Andi Kleen <[EMAIL PROTECTED]> wrote:
>
> After thinking about it more I agree. Just replacing TASK_SIZE with
> something that depends on the mm is the best solution here.
OK.
Here's what I currently have:
From: Anton Blanchard <[EMAIL PROTECTED]>
The 4 level pagetable code changed the exit_mmap code to rely on TASK_SIZE.
On some architectures (eg ppc64 and ia64), this is a per task property and
bad things can happen in certain circumstances when using it.
It is possible for one task to end up "owning" an mm from another - we have
seen this with the procfs code when process 1 accesses /proc/pid/cmdline of
process 2 while it is exiting. Process 2 exits but does not tear its mm
down. Later on process 1 finishes with the proc file and the mm gets torn
down at this point.
Now if process 1 was 32bit and process 2 was 64bit then we end up using a
bad value for TASK_SIZE in exit_mmap. We only tear down part of the
address space and leave half initialised pagetables and entries in the MMU
etc.
MM_VM_SIZE() was created for this purpose (and is used in the next line for
tlb_finish_mmu), so use it. I moved the PGD round up of TASK_SIZE into the
default MM_VM_SIZE.
As an aside, all architectures except one define FIRST_USER_PGD_NR as 0:
include/asm-arm26/pgtable.h:#define FIRST_USER_PGD_NR 1
It would be nice to get rid of one more magic constant and just clear from
0 ... MM_VM_SIZE(). That would make it consistent with the tlb_flush_mmu
call below it too.
Signed-off-by: Anton Blanchard <[EMAIL PROTECTED]>
Signed-off-by: Andrew Morton <[EMAIL PROTECTED]>
---
25-akpm/include/linux/mm.h | 2 +-
25-akpm/mm/mmap.c | 3 +--
2 files changed, 2 insertions(+), 3 deletions(-)
diff -puN include/linux/mm.h~use-mm_vm_size-in-exit_mmap include/linux/mm.h
--- 25/include/linux/mm.h~use-mm_vm_size-in-exit_mmap 2005-01-25
21:41:44.365536624 -0800
+++ 25-akpm/include/linux/mm.h 2005-01-25 21:41:44.371535712 -0800
@@ -38,7 +38,7 @@ extern int sysctl_legacy_va_layout;
#include <asm/atomic.h>
#ifndef MM_VM_SIZE
-#define MM_VM_SIZE(mm) TASK_SIZE
+#define MM_VM_SIZE(mm) ((TASK_SIZE + PGDIR_SIZE - 1) & PGDIR_MASK)
#endif
#define nth_page(page,n) pfn_to_page(page_to_pfn((page)) + (n))
diff -puN mm/mmap.c~use-mm_vm_size-in-exit_mmap mm/mmap.c
--- 25/mm/mmap.c~use-mm_vm_size-in-exit_mmap 2005-01-25 21:41:44.367536320
-0800
+++ 25-akpm/mm/mmap.c 2005-01-25 21:41:44.373535408 -0800
@@ -1980,8 +1980,7 @@ void exit_mmap(struct mm_struct *mm)
~0UL, &nr_accounted, NULL);
vm_unacct_memory(nr_accounted);
BUG_ON(mm->map_count); /* This is just debugging */
- clear_page_range(tlb, FIRST_USER_PGD_NR * PGDIR_SIZE,
- (TASK_SIZE + PGDIR_SIZE - 1) & PGDIR_MASK);
+ clear_page_range(tlb, FIRST_USER_PGD_NR * PGDIR_SIZE, MM_VM_SIZE(mm));
tlb_finish_mmu(tlb, 0, MM_VM_SIZE(mm));
_
and:
From: David Woodhouse <[EMAIL PROTECTED]>
Bad things can happen if a 32-bit process is the last user of a 64-bit mm.
TASK_SIZE isn't a constant, and we can end up clearing page tables only up
to the 32-bit TASK_SIZE instead of all the way. We should probably
double-check every instance of TASK_SIZE or USER_PTRS_PER_PGD for this kind
of problem.
We should also double-check that MM_VM_SIZE() and other such things are
correctly defined on all architectures. I already fixed ppc64 which let it
stay as TASK_SIZE, and hence dependent on the _current_ context instead of
the mm in the argument.
Signed-off-by: Andrew Morton <[EMAIL PROTECTED]>
---
25-akpm/mm/mmap.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)
diff -puN mm/mmap.c~task_size-is-variable mm/mmap.c
--- 25/mm/mmap.c~task_size-is-variable 2005-01-25 22:08:40.903785456 -0800
+++ 25-akpm/mm/mmap.c 2005-01-25 22:08:40.908784696 -0800
@@ -1612,8 +1612,8 @@ static void free_pgtables(struct mmu_gat
unsigned long last = end + PGDIR_SIZE - 1;
struct mm_struct *mm = tlb->mm;
- if (last > TASK_SIZE || last < end)
- last = TASK_SIZE;
+ if (last > MM_VM_SIZE(mm) || last < end)
+ last = MM_VM_SIZE(mm);
if (!prev) {
prev = mm->mmap;
_
Shall I ship it?
