Re: [PATCH] flush icache before set_pte on ia64 take3

[Zoltan Menyhart <Zoltan.Menyhart@bull.net>]

I do not think changing the semantics of flush_cache_page()
is a good idea. Please have a look at Documentation/cachetlb.txt.
This is for virtual address tagged caches.

The doc. does not say much about flush_icache_page().
It is used in:

install_page(): can release any previously existing mapping
		(similarly: for virtual address tagged I-cache)

On the other hand, I do not see why it is used in these two
cases, I think the virtual address tagged I-cache must have been
flushed when the old PTE of the page was removed:

do_no_page():
	if (pte_none(*page_table)) {
		flush_icache_page(vma, new_page);

do_swap_page():
	flush_icache_page(vma, page);

Anyway, it is a NO-OP on the machines with physical address
tagged caches.

Sure, the names could have been chosen more carefully.

Unless you have a very strong reason, do not change their
semantics (would require updates for the other architectures).

I can agree:
- lazy_mmu_prot_update() is not the right place for a flush
- to flush I-cache before inserting PTE (at least for the
  machines with physical address tagged caches)

Can we have a separate, specific exec_flush_phtag_icache_page()
that is to be called before inserting a PTE with the exec bit on
for physical address tagged I-caches) ?
It is defined in asm-generic/cacheflush.h as NO-OP.

Or exec_flush_icache_page_before_set_pte() ?

Thanks,

Zoltan