Skip to contentSkip to conversation

powerpc: thp: Fix crash on mremap

PATCH In review Aneesh Kumar K.V 3 participants 9 messages Feb 7, 2014 – Feb 14, 2014

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 powerpc: thp: Fix crash on mremap

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

---
 arch/Kconfig                           |  3 +++
 arch/powerpc/platforms/Kconfig.cputype |  1 +
 mm/huge_memory.c                       | 12 ++++++++++++
 3 files changed, 16 insertions(+)
arch/Kconfig +3 -0
index af2cc6eabcc7..bca9e7a18bd2 100644
@@ -365,6 +365,9 @@ config HAVE_ARCH_TRANSPARENT_HUGEPAGE
 config HAVE_ARCH_SOFT_DIRTY
 	bool
 
+config ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW
+	bool
+
 config HAVE_MOD_ARCH_SPECIFIC
 	bool
 	help
arch/powerpc/platforms/Kconfig.cputype +1 -0
index 6704e2e20e6b..0225011231ea 100644
@@ -71,6 +71,7 @@ config PPC_BOOK3S_64
 	select PPC_FPU
 	select PPC_HAVE_PMU_SUPPORT
 	select SYS_SUPPORTS_HUGETLBFS
+	select ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW
 	select HAVE_ARCH_TRANSPARENT_HUGEPAGE if PPC_64K_PAGES
 
 config PPC_BOOK3E_64
mm/huge_memory.c +12 -0
index 292a266e0d42..89b7a647f1cb 100644
@@ -1474,8 +1474,20 @@ int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
 
 	ret = __pmd_trans_huge_lock(old_pmd, vma);
 	if (ret == 1) {
+#ifdef CONFIG_ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW
+		pgtable_t pgtable;
+#endif
 		pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
 		VM_BUG_ON(!pmd_none(*new_pmd));
+#ifdef CONFIG_ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW
+		/*
+		 * Archs like ppc64 use pgtable to store per pmd
+		 * specific information. So when we switch the pmd,
+		 * we should also withdraw and deposit the pgtable
+		 */
+		pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
+		pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
+#endif
 		set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
 		spin_unlock(&mm->page_table_lock);
 	}
Signature · 1 line
1.8.3.2
#1 lore raw
Quoted 18 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

Quoted 6 lines

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

I _REALLY_ dislike patches that are totally different from Linus's tree in stable trees, it has caused nothing but problems in the past.

greg k-h

#2 lore raw
Quoted 26 lines · Greg KH
Quoted 18 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

Quoted 6 lines

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

Quoted 6 lines

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

I _REALLY_ dislike patches that are totally different from Linus's tree in stable trees, it has caused nothing but problems in the past.

I don't think it applies... (I tried on an internal tree) but the affected function changed in 3.13 in various ways. Aneesh, please provide a more details explanation and whether we should backport those other changes too or whether this is not necessary.

BTW. Aneesh, we need a 3.11.x one too

Cheers,
Ben.

#3 lore raw
Quoted 31 lines · Benjamin Herrenschmidt
Quoted 30 lines · Greg KH
Quoted 18 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

Quoted 6 lines

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

Quoted 7 lines
Quoted 2 lines

I _REALLY_ dislike patches that are totally different from Linus's tree in stable trees, it has caused nothing but problems in the past.

I don't think it applies... (I tried on an internal tree) but the affected function changed in 3.13 in various ways. Aneesh, please provide a more details explanation and whether we should backport those other changes too or whether this is not necessary

Yes the affected function added support for split ptl locks for huge pte. I don't think that is a stable material.

.

Quoted 3 lines

BTW. Aneesh, we need a 3.11.x one too

3.11.x it is already applied.

-aneesh

#4 lore raw
In reply to Greg KH
Quoted 23 lines · Greg KH
Quoted 22 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

Quoted 6 lines
Quoted 2 lines

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

Because the code in that function changed in 3.13. Kirill added split ptl locks for huge pte, and we decide whether to withdraw and deposit again based on the ptl locks in 3.13. In 3.12 we do that only for ppc64 using #ifdef

Quoted 4 lines

I _REALLY_ dislike patches that are totally different from Linus's tree in stable trees, it has caused nothing but problems in the past.

-aneesh

#5 lore raw
Quoted 28 lines · Aneesh Kumar K.V
Quoted 26 lines · Greg KH
Quoted 24 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

Quoted 8 lines
Quoted 3 lines

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

Because the code in that function changed in 3.13. Kirill added split ptl locks for huge pte, and we decide whether to withdraw and deposit again based on the ptl locks in 3.13. In 3.12 we do that only for ppc64 using #ifdef

I have no idea what that means...

If you want this patch applied, please be specific as to what is going on, why the code is _very_ different, and all of that. Make it _obvious_ as to what is happening, and why I would be a fool not to take it in the stable tree.

As it is, the code in this patch looks so different that I'm just assuming you got something wrong and are trying to really send me something else, so I'll just ignore it.

greg k-h

#6 lore raw
Quoted 40 lines · Greg KH
Quoted 28 lines · Aneesh Kumar K.V
Quoted 26 lines · Greg KH
Quoted 24 lines · Aneesh Kumar K.V

From: "Aneesh Kumar K.V" <[email protected]>

This patch fix the below crash

NIP [c00000000004cee4] .__hash_page_thp+0x2a4/0x440 LR [c0000000000439ac] .hash_page+0x18c/0x5e0
...

Call Trace:
[c000000736103c40] [00001ffffb000000] 0x1ffffb000000(unreliable)
[437908.479693] [c000000736103d50] [c0000000000439ac] .hash_page+0x18c/0x5e0
[437908.479699] [c000000736103e30] [c00000000000924c] .do_hash_page+0x4c/0x58

On ppc64 we use the pgtable for storing the hpte slot information and store address to the pgtable at a constant offset (PTRS_PER_PMD) from pmd. On mremap, when we switch the pmd, we need to withdraw and deposit the pgtable again, so that we find the pgtable at PTRS_PER_PMD offset from new pmd.

We also want to move the withdraw and deposit before the set_pmd so that, when page fault find the pmd as trans huge we can be sure that pgtable can be located at the offset.

variant of upstream SHA1: b3084f4db3aeb991c507ca774337c7e7893ed04f for 3.12 stable series

Quoted 8 lines
Quoted 3 lines

This doesn't look like a "variant", it looks totally different. Why can't I just take the b3084f4db3aeb991c507ca774337c7e7893ed04f patch (and follow-on fix) for 3.12?

Because the code in that function changed in 3.13. Kirill added split ptl locks for huge pte, and we decide whether to withdraw and deposit again based on the ptl locks in 3.13. In 3.12 we do that only for ppc64 using #ifdef

I have no idea what that means...

Quoted 8 lines

If you want this patch applied, please be specific as to what is going on, why the code is _very_ different, and all of that. Make it _obvious_ as to what is happening, and why I would be a fool not to take it in the stable tree.

As it is, the code in this patch looks so different that I'm just assuming you got something wrong and are trying to really send me something else, so I'll just ignore it.

3.13 we added split huge ptl lock which introduced separate lock at pmd level for hugepage (bf929152e9f6c49b66fad4ebf08cc95b02ce48f5). This required us 3592806cfa08b7cca968f793c33f8e9460bab395. ie, when we move huge page, we need to withdraw and deposit PTE page if we are moving them across different pmd page. We did that by checking spin lock address in 3.13. ie, we have

if (new_ptl != old_ptl) {
.....

       pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
       pgtable_trans_huge_deposit(mm, new_pmd,pgtable);

...
}

ppc64 even without using split ptl had PTE page per pmd entry. The details for that are explained in the commit message above. So when we move huge page we need to withdraw and deposit PTE page always on ppc64.

Now on 3.13 we added a new function which did

static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
spinlock_t *old_pmd_ptl)

{
       /*
        * With split pmd lock we also need to move preallocated
        * PTE page table if new_pmd is on different PMD page table.
        */
       return new_pmd_ptl != old_pmd_ptl;
}

for x86

and on ppc64 we did

static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
spinlock_t *old_pmd_ptl)
{
/*
* Archs like ppc64 use pgtable to store per pmd
* specific information. So when we switch the pmd,
* we should also withdraw and deposit the pgtable */
return true;
}

ie, on ppc64 we always did withdraw and deposit and on x86 we do that only when spin lock address are different.

For 3.12, since we don't have split huge ptl locks yet, we did the below

+#ifdef CONFIG_ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW
+ /*
+ * Archs like ppc64 use pgtable to store per pmd
+ * specific information. So when we switch the pmd,
+ * we should also withdraw and deposit the pgtable
+ */
+ pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
+ pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
+#endif

CONFIG_ARCH_THP_MOVE_PMD_ALWAYS_WITHDRAW is only set for PPC64.

-aneesh

#7 lore raw
In reply to Greg KH
Quoted 11 lines · Greg KH

I have no idea what that means...

If you want this patch applied, please be specific as to what is going on, why the code is _very_ different, and all of that. Make it _obvious_ as to what is happening, and why I would be a fool not to take
it in the stable tree.

As it is, the code in this patch looks so different that I'm just assuming you got something wrong and are trying to really send me something else, so I'll just ignore it.

It looks very different because the function that needs to be fixed changed a lot upstream in 3.13.

In practice it's *not* very different in behaviour. It's just that on powerpc we need to unconditionally call withdraw and deposit when moving PTEs or it will crash, due to how we keep the transparent huge page in sync with the hash table.

With the 3.13 code, due to lock breaking introduced by Kirill in 3.13-rc's, there's already a generic case for doing that (if we dropped the lock). So we just changed the condition to essentially force the condition to true to always do it under control of an arch helper.

The pre-3.13 code didn't do the withdraw and deposit at all in that function however, so in that case, the patch (this 3.12 one) basically just adds the calls to withdraw and deposit under control of an ifdef which is only enabled for powerpc64.

So you are taking 0 risk with other architecture and as the powerpc maintainer I'm happy with the patch.

Cheers,
Ben.

#8 lore raw
Quoted 2 lines · Benjamin Herrenschmidt

It looks very different because the function that needs to be fixed changed a lot upstream in 3.13.

.../...

Hi Greg !

You didn't say if that explanation was to your liking :-)

If it is, do you want Aneesh to re-submit the patch with such an explanation in the changelog ?

Cheers,
Ben.

Quoted 22 lines

In practice it's *not* very different in behaviour. It's just that on powerpc we need to unconditionally call withdraw and deposit when moving PTEs or it will crash, due to how we keep the transparent huge page in sync with the hash table.

With the 3.13 code, due to lock breaking introduced by Kirill in 3.13-rc's, there's already a generic case for doing that (if we dropped the lock). So we just changed the condition to essentially force the condition to true to always do it under control of an arch helper.

The pre-3.13 code didn't do the withdraw and deposit at all in that function however, so in that case, the patch (this 3.12 one) basically just adds the calls to withdraw and deposit under control of an ifdef which is only enabled for powerpc64.

So you are taking 0 risk with other architecture and as the powerpc maintainer I'm happy with the patch.

Cheers,
Ben.