Kernel 4.15 lost set_robust_list support on POWER 9

I get this:

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)

The system call works on 4.14.  Looks like the probing for 
futex_cmpxchg_enabled goes wrong.

Sorry, I have no idea where to start digging.  I don't see anything 
obvious in dmesg.

I'm trying to revert

commit 30d6e0a4190d37740e9447e4e4815f06992dd8c3
Author: Jiri Slaby [off-list ref]
Date:   Thu Aug 24 09:31:05 2017 +0200

     futex: Remove duplicated code and fix undefined behaviour

to see if it makes a difference.

Thanks,
Florian

On 02/05/2018 01:48 PM, Florian Weimer wrote:

I get this:

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)

The system call works on 4.14.  Looks like the probing for 
futex_cmpxchg_enabled goes wrong.

Sorry, I have no idea where to start digging.  I don't see anything 
obvious in dmesg.

I'm trying to revert

commit 30d6e0a4190d37740e9447e4e4815f06992dd8c3
Author: Jiri Slaby [off-list ref]
Date:   Thu Aug 24 09:31:05 2017 +0200

     futex: Remove duplicated code and fix undefined behaviour

to see if it makes a difference

Never mind, it must be something else because that commit is in 4.14, 
but set_robust_list is still working there.  (Let's hope that the 
probing doesn't fail randomly …)

Thanks,
Florian

Nick, Michael,

On 02/05/2018 10:48 AM, Florian Weimer wrote:

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)

The regression was introduced by commit 371b8044 ("powerpc/64s: 
Initialize ISAv3 MMU registers before setting partition table").

The problem is Radix MMU specific (does not occur with 'disable_radix'),
and does not occur with that code reverted (ie do not set PIDR to zero).

Do you see any reasons why?
(wondering if at all related to access_ok() in include/asm/uaccess.h)

with:

     # strace -e set_robust_list -f ./test
     set_robust_list(0x7fffa4b03910, 24)     = -1 ENOSYS (Function not 
implemented)
     +++ exited with 1 +++

     # uname -r
     4.15.0

without:

     # strace -e set_robust_list -f ./test
     set_robust_list(0x7fff889c3910, 24)     = 0
     +++ exited with 0 +++

     # uname -r
     4.15.0.nopidr

On Mon, 2018-02-05 at 19:14 -0200, Mauricio Faria de Oliveira wrote:

Nick, Michael,

+Aneesh.

On 02/05/2018 10:48 AM, Florian Weimer wrote:

quoted

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)

The regression was introduced by commit 371b8044 ("powerpc/64s: 
Initialize ISAv3 MMU registers before setting partition table").

The problem is Radix MMU specific (does not occur with 'disable_radix'),
and does not occur with that code reverted (ie do not set PIDR to zero).

Do you see any reasons why?
(wondering if at all related to access_ok() in include/asm/uaccess.h)

with:

     # strace -e set_robust_list -f ./test
     set_robust_list(0x7fffa4b03910, 24)     = -1 ENOSYS (Function not 
implemented)
     +++ exited with 1 +++

     # uname -r
     4.15.0

without:

     # strace -e set_robust_list -f ./test
     set_robust_list(0x7fff889c3910, 24)     = 0
     +++ exited with 0 +++

     # uname -r
     4.15.0.nopidr

On Tue, 06 Feb 2018 08:55:31 +1100
Benjamin Herrenschmidt [off-list ref] wrote:

On Mon, 2018-02-05 at 19:14 -0200, Mauricio Faria de Oliveira wrote:

quoted

Nick, Michael,  

+Aneesh.

quoted

On 02/05/2018 10:48 AM, Florian Weimer wrote:  

quoted

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)  

The regression was introduced by commit 371b8044 ("powerpc/64s: 
Initialize ISAv3 MMU registers before setting partition table").

The problem is Radix MMU specific (does not occur with 'disable_radix'),
and does not occur with that code reverted (ie do not set PIDR to zero).

Do you see any reasons why?
(wondering if at all related to access_ok() in include/asm/uaccess.h)

Does this help?

powerpc/64s/radix: allocate guard-PID for kernel contexts at boot

64s/radix uses PID 0 for its kernel mapping at the 0xCxxx (quadrant 3)
address. This mapping is also accessible at 0x0xxx when PIDR=0 -- the
top 2 bits just selects the addressing mode, which is effectively the
same when PIDR=0 -- so address 0 translates to physical address 0 by
the kernel's linear map.

Commit 371b8044 ("powerpc/64s: Initialize ISAv3 MMU registers before
setting partition table"), which zeroes PIDR at boot, caused this
situation, and that stops kernel access to NULL from faulting in boot.
Before this, we inherited what firmware or kexec gave, which is almost
always non-zero.

futex_atomic_cmpxchg detection is done in boot, by testing if it
returns -EFAULT on a NULL address. This breaks when kernel access to
NULL during boot does not fault.

This patch allocates a non-zero guard PID for init_mm, and switches
kernel context to the guard PID at boot. This disallows access to the
kernel mapping from quadrant 0 at boot.

The effectiveness of this protection will be diminished a little after
boot when kernel threads inherit the last context, but those should
have NULL guard areas, and it's possible we will actually prefer to do
a non-lazy switch back to the guard PID in a future change. For now,
this gives a minimal fix, and gives NULL pointer protection for boot.
---
 arch/powerpc/mm/pgtable-radix.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/mm/pgtable-radix.c b/arch/powerpc/mm/pgtable-radix.c
index 573a9a2ee455..6389a8527e4a 100644
--- a/arch/powerpc/mm/pgtable-radix.c
+++ b/arch/powerpc/mm/pgtable-radix.c

@@ -20,6 +20,7 @@
 
 #include <asm/pgtable.h>
 #include <asm/pgalloc.h>
+#include <asm/mmu_context.h>
 #include <asm/dma.h>
 #include <asm/machdep.h>
 #include <asm/mmu.h>

@@ -333,6 +334,9 @@ static void __init radix_init_pgtable(void)
 		     "r" (TLBIEL_INVAL_SET_LPID), "r" (0));
 	asm volatile("eieio; tlbsync; ptesync" : : : "memory");
 	trace_tlbie(0, 0, TLBIEL_INVAL_SET_LPID, 0, 2, 1, 1);
+
+	init_mm.context.id = mmu_base_pid;
+	mmu_base_pid++;
 }
 
 static void __init radix_init_partition_table(void)

@@ -579,7 +583,7 @@ void __init radix__early_init_mmu(void)
 
 	radix_init_iamr();
 	radix_init_pgtable();
-
+	radix__switch_mmu_context(NULL, &init_mm);
 	if (cpu_has_feature(CPU_FTR_HVMODE))
 		tlbiel_all();
 }

@@ -604,6 +608,7 @@ void radix__early_init_mmu_secondary(void)
 	}
 	radix_init_iamr();
 
+	radix__switch_mmu_context(NULL, &init_mm);
 	if (cpu_has_feature(CPU_FTR_HVMODE))
 		tlbiel_all();
 }

-- 
2.15.1

Nicholas Piggin [off-list ref] writes:

On Tue, 06 Feb 2018 08:55:31 +1100
Benjamin Herrenschmidt [off-list ref] wrote:

quoted

On Mon, 2018-02-05 at 19:14 -0200, Mauricio Faria de Oliveira wrote:

quoted

Nick, Michael,  

+Aneesh.

quoted

On 02/05/2018 10:48 AM, Florian Weimer wrote:  

quoted

7041  set_robust_list(0x7fff93dc3980, 24) = -1 ENOSYS (Function not 
implemented)  

The regression was introduced by commit 371b8044 ("powerpc/64s: 
Initialize ISAv3 MMU registers before setting partition table").

The problem is Radix MMU specific (does not occur with 'disable_radix'),
and does not occur with that code reverted (ie do not set PIDR to zero).

Do you see any reasons why?
(wondering if at all related to access_ok() in include/asm/uaccess.h)

Does this help?

powerpc/64s/radix: allocate guard-PID for kernel contexts at boot

64s/radix uses PID 0 for its kernel mapping at the 0xCxxx (quadrant 3)
address. This mapping is also accessible at 0x0xxx when PIDR=0 -- the
top 2 bits just selects the addressing mode, which is effectively the
same when PIDR=0 -- so address 0 translates to physical address 0 by
the kernel's linear map.

Commit 371b8044 ("powerpc/64s: Initialize ISAv3 MMU registers before
setting partition table"), which zeroes PIDR at boot, caused this
situation, and that stops kernel access to NULL from faulting in boot.
Before this, we inherited what firmware or kexec gave, which is almost
always non-zero.

futex_atomic_cmpxchg detection is done in boot, by testing if it
returns -EFAULT on a NULL address. This breaks when kernel access to
NULL during boot does not fault.

This patch allocates a non-zero guard PID for init_mm, and switches
kernel context to the guard PID at boot. This disallows access to the
kernel mapping from quadrant 0 at boot.

The effectiveness of this protection will be diminished a little after
boot when kernel threads inherit the last context, but those should
have NULL guard areas, and it's possible we will actually prefer to do
a non-lazy switch back to the guard PID in a future change. For now,
this gives a minimal fix, and gives NULL pointer protection for boot.

I also have this as a part of another patch series. Since we already
support cmpxchg(), i would suggest we avoid the runtime check.

I needed this w.r.t hash so that we don't detect a NULL access as bad
slb address because we don't have PACA slb_addr_limit initialized
correctly that early.

commit c42b0fb10027af0c44fc9e2f6f9586203c38f99b
Author: Aneesh Kumar K.V [off-list ref]
Date:   Wed Jan 24 13:54:22 2018 +0530

    Don't do futext cmp test.
    
    It access NULL address early in the boot and we want to avoid that to simplify
    the fault handling.
    futex_detect_cmpxchg() does a cmpxchg_futex_value_locked on a NULL user addr
    to runtime detect whether architecture implements atomic cmpxchg for futex.

diff --git a/arch/powerpc/platforms/Kconfig.cputype b/arch/powerpc/platforms/Kconfig.cputype
index a429d859f15d..31bc2bd5dfd1 100644
--- a/arch/powerpc/platforms/Kconfig.cputype
+++ b/arch/powerpc/platforms/Kconfig.cputype

@@ -75,6 +75,7 @@ config PPC_BOOK3S_64
 	select ARCH_SUPPORTS_NUMA_BALANCING
 	select IRQ_WORK
 	select HAVE_KERNEL_XZ
+	select HAVE_FUTEX_CMPXCHG if FUTEX
 
 config PPC_BOOK3E_64
 	bool "Embedded processors"
 

On Tue, 06 Feb 2018 08:47:03 +0530
"Aneesh Kumar K.V" [off-list ref] wrote:

I think that's okay, but what I'd prefer is to set up the hash context
sufficiently that it will cope with a userspace access (and preferably
fault) before we switch on the MMU at boot.

We can do this patch as well, as a "don't bother testing because we always
support it" cleanup.

Thanks,
Nick

On 02/05/2018 11:06 PM, Nicholas Piggin wrote:

Does this help?

powerpc/64s/radix: allocate guard-PID for kernel contexts at boot

Yes, the test-case passes:

     # strace -e set_robust_list -f ./test
     set_robust_list(0x7fff8d453910, 24)     = 0
     +++ exited with 0 +++

     # uname -r
     4.15.0.guardpid

Thanks!

cheers,
Mauricio