Re: kswapd deadlock 2.4.3-pre6

[David S. Miller]

Marcelo Tosatti writes:
 > Are you sure flush_page_to_ram()/flush_icache_page() without
 > page_table_lock held is ok for all archs? 

This is actually a sticky area.  For example, I remember that a long
time ago I noticed that it wasn't necessarily guarenteed that even
all the flush_tlb_{page,mm,range}() stuff was done under the page
table lock.

Maybe things have changed since then, but...

Furthermore I did not specify that flush_page_to_ram/flush_icache_page
can assume this, see Documentation/cachetlb.txt

I have no problem adding this invariant to some of the interfaces, but
we have to audit things.

 > Looking at arch/mips/mm/r2300.c:

The r2300 port is UP-only btw...

Later,
David S. Miller
[e-mail]