[Qemu-devel] Re: [RFC][PATCH 1/2] machine-specific command line switches

[Jan Kiszka]

Glauber Costa wrote:
> On Sat, May 17, 2008 at 10:34 AM, Jan Kiszka <address@hidden> wrote:
>> For a different project, I once wrote a patch to organize purely
>> machine-specific command line switches under the hood of the respective
>> machine implementations. Now the MusicPal has precisely that need as
>> well. So I reanimated the patch, and here we go:
>>
>> The idea is to add two fields to QEMUMachine and process them:
>>  o options_help - a string that is inserted under a separate section of
>>   the "qemu -h" output.
>>  o parse_option - a callback invoked if a given option was not handled
>>   by the generic code. It returns -1 if the option is unkown, 0 if it
>>   is know but comes without an argument, and 1 when the argument was
>>   consumed.
> 
> This would be quite useful for QEMUAccel too.
> 
> So...
> 
>> +typedef int QEMUMachineParseOption(const char *optname, const char *optarg);
>> +
>>  typedef struct QEMUMachine {
>>     const char *name;
>>     const char *desc;
>>     QEMUMachineInitFunc *init;
>>  #define RAMSIZE_FIXED  (1 << 0)
>>     ram_addr_t ram_require;
>> +    const char *options_help;
>> +    QEMUMachineParseOption *parse_option;
>>     struct QEMUMachine *next;
>>  } QEMUMachine;
> 
> Why don't turn the naming into a more generic one? Maybe QEMUOpts, or
> smth like this.
> 
>> Index: b/vl.c
>> ===================================================================
>> --- a/vl.c
>> +++ b/vl.c
>> @@ -7141,6 +7141,8 @@ static int main_loop(void)
>>
>>  static void help(int exitcode)
>>  {
>> +    QEMUMachine *m;
>> +
>>     printf("QEMU PC emulator version " QEMU_VERSION ", Copyright (c) 
>> 2003-2008 Fabrice Bellard\n"
>>            "usage: %s [options] [disk_image]\n"
>>            "\n"
>> @@ -7275,14 +7277,7 @@ static void help(int exitcode)
>>            "-clock          force the use of the given methods for timer 
>> alarm.\n"
>>            "                To see what timers are available use -clock ?\n"
>>            "-startdate      select initial date of the clock\n"
>> -           "\n"
>> -           "During emulation, the following keys are useful:\n"
>> -           "ctrl-alt-f      toggle full screen\n"
>> -           "ctrl-alt-n      switch to virtual console 'n'\n"
>> -           "ctrl-alt        toggle mouse and keyboard grab\n"
>> -           "\n"
>> -           "When using -nographic, press 'ctrl-a h' to get some help.\n"
>> -           ,
>> +           "\n",
>>            "qemu",
>>            DEFAULT_RAM_SIZE,
>>  #ifndef _WIN32
>> @@ -7291,6 +7286,17 @@ static void help(int exitcode)
>>  #endif
>>            DEFAULT_GDBSTUB_PORT,
>>            "/tmp/qemu.log");
>> +    for (m = first_machine; m != NULL; m = m->next) {
>> +        if (m->options_help)
>> +            printf("Options specific to %s machine:\n%s\n",
>> +                   m->name, m->options_help);
>> +    }
> So, If I understand correctly what you mean here, This will print out
> specific options for every registered machine, right?

Right.

> It does not sound correct, since we won't have support for more than
> one in the same binary anyway. It looks correct, tough, if we think

I was thinking of such examples:

$ arm-softmmu/qemu-system-arm -M ?
Supported machines are:
integratorcp ARM Integrator/CP (ARM926EJ-S) (default)
versatilepb ARM Versatile/PB (ARM926EJ-S)
versatileab ARM Versatile/AB (ARM926EJ-S)
realview   ARM RealView Emulation Baseboard (ARM926EJ-S)
akita      Akita PDA (PXA270)
spitz      Spitz PDA (PXA270)
borzoi     Borzoi PDA (PXA270)
terrier    Terrier PDA (PXA270)
cheetah    Palm Tungsten|E aka. Cheetah PDA (OMAP310)
n800       Nokia N800 tablet aka. RX-34 (OMAP2420)
n810       Nokia N810 tablet aka. RX-44 (OMAP2420)
lm3s811evb Stellaris LM3S811EVB
lm3s6965evb Stellaris LM3S6965EVB
connex     Gumstix Connex (PXA255)
verdex     Gumstix Verdex (PXA270)
mainstone  Mainstone II (PXA27x)
musicpal   Marvell 88w8618 / MusicPal (ARM926EJ-S)

And only the latter one needs a special switch here. Thus the user
should know that this switch is not interpreted if some other machine is
picked.

> that we're not representing 'machines', but anything dumping specific
> options here. But in this case, wouldn't it be better to leave the
> whole help string
> to the user of the interface, instead of just using m->name and m->help?

Nevertheless, I do agree that a non-machine oriented abstraction would
be even nicer in order to organize all those specific options without
the help of increasing #ifdef'ery. However, yet no smart idea came to my
mind to handle all cases (per-arch, per-machine, per-accelerator,
per-whatever - and all this combined).

> 
>> @@ -7673,7 +7679,7 @@ int main(int argc, char **argv)
>>     const char *gdbstub_port;
>>  #endif
>>     uint32_t boot_devices_bitmap = 0;
>> -    int i;
>> +    int i, result;
>>     int snapshot, linux_boot, net_boot;
>>     const char *initrd_filename;
>>     const char *kernel_filename, *kernel_cmdline;
>> @@ -7692,7 +7698,7 @@ int main(int argc, char **argv)
>>     const char *parallel_devices[MAX_PARALLEL_PORTS];
>>     int parallel_device_index;
>>     const char *loadvm = NULL;
>> -    QEMUMachine *machine;
>> +    QEMUMachine *machine, *m;
>>     const char *cpu_model;
>>     const char *usb_devices[MAX_USB_CMDLINE];
>>     int usb_devices_index;
>> @@ -7784,6 +7790,21 @@ int main(int argc, char **argv)
>>             /* Treat --foo the same as -foo.  */
>>             if (r[1] == '-')
>>                 r++;
>> +
>> +            result = -1;
>> +            for (m = first_machine; m != NULL; m = m->next) {
>> +                if (m->parse_option) {
> I don't like this very much. There's no point in having specific
> options without this parse_option anyway. So it would be better to
> check for it before displaying the options at all, and simplify the
> code here.

Don't understand your concern yet. If machine provides options_help, it
is supposed to provide parse_options and vice versa. Thus you don't dump
help about non-existent services - unless someone messes up the machine
definition.

> 
>> +                    result = m->parse_option(r,
>> +                        (optind < argc) ? argv[optind] : NULL);
>> +                    if (result >= 0)
>> +                        break;
>> +                }
>> +            }
>> +            if (result >= 0) {
>> +                optind += result;
>> +                continue;
>> +            }
>> +
> 
> Other than the commented, it all looks very good.

OK, let's try again and do some brainstorming about more flexible option
handling. My initial idea was machine focused (both for the MusicPal and
here @work), thus I stuffed the information into the machine descriptor.
However, my scenario would be fine as well when the machines register
some additional options (in what ever form). The same could be done by
arch-common initialization functions for per-arch options or by an
accelerator initializer for its special switches. Sounds feasible?

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux