svn commit: r279325 - head/sys/x86/xen
Roger Pau Monné
royger at FreeBSD.org
Thu Feb 26 16:05:11 UTC 2015
Author: royger
Date: Thu Feb 26 16:05:09 2015
New Revision: 279325
URL: https://svnweb.freebsd.org/changeset/base/279325
Log:
xen/intr: fix fallout from r278854
r278854 introduced a race in the event channel handling code. We must make
sure that the pending bit is cleared before executing the filter, or else we
might miss other events that would be injected after the filter has ran but
before the pending bit is cleared.
While there also mask event channels while FreeBSD executes the ithread
bound to that event channel. This refrains Xen from injecting more
interrupts while the ithread has not finished it's work.
Sponsored by: Citrix Systems R&D
Reported by: sbruno, robak
Tested by: robak
Modified:
head/sys/x86/xen/xen_intr.c
Modified: head/sys/x86/xen/xen_intr.c
==============================================================================
--- head/sys/x86/xen/xen_intr.c Thu Feb 26 15:59:45 2015 (r279324)
+++ head/sys/x86/xen/xen_intr.c Thu Feb 26 16:05:09 2015 (r279325)
@@ -128,6 +128,7 @@ struct xenisrc {
u_int xi_close:1; /* close on unbind? */
u_int xi_activehi:1;
u_int xi_edgetrigger:1;
+ u_int xi_masked:1;
};
#define ARRAY_SIZE(a) (sizeof(a) / sizeof(a[0]))
@@ -578,12 +579,11 @@ xen_intr_handle_upcall(struct trapframe
/* process port */
port = (l1i * LONG_BIT) + l2i;
+ synch_clear_bit(port, &s->evtchn_pending[0]);
isrc = xen_intr_port_to_isrc[port];
- if (__predict_false(isrc == NULL)) {
- synch_clear_bit(port, &s->evtchn_pending[0]);
+ if (__predict_false(isrc == NULL))
continue;
- }
/* Make sure we are firing on the right vCPU */
KASSERT((isrc->xi_cpu == PCPU_GET(cpuid)),
@@ -930,11 +930,21 @@ out:
* acknowledgements.
*/
static void
-xen_intr_disable_source(struct intsrc *isrc, int eoi)
+xen_intr_disable_source(struct intsrc *base_isrc, int eoi)
{
+ struct xenisrc *isrc;
- if (eoi == PIC_EOI)
- xen_intr_eoi_source(isrc);
+ isrc = (struct xenisrc *)base_isrc;
+
+ /*
+ * NB: checking if the event channel is already masked is
+ * needed because the event channel user-space device
+ * masks event channels on it's filter as part of it's
+ * normal operation, and those shouldn't be automatically
+ * unmasked by the generic interrupt code. The event channel
+ * device will unmask them when needed.
+ */
+ isrc->xi_masked = !!evtchn_test_and_set_mask(isrc->xi_port);
}
/*
@@ -943,8 +953,14 @@ xen_intr_disable_source(struct intsrc *i
* \param isrc The interrupt source to unmask (if necessary).
*/
static void
-xen_intr_enable_source(struct intsrc *isrc)
+xen_intr_enable_source(struct intsrc *base_isrc)
{
+ struct xenisrc *isrc;
+
+ isrc = (struct xenisrc *)base_isrc;
+
+ if (isrc->xi_masked == 0)
+ evtchn_unmask_port(isrc->xi_port);
}
/*
@@ -955,11 +971,6 @@ xen_intr_enable_source(struct intsrc *is
static void
xen_intr_eoi_source(struct intsrc *base_isrc)
{
- struct xenisrc *isrc;
-
- isrc = (struct xenisrc *)base_isrc;
- synch_clear_bit(isrc->xi_port,
- &HYPERVISOR_shared_info->evtchn_pending[0]);
}
/*
@@ -1025,8 +1036,6 @@ xen_intr_pirq_eoi_source(struct intsrc *
isrc = (struct xenisrc *)base_isrc;
- synch_clear_bit(isrc->xi_port,
- &HYPERVISOR_shared_info->evtchn_pending[0]);
if (test_bit(isrc->xi_pirq, xen_intr_pirq_eoi_map)) {
struct physdev_eoi eoi = { .irq = isrc->xi_pirq };
More information about the svn-src-head
mailing list