Re: [Qemu-devel] Re: [RFC][PATCH 4/4] Add support for Marvell 88w8618 / MusicPal

[andrzej zaborowski]

On 20/04/2008, Jan Kiszka <address@hidden> wrote:
> andrzej zaborowski wrote:
>  > On 19/04/2008, Jan Kiszka <address@hidden> wrote:
>  >> andrzej zaborowski wrote:
>  >>  > On 18/04/2008, Jan Kiszka <address@hidden> wrote:
>  >>  >>  Andrzej, as you have written the wm8750, do you already know where 
> which
>  >>  >>  volume level would have to be applied (open-coded or via some
>  >>  >>  AUD_set_volume)? I'm currently only using LOUT2VOL, and I'm a bit 
> lazy
>  >>  >>  to study the datasheet /wrt all the mixer details.
>  >>  >
>  >>  > My idea was to open
>  >>  > http://www.wolfsonmicro.com/uploads/documents/en/WM8750.pdf and on the
>  >>  > first page every Wolfson datasheet has its diagram of all audio paths
>  >>  > (of which there are always too many) and then trace with my finger the
>  >>  > path between the source (the I2C or I2S interfaces) and the sink (the
>  >>  > analog output), and then multiply all the volume values that are
>  >>  > applied there (both analog and digital) and pass that to host mixer
>  >>  > through some functions in audio/ for the given SWVoice - but we don't
>  >>  > have any such functions and I'm ok with using the host mixer manually.
>  >>  >  (VirtualBox has them implemented iirc)  So yes, maybe it makes sense
>  >>  > to multiply the samples for the moment and use only LOUTnVOL /
>  >>  > ROUTnVOL values as these are used by the guests we're interested in.
>  >>
>  >>
>  >> Done, and it finally works. One of the two quirks I found in wm8750 made
>  >>  the switch a bit hairy. Patches will follow.
>  >
>  > Thanks.  I pushed the patch with fixes.  Regarding the wm8750_fini
>  > patch, I'll #if 0 it because it's possible that a board will have this
>  > chip on something hotpluggable and will need to create and destroy it
>  > various times and it's easy to miss something in the clean-up.
>  > Regarding the volume patch, I'll make a look-up table at one point,
>
> Don't understand yet why (are you afraid of pow, libm, or float in
>  general?), but if it helps to get things merged... ;)

I think the idea (not mine) was to not depend on it if not necessary,
like in this case, pow() is a rather heavy tool for the simple issue.

>
>
>  > and then merge.  Also, if we have 16-bit data and 7-bit volume scale
>  > maybe we're fine with scalling only the most-significant-byte and
>
>
> Hmm, wasn't endianness about finding out which byte is most-significant
>  and which not? :->

Yes, the thing is that WM8750's input endianness is known and the MSB
is always the same one.  The host's endianness varies and to do the
multiplication in C we need the samples to be host-endian (but audio
functions do accept guest-endian).

>
>
>  > avoiding endianness headaches (or maybe not).  Nevertheless the
>  > MusicPal emulator should be bootable without that.
>  >
>  >>
>  >>  >
>  >>  >>
>  >>  >>  >>>   - 128×64 display with brightness control
>  >>  >>  >>>   - all input buttons
>  >>  >>  >>>
>  >>  >>  >>>  Using up to 32 MB flash, I hit a limit /wrt phys_ram_size. I 
> worked
>  >>  >>  >>>  around this for now by extending MAX_BIOS_SIZE to 32 MB, surely 
> not a
>  >>  >>  >>>  nice solution.
>  >>  >>  >> You can use -m 150 or similar.
>  >>  >>  >>
>  >>  >>  >> Please also format the code similarly to rest of Qemu.
>  >>  >>  >
>  >>  >>  > That would just increase ram_size, thus won't help as I need memory
>  >>  >>  > beyond it (here for the pflash in R/W mode).
>  >>  >
>  >>  > Yes, I had not looked at how ram_size was used in the musicpal board
>  >>  > initialisation, sorry.
>  >>  >
>  >>  >>
>  >>  >> OK, I see what you mean after looking at your N800 patches: You apply 
> a
>  >>  >>  fixed RAM size, leaving the rest of what the user provided via -m to
>  >>  >>  SRAM and flash. Not optimal IMHO, you may sometimes also want to play
>  >>  >>  with the RAM size even if the real devices has a fixed amount. And 
> it is
>  >>  >>  far from being intuitive as well.
>  >>  >
>  >>  > Yes, although you allow the user to set also a smaller RAM than what
>  >>  > the virtual machine expects.
>  >>
>  >>
>  >> That's indeed something the machine should take of (if there are such
>  >>  hard limits).
>  >>
>  >>
>  >>  >
>  >>  >>  The only true solution I see right now is moving qemu_vmalloc into 
> the
>  >>  >>  machine initialization code. Is there anything between current
>  >>  >>  qemu_vmalloc and machine->init that relies on phys_ram_base being 
> valid
>  >>  >>  (and which can't be moved after the machine init) and thus prevents 
> this?
>  >>  >
>  >>  > I had a different idea: add a field ram_constraint in struct
>  >>  > QEMUMachine, which would hold the amount of RAM the machine always
>  >>  > needs (e.g. bios and video RAM), and the low bit could hold a flag
>  >>  > RAM_SIZE_FIXED for machines that have only such RAM (basically the
>  >>  > criteria should be whether it's possible for the guest to detect the
>  >>  > memory size there is on board - on machines like Spitz there's no way)
>  >>
>  >>
>  >> IIRC, embedded boards let the boot loader "detect" this. I see valid
>  >>  scenarios where one wants to play with different sizes and may therefore
>  >>  patch U-Boot - or load the kernel directly which should make QEMU set
>  >>  the related ATAG field appropriately, no?
>  >
>  > Yes, in case of a standard firmware like Linux or U-boot - but we
>  > probably don't need to provide options for everything one may want to
>  > play with unless it's a valid hardware configuration (like in the PC
>  > case where you can add and take away RAM sticks), at some point the
>  > user needs to edit the source either way.
>  >
>  > Anyway almost half of the boards in qemu ignored ram_size until now
>  > and risked the provided size being too low and segfaulting, so with
>  > the patch I sent in another mail at least there's a check, and the
>  > check is only done once for all boards so it can be removed from the
>  > few boards that did it.
>  >
>  >>
>  >>  > and for such machines the -m parameter would be invalid.  I'll try to
>  >>  > come up with a patch.
>  >>
>  >>
>  >> I originally had the same idea but I dropped it because it would still
>  >>  overload -m with semantics that don't belong there. IMHO -m should only
>  >>  describe the main RAM size, not any additionally by QEMU required memory
>  >>  for establishing fixed SRAM or even for backing up flash devices. That's
>  >>  at least what I would expect from this switch and what the documentation
>  >>  suggests as well so far.
>  >
>  > This property is not changed by the patch (I hope).
>
> Yes, it restores the original semantic, at least as long as
>  RAMSIZE_FIXED is not set. That case is still a bit suboptimal as you
>  have to provide pessimistic values, e.g. the maximum flash size that can
>  be used. But I can live with it I guess.

You absolutely don't have to set RAMSIZE_FIXED if you want to use -m.
In the board init you can for example decide to have fixed ram size
and flash size decided by ram_size if that's what you need.  As long
as there's only one -m allowed one of the sizes has to be hardcoded.

Thanks for the rework.
-- 
Please do not print this email unless absolutely necessary. Spread
environmental awareness.