Re: [Qemu-devel] [PATCH 2/2] vl: convert -m to QemuOpts

[Igor Mammedov]

On Mon, 10 Feb 2014 18:38:59 +0100
Laszlo Ersek <address@hidden> wrote:

> comments below
> 
> On 02/06/14 09:16, Igor Mammedov wrote:
> > Adds option to -m
> >  "mem" - startup memory amount
> > 
> > For compatibility with legacy CLI if suffix-less number is passed,
> > it assumes amount in Mb.
> > 
> > Otherwise user is free to use suffixed number using suffixes b,k/K,M,G
> > 
> > Signed-off-by: Igor Mammedov <address@hidden>
> > Signed-off-by: Paolo Bonzini <address@hidden>
> > ---
> >  qemu-options.hx |    7 +++++--
> >  vl.c            |   53 
> > ++++++++++++++++++++++++++++++++++++++++++-----------
> >  2 files changed, 47 insertions(+), 13 deletions(-)
> > 
> > diff --git a/qemu-options.hx b/qemu-options.hx
> > index 56e5fdf..4d7ef52 100644
> > --- a/qemu-options.hx
> > +++ b/qemu-options.hx
> > @@ -210,8 +210,11 @@ use is discouraged as it may be removed from future 
> > versions.
> >  ETEXI
> >  
> >  DEF("m", HAS_ARG, QEMU_OPTION_m,
> > -    "-m megs         set virtual RAM size to megs MB [default="
> > -    stringify(DEFAULT_RAM_SIZE) "]\n", QEMU_ARCH_ALL)
> > +    "-m [mem=]megs\n"
> > +    "                configure guest RAM\n"
> 
> "configure guest RAM size"
fixed

> 
> > +    "                mem: initial amount of guest memory (default: "
> > +    stringify(DEFAULT_RAM_SIZE) "Mb)\n",
> 
> I wonder if it should rather say "MB" -- small "b" has this "bits"
> connotation for me. But I could be wrong.
fixed

> 
> Also, again, I believe explaining the default used to mean something
> else, but I'm OK with that part as-is.
> 
> > +    QEMU_ARCH_ALL)
> >  STEXI
> >  @item -m @var{megs}
> >  @findex -m
> > diff --git a/vl.c b/vl.c
> > index 7f2595c..fe5dae3 100644
> > --- a/vl.c
> > +++ b/vl.c
> > @@ -532,6 +532,20 @@ static QemuOptsList qemu_msg_opts = {
> >      },
> >  };
> 
> (this could conflict with Alan's series -- modifies the same spot)
I'll put it in the beginning of list to avoid conflict.

> > +static QemuOptsList qemu_mem_opts = {
> > +    .name = "memory-opts",
> > +    .implied_opt_name = "mem",
> > +    .head = QTAILQ_HEAD_INITIALIZER(qemu_mem_opts.head),
> > +    .merge_lists = true,
> 
> OK, so we've set merge_list to true here as well, same as for "machine".
> Further support for simplifying qemu_find_opts_singleton(); see patch #1.
I believe it's not good for generic function to depend on merge_lists.

> 
> > +    .desc = {
> > +        {
> > +            .name = "mem",
> > +            .type = QEMU_OPT_SIZE,
> 
> QEMU_OPT_SIZE implies (in parse_option_size()) that "no suffix" means
> "unit==byte" (); I'll check lower down how that's solved.
it's(i.e. legacy behavior) taken care of later at -m parsing time

> 
> > +        },
> > +        { /* end of list */ }
> > +    },
> > +};
> > +
> >  /**
> >   * Get machine options
> >   *
> > @@ -2868,6 +2882,7 @@ int main(int argc, char **argv, char **envp)
> >      };
> >      const char *trace_events = NULL;
> >      const char *trace_file = NULL;
> > +    const ram_addr_t default_ram_size = DEFAULT_RAM_SIZE * 1024 * 1024;
> 
> I'd feel safer if the multiplications were done in ram_addr_t. Currently
> they are done in "int". It's unlikely that we'll make 2GB+ the default
> ram size, but still.
fixed

> 
> >  
> >      atexit(qemu_run_exit_notifiers);
> >      error_set_progname(argv[0]);
> > @@ -2906,6 +2921,7 @@ int main(int argc, char **argv, char **envp)
> >      qemu_add_opts(&qemu_tpmdev_opts);
> >      qemu_add_opts(&qemu_realtime_opts);
> >      qemu_add_opts(&qemu_msg_opts);
> > +    qemu_add_opts(&qemu_mem_opts);
> >  
> >      runstate_init();
> >  
> > @@ -2921,7 +2937,7 @@ int main(int argc, char **argv, char **envp)
> >      module_call_init(MODULE_INIT_MACHINE);
> >      machine = find_default_machine();
> >      cpu_model = NULL;
> > -    ram_size = 0;
> > +    ram_size = default_ram_size;
> >      snapshot = 0;
> >      cyls = heads = secs = 0;
> >      translation = BIOS_ATA_TRANSLATION_AUTO;
> > @@ -3198,16 +3214,32 @@ int main(int argc, char **argv, char **envp)
> >                  exit(0);
> >                  break;
> >              case QEMU_OPTION_m: {
> > -                int64_t value;
> >                  uint64_t sz;
> > -                char *end;
> > +                const char *mem_str;
> >  
> > -                value = strtosz(optarg, &end);
> > -                if (value < 0 || *end) {
> > -                    fprintf(stderr, "qemu: invalid ram size: %s\n", 
> > optarg);
> > +                opts = qemu_opts_parse(qemu_find_opts("memory-opts"),
> > +                                       optarg, 1);
> 
> This can set "opts" to NULL if parsing fails, and then the
> qemu_opt_get() just below will SIGSEGV. You need to check if "opts"
> becomes NULL here, and exit if so (see other calls to qemu_opts_parse()
> in main()).
fixed

> 
> In particular, see commit f46e720a.
> 
> Also, unfortunately, this conversion kind of relaxes the error checking
> that happens during parsing. The pre-patch version ends up in
> strtosz_suffix_unit(), which rejects the empty string, for example. The
> new version, which ends up in parse_option_size(), is not that smart
> about strtod(). I think it will simply return zero for
> 
>   -m mem=""
> 
> However it's not the fault of this patch.
parse_option_size() does return 0, but following code fixes it,
look for comment: /* backward compatibility behavior for case "-m 0" */
but as you pointed below empty mem option causes undefined behavior later
so fix by checking that string is not empty.

> 
> > +
> > +                mem_str = qemu_opt_get(opts, "mem");
> > +                if (!mem_str) {
> > +                    fprintf(stderr, "qemu: invalid -m option, missing "
> > +                            " 'mem' option\n");
> 
> Double space. (There's one at the end of the first string literal, and
> another at the beginning of the second literal.)
fixed

> 
> >                      exit(1);
> >                  }
> > -                sz = QEMU_ALIGN_UP((uint64_t)value, 8192);
> > +
> > +                sz = qemu_opt_get_size(opts, "mem", ram_size);
> > +
> > +                /* Fix up legacy suffix-less format */
> > +                if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
> 
> Undefined behavior if mem_str is the emptry string. (I think it is
> possible, but I didn't test it.)
indeed, fixed.

> 
> > +                    sz <<= 20;
> > +                }
> 
> We could check for overflow here, if we wanted.
fixed

> 
> > +
> > +                /* backward compatibility behaviour for case "-m 0" */
> > +                if (sz == 0) {
> > +                    sz = default_ram_size;
> > +                }
> > +
> > +                sz = QEMU_ALIGN_UP(sz, 8192);
> >                  ram_size = sz;
> >                  if (ram_size != sz) {
> >                      fprintf(stderr, "qemu: ram size too large\n");
> > @@ -4056,10 +4088,9 @@ int main(int argc, char **argv, char **envp)
> >          exit(1);
> >      }
> >  
> > -    /* init the memory */
> > -    if (ram_size == 0) {
> > -        ram_size = DEFAULT_RAM_SIZE * 1024 * 1024;
> > -    }
> > +    /* store value for the future use */
> > +    qemu_opt_set_number(qemu_find_opts_singleton("memory-opts"),
> > +                        "mem", ram_size);
> 
> Slight possibility here to overflow the int64_t "val" parameter with the
> potentially uint64_t "ram_size" argument. I guess we don't care.
yep, everywhere in API *_number is treated as uint64_t and only
qemu_opt_set_number() treats it as int64_t, perhaps this function should be
fixed.

> 
> Also, I wonder what happens when we have passed a non-default memory
> size on the command line. In that case, qemu_opt_set_number() seems to
> create a second QemuOpt here. I guess that's maybe expected though?
which appends it to the tail, and following lookup finds it first.

> 
> >  
> >      if (qemu_opts_foreach(qemu_find_opts("device"), device_help_func, 
> > NULL, 0)
> >          != 0) {
> > 
> 
> It's your call what you'd like to address from the above.
> 
> Thanks
> Laszlo