Commit a82a7fcd authored by Mike Marciniszyn's avatar Mike Marciniszyn Committed by Doug Ledford

IB/hfi1: Process qp wait list in IRQ thread periodically

In the event that the IRQ thread is extremely busy, the
processing of an rcd wait list can be delayed by quite
a bit until the IRQ thread completes its work.

The QP reset reference count wait can then appear to be stuck, thus
causing up a QP destroy to emit the hung task diagnostic.

Fix by processing the qp wait list periodically from the thread.  The
interval is a multiple (currently 4) of the MAX_PKT_RECV.

Also, reduce some of the excessive inlining.   The guidelines
are per packet is ok inline, otherwise the choice is based on
likelyhood of execution.
Reviewed-by: default avatarSebastian Sanchez <sebastian.sanchez@intel.com>
Reviewed-by: default avatarDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: default avatarMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: default avatarDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: default avatarDoug Ledford <dledford@redhat.com>
parent 4fcf1de5
......@@ -100,6 +100,11 @@ MODULE_VERSION(HFI1_DRIVER_VERSION);
* MAX_PKT_RCV is the max # if packets processed per receive interrupt.
*/
#define MAX_PKT_RECV 64
/*
* MAX_PKT_THREAD_RCV is the max # of packets processed before
* the qp_wait_list queue is flushed.
*/
#define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4)
#define EGR_HEAD_UPDATE_THRESHOLD 16
struct hfi1_ib_stats hfi1_stats;
......@@ -259,7 +264,7 @@ static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf,
* allowed size ranges for the respective type and, optionally,
* return the proper encoding.
*/
inline int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)
int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)
{
if (unlikely(!PAGE_ALIGNED(size)))
return 0;
......@@ -654,24 +659,68 @@ static void __prescan_rxq(struct hfi1_packet *packet)
}
}
static inline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
static void process_rcv_qp_work(struct hfi1_ctxtdata *rcd)
{
struct rvt_qp *qp, *nqp;
/*
* Iterate over all QPs waiting to respond.
* The list won't change since the IRQ is only run on one CPU.
*/
list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
list_del_init(&qp->rspwait);
if (qp->r_flags & RVT_R_RSP_NAK) {
qp->r_flags &= ~RVT_R_RSP_NAK;
hfi1_send_rc_ack(rcd, qp, 0);
}
if (qp->r_flags & RVT_R_RSP_SEND) {
unsigned long flags;
qp->r_flags &= ~RVT_R_RSP_SEND;
spin_lock_irqsave(&qp->s_lock, flags);
if (ib_rvt_state_ops[qp->state] &
RVT_PROCESS_OR_FLUSH_SEND)
hfi1_schedule_send(qp);
spin_unlock_irqrestore(&qp->s_lock, flags);
}
rvt_put_qp(qp);
}
}
static noinline int max_packet_exceeded(struct hfi1_packet *packet, int thread)
{
if (thread) {
if ((packet->numpkt & (MAX_PKT_RECV_THREAD - 1)) == 0)
/* allow defered processing */
process_rcv_qp_work(packet->rcd);
cond_resched();
return RCV_PKT_OK;
} else {
this_cpu_inc(*packet->rcd->dd->rcv_limit);
return RCV_PKT_LIMIT;
}
}
static inline int check_max_packet(struct hfi1_packet *packet, int thread)
{
int ret = RCV_PKT_OK;
if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0))
ret = max_packet_exceeded(packet, thread);
return ret;
}
static noinline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
{
int ret;
/* Set up for the next packet */
packet->rhqoff += packet->rsize;
if (packet->rhqoff >= packet->maxcnt)
packet->rhqoff = 0;
packet->numpkt++;
if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0)) {
if (thread) {
cond_resched();
} else {
ret = RCV_PKT_LIMIT;
this_cpu_inc(*packet->rcd->dd->rcv_limit);
}
}
ret = check_max_packet(packet, thread);
packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
packet->rcd->dd->rhf_offset;
......@@ -682,7 +731,7 @@ static inline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)
{
int ret = RCV_PKT_OK;
int ret;
packet->hdr = hfi1_get_msgheader(packet->rcd->dd,
packet->rhf_addr);
......@@ -723,14 +772,7 @@ static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)
if (packet->rhqoff >= packet->maxcnt)
packet->rhqoff = 0;
if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0)) {
if (thread) {
cond_resched();
} else {
ret = RCV_PKT_LIMIT;
this_cpu_inc(*packet->rcd->dd->rcv_limit);
}
}
ret = check_max_packet(packet, thread);
packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
packet->rcd->dd->rhf_offset;
......@@ -767,38 +809,6 @@ static inline void finish_packet(struct hfi1_packet *packet)
packet->etail, rcv_intr_dynamic, packet->numpkt);
}
static inline void process_rcv_qp_work(struct hfi1_packet *packet)
{
struct hfi1_ctxtdata *rcd;
struct rvt_qp *qp, *nqp;
rcd = packet->rcd;
rcd->head = packet->rhqoff;
/*
* Iterate over all QPs waiting to respond.
* The list won't change since the IRQ is only run on one CPU.
*/
list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
list_del_init(&qp->rspwait);
if (qp->r_flags & RVT_R_RSP_NAK) {
qp->r_flags &= ~RVT_R_RSP_NAK;
hfi1_send_rc_ack(rcd, qp, 0);
}
if (qp->r_flags & RVT_R_RSP_SEND) {
unsigned long flags;
qp->r_flags &= ~RVT_R_RSP_SEND;
spin_lock_irqsave(&qp->s_lock, flags);
if (ib_rvt_state_ops[qp->state] &
RVT_PROCESS_OR_FLUSH_SEND)
hfi1_schedule_send(qp);
spin_unlock_irqrestore(&qp->s_lock, flags);
}
rvt_put_qp(qp);
}
}
/*
* Handle receive interrupts when using the no dma rtail option.
*/
......@@ -826,7 +836,8 @@ int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread)
last = RCV_PKT_DONE;
process_rcv_update(last, &packet);
}
process_rcv_qp_work(&packet);
process_rcv_qp_work(rcd);
rcd->head = packet.rhqoff;
bail:
finish_packet(&packet);
return last;
......@@ -854,7 +865,8 @@ int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread)
last = RCV_PKT_DONE;
process_rcv_update(last, &packet);
}
process_rcv_qp_work(&packet);
process_rcv_qp_work(rcd);
rcd->head = packet.rhqoff;
bail:
finish_packet(&packet);
return last;
......@@ -1024,7 +1036,8 @@ int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread)
process_rcv_update(last, &packet);
}
process_rcv_qp_work(&packet);
process_rcv_qp_work(rcd);
rcd->head = packet.rhqoff;
bail:
/*
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment