Date: Sun, 27 Oct 2002 18:00:42 +0000 From: Ian Dowse <iedowse@maths.tcd.ie> To: freebsd-mobile@freebsd.org Subject: Patch to fix/shorten "wi" freezes Message-ID: <200210271800.aa11251@salmon.maths.tcd.ie>
next in thread | raw e-mail | index | archive | help
The wi driver causes quite long system freezes both when the pccard
is removed, and also if the hardware becomes confused. I've found
on -current that sometimes the whole machine can become unresponsive
for a period of minutes with messages such as:
wi0: timeout in wi_cmd 0x0002; event status 0x8080
wi0: timeout in wi_cmd 0x0000; event status 0x8080
wi0: wi_cmd: busy bit won't clear.
wi0: wi_cmd: busy bit won't clear.
wi0: init failed
wi0: wi_cmd: busy bit won't clear.
... <repeated 20 times>
wi0: wi_cmd: busy bit won't clear.
wi0: failed to allocate 1594 bytes on NIC
wi0: tx buffer allocation failed
wi0: wi_cmd: busy bit won't clear.
wi0: failed to allocate 1594 bytes on NIC
The "wi_cmd 0x0002" is WI_CMD_DISABLE from wi_stop(). Each of the
"busy bit won't clear" messages comes after a 5-second busy-loop
delay in wi_cmd(), so the above takes 2-3 minutes to complete.
This, BTW, is a Lucent silver card, probed as:
wi0: <WaveLAN/IEEE> at port 0x100-0x13f irq 9 function 0
config 1 on pccard0
wi0: 802.11 address: 00:02:2d:21:57:d3
wi0: using Lucent Technologies, WaveLAN/IEEE
wi0: Lucent Firmware: Station 6.16.01
The patch below does a few things:
- It adds a 20ms delay at the end of wi_init(), which seems to fix
the above timeouts in wi_stop(), as it seems that calling wi_stop()
too soon after wi_init() can cause these.
- The busy-bit loop timeout is reduced from 5 seconds to 500ms.
- When a status of 0xffff is returned or the "busy bit won't clear"
error occurs, sc->wi_gone is set to 1, so that other operations
will fail immediately instead of going back into the long busy
loops. Since sc->wi_gone had been used as a sanity test in
wi_generic_detach() to make sure devices are not detached twice,
this has been changed to use the previously unused WI_FLAGS_ATTACHED
flag. We also need to remove the wi_gone test in wi_stop(), since
otherwise the untimeout() calls will be missed if wi_gone is set
by something other than wi_generic_detach().
- The functions wi_cmd() and wi_seek() now test wi_gone, and return
immediately if it is set.
For me this makes the card work much more reliably and it reduces
the length of any hangs to less than 1 second. I guess it will take
testing on other cards and configurations to see if this improves
things in general or causes problems with some combinations.
Ian
Index: if_wi.c
===================================================================
RCS file: /dump/FreeBSD-CVS/src/sys/dev/wi/if_wi.c,v
retrieving revision 1.117
diff -u -r1.117 if_wi.c
--- if_wi.c 14 Oct 2002 01:59:57 -0000 1.117
+++ if_wi.c 27 Oct 2002 16:57:04 -0000
@@ -200,7 +200,7 @@
WI_LOCK(sc, s);
ifp = &sc->arpcom.ac_if;
- if (sc->wi_gone) {
+ if ((sc->wi_flags & WI_FLAGS_ATTACHED) == 0) {
device_printf(dev, "already unloaded\n");
WI_UNLOCK(sc, s);
return(ENODEV);
@@ -214,6 +214,7 @@
ether_ifdetach(ifp, ETHER_BPF_SUPPORTED);
bus_teardown_intr(dev, sc->irq, sc->wi_intrhand);
wi_free(dev);
+ sc->wi_flags &= ~WI_FLAGS_ATTACHED;
sc->wi_gone = 1;
WI_UNLOCK(sc, s);
@@ -471,6 +472,7 @@
*/
ether_ifattach(ifp, ETHER_BPF_SUPPORTED);
callout_handle_init(&sc->wi_stat_ch);
+ sc->wi_flags |= WI_FLAGS_ATTACHED;
WI_UNLOCK(sc, s);
return(0);
@@ -1002,20 +1004,24 @@
{
int i, s = 0;
static volatile int count = 0;
+
+ if (sc->wi_gone)
+ return (ENODEV);
if (count > 1)
panic("Hey partner, hold on there!");
count++;
/* wait for the busy bit to clear */
- for (i = 500; i > 0; i--) { /* 5s */
+ for (i = 500; i > 0; i--) { /* 500ms */
if (!(CSR_READ_2(sc, WI_COMMAND) & WI_CMD_BUSY)) {
break;
}
- DELAY(10*1000); /* 10 m sec */
+ DELAY(1000); /* 1ms */
}
if (i == 0) {
device_printf(sc->dev, "wi_cmd: busy bit won't clear.\n" );
+ sc->wi_gone = 1;
count--;
return(ETIMEDOUT);
}
@@ -1052,6 +1058,8 @@
if (i == WI_TIMEOUT) {
device_printf(sc->dev,
"timeout in wi_cmd 0x%04x; event status 0x%04x\n", cmd, s);
+ if (s == 0xffff)
+ sc->wi_gone = 1;
return(ETIMEDOUT);
}
return(0);
@@ -1364,6 +1372,9 @@
int selreg, offreg;
int status;
+ if (sc->wi_gone)
+ return (ENODEV);
+
switch (chan) {
case WI_BAP0:
selreg = WI_SEL0;
@@ -1391,6 +1402,8 @@
if (i == WI_TIMEOUT) {
device_printf(sc->dev, "timeout in wi_seek to %x/%x; last status %x\n",
id, off, status);
+ if (status == 0xffff)
+ sc->wi_gone = 1;
return(ETIMEDOUT);
}
@@ -2196,6 +2209,13 @@
sc->wi_stat_ch = timeout(wi_inquire, sc, hz * 60);
WI_UNLOCK(sc, s);
+ /*
+ * A 10ms or greater delay here seems to avoid a problem that
+ * causes some Lucent orinoco cards to time out in wi_stop()
+ * if called immediately after wi_init(). Use 20ms to be safe.
+ */
+ DELAY(20000);
+
return;
}
@@ -2471,11 +2491,11 @@
int s;
WI_LOCK(sc, s);
-
- if (sc->wi_gone) {
- WI_UNLOCK(sc, s);
- return;
- }
+ /*
+ * Ignore wi_gone here, as we still need to do the untimeout calls.
+ * Currently everything here should be safe to do even if the
+ * hardware is gone.
+ */
wihap_shutdown(sc);
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-mobile" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi? <200210271800.aa11251>