Under heavy IO (e.g. fio) the queue is not checked frequently enough for
pending commands. As a result some pending commands are timed out by the
linux sym53c8xx driver, which sends SCSI Abort messages for the timed out
commands. The SCSI Abort messages result in linux errors, which show up
on the console and in /var/log/messages.
e.g.
sd 0:0:3:0: [sdd] tag#33 ABORT operation started
scsi target0:0:3: control msgout:
80 20 47 d
sd 0:0:3:0: ABORT operation complete.
scsi target0:0:4: message d sent on bad reselection
When the current command completes, check if there is a pending command
on the queue and if a pending command exists, set a flag indicating
that a call to lsi_wait_reselect() is needed to handle a queued pending
command. The flag is checked following a WAIT DISCONNECT Script
instruction and if set, lsi_wait_reselect() is called to handle a
queued pending command.
Signed-off-by: George Kennedy <george.kennedy@oracle.com>
---
Thanks Paolo,
Your latest suggestion was the "missing link". Calling lsi_wait_reselect()
after a WAIT DISCONNECT Script instruction when there are commands on
the pending queue is all the is needed. The patch has been greatly
reduced in size and complexity.
hw/scsi/lsi53c895a.c | 43 ++++++++++++++++++++++++++++++++-----------
1 file changed, 32 insertions(+), 11 deletions(-)
diff --git a/hw/scsi/lsi53c895a.c b/hw/scsi/lsi53c895a.c
index d1e6534..7f9ed2f 100644
--- a/hw/scsi/lsi53c895a.c
+++ b/hw/scsi/lsi53c895a.c
@@ -219,6 +219,7 @@ typedef struct {
int command_complete;
QTAILQ_HEAD(, lsi_request) queue;
lsi_request *current;
+ bool handle_pending; /* handle queued pending commands */
uint32_t dsa;
uint32_t temp;
@@ -298,6 +299,18 @@ static inline int lsi_irq_on_rsl(LSIState *s)
return (s->sien0 & LSI_SIST0_RSL) && (s->scid & LSI_SCID_RRE);
}
+static lsi_request *get_pending_req(LSIState *s)
+{
+ lsi_request *p;
+
+ QTAILQ_FOREACH(p, &s->queue, next) {
+ if (p->pending) {
+ return p;
+ }
+ }
+ return NULL;
+}
+
static void lsi_soft_reset(LSIState *s)
{
trace_lsi_reset();
@@ -446,7 +459,6 @@ static void lsi_update_irq(LSIState *s)
{
int level;
static int last_level;
- lsi_request *p;
/* It's unclear whether the DIP/SIP bits should be cleared when the
Interrupt Status Registers are cleared or when istat0 is read.
@@ -477,12 +489,12 @@ static void lsi_update_irq(LSIState *s)
lsi_set_irq(s, level);
if (!level && lsi_irq_on_rsl(s) && !(s->scntl1 & LSI_SCNTL1_CON)) {
+ lsi_request *p;
+
trace_lsi_update_irq_disconnected();
- QTAILQ_FOREACH(p, &s->queue, next) {
- if (p->pending) {
- lsi_reselect(s, p);
- break;
- }
+ p = get_pending_req(s);
+ if (p) {
+ lsi_reselect(s, p);
}
}
}
@@ -759,6 +771,8 @@ static void lsi_command_complete(SCSIRequest *req, uint32_t status, size_t resid
lsi_request_free(s, s->current);
scsi_req_unref(req);
}
+ s->handle_pending = get_pending_req(s) ? true : false;
+
lsi_resume_script(s);
}
@@ -1064,11 +1078,15 @@ static void lsi_wait_reselect(LSIState *s)
trace_lsi_wait_reselect();
- QTAILQ_FOREACH(p, &s->queue, next) {
- if (p->pending) {
- lsi_reselect(s, p);
- break;
- }
+ s->handle_pending = false;
+
+ if (s->current) {
+ return;
+ }
+
+ p = get_pending_req(s);
+ if (p) {
+ lsi_reselect(s, p);
}
if (s->current == NULL) {
s->waiting = 1;
@@ -1258,6 +1276,9 @@ again:
case 1: /* Disconnect */
trace_lsi_execute_script_io_disconnect();
s->scntl1 &= ~LSI_SCNTL1_CON;
+ if (s->handle_pending) {
+ lsi_wait_reselect(s);
+ }
break;
case 2: /* Wait Reselect */
if (!lsi_irq_on_rsl(s)) {
--
1.8.3.1
On 07/11/2018 23:04, George Kennedy wrote: > Your latest suggestion was the "missing link". Calling lsi_wait_reselect() > after a WAIT DISCONNECT Script instruction when there are commands on > the pending queue is all the is needed. The patch has been greatly > reduced in size and complexity. Yeah, it's mixing initiator and target in slightly unhealthy ways, but it's not something that you introduced. (Can you also send, perhaps offlist, the other patch you were working on which tried to introduce the right sequence of disconnect/reselect before disk accesses and DMA respectively? I'd like to keep it for reference). > hw/scsi/lsi53c895a.c | 43 ++++++++++++++++++++++++++++++++----------- > 1 file changed, 32 insertions(+), 11 deletions(-) > > diff --git a/hw/scsi/lsi53c895a.c b/hw/scsi/lsi53c895a.c > index d1e6534..7f9ed2f 100644 > --- a/hw/scsi/lsi53c895a.c > +++ b/hw/scsi/lsi53c895a.c > @@ -219,6 +219,7 @@ typedef struct { > int command_complete; > QTAILQ_HEAD(, lsi_request) queue; > lsi_request *current; > + bool handle_pending; /* handle queued pending commands */ > > uint32_t dsa; > uint32_t temp; > @@ -298,6 +299,18 @@ static inline int lsi_irq_on_rsl(LSIState *s) > return (s->sien0 & LSI_SIST0_RSL) && (s->scid & LSI_SCID_RRE); > } > > +static lsi_request *get_pending_req(LSIState *s) > +{ > + lsi_request *p; > + > + QTAILQ_FOREACH(p, &s->queue, next) { > + if (p->pending) { > + return p; > + } > + } > + return NULL; > +} > + > static void lsi_soft_reset(LSIState *s) > { > trace_lsi_reset(); > @@ -446,7 +459,6 @@ static void lsi_update_irq(LSIState *s) > { > int level; > static int last_level; > - lsi_request *p; > > /* It's unclear whether the DIP/SIP bits should be cleared when the > Interrupt Status Registers are cleared or when istat0 is read. > @@ -477,12 +489,12 @@ static void lsi_update_irq(LSIState *s) > lsi_set_irq(s, level); > > if (!level && lsi_irq_on_rsl(s) && !(s->scntl1 & LSI_SCNTL1_CON)) { > + lsi_request *p; > + > trace_lsi_update_irq_disconnected(); > - QTAILQ_FOREACH(p, &s->queue, next) { > - if (p->pending) { > - lsi_reselect(s, p); > - break; > - } > + p = get_pending_req(s); > + if (p) { > + lsi_reselect(s, p); > } > } > } > @@ -759,6 +771,8 @@ static void lsi_command_complete(SCSIRequest *req, uint32_t status, size_t resid > lsi_request_free(s, s->current); > scsi_req_unref(req); > } > + s->handle_pending = get_pending_req(s) ? true : false; > + > lsi_resume_script(s); > } > > @@ -1064,11 +1078,15 @@ static void lsi_wait_reselect(LSIState *s) > > trace_lsi_wait_reselect(); > > - QTAILQ_FOREACH(p, &s->queue, next) { > - if (p->pending) { > - lsi_reselect(s, p); > - break; > - } > + s->handle_pending = false; > + > + if (s->current) { > + return; > + } > + > + p = get_pending_req(s); > + if (p) { > + lsi_reselect(s, p); > } > if (s->current == NULL) { > s->waiting = 1; > @@ -1258,6 +1276,9 @@ again: > case 1: /* Disconnect */ > trace_lsi_execute_script_io_disconnect(); > s->scntl1 &= ~LSI_SCNTL1_CON; > + if (s->handle_pending) { Can you explain the if? Is it to avoid that "s->waiting = 1" assignment? Also, could the disconnect case check get_pending_req directly, and call lsi_reselect? That is, replacing this "if" with if (!s->current) { p = get_pending_req(s); if (p) { lsi_reselect(s, p); } } instead of calling lsi_wait_reselect? Thanks, Paolo > + lsi_wait_reselect(s); > + } > break; > case 2: /* Wait Reselect */ > if (!lsi_irq_on_rsl(s)) { >
© 2016 - 2024 Red Hat, Inc.