Skip to contentSkip to conversation

mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

PATCH Boris Ostrovsky 6 participants 27 messages Feb 26, 2013 – Feb 28, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw
Quoted 28 lines · ----- [email protected] wrote:
Quoted 2 lines · Boris Ostrovsky

When CONFIG_DEBUG_PAGEALLOC is set page table updates made by kernel_map_pages() are not made visible (via TLB flush) immediately

if lazy

Quoted 1 line

MMU is on. In environments that support lazy MMU (e.g. Xen) this may

lead to

Quoted 1 line

fatal page faults, for example, when zap_pte_range() needs to

allocate pages

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int

Quoted 12 lines

in __tlb_remove_page() -> tlb_next_batch().

---
 arch/x86/mm/pageattr.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/mm/pageattr.c +0 -0
index ca1f1c2..7b3216e 100644
@@ -1369,6 +1369,8 @@ void kernel_map_pages(struct page *page, int

numpages, int enable)

Quoted 7 lines
arch/x86/mm/pageattr.c +2 -0
 	 * but that can deadlock->flush only current cpu:
 	 */
 	__flush_tlb_all();
+
+	arch_flush_lazy_mmu_mode();
 }
 
Quoted 5 lines

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int · line 1376

Quoted 2 lines
arch/x86/mm/pageattr.c +0 -0
 #ifdef CONFIG_HIBERNATION

This sounds like a critical fix, i.e. a -stable candidate. Am I correct?

I considered copying stable but then I decided that this is a debugging feature --- kernel_map_pages() is only defined if CONFIG_DEBUG_PAGEALLOC is set and my thinking was that stable kernels usually don't do this.

-boris

#1 lore raw

Greg, policy opinion?

-hpa

Quoted 44 lines · Boris Ostrovsky
Quoted 28 lines · ----- [email protected] wrote:
Quoted 2 lines · Boris Ostrovsky

When CONFIG_DEBUG_PAGEALLOC is set page table updates made by kernel_map_pages() are not made visible (via TLB flush) immediately

if lazy

Quoted 1 line

MMU is on. In environments that support lazy MMU (e.g. Xen) this may

lead to

Quoted 1 line

fatal page faults, for example, when zap_pte_range() needs to

allocate pages

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int

Quoted 12 lines

in __tlb_remove_page() -> tlb_next_batch().

---
 arch/x86/mm/pageattr.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/mm/pageattr.c +0 -0
index ca1f1c2..7b3216e 100644
@@ -1369,6 +1369,8 @@ void kernel_map_pages(struct page *page, int

numpages, int enable)

Quoted 7 lines
arch/x86/mm/pageattr.c +2 -0
 	 * but that can deadlock->flush only current cpu:
 	 */
 	__flush_tlb_all();
+
+	arch_flush_lazy_mmu_mode();
 }
 
Quoted 5 lines

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int · line 1376

Quoted 2 lines
arch/x86/mm/pageattr.c +0 -0
 #ifdef CONFIG_HIBERNATION

This sounds like a critical fix, i.e. a -stable candidate. Am I correct?

I considered copying stable but then I decided that this is a debugging feature --- kernel_map_pages() is only defined if CONFIG_DEBUG_PAGEALLOC is set and my thinking was that stable kernels usually don't do this.

-boris

#2 lore raw
In reply to Boris Ostrovsky

Greg, policy opinion?

-hpa

Quoted 44 lines · Boris Ostrovsky
Quoted 28 lines · ----- [email protected] wrote:
Quoted 2 lines · Boris Ostrovsky

When CONFIG_DEBUG_PAGEALLOC is set page table updates made by kernel_map_pages() are not made visible (via TLB flush) immediately

if lazy

Quoted 1 line

MMU is on. In environments that support lazy MMU (e.g. Xen) this may

lead to

Quoted 1 line

fatal page faults, for example, when zap_pte_range() needs to

allocate pages

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int

Quoted 12 lines

in __tlb_remove_page() -> tlb_next_batch().

---
 arch/x86/mm/pageattr.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/mm/pageattr.c +0 -0
index ca1f1c2..7b3216e 100644
@@ -1369,6 +1369,8 @@ void kernel_map_pages(struct page *page, int

numpages, int enable)

Quoted 7 lines
arch/x86/mm/pageattr.c +2 -0
 	 * but that can deadlock->flush only current cpu:
 	 */
 	__flush_tlb_all();
+
+	arch_flush_lazy_mmu_mode();
 }
 
Quoted 5 lines

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int · line 1376

Quoted 2 lines
arch/x86/mm/pageattr.c +0 -0
 #ifdef CONFIG_HIBERNATION

This sounds like a critical fix, i.e. a -stable candidate. Am I correct?

I considered copying stable but then I decided that this is a debugging feature --- kernel_map_pages() is only defined if CONFIG_DEBUG_PAGEALLOC is set and my thinking was that stable kernels usually don't do this.

-boris

#3 lore raw
Quoted 45 lines · H. Peter Anvin

Greg, policy opinion?

-hpa

Quoted 40 lines · Boris Ostrovsky
Quoted 28 lines · ----- [email protected] wrote:
Quoted 2 lines · Boris Ostrovsky

When CONFIG_DEBUG_PAGEALLOC is set page table updates made by kernel_map_pages() are not made visible (via TLB flush) immediately

if lazy

Quoted 1 line

MMU is on. In environments that support lazy MMU (e.g. Xen) this may

lead to

Quoted 1 line

fatal page faults, for example, when zap_pte_range() needs to

allocate pages

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int

Quoted 12 lines

in __tlb_remove_page() -> tlb_next_batch().

---
 arch/x86/mm/pageattr.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/mm/pageattr.c +0 -0
index ca1f1c2..7b3216e 100644
@@ -1369,6 +1369,8 @@ void kernel_map_pages(struct page *page, int

numpages, int enable)

Quoted 7 lines
arch/x86/mm/pageattr.c +2 -0
 	 * but that can deadlock->flush only current cpu:
 	 */
 	__flush_tlb_all();
+
+	arch_flush_lazy_mmu_mode();
 }
 
Quoted 5 lines

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int · line 1376

Quoted 2 lines
arch/x86/mm/pageattr.c +0 -0
 #ifdef CONFIG_HIBERNATION

This sounds like a critical fix, i.e. a -stable candidate. Am I correct?

I considered copying stable but then I decided that this is a debugging feature --- kernel_map_pages() is only defined if CONFIG_DEBUG_PAGEALLOC is set and my thinking was that stable kernels usually don't do this.

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

thanks,

greg k-h

#4 lore raw
In reply to H. Peter Anvin
Quoted 45 lines · H. Peter Anvin

Greg, policy opinion?

-hpa

Quoted 40 lines · Boris Ostrovsky
Quoted 28 lines · ----- [email protected] wrote:
Quoted 2 lines · Boris Ostrovsky

When CONFIG_DEBUG_PAGEALLOC is set page table updates made by kernel_map_pages() are not made visible (via TLB flush) immediately

if lazy

Quoted 1 line

MMU is on. In environments that support lazy MMU (e.g. Xen) this may

lead to

Quoted 1 line

fatal page faults, for example, when zap_pte_range() needs to

allocate pages

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int

Quoted 12 lines

in __tlb_remove_page() -> tlb_next_batch().

---
 arch/x86/mm/pageattr.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/mm/pageattr.c +0 -0
index ca1f1c2..7b3216e 100644
@@ -1369,6 +1369,8 @@ void kernel_map_pages(struct page *page, int

numpages, int enable)

Quoted 7 lines
arch/x86/mm/pageattr.c +2 -0
 	 * but that can deadlock->flush only current cpu:
 	 */
 	__flush_tlb_all();
+
+	arch_flush_lazy_mmu_mode();
 }
 
Quoted 5 lines

arch/x86/mm/pageattr.c void kernel_map_pages(struct page *page, int · line 1376

Quoted 2 lines
arch/x86/mm/pageattr.c +0 -0
 #ifdef CONFIG_HIBERNATION

This sounds like a critical fix, i.e. a -stable candidate. Am I correct?

I considered copying stable but then I decided that this is a debugging feature --- kernel_map_pages() is only defined if CONFIG_DEBUG_PAGEALLOC is set and my thinking was that stable kernels usually don't do this.

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

thanks,

greg k-h

#5 lore raw
Quoted 7 lines · Greg KH

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

OK. I will queue it up in the next fixes (tip:x86/urgent) batch to Linus.

-hpa

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

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

OK. I will queue it up in the next fixes (tip:x86/urgent) batch to Linus.

-hpa

#7 lore raw

Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 10 lines · H. Peter Anvin
Quoted 7 lines · Greg KH

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

OK. I will queue it up in the next fixes (tip:x86/urgent) batch to Linus.

Thank you.

Could you also consider this one (I CC-ed Ingo on it but never got any response):

Quoted 1 line

From a6ed4a88eff4f6329bb4acae3372cccc8a8367d5 Mon Sep 17 00:00:00 2001

From: Samu Kallio <[email protected]>
Date: Sun, 17 Feb 2013 02:35:52 +0000
Subject: [PATCH] x86: mm: Fix vmalloc_fault oops during lazy MMU updates.

In paravirtualized x86_64 kernels, vmalloc_fault may cause an oops when lazy MMU updates are enabled, because set_pgd effects are being deferred.

One instance of this problem is during process mm cleanup with memory cgroups enabled. The chain of events is as follows:

- zap_pte_range enables lazy MMU updates
- zap_pte_range eventually calls mem_cgroup_charge_statistics, which accesses the vmalloc'd mem_cgroup per-cpu stat area
- vmalloc_fault is triggered which tries to sync the corresponding PGD entry with set_pgd, but the update is deferred
- vmalloc_fault oopses due to a mismatch in the PUD entries

The OOPs usually looks as so:

------------[ cut here ]------------
kernel BUG at arch/x86/mm/fault.c:396!

invalid opcode: 0000 [#1] SMP
.. snip ..
CPU 1
Pid: 10866, comm: httpd Not tainted 3.6.10-4.fc18.x86_64 #1

RIP: e030:[<ffffffff816271bf>]  [<ffffffff816271bf>] vmalloc_fault+0x11f/0x208
.. snip ..
Call Trace:
 [<ffffffff81627759>] do_page_fault+0x399/0x4b0
 [<ffffffff81004f4c>] ? xen_mc_extend_args+0xec/0x110
 [<ffffffff81624065>] page_fault+0x25/0x30
 [<ffffffff81184d03>] ? mem_cgroup_charge_statistics.isra.13+0x13/0x50
 [<ffffffff81186f78>] __mem_cgroup_uncharge_common+0xd8/0x350
 [<ffffffff8118aac7>] mem_cgroup_uncharge_page+0x57/0x60
 [<ffffffff8115fbc0>] page_remove_rmap+0xe0/0x150
 [<ffffffff8115311a>] ? vm_normal_page+0x1a/0x80
 [<ffffffff81153e61>] unmap_single_vma+0x531/0x870
 [<ffffffff81154962>] unmap_vmas+0x52/0xa0
 [<ffffffff81007442>] ? pte_mfn_to_pfn+0x72/0x100
 [<ffffffff8115c8f8>] exit_mmap+0x98/0x170
 [<ffffffff810050d9>] ? __raw_callee_save_xen_pmd_val+0x11/0x1e
 [<ffffffff81059ce3>] mmput+0x83/0xf0
 [<ffffffff810624c4>] exit_mm+0x104/0x130
 [<ffffffff8106264a>] do_exit+0x15a/0x8c0
 [<ffffffff810630ff>] do_group_exit+0x3f/0xa0
 [<ffffffff81063177>] sys_exit_group+0x17/0x20
 [<ffffffff8162bae9>] system_call_fastpath+0x16/0x1b

Calling arch_flush_lazy_mmu_mode immediately after set_pgd makes the changes visible to the consistency checks.

RedHat-Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=914737

---
 arch/x86/mm/fault.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/fault.c +4 -2
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();
+	} else {
 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
+	}
 
 	/*
 	 * Below here mismatches are bugs because these lower tables
Signature · 1 line
1.8.0.2
Quoted 4 lines

-hpa

#8 lore raw

Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to H. Peter Anvin
Quoted 10 lines · H. Peter Anvin
Quoted 7 lines · Greg KH

"Stable" kernels are used all over the place, like in distros, which might enable this.

I have no objection to taking this patch in a stable release, as it does fix a real problem.

OK. I will queue it up in the next fixes (tip:x86/urgent) batch to Linus.

Thank you.

Could you also consider this one (I CC-ed Ingo on it but never got any response):

Quoted 1 line

From a6ed4a88eff4f6329bb4acae3372cccc8a8367d5 Mon Sep 17 00:00:00 2001

From: Samu Kallio <[email protected]>
Date: Sun, 17 Feb 2013 02:35:52 +0000
Subject: [PATCH] x86: mm: Fix vmalloc_fault oops during lazy MMU updates.

In paravirtualized x86_64 kernels, vmalloc_fault may cause an oops when lazy MMU updates are enabled, because set_pgd effects are being deferred.

One instance of this problem is during process mm cleanup with memory cgroups enabled. The chain of events is as follows:

- zap_pte_range enables lazy MMU updates
- zap_pte_range eventually calls mem_cgroup_charge_statistics, which accesses the vmalloc'd mem_cgroup per-cpu stat area
- vmalloc_fault is triggered which tries to sync the corresponding PGD entry with set_pgd, but the update is deferred
- vmalloc_fault oopses due to a mismatch in the PUD entries

The OOPs usually looks as so:

------------[ cut here ]------------
kernel BUG at arch/x86/mm/fault.c:396!

invalid opcode: 0000 [#1] SMP
.. snip ..
CPU 1
Pid: 10866, comm: httpd Not tainted 3.6.10-4.fc18.x86_64 #1

RIP: e030:[<ffffffff816271bf>]  [<ffffffff816271bf>] vmalloc_fault+0x11f/0x208
.. snip ..
Call Trace:
 [<ffffffff81627759>] do_page_fault+0x399/0x4b0
 [<ffffffff81004f4c>] ? xen_mc_extend_args+0xec/0x110
 [<ffffffff81624065>] page_fault+0x25/0x30
 [<ffffffff81184d03>] ? mem_cgroup_charge_statistics.isra.13+0x13/0x50
 [<ffffffff81186f78>] __mem_cgroup_uncharge_common+0xd8/0x350
 [<ffffffff8118aac7>] mem_cgroup_uncharge_page+0x57/0x60
 [<ffffffff8115fbc0>] page_remove_rmap+0xe0/0x150
 [<ffffffff8115311a>] ? vm_normal_page+0x1a/0x80
 [<ffffffff81153e61>] unmap_single_vma+0x531/0x870
 [<ffffffff81154962>] unmap_vmas+0x52/0xa0
 [<ffffffff81007442>] ? pte_mfn_to_pfn+0x72/0x100
 [<ffffffff8115c8f8>] exit_mmap+0x98/0x170
 [<ffffffff810050d9>] ? __raw_callee_save_xen_pmd_val+0x11/0x1e
 [<ffffffff81059ce3>] mmput+0x83/0xf0
 [<ffffffff810624c4>] exit_mm+0x104/0x130
 [<ffffffff8106264a>] do_exit+0x15a/0x8c0
 [<ffffffff810630ff>] do_group_exit+0x3f/0xa0
 [<ffffffff81063177>] sys_exit_group+0x17/0x20
 [<ffffffff8162bae9>] system_call_fastpath+0x16/0x1b

Calling arch_flush_lazy_mmu_mode immediately after set_pgd makes the changes visible to the consistency checks.

RedHat-Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=914737

---
 arch/x86/mm/fault.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/fault.c +4 -2
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();
+	} else {
 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
+	}
 
 	/*
 	 * Below here mismatches are bugs because these lower tables
Signature · 1 line
1.8.0.2
Quoted 4 lines

-hpa

maintainer
#9 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

arch/x86/mm/fault.c static noinline __kprobes int vmalloc_fault(unsigned long address) · line 383

Quoted 13 lines · Konrad Rzeszutek Wilk
arch/x86/mm/fault.c +2 -2
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();

Do I understand it correctly that this would cost us a "preempt_disable(); preempt_enable()" needlessly on baremetal when running with CONFIG_PARAVIRT enabled?

Thanks.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
maintainer
#10 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Konrad Rzeszutek Wilk

arch/x86/mm/fault.c static noinline __kprobes int vmalloc_fault(unsigned long address) · line 383

Quoted 13 lines · Konrad Rzeszutek Wilk
arch/x86/mm/fault.c +2 -2
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();

Do I understand it correctly that this would cost us a "preempt_disable(); preempt_enable()" needlessly on baremetal when running with CONFIG_PARAVIRT enabled?

Thanks.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#11 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 9 lines · Borislav Petkov
Quoted 8 lines · Konrad Rzeszutek Wilk
arch/x86/mm/fault.c +0 -0
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
Quoted 10 lines

arch/x86/mm/fault.c static noinline __kprobes int vmalloc_fault(unsigned long address) · line 383

Quoted 5 lines
arch/x86/mm/fault.c +2 -2
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();

Do I understand it correctly that this would cost us a "preempt_disable(); preempt_enable()" needlessly on baremetal when running with CONFIG_PARAVIRT enabled?

OK, this is mind-boggingly crazy and these things that just drives me batty about PV.

arch_flush_lazy_mmu_mode() involves a function call and preemption off *before it even tests for being a noop*. On bare metal this function doesn't do anything, and cannot do anything.

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#12 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Borislav Petkov
Quoted 9 lines · Borislav Petkov
Quoted 8 lines · Konrad Rzeszutek Wilk
arch/x86/mm/fault.c +0 -0
index fb674fd..4f7d793 100644
@@ -378,10 +378,12 @@ static noinline __kprobes int vmalloc_fault(unsigned long address)
 	if (pgd_none(*pgd_ref))
 		return -1;
 
Quoted 10 lines

arch/x86/mm/fault.c static noinline __kprobes int vmalloc_fault(unsigned long address) · line 383

Quoted 5 lines
arch/x86/mm/fault.c +2 -2
-	if (pgd_none(*pgd))
+	if (pgd_none(*pgd)) {
 		set_pgd(pgd, *pgd_ref);
-	else
+		arch_flush_lazy_mmu_mode();

Do I understand it correctly that this would cost us a "preempt_disable(); preempt_enable()" needlessly on baremetal when running with CONFIG_PARAVIRT enabled?

OK, this is mind-boggingly crazy and these things that just drives me batty about PV.

arch_flush_lazy_mmu_mode() involves a function call and preemption off *before it even tests for being a noop*. On bare metal this function doesn't do anything, and cannot do anything.

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
maintainer
#13 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
maintainer
#14 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to H. Peter Anvin
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#15 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 6 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

-boris

#16 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Borislav Petkov
Quoted 6 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

-boris

maintainer
#17 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 9 lines · Boris Ostrovsky
Quoted 6 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
maintainer
#18 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Boris Ostrovsky
Quoted 9 lines · Boris Ostrovsky
Quoted 6 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#19 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 5 lines · Borislav Petkov
Quoted 4 lines · Boris Ostrovsky
Quoted 3 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

That makes more sense, I think... although of course, there is a cost in complexity.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#20 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Borislav Petkov
Quoted 5 lines · Borislav Petkov
Quoted 4 lines · Boris Ostrovsky
Quoted 3 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

... or, patch it out with the alternatives on baremetal, as Steven suggested.

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

That makes more sense, I think... although of course, there is a cost in complexity.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#21 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Borislav Petkov
Quoted 7 lines · Borislav Petkov
Quoted 6 lines · Boris Ostrovsky
Quoted 5 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

What was the suggestion exactly? I don't remember seeing that message.

-boris

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

#22 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Borislav Petkov
Quoted 7 lines · Borislav Petkov
Quoted 6 lines · Boris Ostrovsky
Quoted 5 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

What was the suggestion exactly? I don't remember seeing that message.

-boris

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

maintainer
#23 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

Quoted 1 line · Boris Ostrovsky

What was the suggestion exactly? I don't remember seeing that message.

Use the paravirt patching facility to replace arch_flush_lazy_mmu_mode() with a nop when running on baremetal. I.e., the whole functionality around apply_paravirt() et al.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
maintainer
#24 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Boris Ostrovsky
Quoted 1 line · Boris Ostrovsky

What was the suggestion exactly? I don't remember seeing that message.

Use the paravirt patching facility to replace arch_flush_lazy_mmu_mode() with a nop when running on baremetal. I.e., the whole functionality around apply_paravirt() et al.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#25 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Boris Ostrovsky
Quoted 10 lines · Boris Ostrovsky
Quoted 7 lines · Borislav Petkov
Quoted 6 lines · Boris Ostrovsky
Quoted 5 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

What was the suggestion exactly? I don't remember seeing that message.

Yeah, Borislav talked to me privately on IRC about this and I pointed him to the apply_paravirt() function in arch/x86/kernel/alternative.c where things that apply only to paravirt get patched out on baremetal.

It may add complexity, but there's a method for doing it and I rather not burden baremetal for pravirt nonsense. I know adding a simple call to preempt_count() can show a noticable impact to function tracing. It requires referencing the gs segment register and doing some offset games (as it's stored as a per cpu pointer) to find the stack.

I was actually a bit amazed that it had as big of an impact as it did. I can understand why Christoph Lameter tried hard not to add a preempt_disable() in his code for just a tiny location.

-- Steve

Quoted 8 lines

-boris

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?

#26 lore raw

Re: Is: x86: mm: Fix vmalloc_fault oops during lazy MMU updates Was: Re: [PATCH] mm/x86: Flush lazy MMU when DEBUG_PAGEALLOC is set

In reply to Boris Ostrovsky
Quoted 10 lines · Boris Ostrovsky
Quoted 7 lines · Borislav Petkov
Quoted 6 lines · Boris Ostrovsky
Quoted 5 lines · Borislav Petkov
Quoted 2 lines · H. Peter Anvin

At the very least we should have an early filter for the **COMMON!** case that we are not on a PV platform.

... or, patch it out with the alternatives on baremetal, as Steven suggested.

What was the suggestion exactly? I don't remember seeing that message.

Yeah, Borislav talked to me privately on IRC about this and I pointed him to the apply_paravirt() function in arch/x86/kernel/alternative.c where things that apply only to paravirt get patched out on baremetal.

It may add complexity, but there's a method for doing it and I rather not burden baremetal for pravirt nonsense. I know adding a simple call to preempt_count() can show a noticable impact to function tracing. It requires referencing the gs segment register and doing some offset games (as it's stored as a per cpu pointer) to find the stack.

I was actually a bit amazed that it had as big of an impact as it did. I can understand why Christoph Lameter tried hard not to add a preempt_disable() in his code for just a tiny location.

-- Steve

Quoted 8 lines

-boris

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

I think making a check for paravirt_enabled() is safe enough. I'll send a patch shortly.

Why not make it absolutely for free?