eBPF: reading the result of getaddrinfo

Question

Similar to the gethostlatency.py tool in the BPF Compiler Collection (bcc) I want to trace function calls to getaddrinfo. Additionally, I want to collect the returned values (IP addresses, address families)

However, I seem not be able to get past the BPF verifier with a solution that returns the right result.

The getaddrinfo function:

int getaddrinfo(const char *node, const char *service,
                   const struct addrinfo *hints,
                   struct addrinfo **res);

The result is returned in struct addrinfo **res.

This code sample directly based on gethostlatency.py at least doesn't return any error, but returns the wrong results:

#!/usr/bin/python
#
# Based on (gethostlatency.py) https://github.com/iovisor/bcc/blob/master/tools/gethostlatency.py
# Licensed under the Apache License, Version 2.0 (the "License")

from __future__ import print_function
from bcc import BPF
from time import strftime
import argparse

examples = """examples:
    ./gethostlatency           # trace all TCP accept()s
    ./gethostlatency -p 181    # only trace PID 181
"""
parser = argparse.ArgumentParser(
    description="Show latency for getaddrinfo/gethostbyname[2] calls",
    formatter_class=argparse.RawDescriptionHelpFormatter,
    epilog=examples)
parser.add_argument("-p", "--pid", help="trace this PID only", type=int,
    default=-1)
parser.add_argument("--ebpf", action="store_true",
    help=argparse.SUPPRESS)
args = parser.parse_args()

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

// Copied from: include/netdb.h
struct addrinfo
{
  int ai_flags;         /* Input flags.  */
  int ai_family;        /* Protocol family for socket.  */
  int ai_socktype;      /* Socket type.  */
  int ai_protocol;      /* Protocol for socket.  */
  u32 ai_addrlen;       /* Length of socket address.  */ // CHANGED from socklen_t
  struct sockaddr *ai_addr; /* Socket address for socket.  */
  char *ai_canonname;       /* Canonical name for service location.  */
  struct addrinfo *ai_next; /* Pointer to next in list.  */
};

struct val_t {
    u32 pid;
    char comm[TASK_COMM_LEN];
    char host[80];
    u64 ts;
};

struct data_t {
    u32 pid;
    u64 delta;
    char comm[TASK_COMM_LEN];
    u32 af;
    char host[80];
};

BPF_HASH(start, u32, struct val_t);
BPF_HASH(currres, u32, struct addrinfo *);
BPF_PERF_OUTPUT(events);

int do_entry(struct pt_regs *ctx, const char *node, const char *service,
                       const struct addrinfo *hints,
                       struct addrinfo **res) {
    if (!PT_REGS_PARM1(ctx))
        return 0;

    struct val_t val = {};
    u32 pid = bpf_get_current_pid_tgid();

    if (bpf_get_current_comm(&val.comm, sizeof(val.comm)) == 0) {
        bpf_probe_read(&val.host, sizeof(val.host),
                       (void *)PT_REGS_PARM1(ctx));
        val.pid = bpf_get_current_pid_tgid();
        val.ts = bpf_ktime_get_ns();
        start.update(&pid, &val);
        currres.update(&pid, &res);
    }

    return 0;
}

int do_return(struct pt_regs *ctx) {
    struct val_t *valp;
    struct data_t data = {};
    u64 delta;
    u32 pid = bpf_get_current_pid_tgid();

    u64 tsp = bpf_ktime_get_ns();

    valp = start.lookup(&pid);
    if (valp == 0)
        return 0;       // missed start

    bpf_probe_read(&data.comm, sizeof(data.comm), valp->comm);
    bpf_probe_read(&data.host, sizeof(data.host), (void *)valp->host);

    struct addrinfo **result;
    result = currres.lookup(&pid);
    if (result == 0) {
        return 0;   // missed entry
    }
    
    struct addrinfo* resx = *result;
    bpf_probe_read(&data.af, sizeof(data.af), &resx->ai_family);

    //data.af = resx->ai_family;

    data.pid = valp->pid;
    data.delta = tsp - valp->ts;
    events.perf_submit(ctx, &data, sizeof(data));
    start.delete(&pid);
    return 0;
}
"""
if args.ebpf:
    print(bpf_text)
    exit()

b = BPF(text=bpf_text)
b.attach_uprobe(name="c", sym="getaddrinfo", fn_name="do_entry", pid=args.pid)
b.attach_uretprobe(name="c", sym="getaddrinfo", fn_name="do_return",
                   pid=args.pid)

# header
print("%-9s %-6s %-16s %10s %-10s %s" % ("TIME", "PID", "COMM", "LATms", "AF", "HOST"))

def print_event(cpu, data, size):
    event = b["events"].event(data)
    print("%-9s %-6d %-16s %10.2f %-10d %s" % (strftime("%H:%M:%S"), event.pid,
        event.comm.decode('utf-8', 'replace'), (float(event.delta) / 1000000),
        event.af,
        event.host.decode('utf-8', 'replace')))

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

Things that might be wrong:

• I copied the addrinfo type definition from netdb.h as it seems not accessible in the BCC.
• The way I collect the result: currres.update(&pid, &res);
• The way I read from the result: bpf_probe_read(&data.af, sizeof(data.af), &resx->ai_family);

In this sample the program is just trying to read the address family (AF). So the expected value would be either 2 (AF_INET) or 10 (AF_INET6) instead it shows numbers like 32xxx.

To trigger an output a DNS request has to be made.

This program version above also shows one warning:

/virtual/main.c:52:30: warning: incompatible pointer types passing 'struct addrinfo ***' to parameter of type
    'struct addrinfo **'; remove & [-Wincompatible-pointer-types]
        currres.update(&pid, &res);
                            ^~~~

Resolving the warning results in more mysterious errors in the BPF verifier.

Tested on x64 with Kernel version 4.18.

Answer 1

You want currres to store the kernel pointer you get from getaddrinfo, so its declaration should be:

BPF_HASH(currres, u32, struct addrinfo *);

Then, you'll want to use two probe_reads to access the value from that kernel pointer:

struct addrinfo **resx = *result;
struct addrinfo *resxx;
bpf_probe_read(&resxx, sizeof(resxx), resx);
bpf_probe_read(&data.af, sizeof(data.af), &resxx->ai_family);

Note that with the latest bcc versions (>= v0.6.0), you can also write directly the following and bcc will translate it to the appropriate calls to bpf_probe_read.

struct addrinfo **resx = *result;
struct addrinfo *resxx = *resx;
data.af = resxx->ai_family;

---

Here is the full program:

#!/usr/bin/python
#
# Based on (gethostlatency.py) https://github.com/iovisor/bcc/blob/master/tools/gethostlatency.py
# Licensed under the Apache License, Version 2.0 (the "License")

from __future__ import print_function
from bcc import BPF
from time import strftime
import argparse

examples = """examples:
    ./gethostlatency           # trace all TCP accept()s
    ./gethostlatency -p 181    # only trace PID 181
"""
parser = argparse.ArgumentParser(
    description="Show latency for getaddrinfo/gethostbyname[2] calls",
    formatter_class=argparse.RawDescriptionHelpFormatter,
    epilog=examples)
parser.add_argument("-p", "--pid", help="trace this PID only", type=int,
    default=-1)
parser.add_argument("--ebpf", action="store_true",
    help=argparse.SUPPRESS)
args = parser.parse_args()

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

// Copied from: include/netdb.h
struct addrinfo
{
  int ai_flags;         /* Input flags.  */
  int ai_family;        /* Protocol family for socket.  */
  int ai_socktype;      /* Socket type.  */
  int ai_protocol;      /* Protocol for socket.  */
  u32 ai_addrlen;       /* Length of socket address.  */ // CHANGED from socklen_t
  struct sockaddr *ai_addr; /* Socket address for socket.  */
  char *ai_canonname;       /* Canonical name for service location.  */
  struct addrinfo *ai_next; /* Pointer to next in list.  */
};

struct val_t {
    u32 pid;
    char comm[TASK_COMM_LEN];
    char host[80];
    u64 ts;
};

struct data_t {
    u32 pid;
    u64 delta;
    char comm[TASK_COMM_LEN];
    u32 af;
    char host[80];
};

BPF_HASH(start, u32, struct val_t);
BPF_HASH(currres, u32, struct addrinfo **);
BPF_PERF_OUTPUT(events);

int do_entry(struct pt_regs *ctx, const char *node, const char *service,
                       const struct addrinfo *hints,
                       struct addrinfo **res) {
    if (!PT_REGS_PARM1(ctx))
        return 0;

    struct val_t val = {};
    u32 pid = bpf_get_current_pid_tgid();

    if (bpf_get_current_comm(&val.comm, sizeof(val.comm)) == 0) {
        bpf_probe_read(&val.host, sizeof(val.host),
                       (void *)PT_REGS_PARM1(ctx));
        val.pid = bpf_get_current_pid_tgid();
        val.ts = bpf_ktime_get_ns();
        start.update(&pid, &val);
        currres.update(&pid, &res);
    }

    return 0;
}

int do_return(struct pt_regs *ctx) {
    struct val_t *valp;
    struct data_t data = {};
    u64 delta;
    u32 pid = bpf_get_current_pid_tgid();

    u64 tsp = bpf_ktime_get_ns();

    valp = start.lookup(&pid);
    if (valp == 0)
        return 0;       // missed start

    bpf_probe_read(&data.comm, sizeof(data.comm), valp->comm);
    bpf_probe_read(&data.host, sizeof(data.host), (void *)valp->host);

    struct addrinfo ***result;
    result = currres.lookup(&pid);
    if (!result || !(*result)) {
        return 0;   // missed entry
    }

    struct addrinfo **resx = *result;
    struct addrinfo *resxx = *resx;
    data.af = resxx->ai_family;

    //data.af = resx->ai_family;

    data.pid = valp->pid;
    data.delta = tsp - valp->ts;
    events.perf_submit(ctx, &data, sizeof(data));
    start.delete(&pid);
    return 0;
}
"""
if args.ebpf:
    print(bpf_text)
    exit()

b = BPF(text=bpf_text)
b.attach_uprobe(name="c", sym="getaddrinfo", fn_name="do_entry", pid=args.pid)
b.attach_uretprobe(name="c", sym="getaddrinfo", fn_name="do_return",
                   pid=args.pid)

# header
print("%-9s %-6s %-16s %10s %-10s %s" % ("TIME", "PID", "COMM", "LATms", "AF", "HOST"))

def print_event(cpu, data, size):
    event = b["events"].event(data)
    print("%-9s %-6d %-16s %10.2f %-10d %s" % (strftime("%H:%M:%S"), event.pid,
        event.comm.decode('utf-8', 'replace'), (float(event.delta) / 1000000),
        event.af,
        event.host.decode('utf-8', 'replace')))

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