introduce __ro_mostly_after_init section marker
cover introduce __ro_mostly_after_init section marker
After __ro_after_init marker is included in kernel, many kernel data
objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but __ro_after_init can not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.
__ro_mostly_after_init is basically the same to __ro_after_init. The
section is mapped as read-only after kernel init. The different thing is this section is temporarily mapped as read-write during module_init/exit and de/registration of a subsystem using set_ro_mostly_after_init_rw/ro pair.
- Tested only on arm64.
Description:
0001 patch is __ro_mostly_after_init itself.
0002 patch is to add set_ro_mostly_after_init_rw/ro pair using
set_memory_rw/ro.
0003 patch is to make the section read-write in module_init/exit. 0004 patch is an example for dynamic init/deinit of a subsystem. 0005 patch is an example for __ro_mostly_after_init section modified during
module_init/exit.
0006/0007 patches are fixes for arm64 kernel mapping.
Hoeun Ryu (7):
arch: add __ro_mostly_after_init section marker
init: add set_ro_mostly_after_init_rw/ro function
module: modify memory attrs for __ro_mostly_after_init during
module_init/exit
selinux: mark __ro_mostly_after_init for selinux_hooks/selinux_nf_ops
cpu: mark ro_mostly_after_init for cpuhp_ap/bp_states
arm64: add __map_kernel_segment to accept additional vm flags
arm64: map seperately rodata sections for __ro_mostly_after_init
section
arch/arm64/mm/mmu.c | 44 ++++++++++++++++++++++++++++++++------- include/asm-generic/sections.h | 1 + include/asm-generic/vmlinux.lds.h | 10 +++++++++ include/linux/cache.h | 11 ++++++++++ include/linux/init.h | 6 ++++++ init/main.c | 24 +++++++++++++++++++++ kernel/cpu.c | 4 ++-- kernel/module.c | 10 +++++++-- security/selinux/hooks.c | 8 +++++-- 9 files changed, 105 insertions(+), 13 deletions(-)
Signature · 1 line
2.7.4
Hi Hoeun,
Quoted 5 lines · Hoeun Ryu
After
__ro_after_initmarker is included in kernel, many kernel data
objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but__ro_after_initcan not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.
The argument sounds reasonable, but it would really help if you could describe some use cases in more detail.
Quoted 42 lines
__ro_mostly_after_initis basically the same to__ro_after_init. The
section is mapped as read-only after kernel init. The different thing is this section is temporarily mapped as read-write during module_init/exit and de/registration of a subsystem using set_ro_mostly_after_init_rw/ro pair.- Tested only on arm64.
Description:
0001 patch is__ro_mostly_after_inititself.
0002 patch is to add set_ro_mostly_after_init_rw/ro pair using
set_memory_rw/ro.
0003 patch is to make the section read-write in module_init/exit. 0004 patch is an example for dynamic init/deinit of a subsystem. 0005 patch is an example for__ro_mostly_after_initsection modified during
module_init/exit.
0006/0007 patches are fixes for arm64 kernel mapping.Hoeun Ryu (7):
arch: add__ro_mostly_after_initsection marker
init: add set_ro_mostly_after_init_rw/ro function
module: modify memory attrs for__ro_mostly_after_initduring
module_init/exit
selinux: mark__ro_mostly_after_initfor selinux_hooks/selinux_nf_ops
cpu: markro_mostly_after_initfor cpuhp_ap/bp_states
arm64: add__map_kernel_segmentto accept additional vm flags
arm64: map seperately rodata sections for__ro_mostly_after_init
sectionarch/arm64/mm/mmu.c | 44 ++++++++++++++++++++++++++++++++------- include/asm-generic/sections.h | 1 + include/asm-generic/vmlinux.lds.h | 10 +++++++++ include/linux/cache.h | 11 ++++++++++ include/linux/init.h | 6 ++++++ init/main.c | 24 +++++++++++++++++++++ kernel/cpu.c | 4 ++-- kernel/module.c | 10 +++++++-- security/selinux/hooks.c | 8 +++++-- 9 files changed, 105 insertions(+), 13 deletions(-)--
2.7.4
Quoted 6 lines · Ard Biesheuvel
Hi Hoeun,
On 19 February 2017 at 10:03, Hoeun Ryu <[email protected]> wrote:
Quoted 8 lines
Quoted 5 lines
After
__ro_after_initmarker is included in kernel, many kernel data objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but__ro_after_initcan not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.The argument sounds reasonable, but it would really help if you could describe some use cases in more detail.
Please see RFC 4/7 [1] and RFC 5/7 [2].
RFC 4 shows that cpu_ap/bp_states are marked as __ro_mostly_after_init. cpu_ap/bp_states can not be marked as __ro_after_init because some modules can write to those objects using cpuhp_setup/remove_state api during module_init/exit.
RFC 5 shows that selinux_hooks and selinux_nf_hooks are marked as __ro_mostly_after_init. those hooks can not be marked as __ro_after_init because selinux_disable() function. Those hooks should be writable during selinux_disable to deregister the hooks.
[1] https://lkml.org/lkml/2017/2/19/37
[2] https://lkml.org/lkml/2017/2/19/35
Quoted 43 lines
Quoted 43 lines
__ro_mostly_after_initis basically the same to__ro_after_init. The section is mapped as read-only after kernel init. The different thing is this section is temporarily mapped as read-write during module_init/exit and de/registration of a subsystem using set_ro_mostly_after_init_rw/ro pair.- Tested only on arm64.
Description:
0001 patch is__ro_mostly_after_inititself.
0002 patch is to add set_ro_mostly_after_init_rw/ro pair using
set_memory_rw/ro.
0003 patch is to make the section read-write in module_init/exit. 0004 patch is an example for dynamic init/deinit of a subsystem. 0005 patch is an example for__ro_mostly_after_initsection modified during
module_init/exit.
0006/0007 patches are fixes for arm64 kernel mapping.Hoeun Ryu (7):
arch: add__ro_mostly_after_initsection marker
init: add set_ro_mostly_after_init_rw/ro function
module: modify memory attrs for__ro_mostly_after_initduring
module_init/exit
selinux: mark__ro_mostly_after_initfor selinux_hooks/selinux_nf_ops
cpu: markro_mostly_after_initfor cpuhp_ap/bp_states
arm64: add__map_kernel_segmentto accept additional vm flags
arm64: map seperately rodata sections for__ro_mostly_after_init
sectionarch/arm64/mm/mmu.c | 44 ++++++++++++++++++++++++++++++++------- include/asm-generic/sections.h | 1 + include/asm-generic/vmlinux.lds.h | 10 +++++++++ include/linux/cache.h | 11 ++++++++++ include/linux/init.h | 6 ++++++ init/main.c | 24 +++++++++++++++++++++ kernel/cpu.c | 4 ++-- kernel/module.c | 10 +++++++-- security/selinux/hooks.c | 8 +++++-- 9 files changed, 105 insertions(+), 13 deletions(-)--
2.7.4
Quoted 5 lines · Hoeun Ryu
After
__ro_after_initmarker is included in kernel, many kernel data
objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but__ro_after_initcan not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.
Could you elaborate on this?
For modules, I assume that the __ro_after_init data structures are part of the module, and not part of the "real" kernel image. Is that the case?
Which specific subsystems whish to modify data structures that are __ro_after_init?
This sounds like the proposed mostly-ro/rarely-rw stuff would be a better fit for that case.
Thanks,
Mark.
Quoted 42 lines
__ro_mostly_after_initis basically the same to__ro_after_init. The
section is mapped as read-only after kernel init. The different thing is this section is temporarily mapped as read-write during module_init/exit and de/registration of a subsystem using set_ro_mostly_after_init_rw/ro pair.- Tested only on arm64.
Description:
0001 patch is__ro_mostly_after_inititself.
0002 patch is to add set_ro_mostly_after_init_rw/ro pair using
set_memory_rw/ro.
0003 patch is to make the section read-write in module_init/exit. 0004 patch is an example for dynamic init/deinit of a subsystem. 0005 patch is an example for__ro_mostly_after_initsection modified during
module_init/exit.
0006/0007 patches are fixes for arm64 kernel mapping.Hoeun Ryu (7):
arch: add__ro_mostly_after_initsection marker
init: add set_ro_mostly_after_init_rw/ro function
module: modify memory attrs for__ro_mostly_after_initduring
module_init/exit
selinux: mark__ro_mostly_after_initfor selinux_hooks/selinux_nf_ops
cpu: markro_mostly_after_initfor cpuhp_ap/bp_states
arm64: add__map_kernel_segmentto accept additional vm flags
arm64: map seperately rodata sections for__ro_mostly_after_init
sectionarch/arm64/mm/mmu.c | 44 ++++++++++++++++++++++++++++++++------- include/asm-generic/sections.h | 1 + include/asm-generic/vmlinux.lds.h | 10 +++++++++ include/linux/cache.h | 11 ++++++++++ include/linux/init.h | 6 ++++++ init/main.c | 24 +++++++++++++++++++++ kernel/cpu.c | 4 ++-- kernel/module.c | 10 +++++++-- security/selinux/hooks.c | 8 +++++-- 9 files changed, 105 insertions(+), 13 deletions(-)Signature · 1 line
2.7.4
Quoted 10 lines · Mark Rutland
Quoted 5 lines · Hoeun Ryu
After
__ro_after_initmarker is included in kernel, many kernel data objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but__ro_after_initcan not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.
Quoted 5 lines
Could you elaborate on this?
For modules, I assume that the
__ro_after_initdata structures are part of the module, and not part of the "real" kernel image. Is that the case?
__ro_mostly_after_init is for kernel builtin core subsystems, not for modules themselves. The section can be writable only during kernel init and module_init/exit.
Some hooks (or array of hooks) of a core subsystem can be marked as __ro_mostly_after_init similar to that way of __ro_after_init. After that some modules that may write to those hooks of the subsystem to register/deregister something to the subsystem can safely access those section. Please see RFC 3/7 that makes this section writable.
In addition, some subsystems may use this marker for their (array of) hooks and make them writable only at some point of time via set_ro_mostly_after_init_rw/ro pair.
please read RFC 4/7 for selinux.
Quoted 2 lines
Which specific subsystems whish to modify data structures that are
__ro_after_init?
I’m not intending to make writable __ro_after_init section but introducing new section marker that works mostly like __ro_after_init but can be written to at some points. please see RFC 5/7 for cpuhotplug.
Quoted 49 lines
This sounds like the proposed mostly-ro/rarely-rw stuff would be a better fit for that case.
Thanks,
Mark.Quoted 42 lines
__ro_mostly_after_initis basically the same to__ro_after_init. The section is mapped as read-only after kernel init. The different thing is this section is temporarily mapped as read-write during module_init/exit and de/registration of a subsystem using set_ro_mostly_after_init_rw/ro pair.- Tested only on arm64.
Description:
0001 patch is__ro_mostly_after_inititself.
0002 patch is to add set_ro_mostly_after_init_rw/ro pair using
set_memory_rw/ro.
0003 patch is to make the section read-write in module_init/exit. 0004 patch is an example for dynamic init/deinit of a subsystem. 0005 patch is an example for__ro_mostly_after_initsection modified during
module_init/exit.
0006/0007 patches are fixes for arm64 kernel mapping.Hoeun Ryu (7):
arch: add__ro_mostly_after_initsection marker
init: add set_ro_mostly_after_init_rw/ro function
module: modify memory attrs for__ro_mostly_after_initduring
module_init/exit
selinux: mark__ro_mostly_after_initfor selinux_hooks/selinux_nf_ops
cpu: markro_mostly_after_initfor cpuhp_ap/bp_states
arm64: add__map_kernel_segmentto accept additional vm flags
arm64: map seperately rodata sections for__ro_mostly_after_init
sectionarch/arm64/mm/mmu.c | 44 ++++++++++++++++++++++++++++++++------- include/asm-generic/sections.h | 1 + include/asm-generic/vmlinux.lds.h | 10 +++++++++ include/linux/cache.h | 11 ++++++++++ include/linux/init.h | 6 ++++++ init/main.c | 24 +++++++++++++++++++++ kernel/cpu.c | 4 ++-- kernel/module.c | 10 +++++++-- security/selinux/hooks.c | 8 +++++-- 9 files changed, 105 insertions(+), 13 deletions(-)Signature · 1 line
2.7.4
Quoted 22 lines · Ho-Eun Ryu
Quoted 8 lines · Mark Rutland
Quoted 5 lines · Hoeun Ryu
After
__ro_after_initmarker is included in kernel, many kernel data objects can be read-only-after-init. But there are many other places that would be good to read-only-after-init but__ro_after_initcan not be simply applicable to them because they should be writable at some points, which are during module_init/exit or dynamic de/registration for a specific subsystem.Quoted 4 lines
Could you elaborate on this?
For modules, I assume that the
__ro_after_initdata structures are part of the module, and not part of the "real" kernel image. Is that the case?
__ro_mostly_after_initis for kernel builtin core subsystems, not for modules themselves. The section can be writable only during kernel init and module_init/exit. Some hooks (or array of hooks) of a core subsystem can be marked as__ro_mostly_after_initsimilar to that way of__ro_after_init. After that some modules that may write to those hooks of the subsystem to register/deregister something to the subsystem can safely access those section. Please see RFC 3/7 that makes this section writable.
Quoted 3 lines
In addition, some subsystems may use this marker for their (array of) hooks and make them writable only at some point of time via set_ro_mostly_after_init_rw/ro pair. please read RFC 4/7 for selinux.
Ok.
This sounds like a limited case of __write_rarely, where we only expect to perform writes in specific time windows.
I'd prefer if we could attack __write_rarely, and use the appraoch to cater for this case, but I am not necessarily opposed to this approach if other architectures are happy with it.
Quoted 7 lines
Quoted 2 lines
Which specific subsystems whish to modify data structures that are
__ro_after_init?I’m not intending to make writable
__ro_after_initsection but introducing new section marker that works mostly like__ro_after_initbut can be written to at some points. please see RFC 5/7 for cpuhotplug.
Ah. I clearly had not read this sufficiently thoroughly; sorry about that.
Thanks,
Mark.