I've been hitting a bug on a patch I'm working on and have considered (and
more or less tested with good results) doing this change:
-#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
+#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT))
(and the corresponding thing on other architecture).
In general, the question is whether __P000 and __S000 in protection_map are
ever used except for MAP_POPULATE, and even then if they work well.
I'm seeking for objections to this change and/or anything I'm missing.
This bug showed up while porting remap_file_pages protection support to
2.6.16-rc3. It always existed but couldn't trigger before the PageReserved
changes.
Consider a _PAGE_PROTNONE pte, which has then pte_pfn(pte) == 0 (with
remap_file_pages you need them to exist). Obviously pte_pfn(pte) on such a PTE
doesn't make sense, but since pte_present(pte) gives true the code doesn't
know that.
Consider a call to munmap on this range. We get to zap_pte_range() which (in
condensed source code):
zap_pte_range()
...
if (pte_present(ptent)) {
//This test is passed
struct page *page = vm_normal_page(vma, addr, ptent);
//Now page points to page 0 - which is wrong, page should be NULL
page_remove_rmap(page);
//Which doesn't make any sense.
//If mem_map[0] wasn't mapped we hit a BUG now, if it was we'll hit it later -
//i.e. negative page_mapcount().
Now, since this code doesn't work in this situation, I wonder whether PROTNONE
is indeed used anywhere in the code *at the moment*, since faults on pages
mapped as such are handled with SIGSEGV.
The only possible application, which is only possible in 2.6 and not in 2.4
where _PAGE_PROTNONE still exists, is mmap(MAP_POPULATE) with prot ==
PROT_NONE.
Instead I need to make use of PROTNONE, so the handling of it may need
changes. In particular, I wonder about why:
#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
I see why that _PAGE_PROTNONE can make sense, but in the above code it
doesn't.
--
Inform me of my mistakes, so I can keep imitating Homer Simpson's "Doh!".
Paolo Giarrusso, aka Blaisorblade (Skype ID "PaoloGiarrusso", ICQ 215621894)
http://www.user-mode-linux.org/~blaisorblade
___________________________________
Yahoo! Mail: gratis 1GB per i messaggi e allegati da 10MB
http://mail.yahoo.it
On Mon, 20 Feb 2006, Blaisorblade wrote:
> I've been hitting a bug on a patch I'm working on and have considered (and
> more or less tested with good results) doing this change:
>
> -#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
> +#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT))
>
> (and the corresponding thing on other architecture).
>
> In general, the question is whether __P000 and __S000 in protection_map are
> ever used except for MAP_POPULATE, and even then if they work well.
>
> I'm seeking for objections to this change and/or anything I'm missing.
Objection, your honor.
> This bug showed up while porting remap_file_pages protection support to
> 2.6.16-rc3. It always existed but couldn't trigger before the PageReserved
> changes.
>
> Consider a _PAGE_PROTNONE pte, which has then pte_pfn(pte) == 0 (with
> remap_file_pages you need them to exist). Obviously pte_pfn(pte) on such a PTE
> doesn't make sense, but since pte_present(pte) gives true the code doesn't
> know that.
I didn't fully understand you there, but I think you've got it the wrong
way round: _PAGE_PROTNONE is included in the pte_present() test precisely
because there is a valid page there, pfn is set (it might be pfn 0, yes,
but much more likely to be pfn non-0).
I've never used PROT_NONE myself (beyond testing), but I think the
traditional way it's used is this: mmap(,,PROT_READ|PROT_WRITE,,,),
initialize the pages of that mapping, then mprotect(,,PROT_NONE) -
which retains all those pages but make them generate SIGSEGVs - so
the app can detect accesses and decide if it wants to do something
special with them, other than the obvious mprotect(,,PROT_READ) or
whatever.
PROT_NONE gives you a way of holding the page present (unlike munmap),
yet failing access. And since those pages remain present, they do
need to be freed later when you get to zap_pte_range. They are
normal pages, but user access to them has been restricted.
Hugh
> Consider a call to munmap on this range. We get to zap_pte_range() which (in
> condensed source code):
>
> zap_pte_range()
> ...
> if (pte_present(ptent)) {
> //This test is passed
> struct page *page = vm_normal_page(vma, addr, ptent);
> //Now page points to page 0 - which is wrong, page should be NULL
> page_remove_rmap(page);
> //Which doesn't make any sense.
> //If mem_map[0] wasn't mapped we hit a BUG now, if it was we'll hit it later -
> //i.e. negative page_mapcount().
>
> Now, since this code doesn't work in this situation, I wonder whether PROTNONE
> is indeed used anywhere in the code *at the moment*, since faults on pages
> mapped as such are handled with SIGSEGV.
>
> The only possible application, which is only possible in 2.6 and not in 2.4
> where _PAGE_PROTNONE still exists, is mmap(MAP_POPULATE) with prot ==
> PROT_NONE.
>
> Instead I need to make use of PROTNONE, so the handling of it may need
> changes. In particular, I wonder about why:
>
> #define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
>
> I see why that _PAGE_PROTNONE can make sense, but in the above code it
> doesn't.
On Tuesday 21 February 2006 14:09, Hugh Dickins wrote:
> On Mon, 20 Feb 2006, Blaisorblade wrote:
> > I've been hitting a bug on a patch I'm working on and have considered
> > (and more or less tested with good results) doing this change:
> >
> > -#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
> > +#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT))
> >
> > (and the corresponding thing on other architecture).
> >
> > In general, the question is whether __P000 and __S000 in protection_map
> > are ever used except for MAP_POPULATE, and even then if they work well.
> >
> > I'm seeking for objections to this change and/or anything I'm missing.
> Objection, your honor.
English humor :-) ?
> I didn't fully understand you there, but I think you've got it the wrong
> way round: _PAGE_PROTNONE is included in the pte_present() test precisely
> because there is a valid page there, pfn is set (it might be pfn 0, yes,
> but much more likely to be pfn non-0).
> I've never used PROT_NONE myself (beyond testing), but I think the
> traditional way it's used is this: mmap(,,PROT_READ|PROT_WRITE,,,),
> initialize the pages of that mapping, then mprotect(,,PROT_NONE) -
> which retains all those pages but make them generate SIGSEGVs - so
> the app can detect accesses and decide if it wants to do something
> special with them, other than the obvious mprotect(,,PROT_READ) or
> whatever.
> PROT_NONE gives you a way of holding the page present (unlike munmap),
> yet failing access. And since those pages remain present, they do
> need to be freed later when you get to zap_pte_range. They are
> normal pages, but user access to them has been restricted.
Ok, thanks for the explaination.
The bug is born from the patched install_file_pte(). Before there was no need
to store the protection bits, now it's needed.
So, it sets a pte_file PTE containing no page, and on PROT_NONE it uses
_PAGE_PROTNONE|_PAGE_FILE.
Indeed, what I've actually coded and tested was safer, but I wanted to know if
it could be simpler (and faster). For i386 it should be (I've re-tested only
UML so far):
-#define pte_present(x) ((x).pte_low & (_PAGE_PRESENT | _PAGE_PROTNONE))
+#define pte_present(x) (((x).pte_low & _PAGE_PRESENT) || \
+ (((x).pte_low & (_PAGE_PROTNONE|_PAGE_FILE)) == _PAGE_PROTNONE))
--- linux-2.6.git.orig/include/asm-um/pgtable.h
+++ linux-2.6.git/include/asm-um/pgtable.h
@@
-#define pte_present(x) pte_get_bits(x, (_PAGE_PRESENT | _PAGE_PROTNONE))
+#define pte_present(x) (pte_get_bits(x, (_PAGE_PRESENT)) || (pte_get_bits(x,
(_PAGE_PROTNONE)) && !pte_file(x)))
--
Inform me of my mistakes, so I can keep imitating Homer Simpson's "Doh!".
Paolo Giarrusso, aka Blaisorblade (Skype ID "PaoloGiarrusso", ICQ 215621894)
http://www.user-mode-linux.org/~blaisorblade
___________________________________
Yahoo! Mail: gratis 1GB per i messaggi e allegati da 10MB
http://mail.yahoo.it