Finally some clarity with what is going on inside this Dell CPx650
laptop (TI PCI1225 Cardbus bridge). Yes, it _is_ contact bounce. It
seems to find the 3com NIC particularly offensive - the card can easily
bounce out 150 milliseconds after the first insertion interrupt.
The user-visible effect of contact bounce is the message "cs: socket
xxxxxxxx timed out during reset". What happens is that the insertion
causes an interrupt, we wait for setup_delay to expire, then we apply
power. The card becomes ready about 10 milliseconds later (according
to CB_PWRCYCLE). Then the card bounces in and out of the socket. The
card is auto-powered down and doesn't come back up. Hence the timeout.
On to the patch:
- During the conversion from timer-based to thread-based control,
several of the magical delays in cs.c (setup_delay, resume_delay,
shutdown_delay) got themselves accidentally reduced by a factor of
ten.
These have been set back to what they are in David's latest
release. Everything is now in terms of centiseconds (1 cs = 1/100th
of a second). Not a particularly user-friendly unit, but it
preserves module parm compatibility with pcmcia-cs.
There is one subtle change here: in pcmcia-cs the module parameter:
options cs.o setup_delay=100
means to delay 100 jiffies. With this patch, it means 100
centiseconds. These are equivalent if HZ=100, but the delays in
pcmcia_cs will vary inversely with variations in HZ.
- Capriciously doubled the value of setup_delay to 10 centiseconds.
- Changed the "socket timed out" message to tell people to try
increasing the value of setup_delay.
- Debouncing.
This is a bug in the current implementation. When a card is
inserted, we take an interrupt and the thread is scheduled. If,
during the thread execution, the card bounces out and in (very
likely), an additional interrupt is generated and
socket->events:SS_DETECT is set. When parse_events returns, the
thread sees the buffered insertion event and calls parse_events
again. parse_events has no choice but to power down the card, kick
out the driver, reload the driver, beep, beep, beep.
There are a number of ways of addressing this, and a lot of them
won't work. This patch treats `setup_delay' as the debounce period.
When this delay expires it sends a message to the low-level driver
telling it that the debounce period has ended, and that it should now
forget any buffered insertion events which it had saved up. This was
implemented via the new "one shot" flag socket_state_t.flags:SS_DEBOUNCED.
Now, clearing events:SS_DETECT after the expiry of setup_delay begs
the question "what if the card has really been ejected"? That's OK,
because setup_socket checks this - if the card isn't there then it
takes the "no card" path and everything is happy (I tested this).
Finally, if we _do_ get contact bounce even after the expiry of
setup_delay, we get a reset timeout and the card services state
machine gives up. But there is now a buffered-up insertion event! So
parse_events is again called and everything works out.
BTW: It appears that these Cardbus controllers have protection
circuitry which turns off the voltages when the card is unplugged.
With this controller, the voltage remains off even when the card
bounces back in again (fair enough). So the controller is in a state
where the CB_SOCKET_CONTROL register has the value 0x33, but there is
no voltage at the card. So this test in yenta_set_power():
if (reg != cb_readl(socket, CB_SOCKET_CONTROL))
cb_writel(socket, CB_SOCKET_CONTROL, reg);
prevents us from powering the card back up. If the test is removed,
then the act of writing 0x33 to a register which already has a value of
0x33 _does_ power the card up. But this is just FYI. I left the test
in.
--- linux-2.4.0-test8/include/pcmcia/ss.h Sat Sep 9 16:19:30 2000
+++ linux-akpm/include/pcmcia/ss.h Wed Sep 13 00:13:06 2000
@@ -82,6 +82,7 @@
#define SS_DMA_MODE 0x0080
#define SS_SPKR_ENA 0x0100
#define SS_OUTPUT_ENA 0x0200
+#define SS_DEBOUNCED 0x0400 /* Tell driver that the debounce delay has ended */
/* Flags for I/O port and memory windows */
#define MAP_ACTIVE 0x01
--- linux-2.4.0-test8/drivers/pcmcia/cardbus.c Sat Sep 9 16:19:26 2000
+++ linux-akpm/drivers/pcmcia/cardbus.c Wed Sep 13 00:13:06 2000
@@ -58,11 +58,6 @@
#include <asm/irq.h>
#include <asm/io.h>
-#ifndef PCMCIA_DEBUG
-#define PCMCIA_DEBUG 1
-#endif
-static int pc_debug = PCMCIA_DEBUG;
-
#define IN_CARD_SERVICES
#include <pcmcia/version.h>
#include <pcmcia/cs_types.h>
@@ -72,6 +67,11 @@
#include <pcmcia/cistpl.h>
#include "cs_internal.h"
#include "rsrc_mgr.h"
+
+#ifndef PCMCIA_DEBUG
+#define PCMCIA_DEBUG 1
+#endif
+static int pc_debug = PCMCIA_DEBUG;
/*====================================================================*/
--- linux-2.4.0-test8/drivers/pcmcia/cs.c Sat Sep 9 16:19:26 2000
+++ linux-akpm/drivers/pcmcia/cs.c Wed Sep 13 00:13:06 2000
@@ -103,13 +103,13 @@
#define INT_MODULE_PARM(n, v) static int n = v; MODULE_PARM(n, "i")
-INT_MODULE_PARM(setup_delay, HZ/20); /* ticks */
-INT_MODULE_PARM(resume_delay, HZ/5); /* ticks */
-INT_MODULE_PARM(shutdown_delay, HZ/40); /* ticks */
-INT_MODULE_PARM(vcc_settle, 400); /* msecs */
+INT_MODULE_PARM(setup_delay, 10); /* centiseconds */
+INT_MODULE_PARM(resume_delay, 20); /* centiseconds */
+INT_MODULE_PARM(shutdown_delay, 3); /* centiseconds */
+INT_MODULE_PARM(vcc_settle, 40); /* centiseconds */
INT_MODULE_PARM(reset_time, 10); /* usecs */
-INT_MODULE_PARM(unreset_delay, 100); /* msecs */
-INT_MODULE_PARM(unreset_check, 100); /* msecs */
+INT_MODULE_PARM(unreset_delay, 10); /* centiseconds */
+INT_MODULE_PARM(unreset_check, 10); /* centiseconds */
INT_MODULE_PARM(unreset_limit, 30); /* unreset_check's */
/* Access speed for attribute memory windows */
@@ -446,10 +446,13 @@
static int send_event(socket_info_t *s, event_t event, int priority);
-static void msleep(unsigned int msec)
+/*
+ * Sleep for n_cs centiseconds (1 cs = 1/100th of a second)
+ */
+static void cs_sleep(unsigned int n_cs)
{
current->state = TASK_INTERRUPTIBLE;
- schedule_timeout( (msec * HZ + 999) / 1000);
+ schedule_timeout( (n_cs * HZ + 99) / 100);
}
static void shutdown_socket(socket_info_t *s)
@@ -504,7 +507,7 @@
if (!(val & SS_PENDING))
break;
if (--setup_timeout) {
- msleep(100);
+ cs_sleep(10);
continue;
}
printk(KERN_NOTICE "cs: socket %p voltage interrogation"
@@ -516,7 +519,7 @@
if (val & SS_DETECT) {
DEBUG(1, "cs: setup_socket(%p): applying power\n", s);
s->state |= SOCKET_PRESENT;
- s->socket.flags = 0;
+ s->socket.flags &= SS_DEBOUNCED;
if (val & SS_3VCARD)
s->socket.Vcc = s->socket.Vpp = 33;
else if (!(val & SS_XVCARD))
@@ -533,7 +536,7 @@
#endif
}
set_socket(s, &s->socket);
- msleep(vcc_settle);
+ cs_sleep(vcc_settle);
reset_socket(s);
ret = 1;
} else {
@@ -561,7 +564,7 @@
udelay((long)reset_time);
s->socket.flags &= ~SS_RESET;
set_socket(s, &s->socket);
- msleep(unreset_delay);
+ cs_sleep(unreset_delay/10);
unreset_socket(s);
} /* reset_socket */
@@ -580,11 +583,11 @@
break;
DEBUG(2, "cs: socket %d not ready yet\n", s->sock);
if (--setup_timeout) {
- msleep(unreset_check);
+ cs_sleep(unreset_check);
continue;
}
printk(KERN_NOTICE "cs: socket %p timed out during"
- " reset\n", s);
+ " reset. Try increasing setup_delay.\n", s);
s->state &= ~EVENT_MASK;
return;
}
@@ -656,7 +659,7 @@
DEBUG(0, "cs: flushing pending setup\n");
s->state &= ~EVENT_MASK;
}
- msleep(shutdown_delay);
+ cs_sleep(shutdown_delay);
s->state &= ~SOCKET_PRESENT;
shutdown_socket(s);
}
@@ -679,11 +682,13 @@
}
s->state |= SOCKET_SETUP_PENDING;
if (s->state & SOCKET_SUSPEND)
- msleep(resume_delay);
+ cs_sleep(resume_delay);
else
- msleep(setup_delay);
+ cs_sleep(setup_delay);
+ s->socket.flags |= SS_DEBOUNCED;
if (setup_socket(s) == 0)
s->state &= ~SOCKET_SETUP_PENDING;
+ s->socket.flags &= ~SS_DEBOUNCED;
}
}
if (events & SS_BATDEAD)
--- linux-2.4.0-test8/drivers/pcmcia/yenta.c Sat Sep 9 16:19:26 2000
+++ linux-akpm/drivers/pcmcia/yenta.c Wed Sep 13 00:13:06 2000
@@ -233,6 +233,11 @@
{
u16 bridge;
+ if (state->flags & SS_DEBOUNCED) {
+ /* The insertion debounce period has ended. Clear any pending insertion events */
+ socket->events &= ~SS_DETECT;
+ state->flags &= ~SS_DEBOUNCED; /* SS_DEBOUNCED is oneshot */
+ }
yenta_set_power(socket, state);
socket->io_irq = state->io_irq;
bridge = config_readw(socket, CB_BRIDGE_CONTROL) & ~(CB_BRIDGE_CRST | CB_BRIDGE_INTR);
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri Sep 15 2000 - 21:00:18 EST