Re: [PATCH 1/2] log: Add separate debug option for logging invalid memory accesses

[Peter Xu]

On Mon, Feb 13, 2023 at 07:34:55PM +0100, BALATON Zoltan wrote:
> On Mon, 13 Feb 2023, Peter Xu wrote:
> > On Mon, Feb 13, 2023 at 05:34:04PM +0100, BALATON Zoltan wrote:
> > > On Mon, 13 Feb 2023, Peter Xu wrote:
> > > > On Mon, Feb 13, 2023 at 03:47:42PM +0100, BALATON Zoltan wrote:
> > > > > On Mon, 13 Feb 2023, Peter Xu wrote:
> > > > > > On Mon, Feb 13, 2023 at 12:41:29PM +0100, Thomas Huth wrote:
> > > > > > > On 07/02/2023 17.33, BALATON Zoltan wrote:
> > > > > > > > On Tue, 31 Jan 2023, BALATON Zoltan wrote:
> > > > > > > > > On Thu, 19 Jan 2023, BALATON Zoltan wrote:
> > > > > > > > > > Currently -d guest_errors enables logging of different 
> > > > > > > > > > invalid actions
> > > > > > > > > > by the guest such as misusing hardware, accessing missing 
> > > > > > > > > > features or
> > > > > > > > > > invalid memory areas. The memory access logging can be 
> > > > > > > > > > quite verbose
> > > > > > > > > > which obscures the other messages enabled by this debug 
> > > > > > > > > > switch so
> > > > > > > > > > separate it by adding a new -d memaccess option to make it 
> > > > > > > > > > possible to
> > > > > > > > > > control it independently of other guest error logs.
> > > > > > > > > > 
> > > > > > > > > > Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu>
> > > > > > > > > 
> > > > > > > > > Ping? Could somebody review and pick it up please?
> > > > > > > > 
> > > > > > > > Ping?
> > > > > > > 
> > > > > > > Patch makes sense to me and looks fine, so:
> > > > > > > 
> > > > > > > Reviewed-by: Thomas Huth <thuth@redhat.com>
> > > > > > > 
> > > > > > > ... I think this should go via one of the "Memory API" 
> > > > > > > maintainers branches?
> > > > > > > Paolo? Peter? David?
> > > > > > 
> > > > > > Paolo normally does the pull, I assume that'll still be the case.  
> > > > > > The
> > > > > > patch looks good to me if Phil's comment will be addressed on 
> > > > > > merging with
> > > > > > the old mask, which makes sense to me:
> > > > > 
> > > > > Keeping the old mask kind of defies the purpose. I've tried to 
> > > > > explain that
> > > > > in the commit message but now that two of you did not get it maybe 
> > > > > that
> > > > > message needs to be clarified instead?
> > > > 
> > > > I think it's clear enough.  My fault to not read carefully into the
> > > > message, sorry.
> > > > 
> > > > However, could you explain why a memory_region_access_valid() failure
> > > > shouldn't belong to LOG_GUEST_ERROR?
> > > > 
> > > > commit e54eba1986f6c4bac2951e7f90a849cd842e25e4
> > > > Author: Peter Maydell <peter.maydell@linaro.org>
> > > > Date:   Thu Oct 18 14:11:35 2012 +0100
> > > > 
> > > >    qemu-log: Add new log category for guest bugs
> > > > 
> > > >    Add a new category for device models to log guest behaviour
> > > >    which is likely to be a guest bug of some kind (accessing
> > > >    nonexistent registers, reading 32 bit wide registers with
> > > >    a byte access, etc). Making this its own log category allows
> > > >    those who care (mostly guest OS authors) to see the complaints
> > > >    without bothering most users.
> > > > 
> > > > Such an illegal memory access is definitely a suitable candidate of 
> > > > guest
> > > > misbehave to me.
> > > 
> > > Problem is that a lot of machines have unimplemented hardware that are 
> > > valid
> > > on real machine but we don't model them so running guests which access 
> > > these
> > > generate constant flow of unassigned memory access log which obscures the
> > > actual guest_errors when an modelled device is accessed in unexpected 
> > > ways.
> > > For an example you can try booting MorphOS on mac99,via=pmu as described
> > > here: http://zero.eik.bme.hu/~balaton/qemu/amiga/#morphos
> > > (or the pegasos2 command too). We could add dummy registers to silence 
> > > these
> > > but I think it's better to either implement it correctly or leave it
> > > unimplemented so we don't hide errors by the dummy implementation.
> > > 
> > > > Not to mention Phil always have a good point that you may be violating
> > > > others using guest_error already so what they wanted to capture can
> > > > misterious going away without noticing, even if it may service your 
> > > > goal.
> > > > IOW it's a slight ABI and I think we ned justification to break it.
> > > 
> > > Probably this should be documented in changelog or do we need depracation
> > > for a debug option meant for developers mostly? I did not think so. Also I
> > > can't think of other way to solve this without changing what guest_erorrs 
> > > do
> > > unless we change the name of that flag as well. Also not that when this 
> > > was
> > > originally added it did not contain mem access logs as those were 
> > > controlled
> > > by a define in memory.c until Philippe changed it and added them to
> > > guest_errors. So in a way I want the previous functionality back.
> > 
> > I see, thanks.
> > 
> > Indeed it's only a debug option, so I don't know whether the abi needs the
> > attention here.
> > 
> > I quickly looked at all the masks and afaict this is really a special and
> > very useful one that if I'm a cloud provider I can run some script trying
> > to capture those violations using this bit to identify suspecious guests.
> > 
> > So I think it would still be great to not break it if possible, IMHO.
> > 
> > Since currently I don't see an immediate limitation of having qemu log mask
> > being a single bit for each of the entry, one way to satisfy your need (and
> > also keep the old behavior, iiuc), is to make guest_errors a sugar syntax
> > to cover 2 bits.  It shouldn't be complicated at all, I assume:
> > 
> > +/* This covers the generic guest errors besides memory violations */
> > #define LOG_GUEST_ERROR    (1 << 11)
> > 
> > +/*
> > + * This covers the guest errors on memory violations; see LOG_GUEST_ERROR
> > + * for generic guest errors.
> > + */
> > +#define LOG_GUEST_ERROR_MEM      (1 << 21)
> > +#define LOG_GUEST_ERROR_ALL      (LOG_GUEST_ERROR | LOG_GUEST_ERROR_MEM)
> > 
> > -    { LOG_GUEST_ERROR, "guest_errors",
> > +    { LOG_GUEST_ERROR_ALL, "guest_errors",
> > 
> > Then somehow squashed with your changes.  It'll make "guest_errors" not
> > exactly matching the name of LOG_* but I think it may not be a big concern.
> 
> I'm not sure I understand this. So -d memaccess would give me the unassigned
> logs, that's fine and -d guest_errors are both LOG_GUEST_ERROR and memaccess
> like currently but what option would give me just the guest_Errors before
> mem access started to use this flag too? (I could not locate the commit that
> changed this but I remember previously the unassigned mem logs were enabled
> with a define in memory.c.) Do we need another -d option for just the guest
> errors then? What should that be called?

I forgot to add those two definitions into qemu_log_items just now.  It can
be defined as:

  - "guest_errors_common" for !mem errors
  - "guest_errors_mem" for mem errors
  - "guest_errors" for mem+!mem (compatible to the old code)

With the two lines added:

-    { LOG_GUEST_ERROR, "guest_errors",
+    { LOG_GUEST_ERROR_ALL, "guest_errors",
       "log when the guest OS does something invalid (eg accessing a\n"
       "non-existent register)" },
+    { LOG_GUEST_ERROR, "guest_errors_common", "..." },
+    { LOG_GUEST_ERROR_MEM, "guest_errors_mem", "..." },

I saw that Phil revoked his concern, I don't have a strong opinion
personally, assuming Phil knows better on that since he modified the memory
loggings before.  If all are happy with this, please proceed with either
way.

Thanks,

-- 
Peter Xu