Thread (46 messages) flat view 46 messages, 10 authors, 2013-02-20
STALE4916d

[PATCH] ARM: decompressor: clear SCTLR.A bit for v7 cores

From: Rob Herring <hidden>
Date: 2012-10-11 13:31:47

On 10/11/2012 08:09 AM, Russell King - ARM Linux wrote:
On Thu, Oct 11, 2012 at 07:43:22AM -0500, Rob Herring wrote:
quoted
The contents of this were already reviewed on this thread, so I sent this
to the patch system and this was Russell's reply:
So that's why I couldn't find it - the mailing list thread has a different
subject line to the patch.  Don't do that.  Given the amount of list
traffic we have today, that's as good as not having been posted at all.
quoted
quoted
NAK for two reasons.

1. It hasn't been on the list (I can't find a match for "clear SCTLR.A"
in my mailbox)

2. The behaviour of unaligned accesses vary depending on CPU.  Some
fix-up the access, others load the word and then rotate it.  If we have
decompressors which perform unaligned accesses, we need to fix this
properly to avoid the CPU specific behaviour, rather than tweaking
control bits to hide the problem.
I'm simply matching the behavior of the kernel itself. The A bit is cleared
for v7 kernels and compilers only generate unaligned accesses for v7.
Without this the initial state of the A bit is undefined as a bootloader
could have cleared it already. We should document the required state or set
it to what we want.
Irrespective of this, (2) still stands.  Unaligned accesses in the
decompressor without a fixup (which will be very hard to provide)
will return different data depending on the CPU as I mention in point
2.
This only affects v7 cores. It should not vary for v7 cores as unaligned
access is a required feature. So how is it going to vary on v7 CPUs?
We've got bigger problems if there are v7 cores that don't handle
unaligned accesses.

Rob
So, using unaligned accesses in the decompressor may not give expected
results in all cases, so they're best avoided.

And if they're avoided, then we don't care about the setting of the A
bit.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help