 13 +++--- 8 files changed, 55 insertions(+), 63 deletions(-)
PATCH 1/10 IPoIB: fix module removal race
From: Michael S. Tsirkin <[email protected]>
Since ipoib uses queue_delayed_work to run flush task on port state events, it must flush scheduled work after unregistering the event handler.
---
drivers/infiniband/ulp/ipoib/ipoib_main.c | 2 ++ 1 files changed, 2 insertions(+), 0 deletions(-)
b21607256f3370b9eba48cd0b67e8686c6b51a64
drivers/infiniband/ulp/ipoib/ipoib_main.c +2 -0
@@ -1005,6 +1005,7 @@ debug_failed: register_failed:ib_unregister_event_handler (&priv->event_handler); +flush_scheduled_work (); event_failed:ipoib_dev_cleanup (priv->dev);
@@ -1057,6 +1058,7 @@ static void ipoib_remove_one(struct ib_dlist_for_each_entry_safe (priv, tmp, dev_list, list) {ib_unregister_event_handler (&priv->event_handler); +flush_scheduled_work ();unregister_netdev (priv->dev);ipoib_dev_cleanup (priv->dev);
PATCH 2/10 IB/mthca: assign ACK timeout field correctly
The hardware reads the ACK timeout field from the most significant 5 bits of struct mthca_qp_path's ackto field, not the least significant bits. This fix has the driver put the timeout in the right place. Without this, we get a timeout that is 2^8 times too small.
---
drivers/infiniband/hw/mthca/mthca_qp.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)
6fd9dccd77024ea85b65aa3e8f1cce22caa0d578
drivers/infiniband/hw/mthca/mthca_qp.c +1 -1
@@ -687,7 +687,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, }if (attr_mask &IB_QP_TIMEOUT ) { - qp_context->pri_path.ackto = attr->timeout; + qp_context->pri_path.ackto = attr->timeout <<3 ; qp_param->opt_param_mask |=cpu_to_be32 (MTHCA_QP_OPTPAR_ACK_TIMEOUT ); }
PATCH 3/10 IB/mthca: fix posting of first work request
Fix posting first WQE for mem-free HCAs: we need to link to previous WQE even in that case. While we're at it, simplify code for Tavor-mode HCAs. We don't really need the conditional test there either; we can similarly always link to the previous WQE.
Based on Michael S. Tsirkin's analogous fix for userspace libmthca.
---
drivers/infiniband/hw/mthca/mthca_qp.c | 48 ++++++++++++++----------------- drivers/infiniband/hw/mthca/mthca_srq.c | 14 ++++----- 2 files changed, 28 insertions(+), 34 deletions(-)
bfb454be124d30e9bd140d9c6ef6aec549f4d293
drivers/infiniband/hw/mthca/mthca_qp.c +22 -26
@@ -227,7 +227,6 @@ static void mthca_wq_init(struct mthca_w wq->last_comp = wq->max -1 ; wq->head =0 ; wq->tail =0 ; - wq->last =NULL ; } void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
@@ -1103,6 +1102,9 @@ static int mthca_alloc_qp_common(struct } } + qp->sq.last =get_send_wqe (qp, qp->sq.max -1 ); + qp->rq.last =get_recv_wqe (qp, qp->rq.max -1 ); +return 0 ; }
@@ -1583,15 +1585,13 @@ int mthca_tavor_post_send(struct ib_qp * goto out; } -if (prev_wqe) { - ((struct mthca_next_seg *) prev_wqe)->nda_op = -cpu_to_be32 (((ind << qp->sq.wqe_shift) + - qp->send_wqe_offset) | - mthca_opcode[wr->opcode]); -wmb (); - ((struct mthca_next_seg *) prev_wqe)->ee_nds = -cpu_to_be32 ((size0 ?0 :MTHCA_NEXT_DBD ) | size); - } + ((struct mthca_next_seg *) prev_wqe)->nda_op = +cpu_to_be32 (((ind << qp->sq.wqe_shift) + + qp->send_wqe_offset) | + mthca_opcode[wr->opcode]); +wmb (); + ((struct mthca_next_seg *) prev_wqe)->ee_nds = +cpu_to_be32 ((size0 ?0 :MTHCA_NEXT_DBD ) | size);if (!size0) { size0 = size;
@@ -1688,13 +1688,11 @@ int mthca_tavor_post_receive(struct ib_q qp->wrid[ind] = wr->wr_id; -if (likely (prev_wqe)) { - ((struct mthca_next_seg *) prev_wqe)->nda_op = -cpu_to_be32 ((ind << qp->rq.wqe_shift) |1 ); -wmb (); - ((struct mthca_next_seg *) prev_wqe)->ee_nds = -cpu_to_be32 (MTHCA_NEXT_DBD | size); - } + ((struct mthca_next_seg *) prev_wqe)->nda_op = +cpu_to_be32 ((ind << qp->rq.wqe_shift) |1 ); +wmb (); + ((struct mthca_next_seg *) prev_wqe)->ee_nds = +cpu_to_be32 (MTHCA_NEXT_DBD | size);if (!size0) size0 = size;
@@ -1905,15 +1903,13 @@ int mthca_arbel_post_send(struct ib_qp * goto out; } -if (likely (prev_wqe)) { - ((struct mthca_next_seg *) prev_wqe)->nda_op = -cpu_to_be32 (((ind << qp->sq.wqe_shift) + - qp->send_wqe_offset) | - mthca_opcode[wr->opcode]); -wmb (); - ((struct mthca_next_seg *) prev_wqe)->ee_nds = -cpu_to_be32 (MTHCA_NEXT_DBD | size); - } + ((struct mthca_next_seg *) prev_wqe)->nda_op = +cpu_to_be32 (((ind << qp->sq.wqe_shift) + + qp->send_wqe_offset) | + mthca_opcode[wr->opcode]); +wmb (); + ((struct mthca_next_seg *) prev_wqe)->ee_nds = +cpu_to_be32 (MTHCA_NEXT_DBD | size);if (!size0) { size0 = size;
drivers/infiniband/hw/mthca/mthca_srq.c +6 -8
@@ -189,7 +189,6 @@ int mthca_alloc_srq(struct mthca_dev *de srq->max = attr->max_wr; srq->max_gs = attr->max_sge; - srq->last =NULL ; srq->counter =0 ;if (mthca_is_memfree (dev))
@@ -264,6 +263,7 @@ int mthca_alloc_srq(struct mthca_dev *de srq->first_free =0 ; srq->last_free = srq->max -1 ; + srq->last =get_wqe (srq, srq->max -1 );return 0 ;
@@ -446,13 +446,11 @@ int mthca_tavor_post_srq_recv(struct ib_ ((struct mthca_data_seg *) wqe)->addr =0 ; } -if (likely (prev_wqe)) { - ((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 ); - } + ((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;
PATCH 4/10 IPoIB: Fix SA client retransmission strategy
From: Hal Rosenstock <[email protected]>
We got a little mixed up with what the backoff member holds in the IPoIB multicast group structure: sometimes it was used as a number of seconds, and sometimes it was used as a number of jiffies. Fix the code so that backoff is always in seconds.
---
drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 6 +++--- 1 files changed, 3 insertions(+), 3 deletions(-)
d24ef0519e081774db6a1515ad8dadefd3fcd508
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -3
@@ -145,7 +145,7 @@ static struct ipoib_mcast *ipoib_mcast_a mcast->dev = dev; mcast->created = jiffies; - mcast->backoff =HZ ; + mcast->backoff =1 ; mcast->logcount =0 ;INIT_LIST_HEAD (&mcast->list);
@@ -396,7 +396,7 @@ static void ipoib_mcast_join_complete(in IPOIB_GID_ARG(mcast->mcmember.mgid), status);if (!status && !ipoib_mcast_join_finish (mcast, mcmember)) { - mcast->backoff =HZ ; + mcast->backoff =1 ;down (&mcast_mutex);if (test_bit (IPOIB_MCAST_RUN , &priv->flags))queue_work (ipoib_workqueue, &priv->mcast_task);
@@ -496,7 +496,7 @@ static void ipoib_mcast_join(struct net_if (test_bit (IPOIB_MCAST_RUN , &priv->flags))queue_delayed_work (ipoib_workqueue, &priv->mcast_task, - mcast->backoff); + mcast->backoff *HZ );up (&mcast_mutex); }else mcast->query_id = ret;
PATCH 5/10 IB/mthca: Initialize eq->nent before we use it
In mthca_create_eq(), we call get_eqe() before setting eq->nent. This is wrong, because get_eqe() uses eq->nent. Fix this, and clean up the code a little while we're at it. (We got lucky with the current code, because eq->nent was cleared to 0, which get_eqe() made happen to do the right thing)
Pointed out by Michael S. Tsirkin <[email protected]>
---
drivers/infiniband/hw/mthca/mthca_eq.c | 16 +++++----------- 1 files changed, 5 insertions(+), 11 deletions(-)
8b194835d2f28c793d25d2c41753af2b7ee29f31
drivers/infiniband/hw/mthca/mthca_eq.c +5 -11
@@ -476,12 +476,8 @@ static int __devinit mthca_create_eq(strint i;u8 status; -/* Make sure EQ size is aligned to a power of 2 size. */ -for (i =1 ; i < nent; i <<=1 ) - ;/* nothing */ - nent = i; - - eq->dev = dev; + eq->dev = dev; + eq->nent =roundup_pow_of_two (max (nent,2 )); eq->page_list =kmalloc (npages *sizeof *eq->page_list,GFP_KERNEL );
@@ -512,7 +508,7 @@ static int __devinit mthca_create_eq(strmemset (eq->page_list[i].buf,0 ,PAGE_SIZE ); } -for (i =0 ; i < nent; ++i) +for (i =0 ; i < eq->nent; ++i)set_eqe_hw (get_eqe (eq, i)); eq->eqn =mthca_alloc (&dev->eq_table.alloc);
@@ -528,8 +524,6 @@ static int __devinit mthca_create_eq(strif (err) goto err_out_free_eq; - eq->nent = nent; -memset (eq_context,0 ,sizeof *eq_context); eq_context->flags = cpu_to_be32(MTHCA_EQ_STATUS_OK | MTHCA_EQ_OWNER_HW |
@@ -538,7 +532,7 @@ static int __devinit mthca_create_eq(strif (mthca_is_memfree (dev)) eq_context->flags |=cpu_to_be32 (MTHCA_EQ_STATE_ARBEL ); - eq_context->logsize_usrpage =cpu_to_be32 ((ffs (nent) -1 ) <<24 ); + eq_context->logsize_usrpage =cpu_to_be32 ((ffs (eq->nent) -1 ) <<24 );if (mthca_is_memfree (dev)) { eq_context->arbel_pd =cpu_to_be32 (dev->driver_pd.pd_num); }else {
@@ -569,7 +563,7 @@ static int __devinit mthca_create_eq(str dev->eq_table.arm_mask |= eq->eqn_mask;mthca_dbg (dev,"Allocated EQ %d with %d entries\n" , - eq->eqn, nent); + eq->eqn, eq->nent);return err;
PATCH 6/10 IB/mthca: Fix posting work requests to shared receive queues
The error handling paths in mthca_tavor_post_srq_recv() and mthca_arbel_post_srq_recv() are quite bogus, the result of a screwed up merge. Fix them so they work as intended.
Pointed out by Michael S. Tsirkin <[email protected]>
---
drivers/infiniband/hw/mthca/mthca_srq.c | 10 ++++------ 1 files changed, 4 insertions(+), 6 deletions(-)
b95e7b96cd976a5974fa2ae8f3a1af510cb8d4c9
drivers/infiniband/hw/mthca/mthca_srq.c +4 -6
@@ -409,7 +409,7 @@ int mthca_tavor_post_srq_recv(struct ib_mthca_err (dev,"SRQ %06x full\n" , srq->srqn); err = -ENOMEM ; *bad_wr = wr; -return nreq; +break ; } wqe =get_wqe (srq, ind);
@@ -427,7 +427,7 @@ int mthca_tavor_post_srq_recv(struct ib_ err = -EINVAL ; *bad_wr = wr; srq->last = prev_wqe; -return nreq; +break ; }for (i =0 ; i < wr->num_sge; ++i) {
@@ -456,8 +456,6 @@ int mthca_tavor_post_srq_recv(struct ib_ srq->first_free = next_ind; } -return nreq; -if (likely (nreq)) {__be32 doorbell[2 ];
@@ -501,7 +499,7 @@ int mthca_arbel_post_srq_recv(struct ib_mthca_err (dev,"SRQ %06x full\n" , srq->srqn); err = -ENOMEM ; *bad_wr = wr; -return nreq; +break ; } wqe =get_wqe (srq, ind);
@@ -517,7 +515,7 @@ int mthca_arbel_post_srq_recv(struct ib_if (unlikely (wr->num_sge > srq->max_gs)) { err = -EINVAL ; *bad_wr = wr; -return nreq; +break ; }for (i =0 ; i < wr->num_sge; ++i) {
PATCH 7/10 IB/mthca: Don't try to set srq->last for userspace SRQs
Subject: [PATCH] IB/mthca: Don't try to set srq->last for userspace SRQs
Userspace SRQs don't have a buffer allocated for them in the kernel, so it doesn't make sense to set srq->last during initialization. In fact, this can crash trying to follow a nonexistent buffer pointer.
---
drivers/infiniband/hw/mthca/mthca_srq.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)
6577ae51cf52f5fb0e4a85e673dd7bf2d0074e3e
drivers/infiniband/hw/mthca/mthca_srq.c +2 -1
@@ -172,6 +172,8 @@ static int mthca_alloc_srq_buf(struct mt scatter->lkey =cpu_to_be32 (MTHCA_INVAL_LKEY ); } + srq->last =get_wqe (srq, srq->max -1 ); +return 0 ; }
@@ -263,7 +265,6 @@ int mthca_alloc_srq(struct mthca_dev *de srq->first_free =0 ; srq->last_free = srq->max -1 ; - srq->last =get_wqe (srq, srq->max -1 );return 0 ;
PATCH 8/10 IB: Fix data length for RMPP SA sends
Subject: [PATCH] IB: Fix data length for RMPP SA sends
From: Hal Rosenstock <[email protected]>
Date: 1127163061 -0700
We need to subtract off the header length from our payload length when sending multi-packet SA messages.
---
drivers/infiniband/core/user_mad.c | 5 +++-- 1 files changed, 3 insertions(+), 2 deletions(-)
eff4c654b1a4a5e5493fbdc3affa6dd48765c085
drivers/infiniband/core/user_mad.c +3 -2
@@ -334,10 +334,11 @@ static ssize_t ib_umad_write(struct file ret = -EINVAL ; goto err_ah; } -/* Validate that management class can support RMPP */ + +/* Validate that the management class can support RMPP */ if (rmpp_mad->mad_hdr.mgmt_class ==IB_MGMT_CLASS_SUBN_ADM ) { hdr_len = offsetof(struct ib_sa_mad , data); - data_len = length; + data_len = length - hdr_len; }else if ((rmpp_mad->mad_hdr.mgmt_class >=IB_MGMT_CLASS_VENDOR_RANGE2_START ) && (rmpp_mad->mad_hdr.mgmt_class <=IB_MGMT_CLASS_VENDOR_RANGE2_END )) { hdr_len = offsetof(struct ib_vendor_mad , data);
PATCH 9/10 IPoIB: Don't flush workqueue from within workqueue
Subject: [PATCH] IPoIB: Don't flush workqueue from within workqueue
ipoib_mcast_restart_task() is always called from within the single-threaded IPoIB workqueue, so flushing the workqueue from within the function can lead to a recursion overflow. But since we're running in a single-threaded workqueue, we're already synchronized against other items in the workqueue, so just get rid of the flush in ipoib_mcast_restart_task().
---
drivers/infiniband/ulp/ipoib/ipoib.h | 2 +- drivers/infiniband/ulp/ipoib/ipoib_ib.c | 4 ++-- drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 7 ++++--- 3 files changed, 7 insertions(+), 6 deletions(-)
8d2cae0651502028bf64844508ab18528bbd65c2
drivers/infiniband/ulp/ipoib/ipoib.h +1 -1
@@ -257,7 +257,7 @@ void ipoib_mcast_send(struct net_devicevoid ipoib_mcast_restart_task (void *dev_ptr);int ipoib_mcast_start_thread (struct net_device *dev); -int ipoib_mcast_stop_thread (struct net_device *dev); +int ipoib_mcast_stop_thread (struct net_device *dev,int flush);void ipoib_mcast_dev_down (struct net_device *dev);void ipoib_mcast_dev_flush (struct net_device *dev);
drivers/infiniband/ulp/ipoib/ipoib_ib.c +2 -2
@@ -432,7 +432,7 @@ int ipoib_ib_dev_down(struct net_deviceflush_workqueue (ipoib_workqueue); } -ipoib_mcast_stop_thread (dev); +ipoib_mcast_stop_thread (dev,1 );/* * Flush the multicast groups first so we stop any multicast joins. The
@@ -599,7 +599,7 @@ void ipoib_ib_dev_cleanup(struct net_devipoib_dbg (priv,"cleaning up ib_dev\n" ); -ipoib_mcast_stop_thread (dev); +ipoib_mcast_stop_thread (dev,1 );/* Delete the broadcast address and the local address */ ipoib_mcast_dev_down (dev);
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +4 -3
@@ -598,7 +598,7 @@ int ipoib_mcast_start_thread(struct net_return 0 ; } -int ipoib_mcast_stop_thread (struct net_device *dev) +int ipoib_mcast_stop_thread (struct net_device *dev,int flush) {struct ipoib_dev_priv *priv =netdev_priv (dev);struct ipoib_mcast *mcast;
@@ -610,7 +610,8 @@ int ipoib_mcast_stop_thread(struct net_dcancel_delayed_work (&priv->mcast_task);up (&mcast_mutex); -flush_workqueue (ipoib_workqueue); +if (flush) +flush_workqueue (ipoib_workqueue);if (priv->broadcast && priv->broadcast->query) {ib_sa_cancel_query (priv->broadcast->query_id, priv->broadcast->query);
@@ -832,7 +833,7 @@ void ipoib_mcast_restart_task(void *dev_ipoib_dbg_mcast (priv,"restarting multicast task\n" ); -ipoib_mcast_stop_thread (dev); +ipoib_mcast_stop_thread (dev,0 );spin_lock_irqsave (&priv->lock, flags);
PATCH 10/10 IB/mthca: Fix device removal memory leak
Subject: [PATCH] IB/mthca: Fix device removal memory leak
From: Michael S. Tsirkin <[email protected]>
Date: 1127238888 -0700
Clean up QP table array on device removal.
---
drivers/infiniband/hw/mthca/mthca_qp.c | 1 + 1 files changed, 1 insertions(+), 0 deletions(-)
71eea47d853bb0ce0c6befe11b3e08111263170f
drivers/infiniband/hw/mthca/mthca_qp.c +1 -0
@@ -2123,5 +2123,6 @@ void __devexit mthca_cleanup_qp_table(stfor (i =0 ; i <2 ; ++i)mthca_CONF_SPECIAL_QP (dev, i,0 , &status); +mthca_array_cleanup (&dev->qp_table.qp, dev->limits.num_qps);mthca_alloc_cleanup (&dev->qp_table.alloc); }