Skip to contentSkip to conversation

sparc32: Fix lost pte_present() state while mprotect() system call

PATCH In review Kirill Tkhai 2 participants 4 messages Feb 22, 2013 – Mar 3, 2013 SPARC + UltraSPARC (sparc/sparc64)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sparc32: Fix lost pte_present() state while mprotect() system call

If pte is present before change_pte_range() works on it, the state will be lost after it. pte_modify() zeroes every 0xff bit including SRMMU_ET_MASK while it adds new protection. I receive the stack after this (on a slightly modified kernel):

BUG: Bad page map in process SC_EXE  pte:059d8cfc pmd:00069e01
addr:52d44000 vm_flags:00100070 anon_vma:be6ef3cc mapping:(null) index:52d44
8009e8d0 : zap_pte_range+0x120/0x3bc
8009ed3c : unmap_page_range+0x1d0/0x230
8009ee80 : unmap_vmas+0xe4/0x160
800a5a74 : exit_mmap+0xbc/0x218
8003ed24 : mmput+0x2c/0x100
80044190 : exit_mm+0x108/0x148
800447a0 : do_exit+0x158/0x350
80044a08 : do_group_exit+0x38/0xb8
80044a94 : sys_exit_group+0xc/0x1c
8001232c : syscall_is_too_hard+0x38/0x44

7fd5a7c4 : 0x7fd5a7c4

So, add SRMMU_ET_MASK to SRMMU_CHG_MASK.

---
 arch/sparc/include/asm/pgtsrmmu.h |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/sparc/include/asm/pgtsrmmu.h +1 -1
index 79da178..5d33984 100644
@@ -84,7 +84,7 @@
 
 #define SRMMU_PTE_FILE_SHIFT     8	/* = 32-PTE_FILE_MAX_BITS */
 
-#define SRMMU_CHG_MASK    (0xffffff00 | SRMMU_REF | SRMMU_DIRTY)
+#define SRMMU_CHG_MASK    (0xffffff00 | SRMMU_REF | SRMMU_DIRTY | SRMMU_ET_MASK)
 
 /* SRMMU swap entry encoding
  *
maintainer
#1 lore raw

From: Kirill Tkhai <[email protected]>
Date: Fri, 22 Feb 2013 17:19:07 +0400

Quoted 22 lines

If pte is present before change_pte_range() works on it, the state will be lost after it. pte_modify() zeroes every 0xff bit including SRMMU_ET_MASK while it adds new protection. I receive the stack after this (on a slightly modified kernel):

BUG: Bad page map in process SC_EXE  pte:059d8cfc pmd:00069e01
addr:52d44000 vm_flags:00100070 anon_vma:be6ef3cc mapping:(null) index:52d44
8009e8d0 : zap_pte_range+0x120/0x3bc
8009ed3c : unmap_page_range+0x1d0/0x230
8009ee80 : unmap_vmas+0xe4/0x160
800a5a74 : exit_mmap+0xbc/0x218
8003ed24 : mmput+0x2c/0x100
80044190 : exit_mm+0x108/0x148
800447a0 : do_exit+0x158/0x350
80044a08 : do_group_exit+0x38/0xb8
80044a94 : sys_exit_group+0xc/0x1c
8001232c : syscall_is_too_hard+0x38/0x44

7fd5a7c4 : 0x7fd5a7c4

So, add SRMMU_ET_MASK to SRMMU_CHG_MASK.

...

arch/sparc/include/asm/pgtsrmmu.h line 87

Quoted 2 lines
arch/sparc/include/asm/pgtsrmmu.h +1 -1
-#define SRMMU_CHG_MASK    (0xffffff00 | SRMMU_REF | SRMMU_DIRTY)
+#define SRMMU_CHG_MASK    (0xffffff00 | SRMMU_REF | SRMMU_DIRTY | SRMMU_ET_MASK)

I don't think this is correct, the state is not lost, it must be contained fully in the pgprot_t protections we are changing to.

What if we are trying to change the protections from PAGE_COPY (which has the valid bit set) to PAGE_NONE (which lacks the valid bit)? That no longer works after your change, as we'll leave the valid bit set.

In fact, I think the bug you triggered is exactly this situation, and your patch therefore breaks PAGE_NONE completely.

Present and Valid are two seperate pieces of state, so the bug might really be that the SRMMU code tries to handle them as one. And your trace seems to confirm this, actually. PROT_NONE can be for a present page, but it must have the valid bit cleared.

#2 lore raw
In reply to Kirill Tkhai
Quoted 26 lines

From: Kirill Tkhai <[email protected]>
Date: Fri, 22 Feb 2013 17:19:07 +0400

Quoted 8 lines

If pte is present before change_pte_range() works on it, the state will be lost after it. pte_modify() zeroes every 0xff bit including SRMMU_ET_MASK while it adds new protection. I receive the stack after this (on a slightly modified kernel):

So, add SRMMU_ET_MASK to SRMMU_CHG_MASK.

...

arch/sparc/include/asm/pgtsrmmu.h line 87

Quoted 2 lines
arch/sparc/include/asm/pgtsrmmu.h +1 -1
-#define SRMMU_CHG_MASK (0xffffff00 | SRMMU_REF | SRMMU_DIRTY)
+#define SRMMU_CHG_MASK (0xffffff00 | SRMMU_REF | SRMMU_DIRTY | SRMMU_ET_MASK)

I don't think this is correct, the state is not lost, it must be contained fully in the pgprot_t protections we are changing to.

What if we are trying to change the protections from PAGE_COPY (which has the valid bit set) to PAGE_NONE (which lacks the valid bit)? That no longer works after your change, as we'll leave the valid bit set.

Quoted 7 lines

In fact, I think the bug you triggered is exactly this situation, and your patch therefore breaks PAGE_NONE completely.

Present and Valid are two seperate pieces of state, so the bug might really be that the SRMMU code tries to handle them as one. And your trace seems to confirm this, actually. PROT_NONE can be for a present page, but it must have the valid bit cleared.

Thanks for the explanation, David!

Should we account none pages in pte_present instead of above like this?

static inline int srmmu_pte_present(pte_t pte)
{ return ((pte_val(pte) & SRMMU_ET_MASK) = SRMMU_ET_PTE) ||

          ((pte_val(pte) & SRMMU_PAGE_NONE) = SRMMU_PAGE_NONE); }

--

maintainer
#3 lore raw
In reply to Kirill Tkhai

From: Kirill Tkhai <[email protected]>
Date: Mon, 25 Feb 2013 17:05:18 +0400

Quoted 5 lines

Should we account none pages in pte_present instead of above like this?

static inline int srmmu_pte_present(pte_t pte)
{ return ((pte_val(pte) & SRMMU_ET_MASK) = SRMMU_ET_PTE) ||

          ((pte_val(pte) & SRMMU_PAGE_NONE) = SRMMU_PAGE_NONE); }

No, I don't think that is what we should do.

Instead, we should use the SRMMU_NOREAD bit to encode SRMMU_PAGE_NONE pte's. Then you make SRMMU_PAGE_NONE set SRMMU_ET_PTE but also SRMMU_NOREAD.

Then you'll need no changes at all to srmmu_pte_present() or SRMMU_CHG_MASK.