[PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Mark McLoughlin]

Hi,
        In Fedora rawhide we're currently seeing oopses whenever
a processes dlopen()s a library which prelink() has previously
relocated above 1Gb. Some more details at:

  https://bugzilla.redhat.com/436453

        The issue seems to be that during xen_pgd_pin(), we only
pin PTE pages, but xen's release_pmd() tries to unpin PMD pages
too.

        The following series of patches is against the latest
x86 tree.

Cheers,
Mark.

[PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Mark McLoughlin]

Hi,
        In Fedora rawhide we're currently seeing oopses whenever
a processes dlopen()s a library which prelink() has previously
relocated above 1Gb. Some more details at:

  https://bugzilla.redhat.com/436453

        The issue seems to be that during xen_pgd_pin(), we only
pin PTE pages, but xen's release_pmd() tries to unpin PMD pages
too.

        The following series of patches is against the latest
x86 tree.

Cheers,
Mark.

[PATCH 1/3] xen: Refactor xen_{alloc,release}_{pte,pmd}()

[Mark McLoughlin]

Use the pt_level enum with generic alloc_ptpage()
and release_ptpage() functions to make it a little
more clear what's going on in next commit.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |   27 ++++++++++++++++++++-------
 arch/x86/xen/mmu.c       |    7 -------
 arch/x86/xen/mmu.h       |    7 +++++++
 3 files changed, 27 insertions(+), 14 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 1e5ea9e..a6a1b40 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -669,10 +669,10 @@ static void xen_release_pte_init(u32 pfn)
 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 }
 
-static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
+static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 {
 	struct mmuext_op op;
-	op.cmd = level;
+	op.cmd = cmd;
 	op.arg1.mfn = pfn_to_mfn(pfn);
 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 		BUG();
@@ -689,7 +689,10 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 		if (!PageHighMem(page)) {
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
-			pin_pagetable_pfn(level, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
+			else if (level == PT_PMD)
+				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -699,16 +702,16 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 static void xen_alloc_pte(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L1_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PTE);
 }
 
 static void xen_alloc_pmd(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L2_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PMD);
 }
 
 /* This should never happen until we're OK to use struct page */
-static void xen_release_pte(u32 pfn)
+static void xen_release_ptpage(u32 pfn, unsigned level)
 {
 	struct page *page = pfn_to_page(pfn);
 
@@ -720,6 +723,16 @@ static void xen_release_pte(u32 pfn)
 	}
 }
 
+static void xen_release_pte(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PTE);
+}
+
+static void xen_release_pmd(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PMD);
+}
+
 #ifdef CONFIG_HIGHPTE
 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 {
@@ -840,7 +853,7 @@ static __init void xen_pagetable_setup_done(pgd_t *base)
 	pv_mmu_ops.alloc_pte = xen_alloc_pte;
 	pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
 	pv_mmu_ops.release_pte = xen_release_pte;
-	pv_mmu_ops.release_pmd = xen_release_pte;
+	pv_mmu_ops.release_pmd = xen_release_pmd;
 	pv_mmu_ops.set_pte = xen_set_pte;
 
 	setup_shared_info();
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index f5bb0be..272635a 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -271,13 +271,6 @@ void xen_set_pte(pte_t *ptep, pte_t pte)
 }
 #endif	/* CONFIG_X86_PAE */
 
-enum pt_level {
-	PT_PGD,
-	PT_PUD,
-	PT_PMD,
-	PT_PTE
-};
-
 /*
   (Yet another) pagetable walker.  This one is intended for pinning a
   pagetable.  This means that it walks a pagetable and calls the
diff --git a/arch/x86/xen/mmu.h b/arch/x86/xen/mmu.h
index c9ff27f..b5e189b 100644
--- a/arch/x86/xen/mmu.h
+++ b/arch/x86/xen/mmu.h
@@ -3,6 +3,13 @@
 #include <linux/linkage.h>
 #include <asm/page.h>
 
+enum pt_level {
+	PT_PGD,
+	PT_PUD,
+	PT_PMD,
+	PT_PTE
+};
+
 /*
  * Page-directory addresses above 4GB do not fit into architectural %cr3.
  * When accessing %cr3, or equivalent field in vcpu_guest_context, guests
-- 
1.5.4.1

[PATCH 1/3] xen: Refactor xen_{alloc, release}_{pte, pmd}()

[Mark McLoughlin]

Use the pt_level enum with generic alloc_ptpage()
and release_ptpage() functions to make it a little
more clear what's going on in next commit.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |   27 ++++++++++++++++++++-------
 arch/x86/xen/mmu.c       |    7 -------
 arch/x86/xen/mmu.h       |    7 +++++++
 3 files changed, 27 insertions(+), 14 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 1e5ea9e..a6a1b40 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -669,10 +669,10 @@ static void xen_release_pte_init(u32 pfn)
 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 }
 
-static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
+static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 {
 	struct mmuext_op op;
-	op.cmd = level;
+	op.cmd = cmd;
 	op.arg1.mfn = pfn_to_mfn(pfn);
 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 		BUG();
@@ -689,7 +689,10 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 		if (!PageHighMem(page)) {
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
-			pin_pagetable_pfn(level, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
+			else if (level == PT_PMD)
+				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -699,16 +702,16 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 static void xen_alloc_pte(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L1_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PTE);
 }
 
 static void xen_alloc_pmd(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L2_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PMD);
 }
 
 /* This should never happen until we're OK to use struct page */
-static void xen_release_pte(u32 pfn)
+static void xen_release_ptpage(u32 pfn, unsigned level)
 {
 	struct page *page = pfn_to_page(pfn);
 
@@ -720,6 +723,16 @@ static void xen_release_pte(u32 pfn)
 	}
 }
 
+static void xen_release_pte(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PTE);
+}
+
+static void xen_release_pmd(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PMD);
+}
+
 #ifdef CONFIG_HIGHPTE
 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 {
@@ -840,7 +853,7 @@ static __init void xen_pagetable_setup_done(pgd_t *base)
 	pv_mmu_ops.alloc_pte = xen_alloc_pte;
 	pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
 	pv_mmu_ops.release_pte = xen_release_pte;
-	pv_mmu_ops.release_pmd = xen_release_pte;
+	pv_mmu_ops.release_pmd = xen_release_pmd;
 	pv_mmu_ops.set_pte = xen_set_pte;
 
 	setup_shared_info();
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index f5bb0be..272635a 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -271,13 +271,6 @@ void xen_set_pte(pte_t *ptep, pte_t pte)
 }
 #endif	/* CONFIG_X86_PAE */
 
-enum pt_level {
-	PT_PGD,
-	PT_PUD,
-	PT_PMD,
-	PT_PTE
-};
-
 /*
   (Yet another) pagetable walker.  This one is intended for pinning a
   pagetable.  This means that it walks a pagetable and calls the
diff --git a/arch/x86/xen/mmu.h b/arch/x86/xen/mmu.h
index c9ff27f..b5e189b 100644
--- a/arch/x86/xen/mmu.h
+++ b/arch/x86/xen/mmu.h
@@ -3,6 +3,13 @@
 #include <linux/linkage.h>
 #include <asm/page.h>
 
+enum pt_level {
+	PT_PGD,
+	PT_PUD,
+	PT_PMD,
+	PT_PTE
+};
+
 /*
  * Page-directory addresses above 4GB do not fit into architectural %cr3.
  * When accessing %cr3, or equivalent field in vcpu_guest_context, guests
-- 
1.5.4.1

[PATCH 2/3] xen: Do not pin/unpin PMD pages

[Mark McLoughlin]

During process startup, in xen_pgd_pin(), we pin PTE pages
but not PMD pages. If we then later map and unmap using an
otherwise unused PMD page, the unpin in release_pmd() fails
because it wasn't previously pinned.

i.e. with this simple test case:

    int fd = open("/dev/zero", O_RDONLY);
    munmap(mmap((void *)0x40000000, 0x1000_LEN, PROT_READ, MAP_PRIVATE, fd, 0), 0x1000);
    close(fd);

we currently get:

   kernel BUG at arch/x86/xen/enlighten.c:678!
   ...
   EIP is at xen_release_pt+0x79/0xa9
   ...
   Call Trace:
    [<c041da25>] ? __pmd_free_tlb+0x1a/0x75
    [<c047a192>] ? free_pgd_range+0x1d2/0x2b5
    [<c047a2f3>] ? free_pgtables+0x7e/0x93
    [<c047b272>] ? unmap_region+0xb9/0xf5
    [<c047c1bd>] ? do_munmap+0x193/0x1f5
    [<c047c24f>] ? sys_munmap+0x30/0x3f
    [<c0408cce>] ? syscall_call+0x7/0xb
    =======================

and xen complains:

  (XEN) mm.c:2241:d4 Mfn 1cc37 not pinned

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index a6a1b40..0ce225f 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -691,8 +691,6 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 			if (level == PT_PTE)
 				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
-			else if (level == PT_PMD)
-				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -717,7 +715,8 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 
 	if (PagePinned(page)) {
 		if (!PageHighMem(page)) {
-			pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
 	}
-- 
1.5.4.1

[PATCH 2/3] xen: Do not pin/unpin PMD pages

[Mark McLoughlin]

During process startup, in xen_pgd_pin(), we pin PTE pages
but not PMD pages. If we then later map and unmap using an
otherwise unused PMD page, the unpin in release_pmd() fails
because it wasn't previously pinned.

i.e. with this simple test case:

    int fd = open("/dev/zero", O_RDONLY);
    munmap(mmap((void *)0x40000000, 0x1000_LEN, PROT_READ, MAP_PRIVATE, fd, 0), 0x1000);
    close(fd);

we currently get:

   kernel BUG at arch/x86/xen/enlighten.c:678!
   ...
   EIP is at xen_release_pt+0x79/0xa9
   ...
   Call Trace:
    [<c041da25>] ? __pmd_free_tlb+0x1a/0x75
    [<c047a192>] ? free_pgd_range+0x1d2/0x2b5
    [<c047a2f3>] ? free_pgtables+0x7e/0x93
    [<c047b272>] ? unmap_region+0xb9/0xf5
    [<c047c1bd>] ? do_munmap+0x193/0x1f5
    [<c047c24f>] ? sys_munmap+0x30/0x3f
    [<c0408cce>] ? syscall_call+0x7/0xb
    =======================

and xen complains:

  (XEN) mm.c:2241:d4 Mfn 1cc37 not pinned

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index a6a1b40..0ce225f 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -691,8 +691,6 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 			if (level == PT_PTE)
 				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
-			else if (level == PT_PMD)
-				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -717,7 +715,8 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 
 	if (PagePinned(page)) {
 		if (!PageHighMem(page)) {
-			pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
 	}
-- 
1.5.4.1

[PATCH 3/3] xen: Clear PG_pinned in release_{pte,pmd}()

[Mark McLoughlin]

We don't currently clear PG_pinned in release_pt(),
probably because the page is to be freed anyway,
but it jumps out as an obvious gap in the logic
so add clear the flag for consistency sake.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 0ce225f..8c5ff24 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -719,6 +719,7 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
+		ClearPagePinned(page);
 	}
 }
 
-- 
1.5.4.1

[PATCH 3/3] xen: Clear PG_pinned in release_{pte, pmd}()

[Mark McLoughlin]

We don't currently clear PG_pinned in release_pt(),
probably because the page is to be freed anyway,
but it jumps out as an obvious gap in the logic
so add clear the flag for consistency sake.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 0ce225f..8c5ff24 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -719,6 +719,7 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
+		ClearPagePinned(page);
 	}
 }
 
-- 
1.5.4.1

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Jeremy Fitzhardinge]

Mark McLoughlin wrote:
> Hi,
>         In Fedora rawhide we're currently seeing oopses whenever
> a processes dlopen()s a library which prelink() has previously
> relocated above 1Gb. Some more details at:
>
>   https://bugzilla.redhat.com/436453
>
>         The issue seems to be that during xen_pgd_pin(), we only
> pin PTE pages, but xen's release_pmd() tries to unpin PMD pages
> too.
>
>         The following series of patches is against the latest
> x86 tree.
>   

Hm, OK.  I'm trying to work out what happened here.  I guess its a 
latent bug which is only visible now that its possible to incrementally 
free pmds.

A side note is that I'm planning on pinning pmds in the future, so that 
we can get away replace the full pgd page per mm with a pgd page per cpu.

    J

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Jeremy Fitzhardinge]

Mark McLoughlin wrote:
> Hi,
>         In Fedora rawhide we're currently seeing oopses whenever
> a processes dlopen()s a library which prelink() has previously
> relocated above 1Gb. Some more details at:
>
>   https://bugzilla.redhat.com/436453
>
>         The issue seems to be that during xen_pgd_pin(), we only
> pin PTE pages, but xen's release_pmd() tries to unpin PMD pages
> too.
>
>         The following series of patches is against the latest
> x86 tree.
>   

Hm, OK.  I'm trying to work out what happened here.  I guess its a 
latent bug which is only visible now that its possible to incrementally 
free pmds.

A side note is that I'm planning on pinning pmds in the future, so that 
we can get away replace the full pgd page per mm with a pgd page per cpu.

    J

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Ingo Molnar]

* Mark McLoughlin <markmc@redhat.com> wrote:

>         In Fedora rawhide we're currently seeing oopses whenever a 
> processes dlopen()s a library which prelink() has previously relocated 
> above 1Gb. Some more details at:
> 
>   https://bugzilla.redhat.com/436453
> 
>         The issue seems to be that during xen_pgd_pin(), we only pin 
> PTE pages, but xen's release_pmd() tries to unpin PMD pages too.
> 
>         The following series of patches is against the latest x86 
> tree.

thanks, i've picked them up to get them tested - but it would be nice to 
have an ack from Jeremy as well. This seems like v2.6.25 material as 
well, right?

	Ingo

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Jeremy Fitzhardinge]

Ingo Molnar wrote:
> * Mark McLoughlin <markmc@redhat.com> wrote:
>
>   
>>         In Fedora rawhide we're currently seeing oopses whenever a 
>> processes dlopen()s a library which prelink() has previously relocated 
>> above 1Gb. Some more details at:
>>
>>   https://bugzilla.redhat.com/436453
>>
>>         The issue seems to be that during xen_pgd_pin(), we only pin 
>> PTE pages, but xen's release_pmd() tries to unpin PMD pages too.
>>
>>         The following series of patches is against the latest x86 
>> tree.
>>     
>
> thanks, i've picked them up to get them tested - but it would be nice to 
> have an ack from Jeremy as well. This seems like v2.6.25 material as 
> well, right?

Didn't I reply?  Anyway, the patch looks good.  It's definitely 2.6.25 
material; its fallout from the changes I did to dynamically 
allocate/deallocate pmds rather than always do them with the pgd.

    J

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Mark McLoughlin]

On Mon, 2008-03-31 at 07:55 -0700, Jeremy Fitzhardinge wrote:
> Ingo Molnar wrote:
> > * Mark McLoughlin <markmc@redhat.com> wrote:
> >   
> >>         In Fedora rawhide we're currently seeing oopses whenever a 
> >> processes dlopen()s a library which prelink() has previously relocated 
> >> above 1Gb. Some more details at:
> >>
> >>   https://bugzilla.redhat.com/436453
> >>
> >>         The issue seems to be that during xen_pgd_pin(), we only pin 
> >> PTE pages, but xen's release_pmd() tries to unpin PMD pages too.
> >>
> >>         The following series of patches is against the latest x86 
> >> tree.
> >>     
> >
> > thanks, i've picked them up to get them tested - but it would be nice to 
> > have an ack from Jeremy as well. This seems like v2.6.25 material as 
> > well, right?
> 
> Didn't I reply?  Anyway, the patch looks good.  It's definitely 2.6.25 
> material; its fallout from the changes I did to dynamically 
> allocate/deallocate pmds rather than always do them with the pgd.

Okay, here are the patches rebased to 2.6.25-rc8.

I've reproduced the issue with plain -rc8 and confirmed the patches fix
it.

Cheers,
Mark.

Re: [PATCH 0/3] xen: Fix oops when mapping/unmapping addr above 1Gb

[Mark McLoughlin]

On Mon, 2008-03-31 at 07:55 -0700, Jeremy Fitzhardinge wrote:
> Ingo Molnar wrote:
> > * Mark McLoughlin <markmc@redhat.com> wrote:
> >   
> >>         In Fedora rawhide we're currently seeing oopses whenever a 
> >> processes dlopen()s a library which prelink() has previously relocated 
> >> above 1Gb. Some more details at:
> >>
> >>   https://bugzilla.redhat.com/436453
> >>
> >>         The issue seems to be that during xen_pgd_pin(), we only pin 
> >> PTE pages, but xen's release_pmd() tries to unpin PMD pages too.
> >>
> >>         The following series of patches is against the latest x86 
> >> tree.
> >>     
> >
> > thanks, i've picked them up to get them tested - but it would be nice to 
> > have an ack from Jeremy as well. This seems like v2.6.25 material as 
> > well, right?
> 
> Didn't I reply?  Anyway, the patch looks good.  It's definitely 2.6.25 
> material; its fallout from the changes I did to dynamically 
> allocate/deallocate pmds rather than always do them with the pgd.

Okay, here are the patches rebased to 2.6.25-rc8.

I've reproduced the issue with plain -rc8 and confirmed the patches fix
it.

Cheers,
Mark.

[PATCH 1/3] xen: Refactor xen_{alloc,release}_{pt,pd}()

[Mark McLoughlin]

Use the pt_level enum with generic alloc_ptpage()
and release_ptpage() functions to make it a little
more clear what's going on in next commit.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |   27 ++++++++++++++++++++-------
 arch/x86/xen/mmu.c       |    7 -------
 arch/x86/xen/mmu.h       |    7 +++++++
 3 files changed, 27 insertions(+), 14 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index de4e6f0..16e2f80 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -667,10 +667,10 @@ static void xen_release_pt_init(u32 pfn)
 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 }
 
-static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
+static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 {
 	struct mmuext_op op;
-	op.cmd = level;
+	op.cmd = cmd;
 	op.arg1.mfn = pfn_to_mfn(pfn);
 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 		BUG();
@@ -687,7 +687,10 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 		if (!PageHighMem(page)) {
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
-			pin_pagetable_pfn(level, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
+			else if (level == PT_PMD)
+				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -697,16 +700,16 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L1_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PTE);
 }
 
 static void xen_alloc_pd(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L2_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PMD);
 }
 
 /* This should never happen until we're OK to use struct page */
-static void xen_release_pt(u32 pfn)
+static void xen_release_ptpage(u32 pfn, unsigned level)
 {
 	struct page *page = pfn_to_page(pfn);
 
@@ -718,6 +721,16 @@ static void xen_release_pt(u32 pfn)
 	}
 }
 
+static void xen_release_pt(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PTE);
+}
+
+static void xen_release_pd(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PMD);
+}
+
 #ifdef CONFIG_HIGHPTE
 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 {
@@ -838,7 +851,7 @@ static __init void xen_pagetable_setup_done(pgd_t *base)
 	pv_mmu_ops.alloc_pt = xen_alloc_pt;
 	pv_mmu_ops.alloc_pd = xen_alloc_pd;
 	pv_mmu_ops.release_pt = xen_release_pt;
-	pv_mmu_ops.release_pd = xen_release_pt;
+	pv_mmu_ops.release_pd = xen_release_pd;
 	pv_mmu_ops.set_pte = xen_set_pte;
 
 	setup_shared_info();
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index 0144395..2a054ef 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -310,13 +310,6 @@ pgd_t xen_make_pgd(unsigned long pgd)
 }
 #endif	/* CONFIG_X86_PAE */
 
-enum pt_level {
-	PT_PGD,
-	PT_PUD,
-	PT_PMD,
-	PT_PTE
-};
-
 /*
   (Yet another) pagetable walker.  This one is intended for pinning a
   pagetable.  This means that it walks a pagetable and calls the
diff --git a/arch/x86/xen/mmu.h b/arch/x86/xen/mmu.h
index c9ff27f..b5e189b 100644
--- a/arch/x86/xen/mmu.h
+++ b/arch/x86/xen/mmu.h
@@ -3,6 +3,13 @@
 #include <linux/linkage.h>
 #include <asm/page.h>
 
+enum pt_level {
+	PT_PGD,
+	PT_PUD,
+	PT_PMD,
+	PT_PTE
+};
+
 /*
  * Page-directory addresses above 4GB do not fit into architectural %cr3.
  * When accessing %cr3, or equivalent field in vcpu_guest_context, guests
-- 
1.5.4.5

[PATCH 1/3] xen: Refactor xen_{alloc, release}_{pt, pd}()

[Mark McLoughlin]

Use the pt_level enum with generic alloc_ptpage()
and release_ptpage() functions to make it a little
more clear what's going on in next commit.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |   27 ++++++++++++++++++++-------
 arch/x86/xen/mmu.c       |    7 -------
 arch/x86/xen/mmu.h       |    7 +++++++
 3 files changed, 27 insertions(+), 14 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index de4e6f0..16e2f80 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -667,10 +667,10 @@ static void xen_release_pt_init(u32 pfn)
 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 }
 
-static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
+static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 {
 	struct mmuext_op op;
-	op.cmd = level;
+	op.cmd = cmd;
 	op.arg1.mfn = pfn_to_mfn(pfn);
 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 		BUG();
@@ -687,7 +687,10 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 		if (!PageHighMem(page)) {
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
-			pin_pagetable_pfn(level, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
+			else if (level == PT_PMD)
+				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -697,16 +700,16 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 
 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L1_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PTE);
 }
 
 static void xen_alloc_pd(struct mm_struct *mm, u32 pfn)
 {
-	xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L2_TABLE);
+	xen_alloc_ptpage(mm, pfn, PT_PMD);
 }
 
 /* This should never happen until we're OK to use struct page */
-static void xen_release_pt(u32 pfn)
+static void xen_release_ptpage(u32 pfn, unsigned level)
 {
 	struct page *page = pfn_to_page(pfn);
 
@@ -718,6 +721,16 @@ static void xen_release_pt(u32 pfn)
 	}
 }
 
+static void xen_release_pt(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PTE);
+}
+
+static void xen_release_pd(u32 pfn)
+{
+	xen_release_ptpage(pfn, PT_PMD);
+}
+
 #ifdef CONFIG_HIGHPTE
 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 {
@@ -838,7 +851,7 @@ static __init void xen_pagetable_setup_done(pgd_t *base)
 	pv_mmu_ops.alloc_pt = xen_alloc_pt;
 	pv_mmu_ops.alloc_pd = xen_alloc_pd;
 	pv_mmu_ops.release_pt = xen_release_pt;
-	pv_mmu_ops.release_pd = xen_release_pt;
+	pv_mmu_ops.release_pd = xen_release_pd;
 	pv_mmu_ops.set_pte = xen_set_pte;
 
 	setup_shared_info();
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index 0144395..2a054ef 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -310,13 +310,6 @@ pgd_t xen_make_pgd(unsigned long pgd)
 }
 #endif	/* CONFIG_X86_PAE */
 
-enum pt_level {
-	PT_PGD,
-	PT_PUD,
-	PT_PMD,
-	PT_PTE
-};
-
 /*
   (Yet another) pagetable walker.  This one is intended for pinning a
   pagetable.  This means that it walks a pagetable and calls the
diff --git a/arch/x86/xen/mmu.h b/arch/x86/xen/mmu.h
index c9ff27f..b5e189b 100644
--- a/arch/x86/xen/mmu.h
+++ b/arch/x86/xen/mmu.h
@@ -3,6 +3,13 @@
 #include <linux/linkage.h>
 #include <asm/page.h>
 
+enum pt_level {
+	PT_PGD,
+	PT_PUD,
+	PT_PMD,
+	PT_PTE
+};
+
 /*
  * Page-directory addresses above 4GB do not fit into architectural %cr3.
  * When accessing %cr3, or equivalent field in vcpu_guest_context, guests
-- 
1.5.4.5

[PATCH 2/3] xen: Do not pin/unpin PMD pages

[Mark McLoughlin]

During process startup, in xen_pgd_pin(), we pin PTE pages
but not PMD pages. If we then later map and unmap using an
otherwise unused PMD page, the unpin in release_pd() fails
because it wasn't previously pinned.

i.e. with this simple test case:

    int fd = open("/dev/zero", O_RDONLY);
    munmap(mmap((void *)0x40000000, 0x1000_LEN, PROT_READ, MAP_PRIVATE, fd, 0), 0x1000);
    close(fd);

we currently get:

   kernel BUG at arch/x86/xen/enlighten.c:678!
   ...
   EIP is at xen_release_pt+0x79/0xa9
   ...
   Call Trace:
    [<c041da25>] ? __pmd_free_tlb+0x1a/0x75
    [<c047a192>] ? free_pgd_range+0x1d2/0x2b5
    [<c047a2f3>] ? free_pgtables+0x7e/0x93
    [<c047b272>] ? unmap_region+0xb9/0xf5
    [<c047c1bd>] ? do_munmap+0x193/0x1f5
    [<c047c24f>] ? sys_munmap+0x30/0x3f
    [<c0408cce>] ? syscall_call+0x7/0xb
    =======================

and xen complains:

  (XEN) mm.c:2241:d4 Mfn 1cc37 not pinned

Further details at:

  https://bugzilla.redhat.com/436453

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 16e2f80..f16b056 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -689,8 +689,6 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 			if (level == PT_PTE)
 				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
-			else if (level == PT_PMD)
-				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -715,7 +713,8 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 
 	if (PagePinned(page)) {
 		if (!PageHighMem(page)) {
-			pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
 	}
-- 
1.5.4.5

[PATCH 2/3] xen: Do not pin/unpin PMD pages

[Mark McLoughlin]

During process startup, in xen_pgd_pin(), we pin PTE pages
but not PMD pages. If we then later map and unmap using an
otherwise unused PMD page, the unpin in release_pd() fails
because it wasn't previously pinned.

i.e. with this simple test case:

    int fd = open("/dev/zero", O_RDONLY);
    munmap(mmap((void *)0x40000000, 0x1000_LEN, PROT_READ, MAP_PRIVATE, fd, 0), 0x1000);
    close(fd);

we currently get:

   kernel BUG at arch/x86/xen/enlighten.c:678!
   ...
   EIP is at xen_release_pt+0x79/0xa9
   ...
   Call Trace:
    [<c041da25>] ? __pmd_free_tlb+0x1a/0x75
    [<c047a192>] ? free_pgd_range+0x1d2/0x2b5
    [<c047a2f3>] ? free_pgtables+0x7e/0x93
    [<c047b272>] ? unmap_region+0xb9/0xf5
    [<c047c1bd>] ? do_munmap+0x193/0x1f5
    [<c047c24f>] ? sys_munmap+0x30/0x3f
    [<c0408cce>] ? syscall_call+0x7/0xb
    =======================

and xen complains:

  (XEN) mm.c:2241:d4 Mfn 1cc37 not pinned

Further details at:

  https://bugzilla.redhat.com/436453

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 16e2f80..f16b056 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -689,8 +689,6 @@ static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 			make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 			if (level == PT_PTE)
 				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
-			else if (level == PT_PMD)
-				pin_pagetable_pfn(MMUEXT_PIN_L2_TABLE, pfn);
 		} else
 			/* make sure there are no stray mappings of
 			   this page */
@@ -715,7 +713,8 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 
 	if (PagePinned(page)) {
 		if (!PageHighMem(page)) {
-			pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
+			if (level == PT_PTE)
+				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
 	}
-- 
1.5.4.5

[PATCH 3/3] xen: Clear PG_pinned in release_{pt,pd}()

[Mark McLoughlin]

We don't currently clear PG_pinned in release_pt(),
probably because the page is to be freed anyway,
but it jumps out as an obvious gap in the logic
so add clear the flag for consistency sake.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
---
 arch/x86/xen/enlighten.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index f16b056..27ee26a 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -717,6 +717,7 @@ static void xen_release_ptpage(u32 pfn, unsigned level)
 				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 		}
+		ClearPagePinned(page);
 	}
 }
 
-- 
1.5.4.5