bcc Reference Guide

Intended for search (Ctrl-F) and reference. For tutorials, start with tutorial.md.

This guide is incomplete. If something feels missing, check the bcc and kernel source. And if you confirm we're missing something, please send a pull request to fix it, and help out everyone.

BPF C

This section describes the C part of a bcc program.

Events & Arguments

1. kprobes

Syntax: kprobe__kernel_function_name

kprobe__ is a special prefix that creates a kprobe (dynamic tracing of a kernel function call) for the kernel function name provided as the remainder. You can also use kprobes by declaring a normal C function, then using the Python BPF.attach_kprobe() (covered later) to associate it with a kernel function.

Arguments are specified on the function declaration: kprobe__kernel_function_name(struct pt_regs *ctx [, argument1 ...])

For example:

int kprobe__tcp_v4_connect(struct pt_regs *ctx, struct sock *sk)
    [...]
}

This instruments the tcp_v4_connect() kernel function using a kprobe, with the following arguments:

struct pt_regs *ctx: Registers and BPF context.
struct sock *sk: First argument to tcp_v4_connect().

The first argument is always struct pt_regs *, the remainder are the arguments to the function (they don't need to be specified, if you don't intend to use them).

Examples in situ: code (output), code (output)

2. kretprobes

Syntax: kretprobe__kernel_function_name

kretprobe__ is a special prefix that creates a kretprobe (dynamic tracing of a kernel function return) for the kernel function name provided as the remainder. You can also use kretprobes by declaring a normal C function, then using the Python BPF.attach_kretprobe() (covered later) to associate it with a kernel function.

Return value is available as PT_REGS_RC(ctx), given a function declaration of: kretprobe__kernel_function_name(struct pt_regs *ctx)

For example:

int kretprobe__tcp_v4_connect(struct pt_regs *ctx)
{
    int ret = PT_REGS_RC(ctx);
    [...]
}

This instruments the return of the tcp_v4_connect() kernel function using a kretprobe, and stores the return value in ret.

Examples in situ: code (output)

3. Tracepoints

Syntax: TRACEPOINT_PROBE(category, event)

This is a macro that instruments the tracepoint defined by category:event.

Arguments are available in an args struct, which are the tracepoint arguments. One way to list these is to cat the relevant format file under /sys/kernel/debug/tracing/events/category/event/format.

The args struct can be used in place of ctx in each functions requiring a context as an argument. This includes notably perf_submit().

For example:

TRACEPOINT_PROBE(random, urandom_read) {
    // args is from /sys/kernel/debug/tracing/events/random/urandom_read/format
    bpf_trace_printk("%d\\n", args->got_bits);
    return 0;
}

This instruments the random:urandom_read tracepoint, and prints the tracepoint argument got_bits.

Examples in situ: code (output), search /examples, search /tools

4. uprobes

These are instrumented by declaring a normal function in C, then associating it as a uprobe probe in Python via BPF.attach_uprobe() (covered later).

Arguments can be examined using PT_REGS_PARM macros.

For example:

int count(struct pt_regs *ctx) {
    char buf[64];
    bpf_probe_read_user(&buf, sizeof(buf), (void *)PT_REGS_PARM1(ctx));
    bpf_trace_printk("%s %d", buf, PT_REGS_PARM2(ctx));
    return(0);
}

This reads the first argument as a string, and then prints it with the second argument as an integer.

Examples in situ: code

5. uretprobes

These are instrumented by declaring a normal function in C, then associating it as a uretprobe probe in Python via BPF.attach_uretprobe() (covered later).

Return value is available as PT_REGS_RC(ctx), given a function declaration of: function_name(struct pt_regs *ctx)

For example:

BPF_HISTOGRAM(dist);
int count(struct pt_regs *ctx) {
    dist.increment(PT_REGS_RC(ctx));
    return 0;
}

This increments the bucket in the dist histogram that is indexed by the return value.

Examples in situ: code (output), code (output)

6. USDT probes

These are User Statically-Defined Tracing (USDT) probes, which may be placed in some applications or libraries to provide a user-level equivalent of tracepoints. The primary BPF method provided for USDT support method is enable_probe(). USDT probes are instrumented by declaring a normal function in C, then associating it as a USDT probe in Python via USDT.enable_probe().

Arguments can be read via: bpf_usdt_readarg(index, ctx, &addr)

For example:

int do_trace(struct pt_regs *ctx) {
    uint64_t addr;
    char path[128];
    bpf_usdt_readarg(6, ctx, &addr);
    bpf_probe_read_user(&path, sizeof(path), (void *)addr);
    bpf_trace_printk("path:%s\\n", path);
    return 0;
};

This reads the sixth USDT argument, and then pulls it in as a string to path.

When initializing USDTs via the third argument of BPF::init in the C API, if any USDT fails to init, entire BPF::init will fail. If you're OK with some USDTs failing to init, use BPF::init_usdt before calling BPF::init.

Examples in situ: code, search /examples, search /tools

7. Raw Tracepoints

Syntax: RAW_TRACEPOINT_PROBE(event)

This is a macro that instruments the raw tracepoint defined by event.

The argument is a pointer to struct bpf_raw_tracepoint_args, which is defined in bpf.h. The struct field args contains all parameters of the raw tracepoint where you can found at linux tree include/trace/events directory.

For example:

RAW_TRACEPOINT_PROBE(sched_switch)
{
    // TP_PROTO(bool preempt, struct task_struct *prev, struct task_struct *next)
    struct task_struct *prev = (struct task_struct *)ctx->args[1];
    struct task_struct *next= (struct task_struct *)ctx->args[2];
    s32 prev_tgid, next_tgid;

    bpf_probe_read_kernel(&prev_tgid, sizeof(prev->tgid), &prev->tgid);
    bpf_probe_read_kernel(&next_tgid, sizeof(next->tgid), &next->tgid);
    bpf_trace_printk("%d -> %d\\n", prev_tgid, next_tgid);
}

This instruments the sched:sched_switch tracepoint, and prints the prev and next tgid.

Examples in situ: search /tools

8. system call tracepoints

Syntax: syscall__SYSCALLNAME

syscall__ is a special prefix that creates a kprobe for the system call name provided as the remainder. You can use it by declaring a normal C function, then using the Python BPF.get_syscall_fnname(SYSCALLNAME) and BPF.attach_kprobe() to associate it.

Arguments are specified on the function declaration: syscall__SYSCALLNAME(struct pt_regs *ctx, [, argument1 ...]).

For example:

int syscall__execve(struct pt_regs *ctx,
    const char __user *filename,
    const char __user *const __user *__argv,
    const char __user *const __user *__envp)
{
    [...]
}

This instruments the execve system call.

The first argument is always struct pt_regs *, the remainder are the arguments to the function (they don't need to be specified, if you don't intend to use them).

Corresponding Python code:

b = BPF(text=bpf_text)
execve_fnname = b.get_syscall_fnname("execve")
b.attach_kprobe(event=execve_fnname, fn_name="syscall__execve")

Examples in situ: code (output)

9. kfuncs

Syntax: KFUNC_PROBE(function, typeof(arg1) arg1, typeof(arg2) arge ...)

This is a macro that instruments the kernel function via trampoline before the function is executed. It's defined by function name and the function arguments defined as argX.

For example:

KFUNC_PROBE(do_sys_open, int dfd, const char *filename, int flags, int mode)
{
    ...

This instruments the do_sys_open kernel function and make its arguments accessible as standard argument values.

Examples in situ: search /tools

10. kretfuncs

Syntax: KRETFUNC_PROBE(event, typeof(arg1) arg1, typeof(arg2) arge ..., int ret)

This is a macro that instruments the kernel function via trampoline after the function is executed. It's defined by function name and the function arguments defined as argX.

The last argument of the probe is the return value of the instrumented function.

For example:

KRETFUNC_PROBE(do_sys_open, int dfd, const char *filename, int flags, int mode, int ret)
{
    ...

This instruments the do_sys_open kernel function and make its arguments accessible as standard argument values together with its return value.

Examples in situ: search /tools

11. LSM Probes

Syntax: LSM_PROBE(hook, typeof(arg1) arg1, typeof(arg2) arg2 ...)

This is a macro that instruments an LSM hook as a BPF program. It can be used to audit security events and implement MAC security policies in BPF. It is defined by specifying the hook name followed by its arguments.

Hook names can be found in include/linux/security.h by taking functions like security_hookname and taking just the hookname part. For example, security_bpf would simply become bpf.

Unlike other BPF program types, the return value specified in an LSM probe matters. A return value of 0 allows the hook to succeed, whereas any non-zero return value will cause the hook to fail and deny the security operation.

The following example instruments a hook that denies all future BPF operations:

LSM_PROBE(bpf, int cmd, union bpf_attr *attr, unsigned int size)
{
    return -EPERM;
}

This instruments the security_bpf hook and causes it to return -EPERM. Changing return -EPERM to return 0 would cause the BPF program to allow the operation instead.

LSM probes require at least a 5.7+ kernel with the following configuation options set:

CONFIG_BPF_LSM=y
CONFIG_LSM comma separated string must contain "bpf" (for example, CONFIG_LSM="lockdown,yama,bpf")

Examples in situ: search /tests

12. BPF ITERATORS

Syntax: BPF_ITER(target)

This is a macro to define a program signature for a bpf iterator program. The argument target specifies what to iterate for the program.

Currently, kernel does not have interface to discover what targets are supported. A good place to find what is supported is in tools/testing/selftests/bpf/prog_test/bpf_iter.c and some sample bpf iter programs are in tools/testing/selftests/bpf/progs with file name prefix bpf_iter.

The following example defines a program for target task, which traverses all tasks in the kernel.

BPF_ITER(task)
{
  struct seq_file *seq = ctx->meta->seq;
  struct task_struct *task = ctx->task;

  if (task == (void *)0)
    return 0;

  ... task->pid, task->tgid, task->comm, ...
  return 0;
}

BPF iterators are introduced in 5.8 kernel for task, task_file, bpf_map, netlink_sock and ipv6_route . In 5.9, support is added to tcp/udp sockets and bpf map element (hashmap, arraymap and sk_local_storage_map) traversal.

Data

1. bpf_probe_read_kernel()

Syntax: int bpf_probe_read_kernel(void *dst, int size, const void *src)

Return: 0 on success

This copies size bytes from kernel address space to the BPF stack, so that BPF can later operate on it. For safety, all kernel memory reads must pass through bpf_probe_read_kernel(). This happens automatically in some cases, such as dereferencing kernel variables, as bcc will rewrite the BPF program to include the necessary bpf_probe_read_kernel().

Examples in situ: search /examples, search /tools

2. bpf_probe_read_kernel_str()

Syntax: int bpf_probe_read_kernel_str(void *dst, int size, const void *src)

Return:

> 0 length of the string including the trailing NULL on success
< 0 error

This copies a NULL terminated string from kernel address space to the BPF stack, so that BPF can later operate on it. In case the string length is smaller than size, the target is not padded with further NULL bytes. In case the string length is larger than size, just size - 1 bytes are copied and the last byte is set to NULL.

Examples in situ: search /examples, search /tools

3. bpf_ktime_get_ns()

Syntax: u64 bpf_ktime_get_ns(void)

Return: u64 number of nanoseconds. Starts at system boot time but stops during suspend.

Examples in situ: search /examples, search /tools

4. bpf_get_current_pid_tgid()

Syntax: u64 bpf_get_current_pid_tgid(void)

Return: current->tgid << 32 | current->pid

Returns the process ID in the lower 32 bits (kernel's view of the PID, which in user space is usually presented as the thread ID), and the thread group ID in the upper 32 bits (what user space often thinks of as the PID). By directly setting this to a u32, we discard the upper 32 bits.

Examples in situ: search /examples, search /tools

5. bpf_get_current_uid_gid()

Syntax: u64 bpf_get_current_uid_gid(void)

Return: current_gid << 32 | current_uid

Returns the user ID and group IDs.

Examples in situ: search /examples, search /tools

6. bpf_get_current_comm()

Syntax: bpf_get_current_comm(char *buf, int size_of_buf)

Return: 0 on success

Populates the first argument address with the current process name. It should be a pointer to a char array of at least size TASK_COMM_LEN, which is defined in linux/sched.h. For example:

#include <linux/sched.h>

int do_trace(struct pt_regs *ctx) {
    char comm[TASK_COMM_LEN];
    bpf_get_current_comm(&comm, sizeof(comm));
[...]

Examples in situ: search /examples, search /tools

7. bpf_get_current_task()

Syntax: bpf_get_current_task()

Return: current task as a pointer to struct task_struct.

Returns a pointer to the current task's task_struct object. This helper can be used to compute the on-CPU time for a process, identify kernel threads, get the current CPU's run queue, or retrieve many other pieces of information.

With Linux 4.13, due to issues with field randomization, you may need two #define directives before the includes:

#define randomized_struct_fields_start  struct {
#define randomized_struct_fields_end    };
#include <linux/sched.h>

int do_trace(void *ctx) {
    struct task_struct *t = (struct task_struct *)bpf_get_current_task();
[...]

Examples in situ: search /examples, search /tools

8. bpf_log2l()

Syntax: unsigned int bpf_log2l(unsigned long v)

Returns the log-2 of the provided value. This is often used to create indexes for histograms, to construct power-of-2 histograms.

Examples in situ: search /examples, search /tools

9. bpf_get_prandom_u32()

Syntax: u32 bpf_get_prandom_u32()

Returns a pseudo-random u32.

Example in situ: search /examples, search /tools

10. bpf_probe_read_user()

Syntax: int bpf_probe_read_user(void *dst, int size, const void *src)

Return: 0 on success

This attempts to safely read size bytes from user address space to the BPF stack, so that BPF can later operate on it. For safety, all user address space memory reads must pass through bpf_probe_read_user().

Examples in situ: search /examples, search /tools

11. bpf_probe_read_user_str()

Syntax: int bpf_probe_read_user_str(void *dst, int size, const void *src)

Return:

> 0 length of the string including the trailing NULL on success
< 0 error

This copies a NULL terminated string from user address space to the BPF stack, so that BPF can later operate on it. In case the string length is smaller than size, the target is not padded with further NULL bytes. In case the string length is larger than size, just size - 1 bytes are copied and the last byte is set to NULL.

Examples in situ: search /examples, search /tools

12. bpf_get_ns_current_pid_tgid()

Syntax: u32 bpf_get_ns_current_pid_tgid(u64 dev, u64 ino, struct bpf_pidns_info* nsdata, u32 size)

Values for pid and tgid as seen from the current namespace will be returned in nsdata.

Return 0 on success, or one of the following in case of failure:

-EINVAL if dev and inum supplied don't match dev_t and inode number with nsfs of current task, or if dev conversion to dev_t lost high bits.
-ENOENT if pidns does not exists for the current task.

Examples in situ: search /examples, search /tools

Debugging

1. bpf_override_return()

Syntax: int bpf_override_return(struct pt_regs *, unsigned long rc)

Return: 0 on success

When used in a program attached to a function entry kprobe, causes the execution of the function to be skipped, immediately returning rc instead. This is used for targeted error injection.

bpf_override_return will only work when the kprobed function is whitelisted to allow error injections. Whitelisting entails tagging a function with ALLOW_ERROR_INJECTION() in the kernel source tree; see io_ctl_init for an example. If the kprobed function is not whitelisted, the bpf program will fail to attach with ioctl(PERF_EVENT_IOC_SET_BPF): Invalid argument

int kprobe__io_ctl_init(void *ctx) {
	bpf_override_return(ctx, -ENOMEM);
	return 0;
}

Output

1. bpf_trace_printk()

Syntax: int bpf_trace_printk(const char *fmt, ...)

Return: 0 on success

A simple kernel facility for printf() to the common trace_pipe (/sys/kernel/debug/tracing/trace_pipe). This is ok for some quick examples, but has limitations: 3 args max, 1 %s only, and trace_pipe is globally shared, so concurrent programs will have clashing output. A better interface is via BPF_PERF_OUTPUT(). Note that calling this helper is made simpler than the original kernel version, which has fmt_size as the second parameter.

Examples in situ: search /examples, search /tools

2. BPF_PERF_OUTPUT

Syntax: BPF_PERF_OUTPUT(name)

Creates a BPF table for pushing out custom event data to user space via a perf ring buffer. This is the preferred method for pushing per-event data to user space.

For example:

struct data_t {
    u32 pid;
    u64 ts;
    char comm[TASK_COMM_LEN];
};
BPF_PERF_OUTPUT(events);

int hello(struct pt_regs *ctx) {
    struct data_t data = {};

    data.pid = bpf_get_current_pid_tgid();
    data.ts = bpf_ktime_get_ns();
    bpf_get_current_comm(&data.comm, sizeof(data.comm));

    events.perf_submit(ctx, &data, sizeof(data));

    return 0;
}

The output table is named events, and data is pushed to it via events.perf_submit().

Examples in situ: search /examples, search /tools

3. perf_submit()

Syntax: int perf_submit((void *)ctx, (void *)data, u32 data_size)

Return: 0 on success

A method of a BPF_PERF_OUTPUT table, for submitting custom event data to user space. See the BPF_PERF_OUTPUT entry. (This ultimately calls bpf_perf_event_output().)

The ctx parameter is provided in kprobes or kretprobes. For SCHED_CLS or SOCKET_FILTER programs, the struct __sk_buff *skb must be used instead.

Examples in situ: search /examples, search /tools

4. BPF_RINGBUF_OUTPUT

Syntax: BPF_RINGBUF_OUTPUT(name, page_cnt)

Creates a BPF table for pushing out custom event data to user space via a ringbuf ring buffer. BPF_RINGBUF_OUTPUT has several advantages over BPF_PERF_OUTPUT, summarized as follows:

Buffer is shared across all CPUs, meaning no per-CPU allocation
Supports two APIs for BPF programs
- map.ringbuf_output() works like map.perf_submit() (covered in ringbuf_output)
- map.ringbuf_reserve()/map.ringbuf_submit()/map.ringbuf_discard() split the process of reserving buffer space and submitting events into two steps (covered in ringbuf_reserve, ringbuf_submit, ringbuf_discard)
BPF APIs do not require access to a CPU ctx argument
Superior performance and latency in userspace thanks to a shared ring buffer manager
Supports two ways of consuming data in userspace

Starting in Linux 5.8, this should be the preferred method for pushing per-event data to user space.

Example of both APIs:

struct data_t {
    u32 pid;
    u64 ts;
    char comm[TASK_COMM_LEN];
};

// Creates a ringbuf called events with 8 pages of space, shared across all CPUs
BPF_RINGBUF_OUTPUT(events, 8);

int first_api_example(struct pt_regs *ctx) {
    struct data_t data = {};

    data.pid = bpf_get_current_pid_tgid();
    data.ts = bpf_ktime_get_ns();
    bpf_get_current_comm(&data.comm, sizeof(data.comm));

    events.ringbuf_output(&data, sizeof(data), 0 /* flags */);

    return 0;
}

int second_api_example(struct pt_regs *ctx) {
    struct data_t *data = events.ringbuf_reserve(sizeof(struct data_t));
    if (!data) { // Failed to reserve space
        return 1;
    }

    data->pid = bpf_get_current_pid_tgid();
    data->ts = bpf_ktime_get_ns();
    bpf_get_current_comm(&data->comm, sizeof(data->comm));

    events.ringbuf_submit(data, 0 /* flags */);

    return 0;
}

The output table is named events. Data is allocated via events.ringbuf_reserve() and pushed to it via events.ringbuf_submit().

Examples in situ: search /examples,

5. ringbuf_output()

Syntax: int ringbuf_output((void *)data, u64 data_size, u64 flags)

Return: 0 on success

Flags:

BPF_RB_NO_WAKEUP: Do not sent notification of new data availability
BPF_RB_FORCE_WAKEUP: Send notification of new data availability unconditionally

A method of the BPF_RINGBUF_OUTPUT table, for submitting custom event data to user space. This method works like perf_submit(), although it does not require a ctx argument.

Examples in situ: search /examples,

6. ringbuf_reserve()

Syntax: void* ringbuf_reserve(u64 data_size)

Return: Pointer to data struct on success, NULL on failure

A method of the BPF_RINGBUF_OUTPUT table, for reserving space in the ring buffer and simultaenously allocating a data struct for output. Must be used with one of ringbuf_submit or ringbuf_discard.

Examples in situ: search /examples,

7. ringbuf_submit()

Syntax: void ringbuf_submit((void *)data, u64 flags)

Return: Nothing, always succeeds

Flags:

BPF_RB_NO_WAKEUP: Do not sent notification of new data availability
BPF_RB_FORCE_WAKEUP: Send notification of new data availability unconditionally

A method of the BPF_RINGBUF_OUTPUT table, for submitting custom event data to user space. Must be preceded by a call to ringbuf_reserve() to reserve space for the data.

Examples in situ: search /examples,

8. ringbuf_discard()

Syntax: void ringbuf_discard((void *)data, u64 flags)

Return: Nothing, always succeeds

Flags:

BPF_RB_NO_WAKEUP: Do not sent notification of new data availability
BPF_RB_FORCE_WAKEUP: Send notification of new data availability unconditionally

A method of the BPF_RINGBUF_OUTPUT table, for discarding custom event data; userspace ignores the data associated with the discarded event. Must be preceded by a call to ringbuf_reserve() to reserve space for the data.

Examples in situ: search /examples,

Maps

Maps are BPF data stores, and are the basis for higher level object types including tables, hashes, and histograms.

1. BPF_TABLE

Syntax: BPF_TABLE(_table_type, _key_type, _leaf_type, _name, _max_entries)

Creates a map named _name. Most of the time this will be used via higher-level macros, like BPF_HASH, BPF_HIST, etc.

Methods (covered later): map.lookup(), map.lookup_or_try_init(), map.delete(), map.update(), map.insert(), map.increment().

Examples in situ: search /examples, search /tools

Pinned Maps

Maps that were pinned to the BPF filesystem can be accessed through an extended syntax: BPF_TABLE_PINNED(_table_type, _key_type, _leaf_type, _name, _max_entries, "/sys/fs/bpf/xyz") The type information is not enforced and the actual map type depends on the map that got pinned to the location.

For example:

BPF_TABLE_PINNED("hash", u64, u64, ids, 1024, "/sys/fs/bpf/ids");

2. BPF_HASH

Syntax: BPF_HASH(name [, key_type [, leaf_type [, size]]])

Creates a hash map (associative array) named name, with optional parameters.

Defaults: BPF_HASH(name, key_type=u64, leaf_type=u64, size=10240)

For example:

BPF_HASH(start, struct request *);

This creates a hash named start where the key is a struct request *, and the value defaults to u64. This hash is used by the disksnoop.py example for saving timestamps for each I/O request, where the key is the pointer to struct request, and the value is the timestamp.

Methods (covered later): map.lookup(), map.lookup_or_try_init(), map.delete(), map.update(), map.insert(), map.increment().

Examples in situ: search /examples, search /tools

3. BPF_ARRAY

Syntax: BPF_ARRAY(name [, leaf_type [, size]])

Creates an int-indexed array which is optimized for fastest lookup and update, named name, with optional parameters.

Defaults: BPF_ARRAY(name, leaf_type=u64, size=10240)

For example:

BPF_ARRAY(counts, u64, 32);

This creates an array named counts where with 32 buckets and 64-bit integer values. This array is used by the funccount.py example for saving call count of each function.

Methods (covered later): map.lookup(), map.update(), map.increment(). Note that all array elements are pre-allocated with zero values and can not be deleted.

Examples in situ: search /examples, search /tools

4. BPF_HISTOGRAM

Syntax: BPF_HISTOGRAM(name [, key_type [, size ]])

Creates a histogram map named name, with optional parameters.

Defaults: BPF_HISTOGRAM(name, key_type=int, size=64)

For example:

BPF_HISTOGRAM(dist);

This creates a histogram named dist, which defaults to 64 buckets indexed by keys of type int.

Methods (covered later): map.increment().

Examples in situ: search /examples, search /tools

5. BPF_STACK_TRACE

Syntax: BPF_STACK_TRACE(name, max_entries)

Creates stack trace map named name, with a maximum entry count provided. These maps are used to store stack traces.

For example:

BPF_STACK_TRACE(stack_traces, 1024);

This creates stack trace map named stack_traces, with a maximum number of stack trace entries of 1024.

Methods (covered later): map.get_stackid().

Examples in situ: search /examples, search /tools

6. BPF_PERF_ARRAY

Syntax: BPF_PERF_ARRAY(name, max_entries)

Creates perf array named name, with a maximum entry count provided, which must be equal to the number of system cpus. These maps are used to fetch hardware performance counters.

For example:

text="""
BPF_PERF_ARRAY(cpu_cycles, NUM_CPUS);
"""
b = bcc.BPF(text=text, cflags=["-DNUM_CPUS=%d" % multiprocessing.cpu_count()])
b["cpu_cycles"].open_perf_event(b["cpu_cycles"].HW_CPU_CYCLES)

This creates a perf array named cpu_cycles, with number of entries equal to the number of cpus/cores. The array is configured so that later calling map.perf_read() will return a hardware-calculated counter of the number of cycles elapsed from some point in the past. Only one type of hardware counter may be configured per table at a time.

Methods (covered later): map.perf_read().

Examples in situ: search /tests

7. BPF_PERCPU_ARRAY

Syntax: BPF_PERCPU_ARRAY(name [, leaf_type [, size]])

Creates NUM_CPU int-indexed arrays which are optimized for fastest lookup and update, named name, with optional parameters. Each CPU will have a separate copy of this array. The copies are not kept synchronized in any way.

Note that due to limits defined in the kernel (in linux/mm/percpu.c), the leaf_type cannot have a size of more than 32KB. In other words, BPF_PERCPU_ARRAY elements cannot be larger than 32KB in size.

Defaults: BPF_PERCPU_ARRAY(name, leaf_type=u64, size=10240)

For example:

BPF_PERCPU_ARRAY(counts, u64, 32);

This creates NUM_CPU arrays named counts where with 32 buckets and 64-bit integer values.

Methods (covered later): map.lookup(), map.update(), map.increment(). Note that all array elements are pre-allocated with zero values and can not be deleted.

Examples in situ: search /examples, search /tools

8. BPF_LPM_TRIE

Syntax: BPF_LPM_TRIE(name [, key_type [, leaf_type [, size]]])

Creates a longest prefix match trie map named name, with optional parameters.

Defaults: BPF_LPM_TRIE(name, key_type=u64, leaf_type=u64, size=10240)

For example:

BPF_LPM_TRIE(trie, struct key_v6);

This creates an LPM trie map named trie where the key is a struct key_v6, and the value defaults to u64.

Methods (covered later): map.lookup(), map.lookup_or_try_init(), map.delete(), map.update(), map.insert(), map.increment().

Examples in situ: search /examples, search /tools

9. BPF_PROG_ARRAY

Syntax: BPF_PROG_ARRAY(name, size)

This creates a program array named name with size entries. Each entry of the array is either a file descriptor to a bpf program or NULL. The array acts as a jump table so that bpf programs can "tail-call" other bpf programs.

Methods (covered later): map.call().

Examples in situ: search /examples, search /tests, assign fd

10. BPF_DEVMAP

Syntax: BPF_DEVMAP(name, size)

This creates a device map named name with size entries. Each entry of the map is an ifindex to a network interface. This map is only used in XDP.

For example:

BPF_DEVMAP(devmap, 10);

Methods (covered later): map.redirect_map().

Examples in situ: search /examples,

11. BPF_CPUMAP

Syntax: BPF_CPUMAP(name, size)

This creates a cpu map named name with size entries. The index of the map represents the CPU id and each entry is the size of the ring buffer allocated for the CPU. This map is only used in XDP.

For example:

BPF_CPUMAP(cpumap, 16);

Methods (covered later): map.redirect_map().

Examples in situ: search /examples,

12. BPF_XSKMAP

Syntax: BPF_XSKMAP(name, size)

This creates a xsk map named name with size entries. Each entry represents one NIC's queue id. This map is only used in XDP to redirect packet to an AF_XDP socket. If the AF_XDP socket is binded to a queue which is different than the current packet's queue id, the packet will be dropped. For kernel v5.3 and latter, lookup method is available and can be used to check whether and AF_XDP socket is available for the current packet's queue id. More details at AF_XDP.

For example:

BPF_XSKMAP(xsks_map, 8);

Methods (covered later): map.redirect_map(). map.lookup()

Examples in situ: search /examples,

13. BPF_ARRAY_OF_MAPS

Syntax: BPF_ARRAY_OF_MAPS(name, inner_map_name, size)

This creates an array map with a map-in-map type (BPF_MAP_TYPE_HASH_OF_MAPS) map named name with size entries. The inner map meta data is provided by map inner_map_name and can be most of array or hash maps except BPF_MAP_TYPE_PROG_ARRAY, BPF_MAP_TYPE_CGROUP_STORAGE and BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE.

For example:

BPF_TABLE("hash", int, int, ex1, 1024);
BPF_TABLE("hash", int, int, ex2, 1024);
BPF_ARRAY_OF_MAPS(maps_array, "ex1", 10);

14. BPF_HASH_OF_MAPS

Syntax: BPF_HASH_OF_MAPS(name, inner_map_name, size)

This creates a hash map with a map-in-map type (BPF_MAP_TYPE_HASH_OF_MAPS) map named name with size entries. The inner map meta data is provided by map inner_map_name and can be most of array or hash maps except BPF_MAP_TYPE_PROG_ARRAY, BPF_MAP_TYPE_CGROUP_STORAGE and BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE.

For example:

BPF_ARRAY(ex1, int, 1024);
BPF_ARRAY(ex2, int, 1024);
BPF_HASH_OF_MAPS(maps_hash, "ex1", 10);

15. BPF_STACK

Syntax: BPF_STACK(name, leaf_type, max_entries[, flags])

Creates a stack named name with value type leaf_type and max entries max_entries. Stack and Queue maps are only available from Linux 4.20+.

For example:

BPF_STACK(stack, struct event, 10240);

This creates a stack named stack where the value type is struct event, that holds up to 10240 entries.

Methods (covered later): map.push(), map.pop(), map.peek().

Examples in situ: search /tests,

16. BPF_QUEUE

Syntax: BPF_QUEUE(name, leaf_type, max_entries[, flags])

Creates a queue named name with value type leaf_type and max entries max_entries. Stack and Queue maps are only available from Linux 4.20+.

For example:

BPF_QUEUE(queue, struct event, 10240);

This creates a queue named queue where the value type is struct event, that holds up to 10240 entries.

Methods (covered later): map.push(), map.pop(), map.peek().

Examples in situ: search /tests,

17. map.lookup()

Syntax: *val map.lookup(&key)

Lookup the key in the map, and return a pointer to its value if it exists, else NULL. We pass the key in as an address to a pointer.

Examples in situ: search /examples, search /tools

18. map.lookup_or_try_init()

Syntax: *val map.lookup_or_try_init(&key, &zero)

Lookup the key in the map, and return a pointer to its value if it exists, else initialize the key's value to the second argument. This is often used to initialize values to zero. If the key cannot be inserted (e.g. the map is full) then NULL is returned.

Examples in situ: search /examples, search /tools

Note: The old map.lookup_or_init() may cause return from the function, so lookup_or_try_init() is recommended as it does not have this side effect.

19. map.delete()

Syntax: map.delete(&key)

Delete the key from the hash.

Examples in situ: search /examples, search /tools

20. map.update()

Syntax: map.update(&key, &val)

Associate the value in the second argument to the key, overwriting any previous value.

Examples in situ: search /examples, search /tools

21. map.insert()

Syntax: map.insert(&key, &val)

Associate the value in the second argument to the key, only if there was no previous value.

Examples in situ: search /examples, search /tools

22. map.increment()

Syntax: map.increment(key[, increment_amount])

Increments the key's value by increment_amount, which defaults to 1. Used for histograms.

Examples in situ: search /examples, search /tools

23. map.get_stackid()

Syntax: int map.get_stackid(void *ctx, u64 flags)

This walks the stack found via the struct pt_regs in ctx, saves it in the stack trace map, and returns a unique ID for the stack trace.

Examples in situ: search /examples, search /tools

24. map.perf_read()

Syntax: u64 map.perf_read(u32 cpu)

This returns the hardware performance counter as configured in 5. BPF_PERF_ARRAY

Examples in situ: search /tests

25. map.call()

Syntax: void map.call(void *ctx, int index)

This invokes bpf_tail_call() to tail-call the bpf program which the index entry in 9. BPF_PROG_ARRAY points to. A tail-call is different from the normal call. It reuses the current stack frame after jumping to another bpf program and never goes back. If the index entry is empty, it won't jump anywhere and the program execution continues as normal.

For example:

BPF_PROG_ARRAY(prog_array, 10);

int tail_call(void *ctx) {
    bpf_trace_printk("Tail-call\n");
    return 0;
}

int do_tail_call(void *ctx) {
    bpf_trace_printk("Original program\n");
    prog_array.call(ctx, 2);
    return 0;
}

b = BPF(src_file="example.c")
tail_fn = b.load_func("tail_call", BPF.KPROBE)
prog_array = b.get_table("prog_array")
prog_array[c_int(2)] = c_int(tail_fn.fd)
b.attach_kprobe(event="some_kprobe_event", fn_name="do_tail_call")

This assigns tail_call() to prog_array[2]. In the end of do_tail_call(), prog_array.call(ctx, 2) tail-calls tail_call() and executes it.

NOTE: To prevent infinite loop, the maximum number of tail-calls is 32 (MAX_TAIL_CALL_CNT).

Examples in situ: search /examples, search /tests

26. map.redirect_map()

Syntax: int map.redirect_map(int index, int flags)

This redirects the incoming packets based on the index entry. If the map is 10. BPF_DEVMAP, the packet will be sent to the transmit queue of the network interface that the entry points to. If the map is 11. BPF_CPUMAP, the packet will be sent to the ring buffer of the index CPU and be processed by the CPU later. If the map is 12. BPF_XSKMAP, the packet will be sent to the AF_XDP socket attached to the queue.

If the packet is redirected successfully, the function will return XDP_REDIRECT. Otherwise, it will return XDP_ABORTED to discard the packet.

For example:

BPF_DEVMAP(devmap, 1);

int redirect_example(struct xdp_md *ctx) {
    return devmap.redirect_map(0, 0);
}
int xdp_dummy(struct xdp_md *ctx) {
    return XDP_PASS;
}

ip = pyroute2.IPRoute()
idx = ip.link_lookup(ifname="eth1")[0]

b = bcc.BPF(src_file="example.c")

devmap = b.get_table("devmap")
devmap[c_uint32(0)] = c_int(idx)

in_fn = b.load_func("redirect_example", BPF.XDP)
out_fn = b.load_func("xdp_dummy", BPF.XDP)
b.attach_xdp("eth0", in_fn, 0)
b.attach_xdp("eth1", out_fn, 0)

Examples in situ: search /examples,

27. map.push()

Syntax: int map.push(&val, int flags)

Push an element onto a Stack or Queue table. Passing BPF_EXIST as a flag causes the Queue or Stack to discard the oldest element if it is full. Returns 0 on success, negative error on failure.

Examples in situ: search /tests,

28. map.pop()

Syntax: int map.pop(&val)

Pop an element from a Stack or Queue table. *val is populated with the result. Unlike peeking, popping removes the element. Returns 0 on success, negative error on failure.

Examples in situ: search /tests,

29. map.peek()

Syntax: int map.peek(&val)

Peek an element at the head of a Stack or Queue table. *val is populated with the result. Unlike popping, peeking does not remove the element. Returns 0 on success, negative error on failure.

Examples in situ: search /tests,

Licensing

Depending on which BPF helpers are used, a GPL-compatible license is required.

The special BCC macro BPF_LICENSE specifies the license of the BPF program. You can set the license as a comment in your source code, but the kernel has a special interface to specify it programmatically. If you need to use GPL-only helpers, it is recommended to specify the macro in your C code so that the kernel can understand it:

// SPDX-License-Identifier: GPL-2.0+
#define BPF_LICENSE GPL

Otherwise, the kernel may reject loading your program (see the error description below). Note that it supports multiple words and quotes are not necessary:

// SPDX-License-Identifier: GPL-2.0+ OR BSD-2-Clause
#define BPF_LICENSE Dual BSD/GPL

Check the BPF helpers reference to see which helpers are GPL-only and what the kernel understands as GPL-compatible.

If the macro is not specified, BCC will automatically define the license of the program as GPL.

Rewriter

One of jobs for rewriter is to turn implicit memory accesses to explicit ones using kernel helpers. Recent kernel introduced a config option ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE which will be set for architectures who user address space and kernel address are disjoint. x86 and arm has this config option set while s390 does not. If ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE is not set, the bpf old helper bpf_probe_read() will not be available. Some existing users may have implicit memory accesses to access user memory, so using bpf_probe_read_kernel() will cause their application to fail. Therefore, for non-s390, the rewriter will use bpf_probe_read() for these implicit memory accesses. For s390, bpf_probe_read_kernel() is used as default and users should use bpf_probe_read_user() explicitly when accessing user memories.

bcc Python

Initialization

Constructors.

1. BPF

Syntax: BPF({text=BPF_program | src_file=filename} [, usdt_contexts=[USDT_object, ...]] [, cflags=[arg1, ...]] [, debug=int])

Creates a BPF object. This is the main object for defining a BPF program, and interacting with its output.

Exactly one of text or src_file must be supplied (not both).

The cflags specifies additional arguments to be passed to the compiler, for example -DMACRO_NAME=value or -I/include/path. The arguments are passed as an array, with each element being an additional argument. Note that strings are not split on whitespace, so each argument must be a different element of the array, e.g. ["-include", "header.h"].

The debug flags control debug output, and can be or'ed together:

DEBUG_LLVM_IR = 0x1 compiled LLVM IR
DEBUG_BPF = 0x2 loaded BPF bytecode and register state on branches
DEBUG_PREPROCESSOR = 0x4 pre-processor result
DEBUG_SOURCE = 0x8 ASM instructions embedded with source
DEBUG_BPF_REGISTER_STATE = 0x10 register state on all instructions in addition to DEBUG_BPF

Examples:

# define entire BPF program in one line:
BPF(text='int do_trace(void *ctx) { bpf_trace_printk("hit!\\n"); return 0; }');

# define program as a variable:
prog = """
int hello(void *ctx) {
    bpf_trace_printk("Hello, World!\\n");
    return 0;
}
"""
b = BPF(text=prog)

# source a file:
b = BPF(src_file = "vfsreadlat.c")

# include a USDT object:
u = USDT(pid=int(pid))
[...]
b = BPF(text=bpf_text, usdt_contexts=[u])

# add include paths:
u = BPF(text=prog, cflags=["-I/path/to/include"])

Examples in situ: search /examples, search /tools

2. USDT

Syntax: USDT({pid=pid | path=path})

Creates an object to instrument User Statically-Defined Tracing (USDT) probes. Its primary method is enable_probe().

Arguments:

pid: attach to this process ID.
path: instrument USDT probes from this binary path.

Examples:

# include a USDT object:
u = USDT(pid=int(pid))
[...]
b = BPF(text=bpf_text, usdt_contexts=[u])

Examples in situ: search /examples, search /tools

Events

1. attach_kprobe()

Syntax: BPF.attach_kprobe(event="event", fn_name="name")

Instruments the kernel function event() using kernel dynamic tracing of the function entry, and attaches our C defined function name() to be called when the kernel function is called.

For example:

b.attach_kprobe(event="sys_clone", fn_name="do_trace")

This will instrument the kernel sys_clone() function, which will then run our BPF defined do_trace() function each time it is called.

You can call attach_kprobe() more than once, and attach your BPF function to multiple kernel functions.

See the previous kprobes section for how to instrument arguments from BPF.

Examples in situ: search /examples, search /tools

2. attach_kretprobe()

Syntax: BPF.attach_kretprobe(event="event", fn_name="name" [, maxactive=int])

Instruments the return of the kernel function event() using kernel dynamic tracing of the function return, and attaches our C defined function name() to be called when the kernel function returns.

For example:

b.attach_kretprobe(event="vfs_read", fn_name="do_return")

This will instrument the kernel vfs_read() function, which will then run our BPF defined do_return() function each time it is called.

You can call attach_kretprobe() more than once, and attach your BPF function to multiple kernel function returns.

When a kretprobe is installed on a kernel function, there is a limit on how many parallel calls it can catch. You can change that limit with maxactive. See the kprobes documentation for its default value.

See the previous kretprobes section for how to instrument the return value from BPF.

Examples in situ: search /examples, search /tools

3. attach_tracepoint()

Syntax: BPF.attach_tracepoint(tp="tracepoint", fn_name="name")

Instruments the kernel tracepoint described by tracepoint, and when hit, runs the BPF function name().

This is an explicit way to instrument tracepoints. The TRACEPOINT_PROBE syntax, covered in the earlier tracepoints section, is an alternate method with the advantage of auto-declaring an args struct containing the tracepoint arguments. With attach_tracepoint(), the tracepoint arguments need to be declared in the BPF program.

For example:

# define BPF program
bpf_text = """
#include <uapi/linux/ptrace.h>

struct urandom_read_args {
    // from /sys/kernel/debug/tracing/events/random/urandom_read/format
    u64 __unused__;
    u32 got_bits;
    u32 pool_left;
    u32 input_left;
};

int printarg(struct urandom_read_args *args) {
    bpf_trace_printk("%d\\n", args->got_bits);
    return 0;
};
"""

# load BPF program
b = BPF(text=bpf_text)
b.attach_tracepoint("random:urandom_read", "printarg")

Notice how the first argument to printarg() is now our defined struct.

Examples in situ: code, search /examples, search /tools

4. attach_uprobe()

Syntax: BPF.attach_uprobe(name="location", sym="symbol", fn_name="name" [, sym_off=int]), BPF.attach_uprobe(name="location", sym_re="regex", fn_name="name"), BPF.attach_uprobe(name="location", addr=int, fn_name="name")

Instruments the user-level function symbol() from either the library or binary named by location using user-level dynamic tracing of the function entry, and attach our C defined function name() to be called whenever the user-level function is called. If sym_off is given, the function is attached to the offset within the symbol.

The real address addr may be supplied in place of sym, in which case sym must be set to its default value. If the file is a non-PIE executable, addr must be a virtual address, otherwise it must be an offset relative to the file load address.

Instead of a symbol name, a regular expression can be provided in sym_re. The uprobe will then attach to symbols that match the provided regular expression.

Libraries can be given in the name argument without the lib prefix, or with the full path (/usr/lib/...). Binaries can be given only with the full path (/bin/sh).

For example:

b.attach_uprobe(name="c", sym="strlen", fn_name="count")

This will instrument strlen() function from libc, and call our BPF function count() when it is called. Note how the "lib" in "libc" is not necessary to specify.

Other examples:

b.attach_uprobe(name="c", sym="getaddrinfo", fn_name="do_entry")
b.attach_uprobe(name="/usr/bin/python", sym="main", fn_name="do_main")

You can call attach_uprobe() more than once, and attach your BPF function to multiple user-level functions.

See the previous uprobes section for how to instrument arguments from BPF.

Examples in situ: search /examples, search /tools

5. attach_uretprobe()

Syntax: BPF.attach_uretprobe(name="location", sym="symbol", fn_name="name")

Instruments the return of the user-level function symbol() from either the library or binary named by location using user-level dynamic tracing of the function return, and attach our C defined function name() to be called whenever the user-level function returns.

For example:

b.attach_uretprobe(name="c", sym="strlen", fn_name="count")

This will instrument strlen() function from libc, and call our BPF function count() when it returns.

Other examples:

b.attach_uretprobe(name="c", sym="getaddrinfo", fn_name="do_return")
b.attach_uretprobe(name="/usr/bin/python", sym="main", fn_name="do_main")

You can call attach_uretprobe() more than once, and attach your BPF function to multiple user-level functions.

See the previous uretprobes section for how to instrument the return value from BPF.

Examples in situ: search /examples, search /tools

6. USDT.enable_probe()

Syntax: USDT.enable_probe(probe=probe, fn_name=name)

Attaches a BPF C function name to the USDT probe probe.

Example:

# enable USDT probe from given PID
u = USDT(pid=int(pid))
u.enable_probe(probe="http__server__request", fn_name="do_trace")

To check if your binary has USDT probes, and what they are, you can run readelf -n binary and check the stap debug section.

Examples in situ: search /examples, search /tools

7. attach_raw_tracepoint()

Syntax: BPF.attach_raw_tracepoint(tp="tracepoint", fn_name="name")

Instruments the kernel raw tracepoint described by tracepoint (event only, no category), and when hit, runs the BPF function name().

This is an explicit way to instrument tracepoints. The RAW_TRACEPOINT_PROBE syntax, covered in the earlier raw tracepoints section, is an alternate method.

For example:

b.attach_raw_tracepoint("sched_swtich", "do_trace")

Examples in situ: search /tools

Debug Output

1. trace_print()

Syntax: BPF.trace_print(fmt="fields")

This method continually reads the globally shared /sys/kernel/debug/tracing/trace_pipe file and prints its contents. This file can be written to via BPF and the bpf_trace_printk() function, however, that method has limitations, including a lack of concurrent tracing support. The BPF_PERF_OUTPUT mechanism, covered earlier, is preferred.

Arguments:

fmt: optional, and can contain a field formatting string. It defaults to None.

Examples:

# print trace_pipe output as-is:
b.trace_print()

# print PID and message:
b.trace_print(fmt="{1} {5}")

Examples in situ: search /examples, search /tools

2. trace_fields()

Syntax: BPF.trace_fields(nonblocking=False)

This method reads one line from the globally shared /sys/kernel/debug/tracing/trace_pipe file and returns it as fields. This file can be written to via BPF and the bpf_trace_printk() function, however, that method has limitations, including a lack of concurrent tracing support. The BPF_PERF_OUTPUT mechanism, covered earlier, is preferred.

Arguments:

nonblocking: optional, defaults to False. When set to True, the program will not block waiting for input.

Examples:

while 1:
    try:
        (task, pid, cpu, flags, ts, msg) = b.trace_fields()
    except ValueError:
        continue
    [...]

Examples in situ: search /examples, search /tools

Output

Normal output from a BPF program is either:

per-event: using PERF_EVENT_OUTPUT, open_perf_buffer(), and perf_buffer_poll().
map summary: using items(), or print_log2_hist(), covered in the Maps section.

1. perf_buffer_poll()

Syntax: BPF.perf_buffer_poll(timeout=T)

This polls from all open perf ring buffers, calling the callback function that was provided when calling open_perf_buffer for each entry.

The timeout parameter is optional and measured in milliseconds. In its absence, polling continues indefinitely.

Example:

# loop with callback to print_event
b["events"].open_perf_buffer(print_event)
while 1:
    try:
        b.perf_buffer_poll()
    except KeyboardInterrupt:
        exit();

Examples in situ: code, search /examples, search /tools

2. ring_buffer_poll()

Syntax: BPF.ring_buffer_poll(timeout=T)

This polls from all open ringbuf ring buffers, calling the callback function that was provided when calling open_ring_buffer for each entry.

The timeout parameter is optional and measured in milliseconds. In its absence, polling continues until there is no more data or the callback returns a negative value.

Example:

# loop with callback to print_event
b["events"].open_ring_buffer(print_event)
while 1:
    try:
        b.ring_buffer_poll(30)
    except KeyboardInterrupt:
        exit();

Examples in situ: search /examples,

3. ring_buffer_consume()

Syntax: BPF.ring_buffer_consume()

This consumes from all open ringbuf ring buffers, calling the callback function that was provided when calling open_ring_buffer for each entry.

Unlike ring_buffer_poll, this method does not poll for data before attempting to consume. This reduces latency at the expense of higher CPU consumption. If you are unsure which to use, use ring_buffer_poll.

Example:

# loop with callback to print_event
b["events"].open_ring_buffer(print_event)
while 1:
    try:
        b.ring_buffer_consume()
    except KeyboardInterrupt:
        exit();

Examples in situ: search /examples,

Maps

Maps are BPF data stores, and are used in bcc to implement a table, and then higher level objects on top of tables, including hashes and histograms.

1. get_table()

Syntax: BPF.get_table(name)

Returns a table object. This is no longer used, as tables can now be read as items from BPF. Eg: BPF[name].

Examples:

counts = b.get_table("counts")

counts = b["counts"]

These are equivalent.

2. open_perf_buffer()

Syntax: table.open_perf_buffers(callback, page_cnt=N, lost_cb=None)

This operates on a table as defined in BPF as BPF_PERF_OUTPUT(), and associates the callback Python function callback to be called when data is available in the perf ring buffer. This is part of the recommended mechanism for transferring per-event data from kernel to user space. The size of the perf ring buffer can be specified via the page_cnt parameter, which must be a power of two number of pages and defaults to 8. If the callback is not processing data fast enough, some submitted data may be lost. lost_cb will be called to log / monitor the lost count. If lost_cb is the default None value, it will just print a line of message to stderr.

Example:

# process event
def print_event(cpu, data, size):
    event = ct.cast(data, ct.POINTER(Data)).contents
    [...]

# loop with callback to print_event
b["events"].open_perf_buffer(print_event)
while 1:
    try:
        b.perf_buffer_poll()
    except KeyboardInterrupt:
        exit()

Note that the data structure transferred will need to be declared in C in the BPF program. For example:

// define output data structure in C
struct data_t {
    u32 pid;
    u64 ts;
    char comm[TASK_COMM_LEN];
};
BPF_PERF_OUTPUT(events);
[...]

In Python, you can either let bcc generate the data structure from C declaration automatically (recommended):

def print_event(cpu, data, size):
    event = b["events"].event(data)
[...]

or define it manually:

# define output data structure in Python
TASK_COMM_LEN = 16    # linux/sched.h
class Data(ct.Structure):
    _fields_ = [("pid", ct.c_ulonglong),
                ("ts", ct.c_ulonglong),
                ("comm", ct.c_char * TASK_COMM_LEN)]

def print_event(cpu, data, size):
    event = ct.cast(data, ct.POINTER(Data)).contents
[...]

Examples in situ: code, search /examples, search /tools

3. items()

Syntax: table.items()

Returns an array of the keys in a table. This can be used with BPF_HASH maps to fetch, and iterate, over the keys.

Example:

# print output
print("%10s %s" % ("COUNT", "STRING"))
counts = b.get_table("counts")
for k, v in sorted(counts.items(), key=lambda counts: counts[1].value):
    print("%10d \"%s\"" % (v.value, k.c.encode('string-escape')))

This example also uses the sorted() method to sort by value.

Examples in situ: search /examples, search /tools

4. values()

Syntax: table.values()

Returns an array of the values in a table.

5. clear()

Syntax: table.clear()

Clears the table: deletes all entries.

Example:

# print map summary every second:
while True:
    time.sleep(1)
    print("%-8s\n" % time.strftime("%H:%M:%S"), end="")
    dist.print_log2_hist(sym + " return:")
    dist.clear()

Examples in situ: search /examples, search /tools

6. print_log2_hist()

Syntax: table.print_log2_hist(val_type="value", section_header="Bucket ptr", section_print_fn=None)

Prints a table as a log2 histogram in ASCII. The table must be stored as log2, which can be done using the BPF function bpf_log2l().

Arguments:

val_type: optional, column header.
section_header: if the histogram has a secondary key, multiple tables will print and section_header can be used as a header description for each.
section_print_fn: if section_print_fn is not None, it will be passed the bucket value.

Example:

b = BPF(text="""
BPF_HISTOGRAM(dist);

int kprobe__blk_account_io_done(struct pt_regs *ctx, struct request *req)
{
	dist.increment(bpf_log2l(req->__data_len / 1024));
	return 0;
}
""")
[...]

b["dist"].print_log2_hist("kbytes")

Output:

     kbytes          : count     distribution
       0 -> 1        : 3        |                                      |
       2 -> 3        : 0        |                                      |
       4 -> 7        : 211      |**********                            |
       8 -> 15       : 0        |                                      |
      16 -> 31       : 0        |                                      |
      32 -> 63       : 0        |                                      |
      64 -> 127      : 1        |                                      |
     128 -> 255      : 800      |**************************************|

This output shows a multi-modal distribution, with the largest mode of 128->255 kbytes and a count of 800.

This is an efficient way to summarize data, as the summarization is performed in-kernel, and only the count column is passed to user space.

Examples in situ: search /examples, search /tools

6. print_linear_hist()

Syntax: table.print_linear_hist(val_type="value", section_header="Bucket ptr", section_print_fn=None)

Prints a table as a linear histogram in ASCII. This is intended to visualize small integer ranges, eg, 0 to 100.

Arguments:

val_type: optional, column header.
section_header: if the histogram has a secondary key, multiple tables will print and section_header can be used as a header description for each.
section_print_fn: if section_print_fn is not None, it will be passed the bucket value.

Example:

b = BPF(text="""
BPF_HISTOGRAM(dist);

int kprobe__blk_account_io_done(struct pt_regs *ctx, struct request *req)
{
	dist.increment(req->__data_len / 1024);
	return 0;
}
""")
[...]

b["dist"].print_linear_hist("kbytes")

Output:

     kbytes        : count     distribution
        0          : 3        |******                                  |
        1          : 0        |                                        |
        2          : 0        |                                        |
        3          : 0        |                                        |
        4          : 19       |****************************************|
        5          : 0        |                                        |
        6          : 0        |                                        |
        7          : 0        |                                        |
        8          : 4        |********                                |
        9          : 0        |                                        |
        10         : 0        |                                        |
        11         : 0        |                                        |
        12         : 0        |                                        |
        13         : 0        |                                        |
        14         : 0        |                                        |
        15         : 0        |                                        |
        16         : 2        |****                                    |
[...]

This is an efficient way to summarize data, as the summarization is performed in-kernel, and only the values in the count column are passed to user space.

Examples in situ: search /examples, search /tools

8. open_ring_buffer()

Syntax: table.open_ring_buffer(callback, ctx=None)

This operates on a table as defined in BPF as BPF_RINGBUF_OUTPUT(), and associates the callback Python function callback to be called when data is available in the ringbuf ring buffer. This is part of the new (Linux 5.8+) recommended mechanism for transferring per-event data from kernel to user space. Unlike perf buffers, ringbuf sizes are specified within the BPF program, as part of the BPF_RINGBUF_OUTPUT macro. If the callback is not processing data fast enough, some submitted data may be lost. In this case, the events should be polled more frequently and/or the size of the ring buffer should be increased.

Example:

# process event
def print_event(ctx, data, size):
    event = ct.cast(data, ct.POINTER(Data)).contents
    [...]

# loop with callback to print_event
b["events"].open_ring_buffer(print_event)
while 1:
    try:
        b.ring_buffer_poll()
    except KeyboardInterrupt:
        exit()

Note that the data structure transferred will need to be declared in C in the BPF program. For example:

// define output data structure in C
struct data_t {
    u32 pid;
    u64 ts;
    char comm[TASK_COMM_LEN];
};
BPF_RINGBUF_OUTPUT(events, 8);
[...]

In Python, you can either let bcc generate the data structure from C declaration automatically (recommended):

def print_event(ctx, data, size):
    event = b["events"].event(data)
[...]

or define it manually:

# define output data structure in Python
TASK_COMM_LEN = 16    # linux/sched.h
class Data(ct.Structure):
    _fields_ = [("pid", ct.c_ulonglong),
                ("ts", ct.c_ulonglong),
                ("comm", ct.c_char * TASK_COMM_LEN)]

def print_event(ctx, data, size):
    event = ct.cast(data, ct.POINTER(Data)).contents
[...]

Examples in situ: search /examples,

9. push()

Syntax: table.push(leaf, flags=0)

Push an element onto a Stack or Queue table. Raises an exception if the operation does not succeed. Passing QueueStack.BPF_EXIST as a flag causes the Queue or Stack to discard the oldest element if it is full.

Examples in situ: search /tests,

10. pop()

Syntax: leaf = table.pop()

Pop an element from a Stack or Queue table. Unlike peek(), pop() removes the element from the table before returning it. Raises a KeyError exception if the operation does not succeed.

Examples in situ: search /tests,

11. peek()

Syntax: leaf = table.peek()

Peek the element at the head of a Stack or Queue table. Unlike pop(), peek() does not remove the element from the table. Raises an exception if the operation does not succeed.

Examples in situ: search /tests,

Helpers

Some helper methods provided by bcc. Note that since we're in Python, we can import any Python library and their methods, including, for example, the libraries: argparse, collections, ctypes, datetime, re, socket, struct, subprocess, sys, and time.

1. ksym()

Syntax: BPF.ksym(addr)

Translate a kernel memory address into a kernel function name, which is returned.

Example:

print("kernel function: " + b.ksym(addr))

Examples in situ: search /examples, search /tools

2. ksymname()

Syntax: BPF.ksymname(name)

Translate a kernel name into an address. This is the reverse of ksym. Returns -1 when the function name is unknown.

Example:

print("kernel address: %x" % b.ksymname("vfs_read"))

Examples in situ: search /examples, search /tools

3. sym()

Syntax: BPF.sym(addr, pid, show_module=False, show_offset=False)

Translate a memory address into a function name for a pid, which is returned. A pid of less than zero will access the kernel symbol cache. The show_module and show_offset parameters control whether the module in which the symbol lies should be displayed, and whether the instruction offset from the beginning of the symbol should be displayed. These extra parameters default to False.

Example:

print("function: " + b.sym(addr, pid))

Examples in situ: search /examples, search /tools

4. num_open_kprobes()

Syntax: BPF.num_open_kprobes()

Returns the number of open k[ret]probes. Can be useful for scenarios where event_re is used while attaching and detaching probes. Excludes perf_events readers.

Example:

b.attach_kprobe(event_re=pattern, fn_name="trace_count")
matched = b.num_open_kprobes()
if matched == 0:
    print("0 functions matched by \"%s\". Exiting." % args.pattern)
    exit()

Examples in situ: search /examples, search /tools

5. get_syscall_fnname()

Syntax: BPF.get_syscall_fnname(name : str)

Return the corresponding kernel function name of the syscall. This helper function will try different prefixes and use the right one to concatenate with the syscall name. Note that the return value may vary in different versions of linux kernel and sometimes it will causing trouble. (see #2590)

Example:

print("The function name of %s in kernel is %s" % ("clone", b.get_syscall_fnname("clone")))
# sys_clone or __x64_sys_clone or ...

Examples in situ: search /examples, search /tools

BPF Errors

See the "Understanding eBPF verifier messages" section in the kernel source under Documentation/networking/filter.txt.

1. Invalid mem access

This can be due to trying to read memory directly, instead of operating on memory on the BPF stack. All kernel memory reads must be passed via bpf_probe_read_kernel() to copy kernel memory into the BPF stack, which can be automatic by the bcc rewriter in some cases of simple dereferencing. bpf_probe_read_kernel() does all the required checks.

Example:

bpf: Permission denied
0: (bf) r6 = r1
1: (79) r7 = *(u64 *)(r6 +80)
2: (85) call 14
3: (bf) r8 = r0
[...]
23: (69) r1 = *(u16 *)(r7 +16)
R7 invalid mem access 'inv'

Traceback (most recent call last):
  File "./tcpaccept", line 179, in <module>
    b = BPF(text=bpf_text)
  File "/usr/lib/python2.7/dist-packages/bcc/__init__.py", line 172, in __init__
    self._trace_autoload()
  File "/usr/lib/python2.7/dist-packages/bcc/__init__.py", line 612, in _trace_autoload
    fn = self.load_func(func_name, BPF.KPROBE)
  File "/usr/lib/python2.7/dist-packages/bcc/__init__.py", line 212, in load_func
    raise Exception("Failed to load BPF program %s" % func_name)
Exception: Failed to load BPF program kretprobe__inet_csk_accept

2. Cannot call GPL only function from proprietary program

This error happens when a GPL-only helper is called from a non-GPL BPF program. To fix this error, do not use GPL-only helpers from a proprietary BPF program, or relicense the BPF program under a GPL-compatible license. Check which BPF helpers are GPL-only, and what licenses are considered GPL-compatible.

Example calling bpf_get_stackid(), a GPL-only BPF helper, from a proprietary program (#define BPF_LICENSE Proprietary):

bpf: Failed to load program: Invalid argument
[...]
8: (85) call bpf_get_stackid#27
cannot call GPL only function from proprietary program

Environment Variables

1. Kernel source directory

eBPF program compilation needs kernel sources or kernel headers with headers compiled. In case your kernel sources are at a non-standard location where BCC cannot find then, its possible to provide BCC the absolute path of the location by setting BCC_KERNEL_SOURCE to it.

2. Kernel version overriding

By default, BCC stores the LINUX_VERSION_CODE in the generated eBPF object which is then passed along to the kernel when the eBPF program is loaded. Sometimes this is quite inconvenient especially when the kernel is slightly updated such as an LTS kernel release. Its extremely unlikely the slight mismatch would cause any issues with the loaded eBPF program. By setting BCC_LINUX_VERSION_CODE to the version of the kernel that's running, the check for verifying the kernel version can be bypassed. This is needed for programs that use kprobes. This needs to be encoded in the format: (VERSION * 65536) + (PATCHLEVEL * 256) + SUBLEVEL. For example, if the running kernel is 4.9.10, then can set export BCC_LINUX_VERSION_CODE=264458 to override the kernel version check successfully.

Files

reference_guide.md

Latest commit

History

reference_guide.md

File metadata and controls

bcc Reference Guide

Contents

BPF C

Events & Arguments

1. kprobes

2. kretprobes

3. Tracepoints

4. uprobes

5. uretprobes

6. USDT probes

7. Raw Tracepoints

8. system call tracepoints

9. kfuncs

10. kretfuncs

11. LSM Probes

12. BPF ITERATORS

Data

1. bpf_probe_read_kernel()

2. bpf_probe_read_kernel_str()

3. bpf_ktime_get_ns()

4. bpf_get_current_pid_tgid()

5. bpf_get_current_uid_gid()

6. bpf_get_current_comm()

7. bpf_get_current_task()

8. bpf_log2l()

9. bpf_get_prandom_u32()

10. bpf_probe_read_user()

11. bpf_probe_read_user_str()

12. bpf_get_ns_current_pid_tgid()

Debugging

1. bpf_override_return()

Output

1. bpf_trace_printk()

2. BPF_PERF_OUTPUT

3. perf_submit()

4. BPF_RINGBUF_OUTPUT

5. ringbuf_output()

6. ringbuf_reserve()

7. ringbuf_submit()

8. ringbuf_discard()

Maps

1. BPF_TABLE

Pinned Maps

2. BPF_HASH

3. BPF_ARRAY

4. BPF_HISTOGRAM

5. BPF_STACK_TRACE

6. BPF_PERF_ARRAY

7. BPF_PERCPU_ARRAY

8. BPF_LPM_TRIE

9. BPF_PROG_ARRAY

10. BPF_DEVMAP

11. BPF_CPUMAP

12. BPF_XSKMAP

13. BPF_ARRAY_OF_MAPS

14. BPF_HASH_OF_MAPS

15. BPF_STACK

16. BPF_QUEUE

17. map.lookup()

18. map.lookup_or_try_init()

19. map.delete()

20. map.update()

21. map.insert()

22. map.increment()

23. map.get_stackid()

24. map.perf_read()

25. map.call()

26. map.redirect_map()

27. map.push()

28. map.pop()

29. map.peek()

Licensing

Rewriter

bcc Python