On 23/10/2024 17:06, Ayan Kumar Halder wrote:
Hi Luca/Julien,
On 22/10/2024 17:31, Luca Fancellu wrote:
Hi Julien,
On 22 Oct 2024, at 14:13, Julien Grall <[email protected]> wrote:
On 22/10/2024 10:56, Luca Fancellu wrote:
On 22 Oct 2024, at 10:47, Julien Grall <[email protected]> wrote:
Hi Luca,
On 22/10/2024 10:41, Luca Fancellu wrote:
On 21 Oct 2024, at 17:27, Julien Grall <[email protected]> wrote:
2) dsb+isb barrier after enabling the MPU, since we are enabling
the MPU but also because we are disabling the background region
TBH, I don't understand this one. Why would disabling the
background region requires a dsb + isb? Do you have any pointer in
the Armv8-R specification?
I’m afraid this is only my deduction, Section C1.4 of the Arm®
Architecture Reference Manual Supplement Armv8, for R-profile
AArch64 architecture,
(DDI 0600B.a) explains what is the background region, it says it
implements physical address range(s), so when we disable it, we
would like any data
access to complete before changing this implementation defined range
with the ranges defined by us tweaking PRBAR/PRLAR, am I right?
My mental model for the ordering is similar to a TLB flush which is:
1/ dsb nsh
2/ tlbi
3/ dsb nsh
4/ isb
Enabling the MPU is effectively 2. AFAIK, 1 is only necessary to
ensure the update to the page-tables. But it is not a requirement to
ensure any data access are completed (otherwise, we would have needed
a dsb sy because we don't know how far the access are going...).
Uhm… I’m not sure we are on the same page, probably I explained that
wrongly, so my point is that:
FUNC_LOCAL(enable_mpu)
mrs x0, SCTLR_EL2
bic x0, x0, #SCTLR_ELx_BR /* Disable Background region */
orr x0, x0, #SCTLR_Axx_ELx_M /* Enable MPU */
orr x0, x0, #SCTLR_Axx_ELx_C /* Enable D-cache */
orr x0, x0, #SCTLR_Axx_ELx_WXN /* Enable WXN */
dsb sy
^— This data barrier is needed in order to complete any data
access, which guarantees that all explicit memory accesses before
this instruction complete, so we can safely turn ON the
MPU and disable the background region.
I think
I think Julien agreed to this in a previous email as long as we have an
appropriate comment. So we will keep this as it is.
Sorry I didn't manage to answer yet. But yes, I am ok with the barrier
for now, but I am not agree on the reasoning used.
Cheers,
--
Julien Grall
