On Tue, 8 Feb 2005 16:02:12 -0800
Nishanth Aravamudan <[email protected]> wrote:
> Hi,
>
> The following patchset (affecting USB, Bluetooth, IrDA, Sound, DVB and W1 --
> hence the long To: an Cc: lists...) converts the usb_control_msg() and
> usb_bulk_msg() functions to use milliseconds in the timeout parameter instead
> of jiffies. This is not a problem for almost all cases, as the requested delays are usually quite long wrt. these two functions.
As for w1 - this changes do not break anything, but
I'm not sure that it will not break usb core code which can
depend on jiffies and thus arch specific timings.
So I have a question: is this change intentional and thus by design requires
milliseconds, or it just happens that HZ==1000 on the most of the platforms?
> I see several reasons for committing these changes:
>
> 1) Expressing time in milliseconds makes the code cleaner and easier to
> understand. It also makes the delay independent (from a code perspective) of
> HZ's value, which would be important for dynamic-HZ or tickless systems.
>
> 2) Also for dynamic-HZ/tickless systems, it becomes unclear what exactly HZ
> represents. Milliseconds have meaning in the human world beyond the kernel. This
> ties in pretty strongly with 1)'s point about understanding.
>
> 3) As the time/timer subsystem evolves and changes, the overall message I am
> seeing is that jiffies are not a proper expression of time. Thus, I am hoping
> these patches minimize this particular interface's usage of jiffies in this
> way; while, in the end, the call to usb_start_wait_for_urb ends up with the
> same jiffies-based time, if the timer subsystem were to, for instance, convert
> to nanoseconds, we could instead pass on the parameter multiplied by a million.
> It simply makes adjusting to any new time system that much easier.
>
> Exceptions do exist, of course, and the following files do not use a
> HZ-relative timeout variable; thus I was unsure how exactly to proceed.
> Given that HZ==1000 in i386, the constants used technically correspond
> directly to milliseconds. I am unsure if that is the intent of the author
> in all cases, though. I left the hard-coded non-HZ related values alone in all
> cases.
>
> drivers/isdn/hisax/hfc_usb.c
> drivers/usb/media/dabusb.c
> drivers/usb/media/dsbr100.c
> drivers/usb/media/stv680.c
> drivers/usb/media/w9968cf.c
> drivers/usb/misc/emi26.c
> drivers/usb/misc/emi62.c
> drivers/usb/serial/cypress_m8.c
> drivers/usb/serial/ftdi_sio.c
> drivers/usb/serial/io_edgeport.c
> drivers/usb/serial/kobil_sct.c
> drivers/usb/serial/pl2303.c
> drivers/media/dvb/dibusb/dvb-dibusb-usb.c
> drivers/usb/media/dabusb.c
> drivers/w1/dscore.c
>
> As always, any comments or suggestions are greatly appreciated.
>
> -Nish
Evgeniy Polyakov
Only failure makes us experts. -- Theo de Raadt
On Wed, Feb 09, 2005 at 10:41:44PM +0300, Evgeniy Polyakov wrote:
> On Tue, 8 Feb 2005 16:02:12 -0800
> Nishanth Aravamudan <[email protected]> wrote:
>
> > Hi,
> >
> > The following patchset (affecting USB, Bluetooth, IrDA, Sound, DVB and W1 --
> > hence the long To: an Cc: lists...) converts the usb_control_msg() and
> > usb_bulk_msg() functions to use milliseconds in the timeout parameter instead
> > of jiffies. This is not a problem for almost all cases, as the requested delays are usually quite long wrt. these two functions.
>
> As for w1 - this changes do not break anything, but
> I'm not sure that it will not break usb core code which can
> depend on jiffies and thus arch specific timings.
>
> So I have a question: is this change intentional and thus by design requires
> milliseconds, or it just happens that HZ==1000 on the most of the platforms?
So the first check I did, at Greg's suggestion, was to verify the timing
units for usbdevfs, which are milliseconds. So we should be ok there.
Second, all of these patches don't change the fact that jiffies are
used. They only move the usage of jiffies further into the usbcore, thus
making the actual API be in human time-units and relying on correct
translation from HZ-independent units (milliseconds) to HZ-dependent
units (jiffies) in one place.
So the change is very much intentional, but it's an API change only. The
timeout parameter is not even used until usb_start_wait_urb() and
presuming msecs_to_jiffies() does what we want (and I've verified it
does for the most common cases), then the timer in usb_start_wait_urb()
will be set for the same number of jiffies.
Really, this just makes it easier for me later on to go through and
re-audit timer usage :) If we do change the timer core (and I'm working
on figuring out how exactly to do so), then only changing in one place
down the road is better than all of these patches going out then *along
with* a huge rework of core kernel code...
So, in conclusion, these patches shouldn't break anything for anyone, if
I got my conversions correct. Most drivers were passing HZ * __some_var
or HZ / 10, anyways, which is expressible easily in human time units via
1000 * __some_var or 100, respectively. So it shouldn't change anything,
except in what units we are dealing.
Let me know if you see anything that points to the contrary (in USB core
code, or anywhere else).
Thanks for the feedback!
-Nish