[RFC PATCH v2] ftrace: support very early function tracing

From: Abderrahmane Benbachir
Date: Tue Nov 07 2017 - 18:44:56 EST


Hi Steve,

I updated the patch with all the fixes except for update timestamp part.

I used ftrace_export as you suggested, but I don't know how to hook toÂ
ringbuffer to update the timestamp, I used the same way as before to
update the time_delta in function ftrace_vearly_export_write.

static u64 export_delta __initdata;

static void __init ftrace_vearly_export_write(const void *buf, unsigned int len)
{
struct ring_buffer_event *event;
const u32 *entry = buf;

event = container_of(entry, struct ring_buffer_event, array[0]);
event->time_delta = export_delta;
}

Do I have to add a hook somewhere in the kernel/trace/ringbuffer.c ?
And then call it from ftrace_vearly_export_write ?



PATCH v2
--------------------------------------------------------------------

The very early tracing will start from the beginning of start_kernel()
and will stop at ftrace_init().

start_kernel()
{
ftrace_early_init() <--- start very early tracing
...
(calls)
...
ftrace_init() <--- stop very early tracing
early_trace_init(); <--- fill ring buffer, start function tracer if on
...
}

Currently, the events are saved in a temporary static buffer, which they
will be copied to the ring buffer when it will be allocated.

For the timestamp, we use rdtsc for x86, and sched_clock for the other
archs.

Currently dynamic tracing is not implemented with live patching, we use
ftrace_filter and ftrace_notrace to find which functions to be filtered
(traced / not traced), then during the callback from mcount hook, we do
binary search lookup to decide which function to be save or not.

Signed-off-by: Abderrahmane Benbachir <abderrahmane.benbachir@xxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
Cc: linux-kernel@xxxxxxxxxxxxxxx
---
arch/x86/Kconfig | 1 +
arch/x86/kernel/ftrace_32.S | 73 ++++++++++
arch/x86/kernel/ftrace_64.S | 14 ++
include/linux/ftrace.h | 16 ++-
init/main.c | 1 +
kernel/trace/Kconfig | 42 ++++++
kernel/trace/ftrace.c | 341 +++++++++++++++++++++++++++++++++++++++++++-
kernel/trace/trace.c | 3 +
8 files changed, 487 insertions(+), 4 deletions(-)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 971feac13506..03da741c088b 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -137,6 +137,7 @@ config X86
select HAVE_FENTRY if X86_64 || DYNAMIC_FTRACE
select HAVE_FTRACE_MCOUNT_RECORD
select HAVE_FUNCTION_GRAPH_TRACER
+ select HAVE_VERY_EARLY_FTRACE
select HAVE_FUNCTION_TRACER
select HAVE_GCC_PLUGINS
select HAVE_HW_BREAKPOINT
diff --git a/arch/x86/kernel/ftrace_32.S b/arch/x86/kernel/ftrace_32.S
index 722a145b4139..5d93037872b2 100644
--- a/arch/x86/kernel/ftrace_32.S
+++ b/arch/x86/kernel/ftrace_32.S
@@ -29,8 +29,81 @@ EXPORT_SYMBOL(mcount)
# define MCOUNT_FRAME 0 /* using frame = false */
#endif

+.macro save_mcount_regs
+#ifdef USING_FRAME_POINTER
+# ifdef CC_USING_FENTRY
+ /*
+ * Frame pointers are of ip followed by bp.
+ * Since fentry is an immediate jump, we are left with
+ * parent-ip, function-ip. We need to add a frame with
+ * parent-ip followed by ebp.
+ */
+ pushl 4(%esp) /* parent ip */
+ pushl %ebp
+ movl %esp, %ebp
+ pushl 2*4(%esp) /* function ip */
+# endif
+ /* For mcount, the function ip is directly above */
+ pushl %ebp
+ movl %esp, %ebp
+#endif
+ pushl %eax
+ pushl %ecx
+ pushl %edx
+ pushl $0 /* Pass NULL as regs pointer */
+
+#ifdef USING_FRAME_POINTER
+ /* Load parent ebp into edx */
+ movl 4*4(%esp), %edx
+#else
+ /* There's no frame pointer, load the appropriate stack addr instead */
+ lea 4*4(%esp), %edx
+#endif
+
+ movl (MCOUNT_FRAME+4)*4(%esp), %eax /* load the rip */
+ /* Get the parent ip */
+ movl 4(%edx), %edx /* edx has ebp */
+
+ movl function_trace_op, %ecx
+ subl $MCOUNT_INSN_SIZE, %eax
+ .endm
+
+.macro restore_mcount_regs
+ addl $4, %esp /* skip NULL pointer */
+ popl %edx
+ popl %ecx
+ popl %eax
+#ifdef USING_FRAME_POINTER
+ popl %ebp
+# ifdef CC_USING_FENTRY
+ addl $4,%esp /* skip function ip */
+ popl %ebp /* this is the orig bp */
+ addl $4, %esp /* skip parent ip */
+# endif
+#endif
+ .endm
+
+
ENTRY(function_hook)
+#ifdef CONFIG_VERY_EARLY_FUNCTION_TRACER
+ cmpl $__PAGE_OFFSET, %esp
+ jb vearly_stub /* Paging not enabled yet? */
+
+ cmpl $ftrace_stub, ftrace_vearly_trace_function
+ jnz vearly_trace
+
+vearly_stub:
ret
+
+vearly_trace:
+ save_mcount_regs
+ call *ftrace_vearly_trace_function
+ restore_mcount_regs
+
+ jmp vearly_stub
+#else
+ ret
+#endif
END(function_hook)

ENTRY(ftrace_caller)
diff --git a/arch/x86/kernel/ftrace_64.S b/arch/x86/kernel/ftrace_64.S
index 1dfac634bbf7..35f73a9dbbd2 100644
--- a/arch/x86/kernel/ftrace_64.S
+++ b/arch/x86/kernel/ftrace_64.S
@@ -146,7 +146,21 @@ EXPORT_SYMBOL(mcount)
#ifdef CONFIG_DYNAMIC_FTRACE

ENTRY(function_hook)
+#ifdef CONFIG_VERY_EARLY_FUNCTION_TRACER
+ cmpq $ftrace_stub, ftrace_vearly_trace_function
+ jnz vearly_trace
+
+vearly_stub:
retq
+
+vearly_trace:
+ save_mcount_regs
+ call *ftrace_vearly_trace_function
+ restore_mcount_regs
+ jmp vearly_stub
+#else
+ retq
+#endif
END(function_hook)

ENTRY(ftrace_caller)
diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
index 2e028854bac7..c3f2ba246fc9 100644
--- a/include/linux/ftrace.h
+++ b/include/linux/ftrace.h
@@ -272,6 +272,16 @@ static inline void ftrace_kill(void) { }
static inline void ftrace_free_init_mem(void) { }
#endif /* CONFIG_FUNCTION_TRACER */

+#ifdef CONFIG_VERY_EARLY_FUNCTION_TRACER
+extern void ftrace_early_init(char *command_line);
+extern void ftrace_early_shutdown(void);
+extern void ftrace_early_fill_ringbuffer(void *data);
+#else
+static inline void ftrace_early_init(char *command_line) { }
+static inline void ftrace_early_shutdown(void) { }
+static inline void ftrace_early_fill_ringbuffer(void *data) { }
+#endif
+
#ifdef CONFIG_STACK_TRACER

#define STACK_TRACE_ENTRIES 500
@@ -476,6 +486,10 @@ unsigned long ftrace_get_addr_curr(struct dyn_ftrace *rec);

extern ftrace_func_t ftrace_trace_function;

+#if defined(CONFIG_VERY_EARLY_FUNCTION_TRACER) && defined(CONFIG_DYNAMIC_FTRACE)
+extern ftrace_func_t ftrace_vearly_trace_function;
+#endif
+
int ftrace_regex_open(struct ftrace_ops *ops, int flag,
struct inode *inode, struct file *file);
ssize_t ftrace_filter_write(struct file *file, const char __user *ubuf,
@@ -757,7 +771,7 @@ static inline unsigned long get_lock_parent_ip(void)
#ifdef CONFIG_FTRACE_MCOUNT_RECORD
extern void ftrace_init(void);
#else
-static inline void ftrace_init(void) { }
+static inline void ftrace_init(void) { ftrace_early_shutdown(); }
#endif

/*
diff --git a/init/main.c b/init/main.c
index 0ee9c6866ada..c794292b18ef 100644
--- a/init/main.c
+++ b/init/main.c
@@ -511,6 +511,7 @@ asmlinkage __visible void __init start_kernel(void)
char *command_line;
char *after_dashes;

+ ftrace_early_init(boot_command_line);
set_task_stack_end_magic(&init_task);
smp_setup_processor_id();
debug_objects_early_init();
diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
index 434c840e2d82..0e1a9662bb2f 100644
--- a/kernel/trace/Kconfig
+++ b/kernel/trace/Kconfig
@@ -19,6 +19,11 @@ config HAVE_FUNCTION_TRACER
help
See Documentation/trace/ftrace-design.txt

+config HAVE_VERY_EARLY_FTRACE
+ bool
+ help
+ See Documentation/trace/ftrace-design.txt
+
config HAVE_FUNCTION_GRAPH_TRACER
bool
help
@@ -145,6 +150,43 @@ config FUNCTION_TRACER
(the bootup default), then the overhead of the instructions is very
small and not measurable even in micro-benchmarks.

+config VERY_EARLY_FUNCTION_TRACER
+ bool "Very Early Kernel Function Tracer"
+ depends on FUNCTION_TRACER
+ depends on HAVE_VERY_EARLY_FTRACE
+ default y
+ help
+ Enable very early function tracing. When ftrace_init will be called, all
+ mcount hooks will be replaced by nop, but before that (at very early
+ stage) mcounts are still there, and they are used for very early tracing.
+
+config VERY_EARLY_BUF_SHIFT
+ int "Temporary buffer size (17 => 128 KB, 24 => 16 MB)"
+ depends on VERY_EARLY_FUNCTION_TRACER
+ range 0 24
+ default 19
+ help
+ Select the size of the buffer to be used for storing function calls at
+ very early stage.
+ Examples:
+ 20 => 1 MB
+ 19 => 512 KB
+ 17 => 128 KB
+
+config VERY_EARLY_FILTER_SHIFT
+ int "Temporary filter size (filter/ntrace) (17 => 128 KB, 19 => 512 KB)"
+ depends on VERY_EARLY_FUNCTION_TRACER
+ depends on DYNAMIC_FTRACE
+ range 0 19
+ default 17
+ help
+ Select the size of the filter buffer to be used for filtering (trace/
+ no trace) functions at very early stage.
+ Examples:
+ 19 => 512 KB
+ 18 => 256 KB
+ 17 => 128 KB
+
config FUNCTION_GRAPH_TRACER
bool "Kernel Function Graph Tracer"
depends on HAVE_FUNCTION_GRAPH_TRACER
diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
index 8319e09e15b9..5e1b53662bbb 100644
--- a/kernel/trace/ftrace.c
+++ b/kernel/trace/ftrace.c
@@ -33,7 +33,7 @@
#include <linux/list.h>
#include <linux/hash.h>
#include <linux/rcupdate.h>
-
+#include <linux/trace.h>
#include <trace/events/sched.h>

#include <asm/sections.h>
@@ -65,6 +65,8 @@
#define FTRACE_HASH_MAX_BITS 12

#ifdef CONFIG_DYNAMIC_FTRACE
+extern unsigned long __start_mcount_loc[];
+extern unsigned long __stop_mcount_loc[];
#define INIT_OPS_HASH(opsname) \
.func_hash = &opsname.local_hash, \
.local_hash.regex_lock = __MUTEX_INITIALIZER(opsname.local_hash.regex_lock),
@@ -5912,11 +5914,11 @@ void __init ftrace_free_init_mem(void)

void __init ftrace_init(void)
{
- extern unsigned long __start_mcount_loc[];
- extern unsigned long __stop_mcount_loc[];
unsigned long count, flags;
int ret;

+ ftrace_early_shutdown();
+
local_irq_save(flags);
ret = ftrace_dyn_arch_init();
local_irq_restore(flags);
@@ -6882,3 +6884,336 @@ void ftrace_graph_exit_task(struct task_struct *t)
kfree(ret_stack);
}
#endif
+
+
+#ifdef CONFIG_VERY_EARLY_FUNCTION_TRACER
+
+#define VEARLY_BUF_LEN ((1 << CONFIG_VERY_EARLY_BUF_SHIFT) / \
+ sizeof(struct ftrace_vearly_entry))
+/*
+ * In very early stage :
+ * - only CPU0 is running in early, there is no need to store it
+ * - tsc counter is used in x86, otherwise we use sched_clock
+ */
+struct ftrace_vearly_entry {
+ unsigned long ip;
+ unsigned long parent_ip;
+ unsigned long long clock;
+};
+static struct ftrace_vearly_entry ftrace_vearly_entries[VEARLY_BUF_LEN] __initdata;
+static const unsigned long VERY_EARLY_BUF_LEN __initconst = VEARLY_BUF_LEN;
+static struct trace_export ftrace_vearly_export __initdata;
+static char tmp_cmdline[COMMAND_LINE_SIZE] __initdata;
+static unsigned int ftrace_vearly_enabled __initdata;
+static unsigned int vearly_entries_count __initdata;
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+ftrace_func_t ftrace_vearly_trace_function __read_mostly = ftrace_stub;
+#else
+# define ftrace_vearly_trace_function ftrace_trace_function
+#endif
+#define ftrace_vearly_disable() (ftrace_vearly_trace_function = ftrace_stub)
+
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+#define VEARLY_FILTER_SIZE ((1 << CONFIG_VERY_EARLY_FILTER_SHIFT) / \
+ sizeof(unsigned long))
+struct ftrace_vearly_filtering {
+ unsigned long list[VEARLY_FILTER_SIZE];
+ char buf[COMMAND_LINE_SIZE];
+ int size;
+};
+static const unsigned long FILTER_SIZE __initconst = VEARLY_FILTER_SIZE;
+static struct ftrace_vearly_filtering ftrace_data_notrace __initdata;
+static struct ftrace_vearly_filtering ftrace_data_filter __initdata;
+
+static __init int
+ftrace_vearly_match_record(unsigned long ip, struct ftrace_glob *func_g)
+{
+ char str[KSYM_SYMBOL_LEN];
+ char *modname;
+
+ kallsyms_lookup(ip, NULL, NULL, &modname, str);
+ return ftrace_match(str, func_g);
+}
+
+static __init int ftrace_vearly_filter_has_addr(unsigned long addr,
+ unsigned long *filter, int *size)
+{
+ int i;
+
+ for (i = 0; i < *size; i++) {
+ if (filter[i] == addr)
+ return 1;
+ }
+ return 0;
+}
+
+static __init void
+ftrace_vearly_regex(char *func, unsigned long *filter, int *size)
+{
+ struct ftrace_glob func_g = { .type = MATCH_FULL };
+ unsigned long *start = __start_mcount_loc;
+ unsigned long *end = __stop_mcount_loc;
+ unsigned long count;
+ unsigned long addr;
+ unsigned long *p;
+ int clear_filter = 0;
+
+ count = end - start;
+
+ if (!count)
+ return;
+
+ if (func) {
+ func_g.type = filter_parse_regex(func, strlen(func), &func_g.search,
+ &clear_filter);
+ func_g.len = strlen(func_g.search);
+ }
+
+ p = start;
+ while (p < end) {
+ addr = ftrace_call_adjust(*p++);
+ if (!addr)
+ continue;
+
+ if ((*size) > FILTER_SIZE)
+ return;
+
+ if (ftrace_vearly_match_record(addr, &func_g)) {
+ if (!ftrace_vearly_filter_has_addr(addr, filter, size))
+ filter[(*size)++] = addr;
+ }
+ }
+}
+
+static int __init ftrace_addr_compare(const void *a, const void *b)
+{
+ if (*(unsigned long *)a > *(unsigned long *)b)
+ return 1;
+ if (*(unsigned long *)a < *(unsigned long *)b)
+ return -1;
+
+ return 0;
+}
+
+static void __init ftrace_addr_swap(void *a, void *b, int size)
+{
+ unsigned long t = *(unsigned long *)a;
+ *(unsigned long *)a = *(unsigned long *)b;
+ *(unsigned long *)b = t;
+}
+
+static __init int set_ftrace_vearly_filtering(void *data, char *str)
+{
+ struct ftrace_vearly_filtering *ftrace_data = data;
+ char *func;
+ char *buf;
+
+ if (!ftrace_data)
+ return 0;
+ buf = ftrace_data->buf;
+ strlcpy(buf, str, COMMAND_LINE_SIZE);
+
+ while (buf) {
+ func = strsep(&buf, ",");
+ ftrace_vearly_regex(func, ftrace_data->list, &ftrace_data->size);
+ }
+ /* sort filter to use binary search on it */
+ sort(ftrace_data->list, ftrace_data->size,
+ sizeof(unsigned long), ftrace_addr_compare, ftrace_addr_swap);
+
+ return 1;
+}
+
+#define ftrace_vearly_bsearch_addr(addr, data) bsearch(&addr, data.list,\
+ data.size, sizeof(unsigned long), ftrace_addr_compare)
+
+#endif /* CONFIG_DYNAMIC_FTRACE */
+
+
+static __init void
+ftrace_function_vearly_trace_call(unsigned long ip, unsigned long parent_ip,
+ struct ftrace_ops *op, struct pt_regs *regs)
+{
+ struct ftrace_vearly_entry *entry;
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+ if (ftrace_data_notrace.size &&
+ ftrace_vearly_bsearch_addr(ip, ftrace_data_notrace))
+ return;
+
+ if (ftrace_data_filter.size &&
+ !ftrace_vearly_bsearch_addr(ip, ftrace_data_filter))
+ return;
+#endif
+
+ if (vearly_entries_count >= VERY_EARLY_BUF_LEN) {
+ /* stop tracing when buffer is full */
+ ftrace_vearly_disable();
+ return;
+ }
+
+ entry = &ftrace_vearly_entries[vearly_entries_count++];
+ entry->ip = ip;
+ entry->parent_ip = parent_ip;
+#ifdef CONFIG_X86_TSC
+ entry->clock = rdtsc();
+#else
+ entry->clock = trace_clock_local();
+#endif
+}
+
+/*
+ * ns = cycles * (10^6 / cpu_khz)
+ */
+static u64 __init cycles_to_ns(u64 cycles, u64 khz)
+{
+ u64 ns = cycles;
+ u64 dividend, divisor = 0;
+ const u64 constant = 1000000;
+
+ if (khz > constant) {
+ // ns /= (khz / constant);
+ dividend = khz;
+ divisor = constant;
+ do_div(dividend, divisor);
+ do_div(ns, dividend);
+ } else {
+ dividend = constant;
+ divisor = khz;
+ do_div(dividend, divisor);
+ ns *= dividend;
+ }
+ return ns;
+}
+
+/**
+ * delta btw the current and previous event, used in ftrace export to
+ * update entry time delta
+ */
+static u64 export_delta __initdata;
+/**
+ * update entry timestamp
+ */
+static void __init ftrace_vearly_export_write(const void *buf, unsigned int len)
+{
+ struct ring_buffer_event *event;
+ const u32 *entry = buf;
+
+ event = container_of(entry, struct ring_buffer_event, array[0]);
+ event->time_delta = export_delta;
+}
+
+void __init ftrace_early_fill_ringbuffer(void *data)
+{
+ struct ftrace_vearly_entry *entry;
+ struct trace_array *tr = data;
+ unsigned long cpu_khz;
+ u64 ns, prev_ns = 0;
+ int i;
+
+ if (!ftrace_vearly_enabled)
+ return;
+
+#ifdef CONFIG_X86_TSC
+ cpu_khz = native_calibrate_cpu();
+#endif
+
+ ftrace_vearly_export.write = ftrace_vearly_export_write;
+ register_ftrace_export(&ftrace_vearly_export);
+ preempt_disable_notrace();
+ for (i = 0; i < vearly_entries_count; i++) {
+ entry = &ftrace_vearly_entries[i];
+
+#ifdef CONFIG_X86_TSC
+ ns = cycles_to_ns(entry->clock, cpu_khz);
+#else
+ ns = entry->clock;
+#endif
+ export_delta = i > 0 ? ns - prev_ns : 0;
+ prev_ns = ns;
+ trace_function(tr, entry->ip, entry->parent_ip, 0, 0);
+ }
+ preempt_enable_notrace();
+ unregister_ftrace_export(&ftrace_vearly_export);
+}
+
+static __init int set_ftrace_vearly_enable(void *data, char *str)
+{
+ ftrace_vearly_enabled = 1;
+ return 1;
+}
+/*
+ * this will be used as __setup_param
+ */
+struct ftrace_vearly_obs_param {
+ int (*setup_func)(void *, char*);
+ const char *str;
+ void *data;
+};
+static struct ftrace_vearly_obs_param ftrace_vearly_params[] __initdata = {
+ { .str = "ftrace_vearly", .setup_func = set_ftrace_vearly_enable },
+#ifdef CONFIG_DYNAMIC_FTRACE
+ {
+ .str = "ftrace_notrace",
+ .data = &ftrace_data_notrace,
+ .setup_func = set_ftrace_vearly_filtering,
+ },
+ {
+ .str = "ftrace_filter",
+ .data = &ftrace_data_filter,
+ .setup_func = set_ftrace_vearly_filtering,
+ },
+#endif
+};
+
+static int __init ftrace_do_very_early_param(char *param, char *val,
+ const char *unused, void *arg)
+{
+ int size = ARRAY_SIZE(ftrace_vearly_params);
+ struct ftrace_vearly_obs_param *p;
+ int i;
+
+ for (i = 0; i < size; i++) {
+ p = &ftrace_vearly_params[i];
+ if (strcmp(param, p->str) == 0) {
+ p->setup_func(p->data, val);
+ return 0;
+ }
+ }
+ return 0;
+}
+
+void __init ftrace_early_init(char *command_line)
+{
+ /* proceed only if 'ftrace_vearly' exist */
+ if (!strstr(command_line, "ftrace_vearly"))
+ return;
+
+ strlcpy(tmp_cmdline, command_line, COMMAND_LINE_SIZE);
+ parse_args("ftrace vearly options", tmp_cmdline, NULL, 0, 0, 0, NULL,
+ ftrace_do_very_early_param);
+
+ if (ftrace_vearly_enabled)
+ ftrace_vearly_trace_function = ftrace_function_vearly_trace_call;
+}
+
+void __init ftrace_early_shutdown(void)
+{
+ if (!ftrace_vearly_enabled)
+ return;
+
+ ftrace_vearly_disable();
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+ pr_info("ftrace: vearly %u entries, notrace=%d, filter=%d",
+ vearly_entries_count,
+ ftrace_data_notrace.size,
+ ftrace_data_filter.size);
+#else
+ pr_info("ftrace: vearly %u recorded entries", vearly_entries_count);
+#endif
+}
+
+#endif /* CONFIG_VERY_EARLY_FUNCTION_TRACER */
diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 752e5daf0896..c097b97b3025 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -8364,6 +8364,9 @@ __init static int tracer_alloc_buffers(void)
/* Function tracing may start here (via kernel command line) */
init_function_trace();

+ /* Pre-fill the ring buffer with very early events */
+ ftrace_early_fill_ringbuffer(&global_trace);
+
/* All seems OK, enable tracing */
tracing_disabled = 0;

--
2.11.0