fast-get: fix a recent regression: write before mapping RO

[lyakh]

A recent commit introduced a regression: it moved writing data to a buffer after the call, mapping the target buffer as read-only. Fix it by moving the copy operation back up.

Fixes commit cfa5f02 ("fast-get: don't corrupt entries when failing")

[kv2019i]

It is not very obvious that fast_get_access_grant() actually limits the access here, but alas, fixes the regression...

[lyakh]

It is not very obvious that fast_get_access_grant() actually limits the access here, but alas, fixes the regression...

@kv2019i why isn't it obvious?

sof/zephyr/lib/fast-get.c

Line 120 in 3db6d48

.attr = K_MEM_PARTITION_P_RO_U_RO | XTENSA_MMU_CACHED_WB,

clearly maps the memory read-only for kernel and userspace

[kv2019i]

@lyakh wrote:

It is not very obvious that fast_get_access_grant() actually limits the access here, but alas, fixes the regression...
@kv2019i why isn't it obvious?
[...]
clearly maps the memory read-only for kernel and userspace

You have to look inside the grant function implementation to see that. And even then, the function name is surprising. Usually when you grant something, you allow something more, not remove something you already had.

[lyakh]

It is not very obvious that fast_get_access_grant() actually limits the access here, but alas, fixes the regression...
@kv2019i why isn't it obvious?
[...]
clearly maps the memory read-only for kernel and userspace

You have to look inside the grant function implementation to see that. And even then, the function name is surprising. Usually when you grant something, you allow something more, not remove something you already had.

it grants it to userspace. As for the kernel - well, the kernel mode might not need that access at all (after having written to it once). But yes, it isn't immediately obvious without reading the code :-)

[Copilot]

Pull request overview

Fixes a regression in fast_get() where the newly allocated SRAM buffer could be mapped read-only before the data copy occurred, causing the subsequent write to fail.

Changes:

• Move the memcpy_s() into the newly allocated SRAM buffer earlier in fast_get().
• Perform dcache_writeback_region() before any read-only memory-domain mapping occurs.