Syslinux - Mar 2016 - "Tick-counting" vs "Tick-less" timekeeping issues on VMs emulating BIOS PCs

On 03/01/16 21:11, H. Peter Anvin via Syslinux wrote:
> On 03/01/16 07:21, Patrick Masotta via Syslinux wrote:
>>
>> At the moment I'm seeing timing issues on TFTP transfers (lwIP
depends
>> on the new interrupt based timer). I have consistently detected 
>>
>> multiple Requests, double ACKs, etc. All these problems seem to be
>>
>> sourced on the now unreliable timeouts.
>> I have used TFTP on the same (BIOS VM) scenarios with 3.X and TFTP 
>>
>> timing was always rock solid.
>>
>>
>> Sorry if this was discussed before, if not I'd like to ask (Peter)
if
>> there is any reason preventing us from going back to the
"Tick-less"
>> approach. Thanks.
>>
> 
> There are quite a few; primarily we simply cannot make lwIP
> (lpxelinux.0) work without it.  However, pxelinux.0 should not depend on
> this.
> 
Now, tickful timekeeping is wasteful, but this is a bootloader, and
functionality is the main concern.  The "tickless" version still
depended on the BIOS tick.

	-hpa

> 
> There are quite a few; primarily we simply cannot make lwIP
> (lpxelinux.0) work without it.  However, pxelinux.0 should not depend on
> this.

I'm finding most of the problems on lpxelinux.0 
>
>Now, tickful timekeeping is wasteful, but this is a bootloader, and
>functionality is the main concern.  The "tickless" version still
>depended on the BIOS tick.
>
Not really, a virtual environment can easily emulate the BIOS_timer	at 0x046C
but it has problems emulating a "real" timer interrupt to be hooked...
That's the real problem.

Just testing I have replaced the timing functions 


#define BIOS_timer	0x046C	// Timer ticks 
typedef uint32_t jiffies_t;
extern volatile jiffies_t __jiffies, __ms_timer;
static inline jiffies_t jiffies(void)
{
return (*(volatile uint32_t *)BIOS_timer);

}

typedef uint32_t mstime_t;
typedef int32_t  mstimediff_t;
static inline mstime_t ms_timer(void)
{
return ((*(volatile uint32_t *)BIOS_timer) * 55);

}

lpxelinux.0 hangs while legacy works OK.

I think we should not depend on hoking timer interrupts if we want to be
compatible with virtual environments.

Best,
Patrick

On 02/03/16 22:47, Patrick Masotta via Syslinux wrote:
> Not really, a virtual environment can easily emulate the BIOS_timer	at
0x046C
> but it has problems emulating a "real" timer interrupt to be
hooked...
> That's the real problem.
The timer interrupt works fine in at least KVM, Xen, VMware, and 
Hyper-V.  (I've tested iPXE in all of those virtual environments, and 
iPXE relies on the timer interrupt actually triggering a call to an ISR 
within the VM.)

Michael