lkml.org 
[lkml]   [2003]   [Aug]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [parisc-linux] Re: Problems with kernel mmap (failing tst-mmap-eofsync in glibc on parisc)
On 22 Aug 2003 17:27:33 -0500
James Bottomley <James.Bottomley@SteelEye.com> wrote:

> Yes, the issue seems to be that the flush_dcache_page() was implemented
> with the thought that the caches of the shared mappings may contain
> modified data that needs to be flushed to the aliased page.
>
> The opposite property: that the caches of the aliased page need to be
> invalidated because someone else changed data in the aliased page seems
> to work as a byproduct of the above implementation.
>
> But some of the checks for !list_empty(&mapping->i_shared) are going to
> prevent the necessary invalidations on read only shared mappings...which
> was the initial problem.

The theory of operation is that there are two "classes" of mappings
for a page, the implicit kernel mapping and all user mappings. The
goal is to flush out one class from the cache when the other one writes
to such a page.

When a write() system call occurs, the kernel "class" is writing to
the page so all user mappings (shared or not!) need to be flushed
out.

When a read() system call occurs, and shared+writable mmap()'s of
the page exist, we must flush the user mapping "class" before the
kernel "class" tries to read from that page.

You cannot avoid doing things exactly as I've just described without
allowing bad aliases to be in the cache and corrupt data.

Your test case is essentially, annotated with what the kernel
should do at each step:

static const char *test1 = "Line the first\n";
static const char *test2 = "Line the second\n";

temp_fd = open(O_RDWR);
write(temp_fd, test1, sizeof test1 - 1);

No mmaps of this page, so no flush_dcache_page() call.

...
fd = open(O_RDONLY);
p = mmap(fd, ... PROT_READ ...);

Not writable, not added to i_mmap_shared list, instead it is
added to normal i_mmap list.

memcpy(tmp_buf, p, sizeof test1 - 1);
p += sizeof test1 - 1;
if (strcmp(tmp_buf, test1))
BUG();
...
write(temp_fd, test2, sizeof test2 - 1);

Mmaps of this page exist, flush_dcache_page() call is made.

memcpy(tmp_buf, p, sizeof test2 - 1);
if (strcmp(tmp_buf, test2))
BUG();

And thus memcpy() sees correct data.

I think on parisc you are trying to avoid the write() case
of the cache flush for non-shared mmap()s, and sorry you
really can't do this, again this is:

When a write() system call occurs, the kernel "class" is writing to
the page so all user mappings (shared or not!) need to be flushed
out.

If your flush_dcache_page() is not doing this, it's no wonder
the test case fails for you.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:48    [W:0.069 / U:0.300 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site