[PATCH] TRACING: FTRACE: Use xarray structure for ftrace syscalls

[PATCH] TRACING: FTRACE: Use xarray structure for ftrace syscalls

[Hassan Naveed]

From: Hassan Naveed <hnaveed@wavecomp.com>

Currently, a lot of memory is wasted for architectures like MIPS when
init_ftrace_syscalls() allocates the array for syscalls using kcalloc.
This is because syscalls numbers start from 4000, 5000 or 6000 and
array elements up to that point are unused.
Fix this by using a data structure more suited to storing sparsely
populated arrays. The XARRAY data structure, implemented using radix
trees, is much more memory efficient for storing the syscalls in
question.

Signed-off-by: Hassan Naveed <hnaveed@wavecomp.com>
---
 kernel/trace/trace_syscalls.c | 17 ++++-------------
 1 file changed, 4 insertions(+), 13 deletions(-)

diff --git a/kernel/trace/trace_syscalls.c b/kernel/trace/trace_syscalls.c
index f93a56d2db27..1fee710be874 100644
--- a/kernel/trace/trace_syscalls.c
+++ b/kernel/trace/trace_syscalls.c
@@ -7,6 +7,7 @@
 #include <linux/module.h>	/* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
 #include <linux/ftrace.h>
 #include <linux/perf_event.h>
+#include <linux/xarray.h>
 #include <asm/syscall.h>
 
 #include "trace_output.h"
@@ -30,7 +31,7 @@ syscall_get_enter_fields(struct trace_event_call *call)
 extern struct syscall_metadata *__start_syscalls_metadata[];
 extern struct syscall_metadata *__stop_syscalls_metadata[];
 
-static struct syscall_metadata **syscalls_metadata;
+static DEFINE_XARRAY(syscalls_metadata);
 
 #ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
 static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
@@ -101,10 +102,7 @@ find_syscall_meta(unsigned long syscall)
 
 static struct syscall_metadata *syscall_nr_to_meta(int nr)
 {
-	if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
-		return NULL;
-
-	return syscalls_metadata[nr];
+	return xa_load(&syscalls_metadata, (unsigned long)nr);
 }
 
 const char *get_syscall_name(int syscall)
@@ -535,13 +533,6 @@ void __init init_ftrace_syscalls(void)
 	unsigned long addr;
 	int i;
 
-	syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
-				    GFP_KERNEL);
-	if (!syscalls_metadata) {
-		WARN_ON(1);
-		return;
-	}
-
 	for (i = 0; i < NR_syscalls; i++) {
 		addr = arch_syscall_addr(i);
 		meta = find_syscall_meta(addr);
@@ -549,7 +540,7 @@ void __init init_ftrace_syscalls(void)
 			continue;
 
 		meta->syscall_nr = i;
-		syscalls_metadata[i] = meta;
+		xa_store(&syscalls_metadata, i, meta, GFP_KERNEL);
 	}
 }
 
-- 
2.17.1

Re: [PATCH] TRACING: FTRACE: Use xarray structure for ftrace syscalls

[Steven Rostedt]

On Tue, 22 Oct 2019 18:24:25 +0000
Hassan Naveed <hnaveed@wavecomp.com> wrote:


Nit, the subject should simply be:

 "tracing: Use xarray for syscall trace events"


> Signed-off-by: Hassan Naveed <hnaveed@wavecomp.com>
> ---
>  kernel/trace/trace_syscalls.c | 17 ++++-------------
>  1 file changed, 4 insertions(+), 13 deletions(-)
> 
> diff --git a/kernel/trace/trace_syscalls.c b/kernel/trace/trace_syscalls.c
> index f93a56d2db27..1fee710be874 100644
> --- a/kernel/trace/trace_syscalls.c
> +++ b/kernel/trace/trace_syscalls.c
> @@ -7,6 +7,7 @@
>  #include <linux/module.h>	/* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
>  #include <linux/ftrace.h>
>  #include <linux/perf_event.h>
> +#include <linux/xarray.h>
>  #include <asm/syscall.h>
>  
>  #include "trace_output.h"
> @@ -30,7 +31,7 @@ syscall_get_enter_fields(struct trace_event_call *call)
>  extern struct syscall_metadata *__start_syscalls_metadata[];
>  extern struct syscall_metadata *__stop_syscalls_metadata[];
>  
> -static struct syscall_metadata **syscalls_metadata;
> +static DEFINE_XARRAY(syscalls_metadata);
>  
>  #ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
>  static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
> @@ -101,10 +102,7 @@ find_syscall_meta(unsigned long syscall)
>  
>  static struct syscall_metadata *syscall_nr_to_meta(int nr)
>  {
> -	if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
> -		return NULL;
> -
> -	return syscalls_metadata[nr];
> +	return xa_load(&syscalls_metadata, (unsigned long)nr);
>  }
>  
>  const char *get_syscall_name(int syscall)
> @@ -535,13 +533,6 @@ void __init init_ftrace_syscalls(void)
>  	unsigned long addr;
>  	int i;
>  
> -	syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
> -				    GFP_KERNEL);
> -	if (!syscalls_metadata) {
> -		WARN_ON(1);
> -		return;
> -	}
> -
>  	for (i = 0; i < NR_syscalls; i++) {
>  		addr = arch_syscall_addr(i);
>  		meta = find_syscall_meta(addr);
> @@ -549,7 +540,7 @@ void __init init_ftrace_syscalls(void)
>  			continue;
>  
>  		meta->syscall_nr = i;
> -		syscalls_metadata[i] = meta;
> +		xa_store(&syscalls_metadata, i, meta, GFP_KERNEL);

Shouldn't xa_store() return be tested for memory failure?

-- Steve

>  	}
>  }
>  

Re: [PATCH] TRACING: FTRACE: Use xarray structure for ftrace syscalls

[Steven Rostedt]

On Tue, 22 Oct 2019 15:51:04 -0400
Steven Rostedt <rostedt@goodmis.org> wrote:

> >  static struct syscall_metadata *syscall_nr_to_meta(int nr)
> >  {
> > -	if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
> > -		return NULL;
> > -
> > -	return syscalls_metadata[nr];
> > +	return xa_load(&syscalls_metadata, (unsigned long)nr);
> >  }
> >  

There appears to be a slight overhead to this for archs that do not
have a sparse syscall array. I wonder if we should make this only
applicable for archs (via adding a HAVE_SPARSE_SYSCALL_NR define and
checking against it). Then if an arch doesn't have a sparse array of
system calls, it uses a normal lookup, but for archs that do, it can
define this for this type of lookup.

There's not much to this patch, so it wouldn't be too difficult to
support both methods.

Without this patch I ran:

# trace-cmd start -e syscalls
# /work/c/hackbench 50
Time: 15.702
# /work/c/hackbench 50
Time: 15.932
# /work/c/hackbench 50
Time: 15.893
# /work/c/hackbench 50
Time: 16.038
# /work/c/hackbench 50
Time: 15.429


With the patch it had:

# trace-cmd start -e syscalls
# /work/c/hackbench 50
Time: 16.582
# /work/c/hackbench 50
Time: 15.972
# /work/c/hackbench 50
Time: 16.078
# /work/c/hackbench 50
Time: 16.133
# /work/c/hackbench 50
Time: 16.263

-- Steve

Re: [PATCH] TRACING: FTRACE: Use xarray structure for ftrace syscalls

[Steven Rostedt]

ping?

-- Steve


On Tue, 22 Oct 2019 17:45:51 -0400
Steven Rostedt <rostedt@goodmis.org> wrote:

> On Tue, 22 Oct 2019 15:51:04 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > >  static struct syscall_metadata *syscall_nr_to_meta(int nr)
> > >  {
> > > -	if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
> > > -		return NULL;
> > > -
> > > -	return syscalls_metadata[nr];
> > > +	return xa_load(&syscalls_metadata, (unsigned long)nr);
> > >  }
> > >    
> 
> There appears to be a slight overhead to this for archs that do not
> have a sparse syscall array. I wonder if we should make this only
> applicable for archs (via adding a HAVE_SPARSE_SYSCALL_NR define and
> checking against it). Then if an arch doesn't have a sparse array of
> system calls, it uses a normal lookup, but for archs that do, it can
> define this for this type of lookup.
> 
> There's not much to this patch, so it wouldn't be too difficult to
> support both methods.
> 
> Without this patch I ran:
> 
> # trace-cmd start -e syscalls
> # /work/c/hackbench 50
> Time: 15.702
> # /work/c/hackbench 50
> Time: 15.932
> # /work/c/hackbench 50
> Time: 15.893
> # /work/c/hackbench 50
> Time: 16.038
> # /work/c/hackbench 50
> Time: 15.429
> 
> 
> With the patch it had:
> 
> # trace-cmd start -e syscalls
> # /work/c/hackbench 50
> Time: 16.582
> # /work/c/hackbench 50
> Time: 15.972
> # /work/c/hackbench 50
> Time: 16.078
> # /work/c/hackbench 50
> Time: 16.133
> # /work/c/hackbench 50
> Time: 16.263
> 
> -- Steve

[PATCH v2 1/2] tracing: Use xarray for syscall trace events

[Hassan Naveed]

Currently, a lot of memory is wasted for architectures like MIPS when
init_ftrace_syscalls() allocates the array for syscalls using kcalloc.
This is because syscalls numbers start from 4000, 5000 or 6000 and
array elements up to that point are unused.
Fix this by using a data structure more suited to storing sparsely
populated arrays. The XARRAY data structure, implemented using radix
trees, is much more memory efficient for storing the syscalls in
question.

Signed-off-by: Hassan Naveed <hnaveed@wavecomp.com>
Reviewed-by: Paul Burton <paulburton@kernel.org>
---
Changes in v2:
 - New commit message
 - Added HAVE_SPARSE_SYSCALL_NR to arch/Kconfig for archs
   with sparse syscall arrays
 - Changes split up into two patches; one that provides
   xarray support and the second in which MIPS selects it
---
 arch/Kconfig                  |  8 ++++++++
 kernel/trace/trace_syscalls.c | 32 +++++++++++++++++++++++++-------
 2 files changed, 33 insertions(+), 7 deletions(-)

diff --git a/arch/Kconfig b/arch/Kconfig
index 33687dddd86a..e3783d94f919 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -901,6 +901,14 @@ config HAVE_ARCH_PREL32_RELOCATIONS
 config ARCH_USE_MEMREMAP_PROT
 	bool
 
+config HAVE_SPARSE_SYSCALL_NR
+       bool
+       help
+          An architecture should select this if its syscall numbering is sparse
+	  to save space. For example, MIPS architecture has a syscall array with
+	  entries at 4000, 5000 and 6000 locations. This option turns on syscall
+	  related optimizations for a given architecture.
+
 source "kernel/gcov/Kconfig"
 
 source "scripts/gcc-plugins/Kconfig"
diff --git a/kernel/trace/trace_syscalls.c b/kernel/trace/trace_syscalls.c
index f93a56d2db27..ffff52399d97 100644
--- a/kernel/trace/trace_syscalls.c
+++ b/kernel/trace/trace_syscalls.c
@@ -7,6 +7,7 @@
 #include <linux/module.h>	/* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
 #include <linux/ftrace.h>
 #include <linux/perf_event.h>
+#include <linux/xarray.h>
 #include <asm/syscall.h>
 
 #include "trace_output.h"
@@ -30,6 +31,7 @@ syscall_get_enter_fields(struct trace_event_call *call)
 extern struct syscall_metadata *__start_syscalls_metadata[];
 extern struct syscall_metadata *__stop_syscalls_metadata[];
 
+static DEFINE_XARRAY(syscalls_metadata_sparse);
 static struct syscall_metadata **syscalls_metadata;
 
 #ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
@@ -101,6 +103,9 @@ find_syscall_meta(unsigned long syscall)
 
 static struct syscall_metadata *syscall_nr_to_meta(int nr)
 {
+	if (IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR))
+		return xa_load(&syscalls_metadata_sparse, (unsigned long)nr);
+
 	if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
 		return NULL;
 
@@ -534,12 +539,16 @@ void __init init_ftrace_syscalls(void)
 	struct syscall_metadata *meta;
 	unsigned long addr;
 	int i;
-
-	syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
-				    GFP_KERNEL);
-	if (!syscalls_metadata) {
-		WARN_ON(1);
-		return;
+	void *ret;
+
+	if (!IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR)) {
+		syscalls_metadata = kcalloc(NR_syscalls,
+					sizeof(*syscalls_metadata),
+					GFP_KERNEL);
+		if (!syscalls_metadata) {
+			WARN_ON(1);
+			return;
+		}
 	}
 
 	for (i = 0; i < NR_syscalls; i++) {
@@ -549,7 +558,16 @@ void __init init_ftrace_syscalls(void)
 			continue;
 
 		meta->syscall_nr = i;
-		syscalls_metadata[i] = meta;
+
+		if (!IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR)) {
+			syscalls_metadata[i] = meta;
+		} else {
+			ret = xa_store(&syscalls_metadata_sparse, i, meta,
+					GFP_KERNEL);
+			WARN(xa_is_err(ret),
+				"Syscall memory allocation failed\n");
+		}
+
 	}
 }
 
-- 
2.17.1

[PATCH v2 2/2] tracing: enable syscall optimization for MIPS

[Hassan Naveed]

Since MIPS architecture has a sparse syscall array, select the
HAVE_SPARSE_SYSCALL_NR to save space.

Signed-off-by: Hassan Naveed <hnaveed@wavecomp.com>
Reviewed-by: Paul Burton <paulburton@kernel.org>
---
 arch/mips/Kconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 4a5f5b0ee9a9..32421ecff933 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -71,6 +71,7 @@ config MIPS
 	select HAVE_PERF_EVENTS
 	select HAVE_REGS_AND_STACK_ACCESS_API
 	select HAVE_RSEQ
+	select HAVE_SPARSE_SYSCALL_NR
 	select HAVE_STACKPROTECTOR
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_VIRT_CPU_ACCOUNTING_GEN if 64BIT || !SMP
-- 
2.17.1