Skip to contentSkip to conversation

powerpc/mm: Fix spurrious segfaults on radix with Autonuma

PATCH Applied Benjamin Herrenschmidt 4 participants 5 messages Feb 3, 2017 – Feb 9, 2017 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc/mm: Fix spurrious segfaults on radix with Autonuma

When autonuma marks a PTE inaccessible it clears all the protection bits but leave the PTE valid.

With the Radix MMU, an attempt at executing from such a PTE will take a fault with bit 35 of SRR1 set "SRR1_ISI_N_OR_G".

It is thus incorrect to treat all such faults as errors. We should pass them to handle_mm_fault() for autonuma to deal with. The case of pages that are really not executable is handled by the existing test for VM_EXEC further down.

That leaves us with catching the kernel attempts at executing user pages. We can catch that earlier, even before we do find_vma.

It is never valid on powerpc for the kernel to take an exec fault to begin with. So fold that test with the existing test for the kernel faulting on kernel addresses to bail out early.

Fixes 1d18ad0 ("powerpc/mm: Detect instruction fetch denied and report")
Fixes 0ab5171 ("powerpc/mm: Fix no execute fault handling on pre-POWER5")

---

arch/powerpc/mm/fault.c +5 -16
index 6fd30ac..62a50d6 100644
@@ -253,8 +253,11 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	if (unlikely(debugger_fault_handler(regs)))
 		goto bail;
 
-	/* On a kernel SLB miss we can only check for a valid exception entry */
-	if (!user_mode(regs) && (address >= TASK_SIZE)) {
+	/*
+	 * The kernel should never take an execute fault nor should it
+	 * take a page fault to a kernel address.
+	 */
+	if (!user_mode(regs) && (is_exec || (address >= TASK_SIZE))) {
 		rc = SIGSEGV;
 		goto bail;
 	}
@@ -391,20 +394,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 
 	if (is_exec) {
 		/*
-		 * An execution fault + no execute ?
-		 *
-		 * On CPUs that don't have CPU_FTR_COHERENT_ICACHE we
-		 * deliberately create NX mappings, and use the fault to do the
-		 * cache flush. This is usually handled in hash_page_do_lazy_icache()
-		 * but we could end up here if that races with a concurrent PTE
-		 * update. In that case we need to fall through here to the VMA
-		 * check below.
-		 */
-		if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
-			(regs->msr & SRR1_ISI_N_OR_G))
-			goto bad_area;
-
-		/*
 		 * Allow execution from readable areas if the MMU does not
 		 * provide separate controls over reading and executing.
 		 *
Reviewed-by on v1 1/1
#1 lore raw
Quoted 14 lines · Benjamin Herrenschmidt

When autonuma marks a PTE inaccessible it clears all the protection bits but leave the PTE valid.

With the Radix MMU, an attempt at executing from such a PTE will take a fault with bit 35 of SRR1 set "SRR1_ISI_N_OR_G".

It is thus incorrect to treat all such faults as errors. We should pass them to handle_mm_fault() for autonuma to deal with. The case of pages that are really not executable is handled by the existing test for VM_EXEC further down.

That leaves us with catching the kernel attempts at executing user pages. We can catch that earlier, even before we do find_vma.

Quoted 7 lines

It is never valid on powerpc for the kernel to take an exec fault to begin with. So fold that test with the existing test for the kernel faulting on kernel addresses to bail out early.

Fixes 1d18ad0 ("powerpc/mm: Detect instruction fetch denied and report")
Fixes 0ab5171 ("powerpc/mm: Fix no execute fault handling on pre-POWER5")
Quoted 41 lines

---

arch/powerpc/mm/fault.c +5 -16
index 6fd30ac..62a50d6 100644
@@ -253,8 +253,11 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	if (unlikely(debugger_fault_handler(regs)))
 		goto bail;
 
-	/* On a kernel SLB miss we can only check for a valid exception entry */
-	if (!user_mode(regs) && (address >= TASK_SIZE)) {
+	/*
+	 * The kernel should never take an execute fault nor should it
+	 * take a page fault to a kernel address.
+	 */
+	if (!user_mode(regs) && (is_exec || (address >= TASK_SIZE))) {
 		rc = SIGSEGV;
 		goto bail;
 	}
@@ -391,20 +394,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 
 	if (is_exec) {
 		/*
-		 * An execution fault + no execute ?
-		 *
-		 * On CPUs that don't have CPU_FTR_COHERENT_ICACHE we
-		 * deliberately create NX mappings, and use the fault to do the
-		 * cache flush. This is usually handled in hash_page_do_lazy_icache()
-		 * but we could end up here if that races with a concurrent PTE
-		 * update. In that case we need to fall through here to the VMA
-		 * check below.
-		 */
-		if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
-			(regs->msr & SRR1_ISI_N_OR_G))
-			goto bad_area;
-
-		/*
 		 * Allow execution from readable areas if the MMU does not
 		 * provide separate controls over reading and executing.
 		 *
#2 lore raw
In reply to Benjamin Herrenschmidt
Quoted 31 lines · Benjamin Herrenschmidt

When autonuma marks a PTE inaccessible it clears all the protection bits but leave the PTE valid.

With the Radix MMU, an attempt at executing from such a PTE will take a fault with bit 35 of SRR1 set "SRR1_ISI_N_OR_G".

It is thus incorrect to treat all such faults as errors. We should pass them to handle_mm_fault() for autonuma to deal with. The case of pages that are really not executable is handled by the existing test for VM_EXEC further down.

That leaves us with catching the kernel attempts at executing user pages. We can catch that earlier, even before we do find_vma.

It is never valid on powerpc for the kernel to take an exec fault to begin with. So fold that test with the existing test for the kernel faulting on kernel addresses to bail out early.

Fixes 1d18ad0 ("powerpc/mm: Detect instruction fetch denied and report")
Fixes 0ab5171 ("powerpc/mm: Fix no execute fault handling on pre-POWER5")

---

arch/powerpc/mm/fault.c +0 -0
index 6fd30ac..62a50d6 100644
@@ -253,8 +253,11 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	if (unlikely(debugger_fault_handler(regs)))
 		goto bail;
 

arch/powerpc/mm/fault.c int do_page_fault(struct pt_regs *regs, unsigned long address, · line 263

Quoted 10 lines
arch/powerpc/mm/fault.c +5 -2
-	/* On a kernel SLB miss we can only check for a valid exception entry */
-	if (!user_mode(regs) && (address >= TASK_SIZE)) {
+	/*
+	 * The kernel should never take an execute fault nor should it
+	 * take a page fault to a kernel address.
+	 */
+	if (!user_mode(regs) && (is_exec || (address >= TASK_SIZE))) {
 		rc = SIGSEGV;
 		goto bail;
 	}

Aneesh did suggest a check for !user_mode(regs), but we did multiple combinations of potential problems and decided the current check was OK. Not yet tested at my end, I'll do that today.

Balbir Singh

Acked-by on v1 1/1
#3 lore raw
In reply to Benjamin Herrenschmidt
Quoted 14 lines · Benjamin Herrenschmidt

When autonuma marks a PTE inaccessible it clears all the protection bits but leave the PTE valid.

With the Radix MMU, an attempt at executing from such a PTE will take a fault with bit 35 of SRR1 set "SRR1_ISI_N_OR_G".

It is thus incorrect to treat all such faults as errors. We should pass them to handle_mm_fault() for autonuma to deal with. The case of pages that are really not executable is handled by the existing test for VM_EXEC further down.

That leaves us with catching the kernel attempts at executing user pages. We can catch that earlier, even before we do find_vma.

Quoted 9 lines

It is never valid on powerpc for the kernel to take an exec fault to begin with. So fold that test with the existing test for the kernel faulting on kernel addresses to bail out early.

Fixes 1d18ad0 ("powerpc/mm: Detect instruction fetch denied and report")
Fixes 0ab5171 ("powerpc/mm: Fix no execute fault handling on pre-POWER5")

---

#4 lore raw
In reply to Benjamin Herrenschmidt
Quoted 18 lines · Benjamin Herrenschmidt

When autonuma marks a PTE inaccessible it clears all the protection bits but leave the PTE valid.

With the Radix MMU, an attempt at executing from such a PTE will take a fault with bit 35 of SRR1 set "SRR1_ISI_N_OR_G".

It is thus incorrect to treat all such faults as errors. We should pass them to handle_mm_fault() for autonuma to deal with. The case of pages that are really not executable is handled by the existing test for VM_EXEC further down.

That leaves us with catching the kernel attempts at executing user pages. We can catch that earlier, even before we do find_vma.

It is never valid on powerpc for the kernel to take an exec fault to begin with. So fold that test with the existing test for the kernel faulting on kernel addresses to bail out early.

Quoted 5 lines
Fixes 1d18ad0 ("powerpc/mm: Detect instruction fetch denied and report")
Fixes 0ab5171 ("powerpc/mm: Fix no execute fault handling on pre-POWER5")

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/d7df2443cd5f67fc6ee7c05a88e499

cheers