From: "Steven Rostedt (Red Hat)" <[email protected]>
The current method of handling multiple function callbacks is to register
a list function callback that calls all the other callbacks based on
their hash tables and compare it to the function that the callback was
called on. But this is very inefficient.
For example, if you are tracing all functions in the kernel and then
add a kprobe to a function such that the kprobe uses ftrace, the
mcount trampoline will switch from calling the function trace callback
to calling the list callback that will iterate over all registered
ftrace_ops (in this case, the function tracer and the kprobes callback).
That means for every function being traced it checks the hash of the
ftrace_ops for function tracing and kprobes, even though the kprobes
is only set at a single function. The kprobes ftrace_ops is checked
for every function being traced!
Instead of calling the list function for functions that are only being
traced by a single callback, we can call a dynamically allocated
trampoline that calls the callback directly. The function graph tracer
already uses a direct call trampoline when it is being traced by itself
but it is not dynamically allocated. It's trampoline is static in the
kernel core. The infrastructure that called the function graph trampoline
can also be used to call a dynamically allocated one.
For now, only ftrace_ops that are not dynamically allocated can have
a trampoline. That is, users such as function tracer or stack tracer.
kprobes and perf allocate their ftrace_ops, and until there's a safe
way to free the trampoline, it can not be used. The dynamically allocated
ftrace_ops may, although, use the trampoline if the kernel is not
compiled with CONFIG_PREEMPT. But that will come later.
Signed-off-by: Steven Rostedt <[email protected]>
---
arch/x86/kernel/ftrace.c | 195 ++++++++++++++++++++++++++++++++++++++++++--
arch/x86/kernel/mcount_64.S | 25 +++++-
include/linux/ftrace.h | 8 ++
kernel/trace/ftrace.c | 40 ++++++++-
4 files changed, 254 insertions(+), 14 deletions(-)
diff --git a/arch/x86/kernel/ftrace.c b/arch/x86/kernel/ftrace.c
index 3386dc9aa333..e4d48f6cad86 100644
--- a/arch/x86/kernel/ftrace.c
+++ b/arch/x86/kernel/ftrace.c
@@ -17,6 +17,7 @@
#include <linux/ftrace.h>
#include <linux/percpu.h>
#include <linux/sched.h>
+#include <linux/slab.h>
#include <linux/init.h>
#include <linux/list.h>
#include <linux/module.h>
@@ -644,13 +645,8 @@ int __init ftrace_dyn_arch_init(void)
{
return 0;
}
-#endif
-
-#ifdef CONFIG_FUNCTION_GRAPH_TRACER
-
-#ifdef CONFIG_DYNAMIC_FTRACE
-extern void ftrace_graph_call(void);
+#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
{
static union ftrace_code_union calc;
@@ -664,6 +660,193 @@ static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
*/
return calc.code;
}
+#endif
+
+/* Currently only x86_64 supports dynamic trampolines */
+#ifdef CONFIG_X86_64
+
+#ifdef CONFIG_MODULES
+#include <linux/moduleloader.h>
+/* Module allocation simplifies allocating memory for code */
+static inline void *alloc_tramp(unsigned long size)
+{
+ return module_alloc(size);
+}
+static inline void tramp_free(void *tramp)
+{
+ module_free(NULL, tramp);
+}
+#else
+/* Trampolines can only be created if modules are supported */
+static inline void *alloc_tramp(unsigned long size)
+{
+ return NULL;
+}
+static inline void tramp_free(void *tramp) { }
+#endif
+
+/* Defined as markers to the end of the ftrace default trampolines */
+extern void ftrace_caller_end(void);
+extern void ftrace_regs_caller_end(void);
+extern void ftrace_return(void);
+extern void ftrace_caller_op_ptr(void);
+extern void ftrace_regs_caller_op_ptr(void);
+
+/* movq function_trace_op(%rip), %rdx */
+/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
+#define OP_REF_SIZE 7
+
+/*
+ * The ftrace_ops is passed to the function callback. Since the
+ * trampoline only services a single ftrace_ops, we can pass in
+ * that ops directly.
+ *
+ * The ftrace_op_code_union is used to create a pointer to the
+ * ftrace_ops that will be passed to the callback function.
+ */
+union ftrace_op_code_union {
+ char code[OP_REF_SIZE];
+ struct {
+ char op[3];
+ int offset;
+ } __attribute__((packed));
+};
+
+static unsigned long create_trampoline(struct ftrace_ops *ops)
+{
+ unsigned const char *jmp;
+ unsigned long start_offset;
+ unsigned long end_offset;
+ unsigned long op_offset;
+ unsigned long offset;
+ unsigned long size;
+ unsigned long ip;
+ unsigned long *ptr;
+ void *trampoline;
+ /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
+ unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
+ union ftrace_op_code_union op_ptr;
+ int ret;
+
+ if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
+ start_offset = (unsigned long)ftrace_regs_caller;
+ end_offset = (unsigned long)ftrace_regs_caller_end;
+ op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
+ } else {
+ start_offset = (unsigned long)ftrace_caller;
+ end_offset = (unsigned long)ftrace_caller_end;
+ op_offset = (unsigned long)ftrace_caller_op_ptr;
+ }
+
+ size = end_offset - start_offset;
+
+ /*
+ * Allocate enough size to store the ftrace_caller code,
+ * the jmp to ftrace_return, as well as the address of
+ * the ftrace_ops this trampoline is used for.
+ */
+ trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
+ if (!trampoline)
+ return 0;
+
+ /* Copy ftrace_caller onto the trampoline memory */
+ ret = probe_kernel_read(trampoline, (void *)start_offset, size);
+ if (WARN_ON(ret < 0)) {
+ tramp_free(trampoline);
+ return 0;
+ }
+
+ ip = (unsigned long)trampoline + size;
+
+ /* The trampoline ends with a jmp to ftrace_return */
+ jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_return);
+ memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
+
+ /*
+ * The address of the ftrace_ops that is used for this trampoline
+ * is stored at the end of the trampoline. This will be used to
+ * load the third parameter for the callback. Basically, that
+ * location at the end of the trampoline takes the place of
+ * the global function_trace_op variable.
+ */
+
+ ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
+ *ptr = (unsigned long)ops;
+
+ op_offset -= start_offset;
+ memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
+
+ /* Are we pointing to the reference? */
+ if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
+ tramp_free(trampoline);
+ return 0;
+ }
+
+ /* Load the contents of ptr into the callback parameter */
+ offset = (unsigned long)ptr;
+ offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
+
+ op_ptr.offset = offset;
+
+ /* put in the new offset to the ftrace_ops */
+ memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
+
+ /* ALLOC_TRAMP flags lets us know we created it */
+ ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
+
+ return (unsigned long)trampoline;
+}
+
+void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
+{
+ ftrace_func_t func;
+ unsigned char *new;
+ unsigned long start_offset;
+ unsigned long call_offset;
+ unsigned long offset;
+ unsigned long ip;
+ int ret;
+
+ if (ops->trampoline) {
+ /*
+ * The ftrace_ops caller may set up its own trampoline.
+ * In such a case, this code must not modify it.
+ */
+ if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
+ return;
+ } else {
+ ops->trampoline = create_trampoline(ops);
+ if (!ops->trampoline)
+ return;
+ }
+
+ if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
+ start_offset = (unsigned long)ftrace_regs_caller;
+ call_offset = (unsigned long)ftrace_regs_call;
+ } else {
+ start_offset = (unsigned long)ftrace_caller;
+ call_offset = (unsigned long)ftrace_call;
+ }
+
+ offset = call_offset - start_offset;
+ ip = ops->trampoline + offset;
+
+ func = ftrace_ops_get_func(ops);
+
+ /* Do a safe modify in case the trampoline is executing */
+ new = ftrace_call_replace(ip, (unsigned long)func);
+ ret = update_ftrace_func(ip, new);
+
+ /* The update should never fail */
+ WARN_ON(ret);
+}
+#endif /* CONFIG_X86_64 */
+#endif /* CONFIG_DYNAMIC_FTRACE */
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+extern void ftrace_graph_call(void);
static int ftrace_mod_jmp(unsigned long ip, void *func)
{
diff --git a/arch/x86/kernel/mcount_64.S b/arch/x86/kernel/mcount_64.S
index c73aecf10d34..42f0cdd20baf 100644
--- a/arch/x86/kernel/mcount_64.S
+++ b/arch/x86/kernel/mcount_64.S
@@ -28,9 +28,11 @@ ENTRY(function_hook)
END(function_hook)
/* skip is set if stack has been adjusted */
-.macro ftrace_caller_setup skip=0
+.macro ftrace_caller_setup trace_label skip=0
MCOUNT_SAVE_FRAME \skip
+ /* Save this location */
+GLOBAL(\trace_label)
/* Load the ftrace_ops into the 3rd parameter */
movq function_trace_op(%rip), %rdx
@@ -46,7 +48,7 @@ END(function_hook)
.endm
ENTRY(ftrace_caller)
- ftrace_caller_setup
+ ftrace_caller_setup ftrace_caller_op_ptr
/* regs go into 4th parameter (but make it NULL) */
movq $0, %rcx
@@ -54,7 +56,14 @@ GLOBAL(ftrace_call)
call ftrace_stub
MCOUNT_RESTORE_FRAME
-ftrace_return:
+
+ /*
+ * The copied trampoline must call ftrace_return as it
+ * still may need to call the function graph tracer.
+ */
+GLOBAL(ftrace_caller_end)
+
+GLOBAL(ftrace_return)
#ifdef CONFIG_FUNCTION_GRAPH_TRACER
GLOBAL(ftrace_graph_call)
@@ -70,7 +79,7 @@ ENTRY(ftrace_regs_caller)
pushfq
/* skip=8 to skip flags saved in SS */
- ftrace_caller_setup 8
+ ftrace_caller_setup ftrace_regs_caller_op_ptr 8
/* Save the rest of pt_regs */
movq %r15, R15(%rsp)
@@ -122,6 +131,14 @@ GLOBAL(ftrace_regs_call)
/* Restore flags */
popfq
+ /*
+ * As this jmp to ftrace_return can be a short jump
+ * it must not be copied into the trampoline.
+ * The trampoline will add the code to jump
+ * to the return.
+ */
+GLOBAL(ftrace_regs_caller_end)
+
jmp ftrace_return
popfq
diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
index 662697babd48..06e3ca5a5083 100644
--- a/include/linux/ftrace.h
+++ b/include/linux/ftrace.h
@@ -94,6 +94,13 @@ ftrace_func_t ftrace_ops_get_func(struct ftrace_ops *ops);
* ADDING - The ops is in the process of being added.
* REMOVING - The ops is in the process of being removed.
* MODIFYING - The ops is in the process of changing its filter functions.
+ * ALLOC_TRAMP - A dynamic trampoline was allocated by the core code.
+ * The arch specific code sets this flag when it allocated a
+ * trampoline. This lets the arch know that it can update the
+ * trampoline in case the callback function changes.
+ * The ftrace_ops trampoline can be set by the ftrace users, and
+ * in such cases the arch must not modify it. Only the arch ftrace
+ * core code should set this flag.
*/
enum {
FTRACE_OPS_FL_ENABLED = 1 << 0,
@@ -108,6 +115,7 @@ enum {
FTRACE_OPS_FL_ADDING = 1 << 9,
FTRACE_OPS_FL_REMOVING = 1 << 10,
FTRACE_OPS_FL_MODIFYING = 1 << 11,
+ FTRACE_OPS_FL_ALLOC_TRAMP = 1 << 12,
};
#ifdef CONFIG_DYNAMIC_FTRACE
diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
index 31c90fec4158..15f85eac7e95 100644
--- a/kernel/trace/ftrace.c
+++ b/kernel/trace/ftrace.c
@@ -387,6 +387,8 @@ static int remove_ftrace_list_ops(struct ftrace_ops **list,
return ret;
}
+static void ftrace_update_trampoline(struct ftrace_ops *ops);
+
static int __register_ftrace_function(struct ftrace_ops *ops)
{
if (ops->flags & FTRACE_OPS_FL_DELETED)
@@ -419,6 +421,8 @@ static int __register_ftrace_function(struct ftrace_ops *ops)
} else
add_ftrace_ops(&ftrace_ops_list, ops);
+ ftrace_update_trampoline(ops);
+
if (ftrace_enabled)
update_ftrace_function();
@@ -3020,9 +3024,6 @@ ftrace_enabled_open(struct inode *inode, struct file *file)
{
struct ftrace_iterator *iter;
- if (unlikely(ftrace_disabled))
- return -ENODEV;
-
iter = __seq_open_private(file, &show_ftrace_seq_ops, sizeof(*iter));
if (iter) {
iter->pg = ftrace_pages_start;
@@ -3975,6 +3976,9 @@ static char ftrace_graph_buf[FTRACE_FILTER_SIZE] __initdata;
static char ftrace_graph_notrace_buf[FTRACE_FILTER_SIZE] __initdata;
static int ftrace_set_func(unsigned long *array, int *idx, int size, char *buffer);
+static unsigned long save_global_trampoline;
+static unsigned long save_global_flags;
+
static int __init set_graph_function(char *str)
{
strlcpy(ftrace_graph_buf, str, FTRACE_FILTER_SIZE);
@@ -4696,6 +4700,20 @@ void __init ftrace_init(void)
ftrace_disabled = 1;
}
+/* Do nothing if arch does not support this */
+void __weak arch_ftrace_update_trampoline(struct ftrace_ops *ops)
+{
+}
+
+static void ftrace_update_trampoline(struct ftrace_ops *ops)
+{
+ /* Currently, only non dynamic ops can have a trampoline */
+ if (ops->flags & FTRACE_OPS_FL_DYNAMIC)
+ return;
+
+ arch_ftrace_update_trampoline(ops);
+}
+
#else
static struct ftrace_ops global_ops = {
@@ -4738,6 +4756,10 @@ ftrace_ops_test(struct ftrace_ops *ops, unsigned long ip, void *regs)
return 1;
}
+static void ftrace_update_trampoline(struct ftrace_ops *ops)
+{
+}
+
#endif /* CONFIG_DYNAMIC_FTRACE */
__init void ftrace_init_global_array_ops(struct trace_array *tr)
@@ -5522,7 +5544,6 @@ int register_ftrace_graph(trace_func_graph_ret_t retfunc,
update_function_graph_func();
ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
-
out:
mutex_unlock(&ftrace_lock);
return ret;
@@ -5543,6 +5564,17 @@ void unregister_ftrace_graph(void)
unregister_pm_notifier(&ftrace_suspend_notifier);
unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
+#ifdef CONFIG_DYNAMIC_FTRACE
+ /*
+ * Function graph does not allocate the trampoline, but
+ * other global_ops do. We need to reset the ALLOC_TRAMP flag
+ * if one was used.
+ */
+ global_ops.trampoline = save_global_trampoline;
+ if (save_global_flags & FTRACE_OPS_FL_ALLOC_TRAMP)
+ global_ops.flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
+#endif
+
out:
mutex_unlock(&ftrace_lock);
}
--
2.1.1
H. Peter,
Can you give me your acked-by for this.
Thanks!
-- Steve
On Mon, 27 Oct 2014 14:27:03 -0400
Steven Rostedt <[email protected]> wrote:
> From: "Steven Rostedt (Red Hat)" <[email protected]>
>
> The current method of handling multiple function callbacks is to register
> a list function callback that calls all the other callbacks based on
> their hash tables and compare it to the function that the callback was
> called on. But this is very inefficient.
>
> For example, if you are tracing all functions in the kernel and then
> add a kprobe to a function such that the kprobe uses ftrace, the
> mcount trampoline will switch from calling the function trace callback
> to calling the list callback that will iterate over all registered
> ftrace_ops (in this case, the function tracer and the kprobes callback).
> That means for every function being traced it checks the hash of the
> ftrace_ops for function tracing and kprobes, even though the kprobes
> is only set at a single function. The kprobes ftrace_ops is checked
> for every function being traced!
>
> Instead of calling the list function for functions that are only being
> traced by a single callback, we can call a dynamically allocated
> trampoline that calls the callback directly. The function graph tracer
> already uses a direct call trampoline when it is being traced by itself
> but it is not dynamically allocated. It's trampoline is static in the
> kernel core. The infrastructure that called the function graph trampoline
> can also be used to call a dynamically allocated one.
>
> For now, only ftrace_ops that are not dynamically allocated can have
> a trampoline. That is, users such as function tracer or stack tracer.
> kprobes and perf allocate their ftrace_ops, and until there's a safe
> way to free the trampoline, it can not be used. The dynamically allocated
> ftrace_ops may, although, use the trampoline if the kernel is not
> compiled with CONFIG_PREEMPT. But that will come later.
>
> Signed-off-by: Steven Rostedt <[email protected]>
> ---
> arch/x86/kernel/ftrace.c | 195 ++++++++++++++++++++++++++++++++++++++++++--
> arch/x86/kernel/mcount_64.S | 25 +++++-
> include/linux/ftrace.h | 8 ++
> kernel/trace/ftrace.c | 40 ++++++++-
> 4 files changed, 254 insertions(+), 14 deletions(-)
>
> diff --git a/arch/x86/kernel/ftrace.c b/arch/x86/kernel/ftrace.c
> index 3386dc9aa333..e4d48f6cad86 100644
> --- a/arch/x86/kernel/ftrace.c
> +++ b/arch/x86/kernel/ftrace.c
> @@ -17,6 +17,7 @@
> #include <linux/ftrace.h>
> #include <linux/percpu.h>
> #include <linux/sched.h>
> +#include <linux/slab.h>
> #include <linux/init.h>
> #include <linux/list.h>
> #include <linux/module.h>
> @@ -644,13 +645,8 @@ int __init ftrace_dyn_arch_init(void)
> {
> return 0;
> }
> -#endif
> -
> -#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> -
> -#ifdef CONFIG_DYNAMIC_FTRACE
> -extern void ftrace_graph_call(void);
>
> +#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
> static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> {
> static union ftrace_code_union calc;
> @@ -664,6 +660,193 @@ static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> */
> return calc.code;
> }
> +#endif
> +
> +/* Currently only x86_64 supports dynamic trampolines */
> +#ifdef CONFIG_X86_64
> +
> +#ifdef CONFIG_MODULES
> +#include <linux/moduleloader.h>
> +/* Module allocation simplifies allocating memory for code */
> +static inline void *alloc_tramp(unsigned long size)
> +{
> + return module_alloc(size);
> +}
> +static inline void tramp_free(void *tramp)
> +{
> + module_free(NULL, tramp);
> +}
> +#else
> +/* Trampolines can only be created if modules are supported */
> +static inline void *alloc_tramp(unsigned long size)
> +{
> + return NULL;
> +}
> +static inline void tramp_free(void *tramp) { }
> +#endif
> +
> +/* Defined as markers to the end of the ftrace default trampolines */
> +extern void ftrace_caller_end(void);
> +extern void ftrace_regs_caller_end(void);
> +extern void ftrace_return(void);
> +extern void ftrace_caller_op_ptr(void);
> +extern void ftrace_regs_caller_op_ptr(void);
> +
> +/* movq function_trace_op(%rip), %rdx */
> +/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
> +#define OP_REF_SIZE 7
> +
> +/*
> + * The ftrace_ops is passed to the function callback. Since the
> + * trampoline only services a single ftrace_ops, we can pass in
> + * that ops directly.
> + *
> + * The ftrace_op_code_union is used to create a pointer to the
> + * ftrace_ops that will be passed to the callback function.
> + */
> +union ftrace_op_code_union {
> + char code[OP_REF_SIZE];
> + struct {
> + char op[3];
> + int offset;
> + } __attribute__((packed));
> +};
> +
> +static unsigned long create_trampoline(struct ftrace_ops *ops)
> +{
> + unsigned const char *jmp;
> + unsigned long start_offset;
> + unsigned long end_offset;
> + unsigned long op_offset;
> + unsigned long offset;
> + unsigned long size;
> + unsigned long ip;
> + unsigned long *ptr;
> + void *trampoline;
> + /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
> + unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
> + union ftrace_op_code_union op_ptr;
> + int ret;
> +
> + if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
> + start_offset = (unsigned long)ftrace_regs_caller;
> + end_offset = (unsigned long)ftrace_regs_caller_end;
> + op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
> + } else {
> + start_offset = (unsigned long)ftrace_caller;
> + end_offset = (unsigned long)ftrace_caller_end;
> + op_offset = (unsigned long)ftrace_caller_op_ptr;
> + }
> +
> + size = end_offset - start_offset;
> +
> + /*
> + * Allocate enough size to store the ftrace_caller code,
> + * the jmp to ftrace_return, as well as the address of
> + * the ftrace_ops this trampoline is used for.
> + */
> + trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
> + if (!trampoline)
> + return 0;
> +
> + /* Copy ftrace_caller onto the trampoline memory */
> + ret = probe_kernel_read(trampoline, (void *)start_offset, size);
> + if (WARN_ON(ret < 0)) {
> + tramp_free(trampoline);
> + return 0;
> + }
> +
> + ip = (unsigned long)trampoline + size;
> +
> + /* The trampoline ends with a jmp to ftrace_return */
> + jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_return);
> + memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
> +
> + /*
> + * The address of the ftrace_ops that is used for this trampoline
> + * is stored at the end of the trampoline. This will be used to
> + * load the third parameter for the callback. Basically, that
> + * location at the end of the trampoline takes the place of
> + * the global function_trace_op variable.
> + */
> +
> + ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
> + *ptr = (unsigned long)ops;
> +
> + op_offset -= start_offset;
> + memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
> +
> + /* Are we pointing to the reference? */
> + if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
> + tramp_free(trampoline);
> + return 0;
> + }
> +
> + /* Load the contents of ptr into the callback parameter */
> + offset = (unsigned long)ptr;
> + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
> +
> + op_ptr.offset = offset;
> +
> + /* put in the new offset to the ftrace_ops */
> + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
> +
> + /* ALLOC_TRAMP flags lets us know we created it */
> + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> +
> + return (unsigned long)trampoline;
> +}
> +
> +void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
> +{
> + ftrace_func_t func;
> + unsigned char *new;
> + unsigned long start_offset;
> + unsigned long call_offset;
> + unsigned long offset;
> + unsigned long ip;
> + int ret;
> +
> + if (ops->trampoline) {
> + /*
> + * The ftrace_ops caller may set up its own trampoline.
> + * In such a case, this code must not modify it.
> + */
> + if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
> + return;
> + } else {
> + ops->trampoline = create_trampoline(ops);
> + if (!ops->trampoline)
> + return;
> + }
> +
> + if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
> + start_offset = (unsigned long)ftrace_regs_caller;
> + call_offset = (unsigned long)ftrace_regs_call;
> + } else {
> + start_offset = (unsigned long)ftrace_caller;
> + call_offset = (unsigned long)ftrace_call;
> + }
> +
> + offset = call_offset - start_offset;
> + ip = ops->trampoline + offset;
> +
> + func = ftrace_ops_get_func(ops);
> +
> + /* Do a safe modify in case the trampoline is executing */
> + new = ftrace_call_replace(ip, (unsigned long)func);
> + ret = update_ftrace_func(ip, new);
> +
> + /* The update should never fail */
> + WARN_ON(ret);
> +}
> +#endif /* CONFIG_X86_64 */
> +#endif /* CONFIG_DYNAMIC_FTRACE */
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +extern void ftrace_graph_call(void);
>
> static int ftrace_mod_jmp(unsigned long ip, void *func)
> {
> diff --git a/arch/x86/kernel/mcount_64.S b/arch/x86/kernel/mcount_64.S
> index c73aecf10d34..42f0cdd20baf 100644
> --- a/arch/x86/kernel/mcount_64.S
> +++ b/arch/x86/kernel/mcount_64.S
> @@ -28,9 +28,11 @@ ENTRY(function_hook)
> END(function_hook)
>
> /* skip is set if stack has been adjusted */
> -.macro ftrace_caller_setup skip=0
> +.macro ftrace_caller_setup trace_label skip=0
> MCOUNT_SAVE_FRAME \skip
>
> + /* Save this location */
> +GLOBAL(\trace_label)
> /* Load the ftrace_ops into the 3rd parameter */
> movq function_trace_op(%rip), %rdx
>
> @@ -46,7 +48,7 @@ END(function_hook)
> .endm
>
> ENTRY(ftrace_caller)
> - ftrace_caller_setup
> + ftrace_caller_setup ftrace_caller_op_ptr
> /* regs go into 4th parameter (but make it NULL) */
> movq $0, %rcx
>
> @@ -54,7 +56,14 @@ GLOBAL(ftrace_call)
> call ftrace_stub
>
> MCOUNT_RESTORE_FRAME
> -ftrace_return:
> +
> + /*
> + * The copied trampoline must call ftrace_return as it
> + * still may need to call the function graph tracer.
> + */
> +GLOBAL(ftrace_caller_end)
> +
> +GLOBAL(ftrace_return)
>
> #ifdef CONFIG_FUNCTION_GRAPH_TRACER
> GLOBAL(ftrace_graph_call)
> @@ -70,7 +79,7 @@ ENTRY(ftrace_regs_caller)
> pushfq
>
> /* skip=8 to skip flags saved in SS */
> - ftrace_caller_setup 8
> + ftrace_caller_setup ftrace_regs_caller_op_ptr 8
>
> /* Save the rest of pt_regs */
> movq %r15, R15(%rsp)
> @@ -122,6 +131,14 @@ GLOBAL(ftrace_regs_call)
> /* Restore flags */
> popfq
>
> + /*
> + * As this jmp to ftrace_return can be a short jump
> + * it must not be copied into the trampoline.
> + * The trampoline will add the code to jump
> + * to the return.
> + */
> +GLOBAL(ftrace_regs_caller_end)
> +
> jmp ftrace_return
>
> popfq
> diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
> index 662697babd48..06e3ca5a5083 100644
> --- a/include/linux/ftrace.h
> +++ b/include/linux/ftrace.h
> @@ -94,6 +94,13 @@ ftrace_func_t ftrace_ops_get_func(struct ftrace_ops *ops);
> * ADDING - The ops is in the process of being added.
> * REMOVING - The ops is in the process of being removed.
> * MODIFYING - The ops is in the process of changing its filter functions.
> + * ALLOC_TRAMP - A dynamic trampoline was allocated by the core code.
> + * The arch specific code sets this flag when it allocated a
> + * trampoline. This lets the arch know that it can update the
> + * trampoline in case the callback function changes.
> + * The ftrace_ops trampoline can be set by the ftrace users, and
> + * in such cases the arch must not modify it. Only the arch ftrace
> + * core code should set this flag.
> */
> enum {
> FTRACE_OPS_FL_ENABLED = 1 << 0,
> @@ -108,6 +115,7 @@ enum {
> FTRACE_OPS_FL_ADDING = 1 << 9,
> FTRACE_OPS_FL_REMOVING = 1 << 10,
> FTRACE_OPS_FL_MODIFYING = 1 << 11,
> + FTRACE_OPS_FL_ALLOC_TRAMP = 1 << 12,
> };
>
> #ifdef CONFIG_DYNAMIC_FTRACE
> diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
> index 31c90fec4158..15f85eac7e95 100644
> --- a/kernel/trace/ftrace.c
> +++ b/kernel/trace/ftrace.c
> @@ -387,6 +387,8 @@ static int remove_ftrace_list_ops(struct ftrace_ops **list,
> return ret;
> }
>
> +static void ftrace_update_trampoline(struct ftrace_ops *ops);
> +
> static int __register_ftrace_function(struct ftrace_ops *ops)
> {
> if (ops->flags & FTRACE_OPS_FL_DELETED)
> @@ -419,6 +421,8 @@ static int __register_ftrace_function(struct ftrace_ops *ops)
> } else
> add_ftrace_ops(&ftrace_ops_list, ops);
>
> + ftrace_update_trampoline(ops);
> +
> if (ftrace_enabled)
> update_ftrace_function();
>
> @@ -3020,9 +3024,6 @@ ftrace_enabled_open(struct inode *inode, struct file *file)
> {
> struct ftrace_iterator *iter;
>
> - if (unlikely(ftrace_disabled))
> - return -ENODEV;
> -
> iter = __seq_open_private(file, &show_ftrace_seq_ops, sizeof(*iter));
> if (iter) {
> iter->pg = ftrace_pages_start;
> @@ -3975,6 +3976,9 @@ static char ftrace_graph_buf[FTRACE_FILTER_SIZE] __initdata;
> static char ftrace_graph_notrace_buf[FTRACE_FILTER_SIZE] __initdata;
> static int ftrace_set_func(unsigned long *array, int *idx, int size, char *buffer);
>
> +static unsigned long save_global_trampoline;
> +static unsigned long save_global_flags;
> +
> static int __init set_graph_function(char *str)
> {
> strlcpy(ftrace_graph_buf, str, FTRACE_FILTER_SIZE);
> @@ -4696,6 +4700,20 @@ void __init ftrace_init(void)
> ftrace_disabled = 1;
> }
>
> +/* Do nothing if arch does not support this */
> +void __weak arch_ftrace_update_trampoline(struct ftrace_ops *ops)
> +{
> +}
> +
> +static void ftrace_update_trampoline(struct ftrace_ops *ops)
> +{
> + /* Currently, only non dynamic ops can have a trampoline */
> + if (ops->flags & FTRACE_OPS_FL_DYNAMIC)
> + return;
> +
> + arch_ftrace_update_trampoline(ops);
> +}
> +
> #else
>
> static struct ftrace_ops global_ops = {
> @@ -4738,6 +4756,10 @@ ftrace_ops_test(struct ftrace_ops *ops, unsigned long ip, void *regs)
> return 1;
> }
>
> +static void ftrace_update_trampoline(struct ftrace_ops *ops)
> +{
> +}
> +
> #endif /* CONFIG_DYNAMIC_FTRACE */
>
> __init void ftrace_init_global_array_ops(struct trace_array *tr)
> @@ -5522,7 +5544,6 @@ int register_ftrace_graph(trace_func_graph_ret_t retfunc,
> update_function_graph_func();
>
> ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
> -
> out:
> mutex_unlock(&ftrace_lock);
> return ret;
> @@ -5543,6 +5564,17 @@ void unregister_ftrace_graph(void)
> unregister_pm_notifier(&ftrace_suspend_notifier);
> unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
>
> +#ifdef CONFIG_DYNAMIC_FTRACE
> + /*
> + * Function graph does not allocate the trampoline, but
> + * other global_ops do. We need to reset the ALLOC_TRAMP flag
> + * if one was used.
> + */
> + global_ops.trampoline = save_global_trampoline;
> + if (save_global_flags & FTRACE_OPS_FL_ALLOC_TRAMP)
> + global_ops.flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> +#endif
> +
> out:
> mutex_unlock(&ftrace_lock);
> }
Hi Steven,
(2014/10/28 3:27), Steven Rostedt wrote:
> +static unsigned long create_trampoline(struct ftrace_ops *ops)
> +{
> + unsigned const char *jmp;
> + unsigned long start_offset;
> + unsigned long end_offset;
> + unsigned long op_offset;
> + unsigned long offset;
> + unsigned long size;
> + unsigned long ip;
> + unsigned long *ptr;
> + void *trampoline;
> + /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
> + unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
> + union ftrace_op_code_union op_ptr;
> + int ret;
> +
> + if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
> + start_offset = (unsigned long)ftrace_regs_caller;
> + end_offset = (unsigned long)ftrace_regs_caller_end;
> + op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
> + } else {
> + start_offset = (unsigned long)ftrace_caller;
> + end_offset = (unsigned long)ftrace_caller_end;
> + op_offset = (unsigned long)ftrace_caller_op_ptr;
> + }
> +
> + size = end_offset - start_offset;
> +
> + /*
> + * Allocate enough size to store the ftrace_caller code,
> + * the jmp to ftrace_return, as well as the address of
> + * the ftrace_ops this trampoline is used for.
> + */
> + trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
> + if (!trampoline)
> + return 0;
> +
> + /* Copy ftrace_caller onto the trampoline memory */
> + ret = probe_kernel_read(trampoline, (void *)start_offset, size);
> + if (WARN_ON(ret < 0)) {
> + tramp_free(trampoline);
> + return 0;
> + }
> +
> + ip = (unsigned long)trampoline + size;
> +
> + /* The trampoline ends with a jmp to ftrace_return */
> + jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_return);
> + memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
> +
> + /*
> + * The address of the ftrace_ops that is used for this trampoline
> + * is stored at the end of the trampoline. This will be used to
> + * load the third parameter for the callback. Basically, that
> + * location at the end of the trampoline takes the place of
> + * the global function_trace_op variable.
> + */
> +
> + ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
> + *ptr = (unsigned long)ops;
> +
> + op_offset -= start_offset;
> + memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
> +
> + /* Are we pointing to the reference? */
> + if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
> + tramp_free(trampoline);
> + return 0;
> + }
> +
> + /* Load the contents of ptr into the callback parameter */
> + offset = (unsigned long)ptr;
> + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
> +
> + op_ptr.offset = offset;
> +
> + /* put in the new offset to the ftrace_ops */
> + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
Would we better call flush_icache_range here?
Thank you,
> +
> + /* ALLOC_TRAMP flags lets us know we created it */
> + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> +
> + return (unsigned long)trampoline;
> +}
--
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Research Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: [email protected]
On Fri, 31 Oct 2014 14:19:03 +0900
Masami Hiramatsu <[email protected]> wrote:
> > + /* Load the contents of ptr into the callback parameter */
> > + offset = (unsigned long)ptr;
> > + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
> > +
> > + op_ptr.offset = offset;
> > +
> > + /* put in the new offset to the ftrace_ops */
> > + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
>
> Would we better call flush_icache_range here?
Do we need to? We just allocated memory to execute. It shouldn't be in
any CPU cache on the machine.
-- Steve
>
> Thank you,
>
> > +
> > + /* ALLOC_TRAMP flags lets us know we created it */
> > + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> > +
> > + return (unsigned long)trampoline;
> > +}
>
>
(2014/11/01 1:01), Steven Rostedt wrote:
> On Fri, 31 Oct 2014 14:19:03 +0900
> Masami Hiramatsu <[email protected]> wrote:
>
>
>>> + /* Load the contents of ptr into the callback parameter */
>>> + offset = (unsigned long)ptr;
>>> + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
>>> +
>>> + op_ptr.offset = offset;
>>> +
>>> + /* put in the new offset to the ftrace_ops */
>>> + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
>>
>> Would we better call flush_icache_range here?
>
> Do we need to? We just allocated memory to execute. It shouldn't be in
> any CPU cache on the machine.
Indeed, if the trampoline buffer is not re-used (and it seems so),
we don't need to care about it.
OK, I've reviewed that :).
Thank you,
>
> -- Steve
>
>
>>
>> Thank you,
>>
>>> +
>>> + /* ALLOC_TRAMP flags lets us know we created it */
>>> + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
>>> +
>>> + return (unsigned long)trampoline;
>>> +}
>>
>>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
--
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Research Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: [email protected]
On Mon, Oct 27, 2014 at 02:27:03PM -0400, Steven Rostedt wrote:
> From: "Steven Rostedt (Red Hat)" <[email protected]>
>
> The current method of handling multiple function callbacks is to register
> a list function callback that calls all the other callbacks based on
> their hash tables and compare it to the function that the callback was
> called on. But this is very inefficient.
>
> For example, if you are tracing all functions in the kernel and then
> add a kprobe to a function such that the kprobe uses ftrace, the
> mcount trampoline will switch from calling the function trace callback
> to calling the list callback that will iterate over all registered
> ftrace_ops (in this case, the function tracer and the kprobes callback).
> That means for every function being traced it checks the hash of the
> ftrace_ops for function tracing and kprobes, even though the kprobes
> is only set at a single function. The kprobes ftrace_ops is checked
> for every function being traced!
>
> Instead of calling the list function for functions that are only being
> traced by a single callback, we can call a dynamically allocated
> trampoline that calls the callback directly. The function graph tracer
> already uses a direct call trampoline when it is being traced by itself
> but it is not dynamically allocated. It's trampoline is static in the
> kernel core. The infrastructure that called the function graph trampoline
> can also be used to call a dynamically allocated one.
>
> For now, only ftrace_ops that are not dynamically allocated can have
> a trampoline. That is, users such as function tracer or stack tracer.
> kprobes and perf allocate their ftrace_ops, and until there's a safe
> way to free the trampoline, it can not be used. The dynamically allocated
> ftrace_ops may, although, use the trampoline if the kernel is not
> compiled with CONFIG_PREEMPT. But that will come later.
>
> Signed-off-by: Steven Rostedt <[email protected]>
> ---
> arch/x86/kernel/ftrace.c | 195 ++++++++++++++++++++++++++++++++++++++++++--
> arch/x86/kernel/mcount_64.S | 25 +++++-
> include/linux/ftrace.h | 8 ++
> kernel/trace/ftrace.c | 40 ++++++++-
> 4 files changed, 254 insertions(+), 14 deletions(-)
>
> diff --git a/arch/x86/kernel/ftrace.c b/arch/x86/kernel/ftrace.c
> index 3386dc9aa333..e4d48f6cad86 100644
> --- a/arch/x86/kernel/ftrace.c
> +++ b/arch/x86/kernel/ftrace.c
> @@ -17,6 +17,7 @@
> #include <linux/ftrace.h>
> #include <linux/percpu.h>
> #include <linux/sched.h>
> +#include <linux/slab.h>
> #include <linux/init.h>
> #include <linux/list.h>
> #include <linux/module.h>
> @@ -644,13 +645,8 @@ int __init ftrace_dyn_arch_init(void)
> {
> return 0;
> }
> -#endif
> -
> -#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> -
> -#ifdef CONFIG_DYNAMIC_FTRACE
> -extern void ftrace_graph_call(void);
>
> +#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
> static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> {
> static union ftrace_code_union calc;
> @@ -664,6 +660,193 @@ static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> */
> return calc.code;
> }
> +#endif
> +
> +/* Currently only x86_64 supports dynamic trampolines */
> +#ifdef CONFIG_X86_64
> +
> +#ifdef CONFIG_MODULES
> +#include <linux/moduleloader.h>
> +/* Module allocation simplifies allocating memory for code */
... because it makes it EXEC too?
> +static inline void *alloc_tramp(unsigned long size)
Yeah, naming is, hmm, rated R. But we talked about it :-)
> +{
> + return module_alloc(size);
> +}
> +static inline void tramp_free(void *tramp)
> +{
> + module_free(NULL, tramp);
> +}
> +#else
> +/* Trampolines can only be created if modules are supported */
Just for my own understanding: why? You're not loading additional .ko's
right?
> +static inline void *alloc_tramp(unsigned long size)
> +{
> + return NULL;
> +}
> +static inline void tramp_free(void *tramp) { }
> +#endif
> +
> +/* Defined as markers to the end of the ftrace default trampolines */
> +extern void ftrace_caller_end(void);
> +extern void ftrace_regs_caller_end(void);
> +extern void ftrace_return(void);
> +extern void ftrace_caller_op_ptr(void);
> +extern void ftrace_regs_caller_op_ptr(void);
> +
> +/* movq function_trace_op(%rip), %rdx */
> +/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
> +#define OP_REF_SIZE 7
> +
> +/*
> + * The ftrace_ops is passed to the function callback. Since the
> + * trampoline only services a single ftrace_ops, we can pass in
> + * that ops directly.
> + *
> + * The ftrace_op_code_union is used to create a pointer to the
> + * ftrace_ops that will be passed to the callback function.
> + */
> +union ftrace_op_code_union {
> + char code[OP_REF_SIZE];
> + struct {
> + char op[3];
> + int offset;
> + } __attribute__((packed));
> +};
> +
> +static unsigned long create_trampoline(struct ftrace_ops *ops)
> +{
> + unsigned const char *jmp;
> + unsigned long start_offset;
> + unsigned long end_offset;
> + unsigned long op_offset;
> + unsigned long offset;
> + unsigned long size;
> + unsigned long ip;
> + unsigned long *ptr;
> + void *trampoline;
> + /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
> + unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
Ok, this 0x15 is the ModRM which says that the destination register is
hammered to be %rdx. I guess this has something to do with the magic
third parameter thing where ftrace_ops go into...
> + union ftrace_op_code_union op_ptr;
> + int ret;
> +
> + if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
> + start_offset = (unsigned long)ftrace_regs_caller;
> + end_offset = (unsigned long)ftrace_regs_caller_end;
> + op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
> + } else {
> + start_offset = (unsigned long)ftrace_caller;
> + end_offset = (unsigned long)ftrace_caller_end;
> + op_offset = (unsigned long)ftrace_caller_op_ptr;
> + }
> +
> + size = end_offset - start_offset;
> +
> + /*
> + * Allocate enough size to store the ftrace_caller code,
> + * the jmp to ftrace_return, as well as the address of
> + * the ftrace_ops this trampoline is used for.
> + */
> + trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
> + if (!trampoline)
> + return 0;
> +
> + /* Copy ftrace_caller onto the trampoline memory */
> + ret = probe_kernel_read(trampoline, (void *)start_offset, size);
> + if (WARN_ON(ret < 0)) {
> + tramp_free(trampoline);
A tramp-free trampoline? :-P
Btw, "trmp" could be a good, short and politically correct prefix to
this stuff:
trmp_alloc()
trmp_free()
...
:-)
> + return 0;
> + }
> +
> + ip = (unsigned long)trampoline + size;
> +
> + /* The trampoline ends with a jmp to ftrace_return */
> + jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_return);
> + memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
> +
> + /*
> + * The address of the ftrace_ops that is used for this trampoline
> + * is stored at the end of the trampoline. This will be used to
> + * load the third parameter for the callback. Basically, that
> + * location at the end of the trampoline takes the place of
> + * the global function_trace_op variable.
> + */
> +
Spurious newline.
> + ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
Maybe we should call this variable ftrace_ops_ptr to be more clear?
> + *ptr = (unsigned long)ops;
> +
> + op_offset -= start_offset;
> + memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
I guess we can copy only 3 bytes here and not OP_REF_SIZE as we
overwrite op_ptr.offset later anyway.
> +
> + /* Are we pointing to the reference? */
"Are we pointing to the movq" is what we ask here, no?
> + if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
> + tramp_free(trampoline);
> + return 0;
> + }
> +
> + /* Load the contents of ptr into the callback parameter */
> + offset = (unsigned long)ptr;
> + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
> +
> + op_ptr.offset = offset;
> +
> + /* put in the new offset to the ftrace_ops */
> + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
> +
> + /* ALLOC_TRAMP flags lets us know we created it */
> + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> +
> + return (unsigned long)trampoline;
> +}
> +
Yeah, it looks ok as far as I can tell by looking around and trying to
piece it together from staring at ftrace-design.txt and the rest of the
fun.
But don't take my word for it :-)
--
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--
On Wed, 5 Nov 2014 11:28:01 +0100
Borislav Petkov <[email protected]> wrote:
> > --- a/arch/x86/kernel/ftrace.c
> > +++ b/arch/x86/kernel/ftrace.c
> > @@ -17,6 +17,7 @@
> > #include <linux/ftrace.h>
> > #include <linux/percpu.h>
> > #include <linux/sched.h>
> > +#include <linux/slab.h>
> > #include <linux/init.h>
> > #include <linux/list.h>
> > #include <linux/module.h>
> > @@ -644,13 +645,8 @@ int __init ftrace_dyn_arch_init(void)
> > {
> > return 0;
> > }
> > -#endif
> > -
> > -#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> > -
> > -#ifdef CONFIG_DYNAMIC_FTRACE
> > -extern void ftrace_graph_call(void);
> >
> > +#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
> > static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> > {
> > static union ftrace_code_union calc;
> > @@ -664,6 +660,193 @@ static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
> > */
> > return calc.code;
> > }
> > +#endif
> > +
> > +/* Currently only x86_64 supports dynamic trampolines */
> > +#ifdef CONFIG_X86_64
> > +
> > +#ifdef CONFIG_MODULES
> > +#include <linux/moduleloader.h>
> > +/* Module allocation simplifies allocating memory for code */
>
> ... because it makes it EXEC too?
Yeah, I should have specified that.
>
> > +static inline void *alloc_tramp(unsigned long size)
>
> Yeah, naming is, hmm, rated R. But we talked about it :-)
It's used in Disney movies, I doubt it's rated R.
>
> > +{
> > + return module_alloc(size);
> > +}
> > +static inline void tramp_free(void *tramp)
> > +{
> > + module_free(NULL, tramp);
> > +}
> > +#else
> > +/* Trampolines can only be created if modules are supported */
>
> Just for my own understanding: why? You're not loading additional .ko's
> right?
It goes back to the EXEC issue. The module allocation goes through
leaps and bounds to get memory allocated for EXEC correct. I wasn't
about to duplicate that.
Ideally, there should be a generic way to allocate memory that will be
used for executing code, but right now we depend on module code.
>
> > +static inline void *alloc_tramp(unsigned long size)
> > +{
> > + return NULL;
> > +}
> > +static inline void tramp_free(void *tramp) { }
> > +#endif
> > +
> > +/* Defined as markers to the end of the ftrace default trampolines */
> > +extern void ftrace_caller_end(void);
> > +extern void ftrace_regs_caller_end(void);
> > +extern void ftrace_return(void);
> > +extern void ftrace_caller_op_ptr(void);
> > +extern void ftrace_regs_caller_op_ptr(void);
> > +
> > +/* movq function_trace_op(%rip), %rdx */
> > +/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
> > +#define OP_REF_SIZE 7
> > +
> > +/*
> > + * The ftrace_ops is passed to the function callback. Since the
> > + * trampoline only services a single ftrace_ops, we can pass in
> > + * that ops directly.
> > + *
> > + * The ftrace_op_code_union is used to create a pointer to the
> > + * ftrace_ops that will be passed to the callback function.
> > + */
> > +union ftrace_op_code_union {
> > + char code[OP_REF_SIZE];
> > + struct {
> > + char op[3];
> > + int offset;
> > + } __attribute__((packed));
> > +};
> > +
> > +static unsigned long create_trampoline(struct ftrace_ops *ops)
> > +{
> > + unsigned const char *jmp;
> > + unsigned long start_offset;
> > + unsigned long end_offset;
> > + unsigned long op_offset;
> > + unsigned long offset;
> > + unsigned long size;
> > + unsigned long ip;
> > + unsigned long *ptr;
> > + void *trampoline;
> > + /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
> > + unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
>
> Ok, this 0x15 is the ModRM which says that the destination register is
> hammered to be %rdx. I guess this has something to do with the magic
> third parameter thing where ftrace_ops go into...
Yes, the ftrace_ops is passed in as the third parameter.
>
> > + union ftrace_op_code_union op_ptr;
> > + int ret;
> > +
> > + if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
> > + start_offset = (unsigned long)ftrace_regs_caller;
> > + end_offset = (unsigned long)ftrace_regs_caller_end;
> > + op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
> > + } else {
> > + start_offset = (unsigned long)ftrace_caller;
> > + end_offset = (unsigned long)ftrace_caller_end;
> > + op_offset = (unsigned long)ftrace_caller_op_ptr;
> > + }
> > +
> > + size = end_offset - start_offset;
> > +
> > + /*
> > + * Allocate enough size to store the ftrace_caller code,
> > + * the jmp to ftrace_return, as well as the address of
> > + * the ftrace_ops this trampoline is used for.
> > + */
> > + trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
> > + if (!trampoline)
> > + return 0;
> > +
> > + /* Copy ftrace_caller onto the trampoline memory */
> > + ret = probe_kernel_read(trampoline, (void *)start_offset, size);
> > + if (WARN_ON(ret < 0)) {
> > + tramp_free(trampoline);
>
> A tramp-free trampoline? :-P
>
> Btw, "trmp" could be a good, short and politically correct prefix to
> this stuff:
>
> trmp_alloc()
> trmp_free()
I just see "Trump" which to me is less politically correct than
Tramp ;-)
> ...
>
> :-)
>
> > + return 0;
> > + }
> > +
> > + ip = (unsigned long)trampoline + size;
> > +
> > + /* The trampoline ends with a jmp to ftrace_return */
> > + jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_return);
> > + memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
> > +
> > + /*
> > + * The address of the ftrace_ops that is used for this trampoline
> > + * is stored at the end of the trampoline. This will be used to
> > + * load the third parameter for the callback. Basically, that
> > + * location at the end of the trampoline takes the place of
> > + * the global function_trace_op variable.
> > + */
> > +
>
> Spurious newline.
>
> > + ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
>
> Maybe we should call this variable ftrace_ops_ptr to be more clear?
>
> > + *ptr = (unsigned long)ops;
> > +
> > + op_offset -= start_offset;
> > + memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
>
> I guess we can copy only 3 bytes here and not OP_REF_SIZE as we
> overwrite op_ptr.offset later anyway.
Yeah I guess we could.
>
> > +
> > + /* Are we pointing to the reference? */
>
> "Are we pointing to the movq" is what we ask here, no?
Yep. I guess I should have said, "Are we pointing to the correct
label?".
>
> > + if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
> > + tramp_free(trampoline);
> > + return 0;
> > + }
> > +
> > + /* Load the contents of ptr into the callback parameter */
> > + offset = (unsigned long)ptr;
> > + offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
> > +
> > + op_ptr.offset = offset;
> > +
> > + /* put in the new offset to the ftrace_ops */
> > + memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
> > +
> > + /* ALLOC_TRAMP flags lets us know we created it */
> > + ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
> > +
> > + return (unsigned long)trampoline;
> > +}
> > +
>
> Yeah, it looks ok as far as I can tell by looking around and trying to
> piece it together from staring at ftrace-design.txt and the rest of the
> fun.
>
> But don't take my word for it :-)
>
Thanks,
I may update some of the comments and such. It will have to be a
separate patch. I already tested the hell out of these and have a bunch
of commits on top that have also been fully tested. Any rebase would
throw all that away.
-- Steve
On Thu, Nov 06, 2014 at 08:57:15AM -0500, Steven Rostedt wrote:
> > > +/* Trampolines can only be created if modules are supported */
> >
> > Just for my own understanding: why? You're not loading additional .ko's
> > right?
>
> It goes back to the EXEC issue. The module allocation goes through
> leaps and bounds to get memory allocated for EXEC correct. I wasn't
> about to duplicate that.
>
> Ideally, there should be a generic way to allocate memory that will be
> used for executing code, but right now we depend on module code.
Yah, that could be a nice addition (in a separate patch, ontop, as you
say) to explain the dependency on CONFIG_MODULES, which is not trivially
apparent.
--
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--