| Message ID | 20241014105912.3207374-18-ryan.roberts@arm.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Boot-time page size selection for arm64 | expand |
On Mon, 14 Oct 2024 11:58:25 +0100 Ryan Roberts <ryan.roberts@arm.com> wrote: > To prepare for supporting boot-time page size selection, refactor code > to remove assumptions about PAGE_SIZE being compile-time constant. Code > intended to be equivalent when compile-time page size is active. > > Convert BUILD_BUG_ON() BUG_ON() since the argument depends on PAGE_SIZE > and its not trivial to test against a page size limit. > > Redefine FTRACE_KSTACK_ENTRIES so that "struct ftrace_stacks" is always > sized at 32K for 64-bit and 16K for 32-bit. It was previously defined in > terms of PAGE_SIZE (and worked out at the quoted sizes for a 4K page > size). But for 64K pages, the size expanded to 512K. Given the ftrace > stacks should be invariant to page size, this seemed like a waste. As a > side effect, it removes the PAGE_SIZE compile-time constant assumption > from this code. > > Signed-off-by: Ryan Roberts <ryan.roberts@arm.com> > --- > > ***NOTE*** > Any confused maintainers may want to read the cover note here for context: > https://lore.kernel.org/all/20241014105514.3206191-1-ryan.roberts@arm.com/ > > kernel/trace/fgraph.c | 2 +- > kernel/trace/trace.c | 2 +- > 2 files changed, 2 insertions(+), 2 deletions(-) > > diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c > index d7d4fb403f6f0..47aa5c8d8090e 100644 > --- a/kernel/trace/fgraph.c > +++ b/kernel/trace/fgraph.c > @@ -534,7 +534,7 @@ ftrace_push_return_trace(unsigned long ret, unsigned long func, > if (!current->ret_stack) > return -EBUSY; > > - BUILD_BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); > + BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); Absolutely not! BUG_ON() is in no way a substitution of any BUILD_BUG_ON(). BUILD_BUG_ON() is a non intrusive way to see if something isn't lined up correctly, and can fix it before you execute any code. BUG_ON() is the most intrusive way to say something is wrong and you crash the system. Not to mention, when function graph tracing is enabled, this gets triggered for *every* function call! So I do not want any runtime test done. Every nanosecond counts in this code path. If anything, this needs to be moved to initialization and checked once, if it fails, gives a WARN_ON() and disables function graph tracing. -- Steve > > /* Set val to "reserved" with the delta to the new fgraph frame */ > val = (FGRAPH_TYPE_RESERVED << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET; > diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c > index c3b2c7dfadef1..0f2ec3d30579f 100644 > --- a/kernel/trace/trace.c > +++ b/kernel/trace/trace.c > @@ -2887,7 +2887,7 @@ trace_function(struct trace_array *tr, unsigned long ip, unsigned long > /* Allow 4 levels of nesting: normal, softirq, irq, NMI */ > #define FTRACE_KSTACK_NESTING 4 > > -#define FTRACE_KSTACK_ENTRIES (PAGE_SIZE / FTRACE_KSTACK_NESTING) > +#define FTRACE_KSTACK_ENTRIES (SZ_4K / FTRACE_KSTACK_NESTING) > > struct ftrace_stack { > unsigned long calls[FTRACE_KSTACK_ENTRIES];
On 14/10/2024 17:46, Steven Rostedt wrote: > On Mon, 14 Oct 2024 11:58:25 +0100 > Ryan Roberts <ryan.roberts@arm.com> wrote: > >> To prepare for supporting boot-time page size selection, refactor code >> to remove assumptions about PAGE_SIZE being compile-time constant. Code >> intended to be equivalent when compile-time page size is active. >> >> Convert BUILD_BUG_ON() BUG_ON() since the argument depends on PAGE_SIZE >> and its not trivial to test against a page size limit. >> >> Redefine FTRACE_KSTACK_ENTRIES so that "struct ftrace_stacks" is always >> sized at 32K for 64-bit and 16K for 32-bit. It was previously defined in >> terms of PAGE_SIZE (and worked out at the quoted sizes for a 4K page >> size). But for 64K pages, the size expanded to 512K. Given the ftrace >> stacks should be invariant to page size, this seemed like a waste. As a >> side effect, it removes the PAGE_SIZE compile-time constant assumption >> from this code. >> >> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com> >> --- >> >> ***NOTE*** >> Any confused maintainers may want to read the cover note here for context: >> https://lore.kernel.org/all/20241014105514.3206191-1-ryan.roberts@arm.com/ >> >> kernel/trace/fgraph.c | 2 +- >> kernel/trace/trace.c | 2 +- >> 2 files changed, 2 insertions(+), 2 deletions(-) >> >> diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c >> index d7d4fb403f6f0..47aa5c8d8090e 100644 >> --- a/kernel/trace/fgraph.c >> +++ b/kernel/trace/fgraph.c >> @@ -534,7 +534,7 @@ ftrace_push_return_trace(unsigned long ret, unsigned long func, >> if (!current->ret_stack) >> return -EBUSY; >> >> - BUILD_BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); >> + BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); > > Absolutely not! > > BUG_ON() is in no way a substitution of any BUILD_BUG_ON(). BUILD_BUG_ON() > is a non intrusive way to see if something isn't lined up correctly, and > can fix it before you execute any code. BUG_ON() is the most intrusive way > to say something is wrong and you crash the system. Yep, totally agree. I'm afraid this was me being lazy, and there are a couple of other instances where I have done this in other patches that I'll need to fix. Most of the time, I've been able to keep BUILD_BUG_ON() and simply compare against a page size limit. Looking at this again, perhaps the better solution is to define SHADOW_STACK_SIZE as PAGE_SIZE_MIN? Then it remains a compile-time constant. Is there any need for SHADOW_STACK_SIZE to increase with page size? > > Not to mention, when function graph tracing is enabled, this gets triggered > for *every* function call! So I do not want any runtime test done. Every > nanosecond counts in this code path. > > If anything, this needs to be moved to initialization and checked once, if > it fails, gives a WARN_ON() and disables function graph tracing. I'm hoping my suggestion above to decouple SHADOW_STACK_SIZE from PAGE_SIZE is acceptable and simpler? If not, happy to do as you suggest here. Thanks, Ryan > > -- Steve > > >> >> /* Set val to "reserved" with the delta to the new fgraph frame */ >> val = (FGRAPH_TYPE_RESERVED << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET; >> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c >> index c3b2c7dfadef1..0f2ec3d30579f 100644 >> --- a/kernel/trace/trace.c >> +++ b/kernel/trace/trace.c >> @@ -2887,7 +2887,7 @@ trace_function(struct trace_array *tr, unsigned long ip, unsigned long >> /* Allow 4 levels of nesting: normal, softirq, irq, NMI */ >> #define FTRACE_KSTACK_NESTING 4 >> >> -#define FTRACE_KSTACK_ENTRIES (PAGE_SIZE / FTRACE_KSTACK_NESTING) >> +#define FTRACE_KSTACK_ENTRIES (SZ_4K / FTRACE_KSTACK_NESTING) >> >> struct ftrace_stack { >> unsigned long calls[FTRACE_KSTACK_ENTRIES]; >
On Tue, 15 Oct 2024 12:09:38 +0100 Ryan Roberts <ryan.roberts@arm.com> wrote: > > > > Not to mention, when function graph tracing is enabled, this gets triggered > > for *every* function call! So I do not want any runtime test done. Every > > nanosecond counts in this code path. > > > > If anything, this needs to be moved to initialization and checked once, if > > it fails, gives a WARN_ON() and disables function graph tracing. > > I'm hoping my suggestion above to decouple SHADOW_STACK_SIZE from PAGE_SIZE is > acceptable and simpler? If not, happy to do as you suggest here. Yeah, I think we can do that. In fact, I'm thinking it should turn into a kmem_cache item that doesn't have to be a power of two (but must be evenly divisible by the size of long). I'll write up a patch. -- Steve
diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c index d7d4fb403f6f0..47aa5c8d8090e 100644 --- a/kernel/trace/fgraph.c +++ b/kernel/trace/fgraph.c @@ -534,7 +534,7 @@ ftrace_push_return_trace(unsigned long ret, unsigned long func, if (!current->ret_stack) return -EBUSY; - BUILD_BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); + BUG_ON(SHADOW_STACK_SIZE % sizeof(long)); /* Set val to "reserved" with the delta to the new fgraph frame */ val = (FGRAPH_TYPE_RESERVED << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET; diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c index c3b2c7dfadef1..0f2ec3d30579f 100644 --- a/kernel/trace/trace.c +++ b/kernel/trace/trace.c @@ -2887,7 +2887,7 @@ trace_function(struct trace_array *tr, unsigned long ip, unsigned long /* Allow 4 levels of nesting: normal, softirq, irq, NMI */ #define FTRACE_KSTACK_NESTING 4 -#define FTRACE_KSTACK_ENTRIES (PAGE_SIZE / FTRACE_KSTACK_NESTING) +#define FTRACE_KSTACK_ENTRIES (SZ_4K / FTRACE_KSTACK_NESTING) struct ftrace_stack { unsigned long calls[FTRACE_KSTACK_ENTRIES];
To prepare for supporting boot-time page size selection, refactor code to remove assumptions about PAGE_SIZE being compile-time constant. Code intended to be equivalent when compile-time page size is active. Convert BUILD_BUG_ON() BUG_ON() since the argument depends on PAGE_SIZE and its not trivial to test against a page size limit. Redefine FTRACE_KSTACK_ENTRIES so that "struct ftrace_stacks" is always sized at 32K for 64-bit and 16K for 32-bit. It was previously defined in terms of PAGE_SIZE (and worked out at the quoted sizes for a 4K page size). But for 64K pages, the size expanded to 512K. Given the ftrace stacks should be invariant to page size, this seemed like a waste. As a side effect, it removes the PAGE_SIZE compile-time constant assumption from this code. Signed-off-by: Ryan Roberts <ryan.roberts@arm.com> --- ***NOTE*** Any confused maintainers may want to read the cover note here for context: https://lore.kernel.org/all/20241014105514.3206191-1-ryan.roberts@arm.com/ kernel/trace/fgraph.c | 2 +- kernel/trace/trace.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)