Re: [PATCH v2] sysinfo: include availram field in sysinfo struct

[Pintu Agarwal]

On Sun, 9 Jan 2022 at 04:05, David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> From: Pintu Agarwal
> > Sent: 08 January 2022 16:53
> >
> > On Sat, 8 Jan 2022 at 03:52, David Laight <David.Laight@xxxxxxxxxx> wrote:
> > >
> > > From: Pintu Kumar
> > > > Sent: 07 January 2022 18:08
> > > >
> > > > The sysinfo member does not have any "available ram" field and
> > > > the bufferram field is not much helpful either, to get a rough
> > > > estimate of available ram needed for allocation.
> > > >
> > > > One needs to parse MemAvailable field separately from /proc/meminfo
> > > > to get this info instead of directly getting if from sysinfo itself.
> > > >
> > > > Thus, this patch introduce a new field as availram in sysinfo
> > > > so that all the info total/free/available can be retrieved from
> > > > one place itself.
> > > >
> > > ...
> > > > diff --git a/include/uapi/linux/sysinfo.h b/include/uapi/linux/sysinfo.h
> > > > index 435d5c2..fe84c6a 100644
> > > > --- a/include/uapi/linux/sysinfo.h
> > > > +++ b/include/uapi/linux/sysinfo.h
> > > > @@ -19,7 +19,8 @@ struct sysinfo {
> > > >       __kernel_ulong_t totalhigh;     /* Total high memory size */
> > > >       __kernel_ulong_t freehigh;      /* Available high memory size */
> > > >       __u32 mem_unit;                 /* Memory unit size in bytes */
> > > > -     char _f[20-2*sizeof(__kernel_ulong_t)-sizeof(__u32)];   /* Padding: libc5 uses this.. */
> > >
> > > There are 4 pad bytes here on most 64bit architectures.
> > >
> > > > +     __kernel_ulong_t availram;      /* Memory available for allocation */
> > > > +     char _f[20-3*sizeof(__kernel_ulong_t)-sizeof(__u32)];   /* Padding: libc5 uses this.. */
> > > >  };
> > >
> > > You've not compile-time tested the size of the structure.
> > >
> > With "32" instead of "20" in padding I get these size of sysinfo:
> > In x86-64 kernel, with app 64-bit: Size of sysinfo = 128
> > In x86-64 kernel, with app 32-bit:: Size of sysinfo = 76
> > In arm-64 kernel, with app 32-bit: Size of sysinfo = 76
>
> You need to compare the sizes before and after your patch
> to ensure it doesn't change on any architecture.

Without the changes:
On 32-bit, the size of structure = 64
On 64-bit, the size of structure = 112

With the addition of my new field (availram) if I try to fix the size
issue on one arch, the other arch gets disturbed.
I could fix the same size issue on 64-bit with below changes:

        __kernel_ulong_t freeswap;      /* swap space still available */
        __u16 procs;                    /* Number of current processes */
        __u16 pad;                      /* Explicit padding for m68k */
+       __u32 mem_unit;                 /* Memory unit size in bytes
*/        ============> Move the mem_unit up to adjust the padding
        __kernel_ulong_t totalhigh;     /* Total high memory size */
        __kernel_ulong_t freehigh;      /* Available high memory size */
-       __u32 mem_unit;                 /* Memory unit size in bytes */
+       __kernel_ulong_t availram;      /* Memory available for
allocation */   ========> Add the new field here
-        char _f[20-2*sizeof(__kernel_ulong_t)-sizeof(__u32)];   /*
Padding: libc5 uses this.. */
+       char _f[28-3*sizeof(__kernel_ulong_t)-sizeof(__u32)];   /*
Padding: libc5 uses this.. */   ====> Increase the size to 28 (thus _f
becomes 0 like original)
+       //char _f[4];
 };

Output with 64-bit build:
$ gcc test-sysinfo.c ; ./a.out
Total RAM: 32715804 kB
Free RAM: 1111296 kB
Size of sysinfo = 112
Size of sysinfo uptime = 8
Size of sysinfo loads = 24
Size of sysinfo totalram = 8
Size of sysinfo pad = 2
Size of sysinfo memunit = 4
Size of sysinfo _f = 0

Output with 32-bit build:
$ gcc test-sysinfo.c -m32 ; ./a.out
Total RAM: 7987 kB
Free RAM: 271 kB
Size of sysinfo = 72
Size of sysinfo uptime = 4
Size of sysinfo loads = 12
Size of sysinfo totalram = 4
Size of sysinfo pad = 2
Size of sysinfo memunit = 4
Size of sysinfo _f = 12

Are there any more suggestions/ideas to experiment with padding
changes before we give-up ?
Can we handle it using : __arch64__ check ?
Or, the only option is to add one more, say: sysinfo64 ?


> > Okay the sys robot reported some issue in 64-bit build.
> > {{{
> > >> include/uapi/linux/sysinfo.h:23:14: error: size of array '_f' is too large
> > >>    23 |         char _f[20-3*sizeof(__kernel_ulong_t)-sizeof(__u32)];   /* Padding: libc5 uses
> > this.. */
> > >>       |              ^~
> > }}}
> >
> > Also, I got the same issue while building for arm64, so I tried to
> > adjust like this:
> > char _f[32-3*sizeof(__kernel_ulong_t)-sizeof(__u32)];
> >
> > With this the build works on both 32/64 but output fails when running
> > 32-bit program on 64-bit kernel.
> > Also, the free command on 64-bit reports "stack smashing error"..
> >
> > How do we resolve this issue to make it work on both arch ?
> > Also, I don't really understand the significance of that number "20"
> > in padding ?
>
> My guess is that someone added a char _f[20] pad to allow for expansion.
> Then two __kernel_ulong_t and one __u32 field were added, so the
> size of the pad was reduced.
> When __kernel_ulong_t is 64bit then it seems to be char _f[0]
> - which might generate a compile warning since you are supposed
> to use char _f[] to indicate an extensible structure.
> There is, however, 4 bytes of pad after the _f[] on most 64bit
> architectures.
>
Thanks, yes even I guessed the same.

> So actually there isn't enough space to anything useful at all.
>
Is this problem does not exist in other UAPI structures ?
Seems like nothing can be done to allow future expansion without
breaking existing things and without developing the new one.

Thanks,
Pintu