| ============= |
| BPF Iterators |
| ============= |
| |
| |
| ---------- |
| Motivation |
| ---------- |
| |
| There are a few existing ways to dump kernel data into user space. The most |
| popular one is the ``/proc`` system. For example, ``cat /proc/net/tcp6`` dumps |
| all tcp6 sockets in the system, and ``cat /proc/net/netlink`` dumps all netlink |
| sockets in the system. However, their output format tends to be fixed, and if |
| users want more information about these sockets, they have to patch the kernel, |
| which often takes time to publish upstream and release. The same is true for popular |
| tools like `ss <https://man7.org/linux/man-pages/man8/ss.8.html>`_ where any |
| additional information needs a kernel patch. |
| |
| To solve this problem, the `drgn |
| <https://www.kernel.org/doc/html/latest/bpf/drgn.html>`_ tool is often used to |
| dig out the kernel data with no kernel change. However, the main drawback for |
| drgn is performance, as it cannot do pointer tracing inside the kernel. In |
| addition, drgn cannot validate a pointer value and may read invalid data if the |
| pointer becomes invalid inside the kernel. |
| |
| The BPF iterator solves the above problem by providing flexibility on what data |
| (e.g., tasks, bpf_maps, etc.) to collect by calling BPF programs for each kernel |
| data object. |
| |
| ---------------------- |
| How BPF Iterators Work |
| ---------------------- |
| |
| A BPF iterator is a type of BPF program that allows users to iterate over |
| specific types of kernel objects. Unlike traditional BPF tracing programs that |
| allow users to define callbacks that are invoked at particular points of |
| execution in the kernel, BPF iterators allow users to define callbacks that |
| should be executed for every entry in a variety of kernel data structures. |
| |
| For example, users can define a BPF iterator that iterates over every task on |
| the system and dumps the total amount of CPU runtime currently used by each of |
| them. Another BPF task iterator may instead dump the cgroup information for each |
| task. Such flexibility is the core value of BPF iterators. |
| |
| A BPF program is always loaded into the kernel at the behest of a user space |
| process. A user space process loads a BPF program by opening and initializing |
| the program skeleton as required and then invoking a syscall to have the BPF |
| program verified and loaded by the kernel. |
| |
| In traditional tracing programs, a program is activated by having user space |
| obtain a ``bpf_link`` to the program with ``bpf_program__attach()``. Once |
| activated, the program callback will be invoked whenever the tracepoint is |
| triggered in the main kernel. For BPF iterator programs, a ``bpf_link`` to the |
| program is obtained using ``bpf_link_create()``, and the program callback is |
| invoked by issuing system calls from user space. |
| |
| Next, let us see how you can use the iterators to iterate on kernel objects and |
| read data. |
| |
| ------------------------ |
| How to Use BPF iterators |
| ------------------------ |
| |
| BPF selftests are a great resource to illustrate how to use the iterators. In |
| this section, we’ll walk through a BPF selftest which shows how to load and use |
| a BPF iterator program. To begin, we’ll look at `bpf_iter.c |
| <https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/tools/testing/selftests/bpf/prog_tests/bpf_iter.c>`_, |
| which illustrates how to load and trigger BPF iterators on the user space side. |
| Later, we’ll look at a BPF program that runs in kernel space. |
| |
| Loading a BPF iterator in the kernel from user space typically involves the |
| following steps: |
| |
| * The BPF program is loaded into the kernel through ``libbpf``. Once the kernel |
| has verified and loaded the program, it returns a file descriptor (fd) to user |
| space. |
| * Obtain a ``link_fd`` to the BPF program by calling the ``bpf_link_create()`` |
| specified with the BPF program file descriptor received from the kernel. |
| * Next, obtain a BPF iterator file descriptor (``bpf_iter_fd``) by calling the |
| ``bpf_iter_create()`` specified with the ``bpf_link`` received from Step 2. |
| * Trigger the iteration by calling ``read(bpf_iter_fd)`` until no data is |
| available. |
| * Close the iterator fd using ``close(bpf_iter_fd)``. |
| * If needed to reread the data, get a new ``bpf_iter_fd`` and do the read again. |
| |
| The following are a few examples of selftest BPF iterator programs: |
| |
| * `bpf_iter_tcp4.c <https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/tools/testing/selftests/bpf/progs/bpf_iter_tcp4.c>`_ |
| * `bpf_iter_task_vma.c <https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/tools/testing/selftests/bpf/progs/bpf_iter_task_vma.c>`_ |
| * `bpf_iter_task_file.c <https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/tools/testing/selftests/bpf/progs/bpf_iter_task_file.c>`_ |
| |
| Let us look at ``bpf_iter_task_file.c``, which runs in kernel space: |
| |
| Here is the definition of ``bpf_iter__task_file`` in `vmlinux.h |
| <https://facebookmicrosites.github.io/bpf/blog/2020/02/19/bpf-portability-and-co-re.html#btf>`_. |
| Any struct name in ``vmlinux.h`` in the format ``bpf_iter__<iter_name>`` |
| represents a BPF iterator. The suffix ``<iter_name>`` represents the type of |
| iterator. |
| |
| :: |
| |
| struct bpf_iter__task_file { |
| union { |
| struct bpf_iter_meta *meta; |
| }; |
| union { |
| struct task_struct *task; |
| }; |
| u32 fd; |
| union { |
| struct file *file; |
| }; |
| }; |
| |
| In the above code, the field 'meta' contains the metadata, which is the same for |
| all BPF iterator programs. The rest of the fields are specific to different |
| iterators. For example, for task_file iterators, the kernel layer provides the |
| 'task', 'fd' and 'file' field values. The 'task' and 'file' are `reference |
| counted |
| <https://facebookmicrosites.github.io/bpf/blog/2018/08/31/object-lifetime.html#file-descriptors-and-reference-counters>`_, |
| so they won't go away when the BPF program runs. |
| |
| Here is a snippet from the ``bpf_iter_task_file.c`` file: |
| |
| :: |
| |
| SEC("iter/task_file") |
| int dump_task_file(struct bpf_iter__task_file *ctx) |
| { |
| struct seq_file *seq = ctx->meta->seq; |
| struct task_struct *task = ctx->task; |
| struct file *file = ctx->file; |
| __u32 fd = ctx->fd; |
| |
| if (task == NULL || file == NULL) |
| return 0; |
| |
| if (ctx->meta->seq_num == 0) { |
| count = 0; |
| BPF_SEQ_PRINTF(seq, " tgid gid fd file\n"); |
| } |
| |
| if (tgid == task->tgid && task->tgid != task->pid) |
| count++; |
| |
| if (last_tgid != task->tgid) { |
| last_tgid = task->tgid; |
| unique_tgid_count++; |
| } |
| |
| BPF_SEQ_PRINTF(seq, "%8d %8d %8d %lx\n", task->tgid, task->pid, fd, |
| (long)file->f_op); |
| return 0; |
| } |
| |
| In the above example, the section name ``SEC(iter/task_file)``, indicates that |
| the program is a BPF iterator program to iterate all files from all tasks. The |
| context of the program is ``bpf_iter__task_file`` struct. |
| |
| The user space program invokes the BPF iterator program running in the kernel |
| by issuing a ``read()`` syscall. Once invoked, the BPF |
| program can export data to user space using a variety of BPF helper functions. |
| You can use either ``bpf_seq_printf()`` (and BPF_SEQ_PRINTF helper macro) or |
| ``bpf_seq_write()`` function based on whether you need formatted output or just |
| binary data, respectively. For binary-encoded data, the user space applications |
| can process the data from ``bpf_seq_write()`` as needed. For the formatted data, |
| you can use ``cat <path>`` to print the results similar to ``cat |
| /proc/net/netlink`` after pinning the BPF iterator to the bpffs mount. Later, |
| use ``rm -f <path>`` to remove the pinned iterator. |
| |
| For example, you can use the following command to create a BPF iterator from the |
| ``bpf_iter_ipv6_route.o`` object file and pin it to the ``/sys/fs/bpf/my_route`` |
| path: |
| |
| :: |
| |
| $ bpftool iter pin ./bpf_iter_ipv6_route.o /sys/fs/bpf/my_route |
| |
| And then print out the results using the following command: |
| |
| :: |
| |
| $ cat /sys/fs/bpf/my_route |
| |
| |
| ------------------------------------------------------- |
| Implement Kernel Support for BPF Iterator Program Types |
| ------------------------------------------------------- |
| |
| To implement a BPF iterator in the kernel, the developer must make a one-time |
| change to the following key data structure defined in the `bpf.h |
| <https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpf.h>`_ |
| file. |
| |
| :: |
| |
| struct bpf_iter_reg { |
| const char *target; |
| bpf_iter_attach_target_t attach_target; |
| bpf_iter_detach_target_t detach_target; |
| bpf_iter_show_fdinfo_t show_fdinfo; |
| bpf_iter_fill_link_info_t fill_link_info; |
| bpf_iter_get_func_proto_t get_func_proto; |
| u32 ctx_arg_info_size; |
| u32 feature; |
| struct bpf_ctx_arg_aux ctx_arg_info[BPF_ITER_CTX_ARG_MAX]; |
| const struct bpf_iter_seq_info *seq_info; |
| }; |
| |
| After filling the data structure fields, call ``bpf_iter_reg_target()`` to |
| register the iterator to the main BPF iterator subsystem. |
| |
| The following is the breakdown for each field in struct ``bpf_iter_reg``. |
| |
| .. list-table:: |
| :widths: 25 50 |
| :header-rows: 1 |
| |
| * - Fields |
| - Description |
| * - target |
| - Specifies the name of the BPF iterator. For example: ``bpf_map``, |
| ``bpf_map_elem``. The name should be different from other ``bpf_iter`` target names in the kernel. |
| * - attach_target and detach_target |
| - Allows for target specific ``link_create`` action since some targets |
| may need special processing. Called during the user space link_create stage. |
| * - show_fdinfo and fill_link_info |
| - Called to fill target specific information when user tries to get link |
| info associated with the iterator. |
| * - get_func_proto |
| - Permits a BPF iterator to access BPF helpers specific to the iterator. |
| * - ctx_arg_info_size and ctx_arg_info |
| - Specifies the verifier states for BPF program arguments associated with |
| the bpf iterator. |
| * - feature |
| - Specifies certain action requests in the kernel BPF iterator |
| infrastructure. Currently, only BPF_ITER_RESCHED is supported. This means |
| that the kernel function cond_resched() is called to avoid other kernel |
| subsystem (e.g., rcu) misbehaving. |
| * - seq_info |
| - Specifies the set of seq operations for the BPF iterator and helpers to |
| initialize/free the private data for the corresponding ``seq_file``. |
| |
| `Click here |
| <https://lore.kernel.org/bpf/20210212183107.50963-2-songliubraving@fb.com/>`_ |
| to see an implementation of the ``task_vma`` BPF iterator in the kernel. |
| |
| --------------------------------- |
| Parameterizing BPF Task Iterators |
| --------------------------------- |
| |
| By default, BPF iterators walk through all the objects of the specified types |
| (processes, cgroups, maps, etc.) across the entire system to read relevant |
| kernel data. But often, there are cases where we only care about a much smaller |
| subset of iterable kernel objects, such as only iterating tasks within a |
| specific process. Therefore, BPF iterator programs support filtering out objects |
| from iteration by allowing user space to configure the iterator program when it |
| is attached. |
| |
| -------------------------- |
| BPF Task Iterator Program |
| -------------------------- |
| |
| The following code is a BPF iterator program to print files and task information |
| through the ``seq_file`` of the iterator. It is a standard BPF iterator program |
| that visits every file of an iterator. We will use this BPF program in our |
| example later. |
| |
| :: |
| |
| #include <vmlinux.h> |
| #include <bpf/bpf_helpers.h> |
| |
| char _license[] SEC("license") = "GPL"; |
| |
| SEC("iter/task_file") |
| int dump_task_file(struct bpf_iter__task_file *ctx) |
| { |
| struct seq_file *seq = ctx->meta->seq; |
| struct task_struct *task = ctx->task; |
| struct file *file = ctx->file; |
| __u32 fd = ctx->fd; |
| if (task == NULL || file == NULL) |
| return 0; |
| if (ctx->meta->seq_num == 0) { |
| BPF_SEQ_PRINTF(seq, " tgid pid fd file\n"); |
| } |
| BPF_SEQ_PRINTF(seq, "%8d %8d %8d %lx\n", task->tgid, task->pid, fd, |
| (long)file->f_op); |
| return 0; |
| } |
| |
| ---------------------------------------- |
| Creating a File Iterator with Parameters |
| ---------------------------------------- |
| |
| Now, let us look at how to create an iterator that includes only files of a |
| process. |
| |
| First, fill the ``bpf_iter_attach_opts`` struct as shown below: |
| |
| :: |
| |
| LIBBPF_OPTS(bpf_iter_attach_opts, opts); |
| union bpf_iter_link_info linfo; |
| memset(&linfo, 0, sizeof(linfo)); |
| linfo.task.pid = getpid(); |
| opts.link_info = &linfo; |
| opts.link_info_len = sizeof(linfo); |
| |
| ``linfo.task.pid``, if it is non-zero, directs the kernel to create an iterator |
| that only includes opened files for the process with the specified ``pid``. In |
| this example, we will only be iterating files for our process. If |
| ``linfo.task.pid`` is zero, the iterator will visit every opened file of every |
| process. Similarly, ``linfo.task.tid`` directs the kernel to create an iterator |
| that visits opened files of a specific thread, not a process. In this example, |
| ``linfo.task.tid`` is different from ``linfo.task.pid`` only if the thread has a |
| separate file descriptor table. In most circumstances, all process threads share |
| a single file descriptor table. |
| |
| Now, in the userspace program, pass the pointer of struct to the |
| ``bpf_program__attach_iter()``. |
| |
| :: |
| |
| link = bpf_program__attach_iter(prog, &opts); iter_fd = |
| bpf_iter_create(bpf_link__fd(link)); |
| |
| If both *tid* and *pid* are zero, an iterator created from this struct |
| ``bpf_iter_attach_opts`` will include every opened file of every task in the |
| system (in the namespace, actually.) It is the same as passing a NULL as the |
| second argument to ``bpf_program__attach_iter()``. |
| |
| The whole program looks like the following code: |
| |
| :: |
| |
| #include <stdio.h> |
| #include <unistd.h> |
| #include <bpf/bpf.h> |
| #include <bpf/libbpf.h> |
| #include "bpf_iter_task_ex.skel.h" |
| |
| static int do_read_opts(struct bpf_program *prog, struct bpf_iter_attach_opts *opts) |
| { |
| struct bpf_link *link; |
| char buf[16] = {}; |
| int iter_fd = -1, len; |
| int ret = 0; |
| |
| link = bpf_program__attach_iter(prog, opts); |
| if (!link) { |
| fprintf(stderr, "bpf_program__attach_iter() fails\n"); |
| return -1; |
| } |
| iter_fd = bpf_iter_create(bpf_link__fd(link)); |
| if (iter_fd < 0) { |
| fprintf(stderr, "bpf_iter_create() fails\n"); |
| ret = -1; |
| goto free_link; |
| } |
| /* not check contents, but ensure read() ends without error */ |
| while ((len = read(iter_fd, buf, sizeof(buf) - 1)) > 0) { |
| buf[len] = 0; |
| printf("%s", buf); |
| } |
| printf("\n"); |
| free_link: |
| if (iter_fd >= 0) |
| close(iter_fd); |
| bpf_link__destroy(link); |
| return 0; |
| } |
| |
| static void test_task_file(void) |
| { |
| LIBBPF_OPTS(bpf_iter_attach_opts, opts); |
| struct bpf_iter_task_ex *skel; |
| union bpf_iter_link_info linfo; |
| skel = bpf_iter_task_ex__open_and_load(); |
| if (skel == NULL) |
| return; |
| memset(&linfo, 0, sizeof(linfo)); |
| linfo.task.pid = getpid(); |
| opts.link_info = &linfo; |
| opts.link_info_len = sizeof(linfo); |
| printf("PID %d\n", getpid()); |
| do_read_opts(skel->progs.dump_task_file, &opts); |
| bpf_iter_task_ex__destroy(skel); |
| } |
| |
| int main(int argc, const char * const * argv) |
| { |
| test_task_file(); |
| return 0; |
| } |
| |
| The following lines are the output of the program. |
| :: |
| |
| PID 1859 |
| |
| tgid pid fd file |
| 1859 1859 0 ffffffff82270aa0 |
| 1859 1859 1 ffffffff82270aa0 |
| 1859 1859 2 ffffffff82270aa0 |
| 1859 1859 3 ffffffff82272980 |
| 1859 1859 4 ffffffff8225e120 |
| 1859 1859 5 ffffffff82255120 |
| 1859 1859 6 ffffffff82254f00 |
| 1859 1859 7 ffffffff82254d80 |
| 1859 1859 8 ffffffff8225abe0 |
| |
| ------------------ |
| Without Parameters |
| ------------------ |
| |
| Let us look at how a BPF iterator without parameters skips files of other |
| processes in the system. In this case, the BPF program has to check the pid or |
| the tid of tasks, or it will receive every opened file in the system (in the |
| current *pid* namespace, actually). So, we usually add a global variable in the |
| BPF program to pass a *pid* to the BPF program. |
| |
| The BPF program would look like the following block. |
| |
| :: |
| |
| ...... |
| int target_pid = 0; |
| |
| SEC("iter/task_file") |
| int dump_task_file(struct bpf_iter__task_file *ctx) |
| { |
| ...... |
| if (task->tgid != target_pid) /* Check task->pid instead to check thread IDs */ |
| return 0; |
| BPF_SEQ_PRINTF(seq, "%8d %8d %8d %lx\n", task->tgid, task->pid, fd, |
| (long)file->f_op); |
| return 0; |
| } |
| |
| The user space program would look like the following block: |
| |
| :: |
| |
| ...... |
| static void test_task_file(void) |
| { |
| ...... |
| skel = bpf_iter_task_ex__open_and_load(); |
| if (skel == NULL) |
| return; |
| skel->bss->target_pid = getpid(); /* process ID. For thread id, use gettid() */ |
| memset(&linfo, 0, sizeof(linfo)); |
| linfo.task.pid = getpid(); |
| opts.link_info = &linfo; |
| opts.link_info_len = sizeof(linfo); |
| ...... |
| } |
| |
| ``target_pid`` is a global variable in the BPF program. The user space program |
| should initialize the variable with a process ID to skip opened files of other |
| processes in the BPF program. When you parametrize a BPF iterator, the iterator |
| calls the BPF program fewer times which can save significant resources. |
| |
| --------------------------- |
| Parametrizing VMA Iterators |
| --------------------------- |
| |
| By default, a BPF VMA iterator includes every VMA in every process. However, |
| you can still specify a process or a thread to include only its VMAs. Unlike |
| files, a thread can not have a separate address space (since Linux 2.6.0-test6). |
| Here, using *tid* makes no difference from using *pid*. |
| |
| ---------------------------- |
| Parametrizing Task Iterators |
| ---------------------------- |
| |
| A BPF task iterator with *pid* includes all tasks (threads) of a process. The |
| BPF program receives these tasks one after another. You can specify a BPF task |
| iterator with *tid* parameter to include only the tasks that match the given |
| *tid*. |