Re: [XEN][PATCH v4] xen/x86: guest_access: optimize raw_x_guest() for PV and HVM combinations

[Jan Beulich <jbeulich@xxxxxxxx>]

On 18.11.2025 14:08, Grygorii Strashko wrote:
> On 17.11.25 18:43, Jan Beulich wrote:
>> On 14.11.2025 15:01, Grygorii Strashko wrote:
>>> --- a/xen/arch/x86/pv/Makefile
>>> +++ b/xen/arch/x86/pv/Makefile
>>> @@ -14,6 +14,10 @@ obj-y += ro-page-fault.o
>>>   obj-$(CONFIG_PV_SHIM) += shim.o
>>>   obj-$(CONFIG_TRACEBUFFER) += trace.o
>>>   obj-y += traps.o
>>> +obj-$(CONFIG_PV) += usercopy.o
>>
>> Just obj-y with the movement.
>>
>> However, is the movement (and was the adding of $(CONFIG_PV) in the earlier
>> version) actually correct? The file also produces copy_{from,to}_unsafe_ll(),
>> which aren't PV-specific. This may be only a latent issue right now, as we
>> have only a single use site of copy_from_unsafe(), but those functions need
>> to remain available. (We may want to arrange for them to be removed when
>> linking, as long as they're not referenced. But that's a separate topic.)
> 
> It is confusing that none of build cfg combinations have failed
> (HVM=y PV=n, HVM=n PV=n) :(
> 
> copy_to_unsafe_ll()
> - called from copy_to_unsafe()
> - copy_to_unsafe() has no users (unreachable, MISRA 2.1?)
> 
> copy_from_unsafe_ll()
> - called from copy_from_unsafe()
> - copy_from_unsafe() called from one place do_invalid_op() with
>    copy_from_unsafe(,, n = sizeof(bug_insn)).
>    Due to __builtin_constant_p(n) check the copy_from_unsafe() call
>    optimized by compiler to
>    get_unsafe_size(*(uint16_t *)to, from, 2, UA_DROP, ret, 2);
> 
> as result copy_from_unsafe_ll() is unreachable also (?).

Yes, these likely all want to become library-like, so they are linked in only
when actually referenced.

> If those function are not subject to be removed, the
>   usercopy.c can't be moved in "x86/pv", Right?

That's my take, yes.

> Making copy_{from,to}_unsafe_ll() available for !PV means
> rewriting usercopy.c in some way, Right?

"Re-writing" is probably too much, but some adjustments would be needed if
you want to keep the "unsafe" functions but compile out the "guest" ones.
It may be possible to compile the file twice, once from x86/pv/ and once
from x86/, replacing the self-#include near the bottom of the file. The
former would then produce the "guest" functions, the latter the "unsafe"
ones.

Jan