Re: flush_icache_range

[Ralf Baechle <ralf@oss.sgi.com>]

On Mon, Jul 24, 2000 at 09:10:30AM -0700, Kanoj Sarcar wrote:

> > > Can anyone point out the logic of continued existance of
> > > flush_icache_range after the introduction of flush_icache_page()? I
> > > admit that flush_icache_range is still needed in the module loading
> > > code, but do we need it anymore in the a.out loading code? That code
> > > should be incurring page faults, which will do the flush_icache_page
> > > anyway. Seems like double work to me to do flush_icache_range again
> > > after the loading has been done.
> > 
> > binfmt_elf.c:load_aout_interp() uses file->f_op->read to read the
> > interpreter from disk, so actually need to use something else to flush
> > the cache.  Similar for two of three cases in binfmt_aout.c.  For these
> > the page fault won't sufficiently flush cashes.
> 
> Okay, got it. flush_icache_page() can flush the icache, then 
> flush_icache_range() can writeback-invalidate the dcache (for the a.out
> section loading code), and things should work. AFAICS, this would be
> the most optimal way to do things (ie, you don't have to writeback-invalidate
> dcache, and invalidate icache in flush_icache_range(), you can optimize
> out the icache flush relying on flush_icache_page to do the work).

Yep, in other word one of the the four flush_icache_range() calls can die.

> > > This argument to delete the flush_icache_range calls from the a.out
> > > loading code assumes that the f_op->read() code behaves sanely, ie does
> > > not do unexpected things like touch the user address (thus allocating
> > > the page, and doing the icache flush via the page fault handler much
> > > earlier) before it starts reading the a.out sections in ...
> > 
> > There is another MIPS specific problem with this routine.  Originally
> > introduced for kernel modules the various incarnations of flush_icache_range
> > are not protected against access from userspace.  Unable to handle kernel
> > paging request ahead ...
> 
> Could you elaborate? Use mips as an example. Note: for the a.out code,
> there will be one thread, and for the module loading, userspace access 
> to the vmalloced area is not possible.

I think the scenario only affects MIPS >= R4000 caches.  Assume a machine
under memory pressure.  It swaps the pages just loaded by the a.out loader
or the a.out ELF interpreter loader out even before flush_icache_range()
is called.  Next flush_icache_page() gets called and applies a cache
instruction to one of the swapped out pages.  Then the CPU will throw a
TLBL exception for which no exception handler exists, so fault.c will
panic with Unable to handle kernel paging request ...

I've fixed that in current set of MIPS cache patches.

  Ralf
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/