A small related chirp-in from the "lowly Linux userland":
Does the current system-call "cacheflush(2)" works with the newer kernels?
As the "man cacheflush" tells, it was supposed to work only on MIPS
based systems.
In the past I had some problems with it, so used related YAMON functions
(which worked
better).
Maybe it could be made working for MIPS again?
(is it supposed to work with current kernels?)
Cache flushing is a useful feature for some benchmarks (and customers
had requested
related tests in the past).
-zenon
Leonid Yegoshin wrote:
On 02/20/2015 11:17 AM, Kevin Cernekee wrote:
On Thu, Feb 19, 2015 at 8:17 AM, Steven J. Hill
<Steven.Hill@xxxxxxxxxx> wrote:
From: Leonid Yegoshin <Leonid.Yegoshin@xxxxxxxxxx>
Flush the D-cache before the page is given to a process
as an executable (I-cache) page when the backing store
is non-DMA I/O.
Signed-off-by: Leonid Yegoshin <Leonid.Yegoshin@xxxxxxxxxx>
Signed-off-by: Steven J. Hill <Steven.Hill@xxxxxxxxxx>
This patch seems to make several different changes to the cache
maintenance code all at once:
1) Add logic to handle virtually tagged D$
This is needed for 74K/1074K erratas. It is somehow was split from
006a851b10a395955c153a145ad8241494d43688 which was accepted upstream
but erratas not go through.
The HW behaviour is similar to virtually tagged D$ but it is a HW bug.
and perform extra flushes
on TLB updates
As I understand, it is about adding __update_cache() in
update_mmu_cache(), right?
If so, then - this code exists from the first git version of kernel
from Mr. Linus.
It was deleted by mistake, I think.
2) Add new write barriers betwen D$/I$ or D$/L2 flushes
It is required. MIPS32/64 R2 specs say:
"For implementations which implement multiple level of caches without
the inclusion property, the use of a SYNC
instruction after the CACHE instruction is still needed whenever
writeback data has to be resident in the next level of
memory hierarchy."
So, if we need to transfer instruction from D$ to L2 then we should
use SYNC after CACHE D$ before operates with this data in L2. In other
case the CACHE for L2 may go ahead of completion of D$ for the same
line and flush a stale data.
The same is basically for transfer D$ --> I$ because in MIPS it is
done via L2 or memory.
3) Make __flush_anon_page() play nice with HIGHMEM on systems with
cache aliases
and maybe a few more that I missed.
4) It is basically a revert of patch 64f23ab30b1fe which kills a
performance in non Cavium Octeon CPUs, actually - any CPU which has no
D$ cache snooping in I$.
But just revert is incorrect, the another proper stuff is required for
correct operations.
Would it be possible to split this out into individual commits, and
include more comprehensive changelogs for each one describing the
exact problem being solved?
I would ask Steven to do it. The original code dates back to 2.6.32
and it was packed/split in different patches multiple times. After a
lot of split/join following patch acceptance/rejection we have what we
have now.
In my opinion, the process of splitting into individual commits are
something wrong - some patches are accepted and some - not, and we
have a buggy code now. It should be packed into functional patches but
each patch should be a single workable commit. If multiple patches are
needed to fix a problem then result is sometime wrong.
Also, it would be helpful to clarify how this relates to the use of
swap (?) with a backing store that is non-DMA I/O. Do you have an
example of a situation where the existing code broke? A play-by-play
postmortem would make for interesting reading.
I guess, it is a little incorrect - this code is REQUIRED for non-DMA
I/O with root FS or swap but it is not enough. Another patch is still
needed to complete, see http://patchwork.linux-mips.org/patch/8635/
- Leonid.
