[Spice-devel] [RFC qxl-win 2/2] display: handle interrupt handler disappearance (rhbz 721118)

[Alon Levy]

This patch handles the surprise removal whql test. The surprise
removal test sends a surprise removal IRP via a filter driver to
qxl.sys. The handling of that IRP is actually done in videoprt.sys,
and there is no API to get notified. The side effect of the handler
in videoprt.sys is disabling our interrupt handler. This patch takes
care of continuing to work after our interrupt is disabled. This is
in effect the opposite of what the MSDN states that a device should do
after a surprise removal, since the MSDN requires every IO to fail after
the surprise removal IRP has been successfully handled. On the other
hand, this workaround is sufficient to pass the test.

The workaround is to turn the EngWaitForSingleEvent on the io_cmd_event
event into a timed wait.

If it fails, we revert to synchronouse IO, which doesn't require that
event at all.
---
 display/qxldd.h |   39 ++++++++++++++++++++++++++++++++++++++-
 1 files changed, 38 insertions(+), 1 deletions(-)

diff --git a/display/qxldd.h b/display/qxldd.h
index c1a509f..f02a71e 100644
--- a/display/qxldd.h
+++ b/display/qxldd.h
@@ -442,6 +442,25 @@ static _inline RingItem *RingGetTail(PDev *pdev, Ring *ring)
 #define WAIT_FOR_EVENT(pdev, event, timeout) EngWaitForSingleObject(event, timeout)
 #endif
 
+/* Helpers for dealing with ENG_TIME_FIELDS */
+static _inline ULONG64 eng_time_diff_ms(ENG_TIME_FIELDS *b, ENG_TIME_FIELDS *a)
+{
+    ULONG64 ret = 0;
+
+    ret += b->usMilliseconds - a->usMilliseconds;
+    ret += 1000 * (b->usSecond - a->usSecond);
+    ret += 60000 * (b->usMinute - a->usMinute);
+    ret += 3600000L * (b->usHour - a->usHour);
+    // don't get into gregorian calendar, just ignore more then a single day difference
+    if (b->usDay != a->usDay) {
+        ret += (3600L * 24L * 1000L);
+    }
+    return ret;
+}
+
+#define INTERRUPT_NOT_PRESENT_TIMEOUT_MS 2000L
+#define INTERRUPT_NOT_PRESENT_TIMEOUT_100NS (INTERRUPT_NOT_PRESENT_TIMEOUT_MS * 10000L)
+
 /* Write to an ioport. For some operations we support a new port that returns
  * immediatly, and completion is signaled by an interrupt that sets io_cmd_event.
  * If the pci_revision is >= QXL_REVISION_STABLE_V10, we support it, else do
@@ -449,10 +468,28 @@ static _inline RingItem *RingGetTail(PDev *pdev, Ring *ring)
  */
 static _inline void async_io(PDev *pdev, asyncable_t op, UCHAR val)
 {
+    ENG_TIME_FIELDS start, finish;
+    LARGE_INTEGER timeout;                      // 1 => 100 nanoseconds
+    ULONG64 millis;
+
     if (pdev->use_async) {
         EngAcquireSemaphore(pdev->io_sem);
         WRITE_PORT_UCHAR(pdev->asyncable[op][ASYNC], val);
-        WAIT_FOR_EVENT(pdev, pdev->io_cmd_event, NULL);
+        /* Our Interrupt may be taken from us unexpectedly, by a surprise removal.
+         * in which case this event will never be set. This happens only during WHQL
+         * tests (pnpdtest /surprise). So instead: Wait on a timer, if we fail, stop waiting, until
+         * we get reset. We use EngQueryLocalTime because there is no way to differentiate a return on
+         * timeout from a return on event set otherwise. */
+        timeout.QuadPart = -INTERRUPT_NOT_PRESENT_TIMEOUT_100NS; // negative  => relative
+        DEBUG_PRINT((pdev, 15, "WAIT_FOR_EVENT %d\n", (int)op));
+        EngQueryLocalTime(&start);
+        WAIT_FOR_EVENT(pdev, pdev->io_cmd_event, &timeout);
+        EngQueryLocalTime(&finish);
+        millis = eng_time_diff_ms(&finish, &start);
+        if (millis >= INTERRUPT_NOT_PRESENT_TIMEOUT_MS) {
+            pdev->use_async = 0;
+            DEBUG_PRINT((pdev, 0, "%s: timeout reached, disabling async io!\n", __FUNCTION__));
+        }
         EngReleaseSemaphore(pdev->io_sem);
         DEBUG_PRINT((pdev, 3, "finished async %d\n", (int)op));
     } else {
-- 
1.7.6.2