On Tue, Feb 07, 2023 at 03:46:53PM +0800, Zhenhua Huang wrote: > Thanks Marco! > > On 2023/2/7 15:19, Marco Elver wrote: > > On Tue, 7 Feb 2023 at 07:15, Zhenhua Huang <quic_zhenhuah@xxxxxxxxxxx> wrote: > > > > > > Export the variable to ease the judgement of whether kfence enabled > > > at runtime. It should be more precise than through kernel config > > > "CONFIG_KFENCE". > > > > > > For example We can disable kfence at runtime using bootargs > > > "kfence.sample_interval=0" but CONFIG_KFENCE enabled. > > > It was false positive. > > > > > > Signed-off-by: Zhenhua Huang <quic_zhenhuah@xxxxxxxxxxx> > > > --- > > > arch/arm64/mm/pageattr.c | 3 ++- > > > include/linux/kfence.h | 2 ++ > > > mm/kfence/core.c | 2 +- > > > 3 files changed, 5 insertions(+), 2 deletions(-) > > > > > > diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c > > > index 79dd201..208d780 100644 > > > --- a/arch/arm64/mm/pageattr.c > > > +++ b/arch/arm64/mm/pageattr.c > > > @@ -7,6 +7,7 @@ > > > #include <linux/module.h> > > > #include <linux/sched.h> > > > #include <linux/vmalloc.h> > > > +#include <linux/kfence.h> > > > > > > #include <asm/cacheflush.h> > > > #include <asm/set_memory.h> > > > @@ -27,7 +28,7 @@ bool can_set_direct_map(void) > > > * protect/unprotect single pages. > > > */ > > > return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() || > > > - IS_ENABLED(CONFIG_KFENCE); > > > + kfence_enabled; > > > > Unfortunately this won't work, because it's possible to enable KFENCE > > after the kernel has booted with e.g.: echo 100 > > > /sys/module/kfence/parameters/sample_interval > Yeah, got it. Thanks for catching it. > > > > What is the problem you have encountered? Is the page-granular direct > > map causing issues? > We're working on a low memory target, page-granular mapping costed more (2M > per 1GB) memory. Due to GKI constraints, it is not easy to disable > CONFIG_KFENCE. So my intention was to move the judgement to runtime > configurable w/ CONFIG_KFENCE on... > > Do you have any further suggestion/proposal on this? Many Thanks! Just to check, the cost is because we're mapping *all* of memory at page granulatrity, right? If we were to just map the KFENCE region a page granularity, would that be a sufficient saving? We didn't do that so far because it was simpler to just map everything at page granularity (and that's also required by rodata_full, which I though android used?). If it's really important (and rodata_full isn't being used), we could try to do that. To do that we'd need to choose the KFENCE region *before* arm64 creates the fine-grain translation tables, which probable needs an arch_ hook. Thanks, Mark.
