ARM: mm: reduce maximum number of CPUs if DEBUG_KMAP_LOCAL is enabled

The debugging code for kmap_local() doubles the number of fixmap slots allocated for kmap_local(), in order to use half of them as guard regions. This causes the fixmap region to grow downwards and collide with the newly added virtual DT mapping, which is obviously not good.

One manifestation of this is EFI boot on a kernel built with NR_CPUS=32, which may pass the FDT in highmem, resulting in block entries that the fixmap code misidentifies as table entries, and subsequently tries to dereference using a phys-to-virt translation that is only valid for lowmem. This results in a cryptic splat such as the one below.

ftrace: allocating 45548 entries in 89 pages
8<--- cut here ---

  Unable to handle kernel paging request at virtual address fc6006f0
  pgd = (ptrval)
  [fc6006f0] *pgd=80000040207003, *pmd=00000000
  Internal error: Oops: a06 [#1] SMP ARM
  Modules linked in:
  CPU: 0 PID: 0 Comm: swapper Not tainted 5.11.0+ #382
  Hardware name: Generic DT based system
  PC is at cpu_ca15_set_pte_ext+0x24/0x30
  LR is at __set_fixmap+0xe4/0x118
  pc : [<c041ac9c>]    lr : [<c04189d8>]    psr: 400000d3
  sp : c1601ed8  ip : 00400000  fp : 00800000
  r10: 0000071f  r9 : 00421000  r8 : 00c00000
  r7 : 00c00000  r6 : 0000071f  r5 : ffade000  r4 : 4040171f
  r3 : 00c00000  r2 : 4040171f  r1 : c041ac78  r0 : fc6006f0
  Flags: nZcv  IRQs off  FIQs off  Mode SVC_32  ISA ARM  Segment none
  Control: 30c5387d  Table: 40203000  DAC: 00000001
  Process swapper (pid: 0, stack limit = 0x(ptrval))

So let's limit the supported number of CPUs to 16 when CONFIG_DEBUG_KMAP_LOCAL is in effect.

Reported-by Peter Robinson <[email protected]>

Signed-off-by Ard Biesheuvel <[email protected]>

---
 arch/arm/Kconfig | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

arch/arm/Kconfig +7 -2

index 138248999df7..3d2c684eab77 100644
@@ -1310,9 +1310,15 @@ config KASAN_SHADOW_OFFSET
 
 config NR_CPUS
        int "Maximum number of CPUs (2-32)"
-       range 2 32
+       range 2 16 if DEBUG_KMAP_LOCAL
+       range 2 32 if !DEBUG_KMAP_LOCAL
        depends on SMP
        default "4"
+       help
+         The maximum number of CPUs that the kernel can support.
+         Up to 32 CPUs can be supported, or up to 16 if kmap_local()
+         debugging is enabled, which uses half of the per-CPU fixmap
+         slots as guard regions.
 
 config HOTPLUG_CPU
        bool "Support for hot-pluggable CPUs"
--

2.30.0

Quoted 32 lines · Ard Biesheuvel

The debugging code for kmap_local() doubles the number of fixmap slots allocated for kmap_local(), in order to use half of them as guard regions. This causes the fixmap region to grow downwards and collide with the newly added virtual DT mapping, which is obviously not good.

One manifestation of this is EFI boot on a kernel built with NR_CPUS=32, which may pass the FDT in highmem, resulting in block entries that the fixmap code misidentifies as table entries, and subsequently tries to dereference using a phys-to-virt translation that is only valid for lowmem. This results in a cryptic splat such as the one below.

ftrace: allocating 45548 entries in 89 pages
8<--- cut here ---

  Unable to handle kernel paging request at virtual address fc6006f0
  pgd = (ptrval)
  [fc6006f0] *pgd=80000040207003, *pmd=00000000
  Internal error: Oops: a06 [#1] SMP ARM
  Modules linked in:
  CPU: 0 PID: 0 Comm: swapper Not tainted 5.11.0+ #382
  Hardware name: Generic DT based system
  PC is at cpu_ca15_set_pte_ext+0x24/0x30
  LR is at __set_fixmap+0xe4/0x118
  pc : [<c041ac9c>]    lr : [<c04189d8>]    psr: 400000d3
  sp : c1601ed8  ip : 00400000  fp : 00800000
  r10: 0000071f  r9 : 00421000  r8 : 00c00000
  r7 : 00c00000  r6 : 0000071f  r5 : ffade000  r4 : 4040171f
  r3 : 00c00000  r2 : 4040171f  r1 : c041ac78  r0 : fc6006f0
  Flags: nZcv  IRQs off  FIQs off  Mode SVC_32  ISA ARM  Segment none
  Control: 30c5387d  Table: 40203000  DAC: 00000001
  Process swapper (pid: 0, stack limit = 0x(ptrval))

Quoted 5 lines

So let's limit the supported number of CPUs to 16 when CONFIG_DEBUG_KMAP_LOCAL is in effect.

Reported-by Peter Robinson <[email protected]>

Signed-off-by Ard Biesheuvel <[email protected]>

Tested-by Peter Robinson <[email protected]>

Tested with both a LPAE and non LPAE kernel.

Quoted 23 lines

---
 arch/arm/Kconfig | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

arch/arm/Kconfig +7 -1

index 138248999df7..3d2c684eab77 100644
@@ -1310,9 +1310,15 @@ config KASAN_SHADOW_OFFSET
 
 config NR_CPUS
        int "Maximum number of CPUs (2-32)"
-       range 2 32
+       range 2 16 if DEBUG_KMAP_LOCAL
+       range 2 32 if !DEBUG_KMAP_LOCAL
        depends on SMP
        default "4"
+       help
+         The maximum number of CPUs that the kernel can support.
+         Up to 32 CPUs can be supported, or up to 16 if kmap_local()
+         debugging is enabled, which uses half of the per-CPU fixmap
+         slots as guard regions.
 

Quoted 5 lines

arch/arm/Kconfig +0 -1

 config HOTPLUG_CPU
        bool "Support for hot-pluggable CPUs"
--

2.30.0

Quoted 9 lines · Peter Robinson

Quoted 5 lines

So let's limit the supported number of CPUs to 16 when CONFIG_DEBUG_KMAP_LOCAL is in effect.

Reported-by Peter Robinson <[email protected]>

Signed-off-by Ard Biesheuvel <[email protected]>

Tested-by Peter Robinson <[email protected]>

Tested with both a LPAE and non LPAE kernel.

Thanks Peter.

Note that this patch should probably carry

Fixes 2a15ba82fa6ca3f3 ("ARM: highmem: Switch to generic kmap atomic")

given that the ARM specific kmap implementation did not suffer from this issue, and the DT remapping change (7a1be318f579) does not make any difference here.