Hi,
On 20/12/2022 15:24, Ayan Kumar Halder wrote:
On 16/12/2022 10:23, Julien Grall wrote:
Each adaptations are distinct, so I would prefer if they are in in
separate patch.
This will also make clear which components you touched because I would
be surprised if this is really the only place where we need
adaptation. Maybe that's because you didn't compile everything (which
is fine).
On 15/12/2022 19:32, Ayan Kumar Halder wrote:
1. Supersections are supported only for paddr greater than 32 bits.
Two questions:
* Can you outline why we can't keep the code around?
For PA_32, the following bitoperation will overflow :-
*ipa |= (paddr_t)(pte.supersec.extbase1) <<
L1DESC_SUPERSECTION_EXT_BASE1_SHIFT;
*ipa |= (paddr_t)(pte.supersec.extbase2) <<
L1DESC_SUPERSECTION_EXT_BASE2_SHIFT;
Also, pte.supersec.extbase1 and pte.supersec.extbase2 are not valid for
PA_32. Refer xen/arch/arm/include/asm/short-desc.h
You are right about extbase1 and extbase2. See below for the
supersection support.
unsigned int extbase2:4; /* Extended base address, PA[39:36] */
unsigned int extbase1:4; /* Extended base address, PA[35:32] */
* Can you give a pointer to the Arm Arm that says supersection is not
supported?
I could not find any sentence in Arm Arm which says supersection is
**not** supported on 32 bit PA.
However,
Refer"ARM DDI 0487I.a ID081822", G5.4 "The VMSAv8-32 Short-descriptor
translation table format", G5-9163
"Support for Supersections is **optional**, except that an
implementation that supports more than 32 bits of PA must also support
Supersections to provide access to the entire PA space."
Also, G5.1.3 "Address spaces in VMSAv8-32", G5-9149
"AArch32 defines two translation table formats. The Long-descriptor
format gives access to the full 40-bit IPA or PA space at a granularity
of 4KB. The Short-descriptor format:
Gives access to a 32-bit PA space at 4KB granularity.
Gives access to a 40-bit PA space, but only at 16MB granularity, by
the use of Supersections."
from the above 2 snippets, I inferred that supersections are only
supported for PAs greater than 32 bits.
I could not find any evidence of supersections supported for 32 bit PA.
From what you quoted above supersection is optional unless the
processor support more than 32-bit PA. IOW, an implementer is free to
implement the feature even if it is not strictly necessary when the
processor only supports 32-bit PA. This because it is useful to reduce
the TLB contention.
So if we want to #ifdef some code, then only the two lines using
L1DESC_SUPERSECTION_EXT_BASE{1, 2}_SHIFT should be protected. The rest
should stay as-is.
Cheers,
--
Julien Grall
