bpf: implement bpf_send_signal() helper
authorYonghong Song <yhs@fb.com>
Thu, 23 May 2019 21:47:45 +0000 (14:47 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Fri, 24 May 2019 21:26:47 +0000 (23:26 +0200)
This patch tries to solve the following specific use case.

Currently, bpf program can already collect stack traces
through kernel function get_perf_callchain()
when certain events happens (e.g., cache miss counter or
cpu clock counter overflows). But such stack traces are
not enough for jitted programs, e.g., hhvm (jited php).
To get real stack trace, jit engine internal data structures
need to be traversed in order to get the real user functions.

bpf program itself may not be the best place to traverse
the jit engine as the traversing logic could be complex and
it is not a stable interface either.

Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which
it can dump stack traces. When it receives a signal, it will
dump the stack in next such program location.

Such a mechanism can be implemented in the following way:
  . a perf ring buffer is created between bpf program
    and tracing app.
  . once a particular event happens, bpf program writes
    to the ring buffer and the tracing app gets notified.
  . the tracing app sends a signal SIGALARM to the hhvm.

But this method could have large delays and causing profiling
results skewed.

This patch implements bpf_send_signal() helper to send
a signal to hhvm in real time, resulting in intended stack traces.

Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
include/uapi/linux/bpf.h
kernel/trace/bpf_trace.c

index 63e0cf66f01a9698ff6bc41ac5492809bb83d834..68d4470523a0393158c593c76cf38795d5acc4fe 100644 (file)
@@ -2672,6 +2672,20 @@ union bpf_attr {
  *             0 on success.
  *
  *             **-ENOENT** if the bpf-local-storage cannot be found.
+ *
+ * int bpf_send_signal(u32 sig)
+ *     Description
+ *             Send signal *sig* to the current task.
+ *     Return
+ *             0 on success or successfully queued.
+ *
+ *             **-EBUSY** if work queue under nmi is full.
+ *
+ *             **-EINVAL** if *sig* is invalid.
+ *
+ *             **-EPERM** if no permission to send the *sig*.
+ *
+ *             **-EAGAIN** if bpf program can try again.
  */
 #define __BPF_FUNC_MAPPER(FN)          \
        FN(unspec),                     \
@@ -2782,7 +2796,8 @@ union bpf_attr {
        FN(strtol),                     \
        FN(strtoul),                    \
        FN(sk_storage_get),             \
-       FN(sk_storage_delete),
+       FN(sk_storage_delete),          \
+       FN(send_signal),
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
  * function eBPF program intends to call
index f92d6ad5e080671450b8eba785811718bddb1700..70029eafc71f9677940aa9ff714255dbb51e4cb4 100644 (file)
@@ -567,6 +567,63 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = {
        .arg3_type      = ARG_ANYTHING,
 };
 
+struct send_signal_irq_work {
+       struct irq_work irq_work;
+       struct task_struct *task;
+       u32 sig;
+};
+
+static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work);
+
+static void do_bpf_send_signal(struct irq_work *entry)
+{
+       struct send_signal_irq_work *work;
+
+       work = container_of(entry, struct send_signal_irq_work, irq_work);
+       group_send_sig_info(work->sig, SEND_SIG_PRIV, work->task, PIDTYPE_TGID);
+}
+
+BPF_CALL_1(bpf_send_signal, u32, sig)
+{
+       struct send_signal_irq_work *work = NULL;
+
+       /* Similar to bpf_probe_write_user, task needs to be
+        * in a sound condition and kernel memory access be
+        * permitted in order to send signal to the current
+        * task.
+        */
+       if (unlikely(current->flags & (PF_KTHREAD | PF_EXITING)))
+               return -EPERM;
+       if (unlikely(uaccess_kernel()))
+               return -EPERM;
+       if (unlikely(!nmi_uaccess_okay()))
+               return -EPERM;
+
+       if (in_nmi()) {
+               work = this_cpu_ptr(&send_signal_work);
+               if (work->irq_work.flags & IRQ_WORK_BUSY)
+                       return -EBUSY;
+
+               /* Add the current task, which is the target of sending signal,
+                * to the irq_work. The current task may change when queued
+                * irq works get executed.
+                */
+               work->task = current;
+               work->sig = sig;
+               irq_work_queue(&work->irq_work);
+               return 0;
+       }
+
+       return group_send_sig_info(sig, SEND_SIG_PRIV, current, PIDTYPE_TGID);
+}
+
+static const struct bpf_func_proto bpf_send_signal_proto = {
+       .func           = bpf_send_signal,
+       .gpl_only       = false,
+       .ret_type       = RET_INTEGER,
+       .arg1_type      = ARG_ANYTHING,
+};
+
 static const struct bpf_func_proto *
 tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 {
@@ -617,6 +674,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
        case BPF_FUNC_get_current_cgroup_id:
                return &bpf_get_current_cgroup_id_proto;
 #endif
+       case BPF_FUNC_send_signal:
+               return &bpf_send_signal_proto;
        default:
                return NULL;
        }
@@ -1343,5 +1402,18 @@ static int __init bpf_event_init(void)
        return 0;
 }
 
+static int __init send_signal_irq_work_init(void)
+{
+       int cpu;
+       struct send_signal_irq_work *work;
+
+       for_each_possible_cpu(cpu) {
+               work = per_cpu_ptr(&send_signal_work, cpu);
+               init_irq_work(&work->irq_work, do_bpf_send_signal);
+       }
+       return 0;
+}
+
 fs_initcall(bpf_event_init);
+subsys_initcall(send_signal_irq_work_init);
 #endif /* CONFIG_MODULES */