Typing echo -n mem > /sys/power/state has the same effect as
echo 3 >/proc/acpi/sleep, the computer restarts, and does not enter the state

any chance you can capture what the kernel printed when entering S3?
If you have a serial console, that would work, otherwise we'd
probably have to add an infinite loop down in the code to
prevent the system from actually removing power so you can
see the screen...

Sorry, I don't have a serial console, but I will insert the infinite loop myself
into the kernel source. My guess is to insert into file acpi/hardware/hwsleep.c
into the function acpi_enter_sleep_state() but I can't be sure, so please help me.

I am running the official 2.6.4 kernel.

At the end I tried my idea about inserting the loop into
linux-2.6.4/drivers/acpi/hardware/hwsleep.c
This is a part of the file (FUNCTION: acpi_enter_sleep_state) with two marks of
mine:
...
        /* Write #2: SLP_TYP + SLP_EN */

        ACPI_FLUSH_CPU_CACHE (); /* MARK 1*/

        status = acpi_hw_register_write (ACPI_MTX_DO_NOT_LOCK,
ACPI_REGISTER_PM1A_CONTROL, PM1Acontrol);
        if (ACPI_FAILURE (status)) {
                return_ACPI_STATUS (status);
        }
        /* MARK 2 */
        status = acpi_hw_register_write (ACPI_MTX_DO_NOT_LOCK,
ACPI_REGISTER_PM1B_CONTROL, PM1Bcontrol);
        if (ACPI_FAILURE (status)) {
                return_ACPI_STATUS (status);
        }
...

If I insert the loop after Mark 1, the screen will contains:

Stopping tasks: ===========|
hdd: start_power_step (step:0)
hdd: start_power_step (step:1)
hdd: complete_power_step(step:1, stat:50, err:0)
hdd: completing PM request, suspend
hdc: start_power_step (step:0)
hdc: completing PM request, suspend
hda: start_power_step (step:0)
hda: start_power_step (step:1)
hda: complete_power_step(step:1, stat:50, err:0)
hda: completing PM request, suspend
PM: Entering state.
 hwsleep-0300 [13] acpi_enter_sleep_state: Entering sleep state [S3]

This wasn't a cut and paste, I copied it by hand.
At this point the keyboard, mouse and even the power led is turned off, but the
fans are still working.
If I insert the while(1); after Mark 2 the system is already powered down, and
the loop is not executed.

I tried

echo -n mem /sys/power/state

with kernel 2.6.5
but the result is the same (without the ide debug info of course).

My computer too reboots when I try to resume from S3.
The system goes in standby, but when trying to resume (powerbutton), it reboots.
Log files don't show anything usefull.
  
System : Athlon XP1700 on Asrock K7SX8 mainboard (Sis746FX chipset)
OS : Slackware 9.1 + kernel 2.6.7-rc2

still an issue after 2.6.10-rc2?
(it includes the patch in bug 3390)

if so, does this patch from Hiroshi Itoh help?:
http://marc.theaimsgroup.com/?l=linux-kernel&m=110062740314222&w=2

Yes, still an issue on 2.6.10-rc2.
No, the patch does not help.

My machine: Averatec 3225H laptop, Mandrake 10.0.

same with 2.6.10 final?

Sorry for the delay.
The problem persists in 2.6.10 final, 2.6.11-rc1 and 2.6.11-rc2.

Can you track down if your machine actually enters the resume code?
you coulde add 'jmp .' just under '.code16' of 
arch/i386/kernel/acpi/wakeup.S::wakeup_code, and see if your system is frozen. 
Thanks.

The machine reboots as before - it doesn't reach the wakeup code at all.

Could anybody try the debug patch at Bug 3691? Thanks!

*** Bug 2319 has been marked as a duplicate of this bug. ***

It appears many VIA motherboards have such issue. Bug 3691 has more info, mark 
this one as duplicate.

*** This bug has been marked as a duplicate of 3691 ***