tracing: Introduce trace event injection
authorCong Wang <xiyou.wangcong@gmail.com>
Sat, 30 Nov 2019 04:52:18 +0000 (20:52 -0800)
committerSteven Rostedt (VMware) <rostedt@goodmis.org>
Mon, 2 Dec 2019 16:07:00 +0000 (11:07 -0500)
We have been trying to use rasdaemon to monitor hardware errors like
correctable memory errors. rasdaemon uses trace events to monitor
various hardware errors. In order to test it, we have to inject some
hardware errors, unfortunately not all of them provide error
injections. MCE does provide a way to inject MCE errors, but errors
like PCI error and devlink error don't, it is not easy to add error
injection to each of them. Instead, it is relatively easier to just
allow users to inject trace events in a generic way so that all trace
events can be injected.

This patch introduces trace event injection, where a new 'inject' is
added to each tracepoint directory. Users could write into this file
with key=value pairs to specify the value of each fields of the trace
event, all unspecified fields are set to zero values by default.

For example, for the net/net_dev_queue tracepoint, we can inject:

  INJECT=/sys/kernel/debug/tracing/events/net/net_dev_queue/inject
  echo "" > $INJECT
  echo "name='test'" > $INJECT
  echo "name='test' len=1024" > $INJECT
  cat /sys/kernel/debug/tracing/trace
  ...
   <...>-614   [000] ....    36.571483: net_dev_queue: dev= skbaddr=00000000fbf338c2 len=0
   <...>-614   [001] ....   136.588252: net_dev_queue: dev=test skbaddr=00000000fbf338c2 len=0
   <...>-614   [001] .N..   208.431878: net_dev_queue: dev=test skbaddr=00000000fbf338c2 len=1024

Triggers could be triggered as usual too:

  echo "stacktrace if len == 1025" > /sys/kernel/debug/tracing/events/net/net_dev_queue/trigger
  echo "len=1025" > $INJECT
  cat /sys/kernel/debug/tracing/trace
  ...
      bash-614   [000] ....    36.571483: net_dev_queue: dev= skbaddr=00000000fbf338c2 len=0
      bash-614   [001] ....   136.588252: net_dev_queue: dev=test skbaddr=00000000fbf338c2 len=0
      bash-614   [001] .N..   208.431878: net_dev_queue: dev=test skbaddr=00000000fbf338c2 len=1024
      bash-614   [001] .N.1   284.236349: <stack trace>
 => event_inject_write
 => vfs_write
 => ksys_write
 => do_syscall_64
 => entry_SYSCALL_64_after_hwframe

The only thing that can't be injected is string pointers as they
require constant string pointers, this can't be done at run time.

Link: http://lkml.kernel.org/r/20191130045218.18979-1-xiyou.wangcong@gmail.com
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
kernel/trace/Kconfig
kernel/trace/Makefile
kernel/trace/trace.h
kernel/trace/trace_events.c
kernel/trace/trace_events_inject.c [new file with mode: 0644]

index f67620499faa47fc07deee326882278a3ff0edd8..29a9c5058b62fb3743506ca701832aae7d61db27 100644 (file)
@@ -672,6 +672,15 @@ config HIST_TRIGGERS
          See Documentation/trace/histogram.rst.
          If in doubt, say N.
 
+config TRACE_EVENT_INJECT
+       bool "Trace event injection"
+       depends on TRACING
+       help
+         Allow user-space to inject a specific trace event into the ring
+         buffer. This is mainly used for testing purpose.
+
+         If unsure, say N.
+
 config MMIOTRACE_TEST
        tristate "Test module for mmiotrace"
        depends on MMIOTRACE && m
index c2b2148bb1d21ebef8309717a03e0c49ff9df133..0e63db62225fc62fab61558cfefa2d3f5f0c9695 100644 (file)
@@ -69,6 +69,7 @@ obj-$(CONFIG_EVENT_TRACING) += trace_event_perf.o
 endif
 obj-$(CONFIG_EVENT_TRACING) += trace_events_filter.o
 obj-$(CONFIG_EVENT_TRACING) += trace_events_trigger.o
+obj-$(CONFIG_TRACE_EVENT_INJECT) += trace_events_inject.o
 obj-$(CONFIG_HIST_TRIGGERS) += trace_events_hist.o
 obj-$(CONFIG_BPF_EVENTS) += bpf_trace.o
 obj-$(CONFIG_KPROBE_EVENTS) += trace_kprobe.o
index ca7fccafbcbbad8126c058fbe10e31eedc1e498d..63bf60f793987fbebd812fbe5252925879feaad6 100644 (file)
@@ -1601,6 +1601,7 @@ extern struct list_head ftrace_events;
 
 extern const struct file_operations event_trigger_fops;
 extern const struct file_operations event_hist_fops;
+extern const struct file_operations event_inject_fops;
 
 #ifdef CONFIG_HIST_TRIGGERS
 extern int register_trigger_hist_cmd(void);
index 6b3a69e9aa6adc511ddd881219108b24f43ba437..c6de3cebc127fe83842887a300cae6d44dcd8c98 100644 (file)
@@ -2044,6 +2044,12 @@ event_create_dir(struct dentry *parent, struct trace_event_file *file)
        trace_create_file("format", 0444, file->dir, call,
                          &ftrace_event_format_fops);
 
+#ifdef CONFIG_TRACE_EVENT_INJECT
+       if (call->event.type && call->class->reg)
+               trace_create_file("inject", 0200, file->dir, file,
+                                 &event_inject_fops);
+#endif
+
        return 0;
 }
 
diff --git a/kernel/trace/trace_events_inject.c b/kernel/trace/trace_events_inject.c
new file mode 100644 (file)
index 0000000..d437107
--- /dev/null
@@ -0,0 +1,331 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * trace_events_inject - trace event injection
+ *
+ * Copyright (C) 2019 Cong Wang <cwang@twitter.com>
+ */
+
+#include <linux/module.h>
+#include <linux/ctype.h>
+#include <linux/mutex.h>
+#include <linux/slab.h>
+#include <linux/rculist.h>
+
+#include "trace.h"
+
+static int
+trace_inject_entry(struct trace_event_file *file, void *rec, int len)
+{
+       struct trace_event_buffer fbuffer;
+       struct ring_buffer *buffer;
+       int written = 0;
+       void *entry;
+
+       rcu_read_lock_sched();
+       buffer = file->tr->trace_buffer.buffer;
+       entry = trace_event_buffer_reserve(&fbuffer, file, len);
+       if (entry) {
+               memcpy(entry, rec, len);
+               written = len;
+               trace_event_buffer_commit(&fbuffer);
+       }
+       rcu_read_unlock_sched();
+
+       return written;
+}
+
+static int
+parse_field(char *str, struct trace_event_call *call,
+           struct ftrace_event_field **pf, u64 *pv)
+{
+       struct ftrace_event_field *field;
+       char *field_name;
+       int s, i = 0;
+       int len;
+       u64 val;
+
+       if (!str[i])
+               return 0;
+       /* First find the field to associate to */
+       while (isspace(str[i]))
+               i++;
+       s = i;
+       while (isalnum(str[i]) || str[i] == '_')
+               i++;
+       len = i - s;
+       if (!len)
+               return -EINVAL;
+
+       field_name = kmemdup_nul(str + s, len, GFP_KERNEL);
+       if (!field_name)
+               return -ENOMEM;
+       field = trace_find_event_field(call, field_name);
+       kfree(field_name);
+       if (!field)
+               return -ENOENT;
+
+       *pf = field;
+       while (isspace(str[i]))
+               i++;
+       if (str[i] != '=')
+               return -EINVAL;
+       i++;
+       while (isspace(str[i]))
+               i++;
+       s = i;
+       if (isdigit(str[i]) || str[i] == '-') {
+               char *num, c;
+               int ret;
+
+               /* Make sure the field is not a string */
+               if (is_string_field(field))
+                       return -EINVAL;
+
+               if (str[i] == '-')
+                       i++;
+
+               /* We allow 0xDEADBEEF */
+               while (isalnum(str[i]))
+                       i++;
+               num = str + s;
+               c = str[i];
+               if (c != '\0' && !isspace(c))
+                       return -EINVAL;
+               str[i] = '\0';
+               /* Make sure it is a value */
+               if (field->is_signed)
+                       ret = kstrtoll(num, 0, &val);
+               else
+                       ret = kstrtoull(num, 0, &val);
+               str[i] = c;
+               if (ret)
+                       return ret;
+
+               *pv = val;
+               return i;
+       } else if (str[i] == '\'' || str[i] == '"') {
+               char q = str[i];
+
+               /* Make sure the field is OK for strings */
+               if (!is_string_field(field))
+                       return -EINVAL;
+
+               for (i++; str[i]; i++) {
+                       if (str[i] == '\\' && str[i + 1]) {
+                               i++;
+                               continue;
+                       }
+                       if (str[i] == q)
+                               break;
+               }
+               if (!str[i])
+                       return -EINVAL;
+
+               /* Skip quotes */
+               s++;
+               len = i - s;
+               if (len >= MAX_FILTER_STR_VAL)
+                       return -EINVAL;
+
+               *pv = (unsigned long)(str + s);
+               str[i] = 0;
+               /* go past the last quote */
+               i++;
+               return i;
+       }
+
+       return -EINVAL;
+}
+
+static int trace_get_entry_size(struct trace_event_call *call)
+{
+       struct ftrace_event_field *field;
+       struct list_head *head;
+       int size = 0;
+
+       head = trace_get_fields(call);
+       list_for_each_entry(field, head, link) {
+               if (field->size + field->offset > size)
+                       size = field->size + field->offset;
+       }
+
+       return size;
+}
+
+static void *trace_alloc_entry(struct trace_event_call *call, int *size)
+{
+       int entry_size = trace_get_entry_size(call);
+       struct ftrace_event_field *field;
+       struct list_head *head;
+       void *entry = NULL;
+
+       /* We need an extra '\0' at the end. */
+       entry = kzalloc(entry_size + 1, GFP_KERNEL);
+       if (!entry)
+               return NULL;
+
+       head = trace_get_fields(call);
+       list_for_each_entry(field, head, link) {
+               if (!is_string_field(field))
+                       continue;
+               if (field->filter_type == FILTER_STATIC_STRING)
+                       continue;
+               if (field->filter_type == FILTER_DYN_STRING) {
+                       u32 *str_item;
+                       int str_loc = entry_size & 0xffff;
+
+                       str_item = (u32 *)(entry + field->offset);
+                       *str_item = str_loc; /* string length is 0. */
+               } else {
+                       char **paddr;
+
+                       paddr = (char **)(entry + field->offset);
+                       *paddr = "";
+               }
+       }
+
+       *size = entry_size + 1;
+       return entry;
+}
+
+#define INJECT_STRING "STATIC STRING CAN NOT BE INJECTED"
+
+/* Caller is responsible to free the *pentry. */
+static int parse_entry(char *str, struct trace_event_call *call, void **pentry)
+{
+       struct ftrace_event_field *field;
+       unsigned long irq_flags;
+       void *entry = NULL;
+       int entry_size;
+       u64 val;
+       int len;
+
+       entry = trace_alloc_entry(call, &entry_size);
+       *pentry = entry;
+       if (!entry)
+               return -ENOMEM;
+
+       local_save_flags(irq_flags);
+       tracing_generic_entry_update(entry, call->event.type, irq_flags,
+                                    preempt_count());
+
+       while ((len = parse_field(str, call, &field, &val)) > 0) {
+               if (is_function_field(field))
+                       return -EINVAL;
+
+               if (is_string_field(field)) {
+                       char *addr = (char *)(unsigned long) val;
+
+                       if (field->filter_type == FILTER_STATIC_STRING) {
+                               strlcpy(entry + field->offset, addr, field->size);
+                       } else if (field->filter_type == FILTER_DYN_STRING) {
+                               int str_len = strlen(addr) + 1;
+                               int str_loc = entry_size & 0xffff;
+                               u32 *str_item;
+
+                               entry_size += str_len;
+                               *pentry = krealloc(entry, entry_size, GFP_KERNEL);
+                               if (!*pentry) {
+                                       kfree(entry);
+                                       return -ENOMEM;
+                               }
+                               entry = *pentry;
+
+                               strlcpy(entry + (entry_size - str_len), addr, str_len);
+                               str_item = (u32 *)(entry + field->offset);
+                               *str_item = (str_len << 16) | str_loc;
+                       } else {
+                               char **paddr;
+
+                               paddr = (char **)(entry + field->offset);
+                               *paddr = INJECT_STRING;
+                       }
+               } else {
+                       switch (field->size) {
+                       case 1: {
+                               u8 tmp = (u8) val;
+
+                               memcpy(entry + field->offset, &tmp, 1);
+                               break;
+                       }
+                       case 2: {
+                               u16 tmp = (u16) val;
+
+                               memcpy(entry + field->offset, &tmp, 2);
+                               break;
+                       }
+                       case 4: {
+                               u32 tmp = (u32) val;
+
+                               memcpy(entry + field->offset, &tmp, 4);
+                               break;
+                       }
+                       case 8:
+                               memcpy(entry + field->offset, &val, 8);
+                               break;
+                       default:
+                               return -EINVAL;
+                       }
+               }
+
+               str += len;
+       }
+
+       if (len < 0)
+               return len;
+
+       return entry_size;
+}
+
+static ssize_t
+event_inject_write(struct file *filp, const char __user *ubuf, size_t cnt,
+                  loff_t *ppos)
+{
+       struct trace_event_call *call;
+       struct trace_event_file *file;
+       int err = -ENODEV, size;
+       void *entry = NULL;
+       char *buf;
+
+       if (cnt >= PAGE_SIZE)
+               return -EINVAL;
+
+       buf = memdup_user_nul(ubuf, cnt);
+       if (IS_ERR(buf))
+               return PTR_ERR(buf);
+       strim(buf);
+
+       mutex_lock(&event_mutex);
+       file = event_file_data(filp);
+       if (file) {
+               call = file->event_call;
+               size = parse_entry(buf, call, &entry);
+               if (size < 0)
+                       err = size;
+               else
+                       err = trace_inject_entry(file, entry, size);
+       }
+       mutex_unlock(&event_mutex);
+
+       kfree(entry);
+       kfree(buf);
+
+       if (err < 0)
+               return err;
+
+       *ppos += err;
+       return cnt;
+}
+
+static ssize_t
+event_inject_read(struct file *file, char __user *buf, size_t size,
+                 loff_t *ppos)
+{
+       return -EPERM;
+}
+
+const struct file_operations event_inject_fops = {
+       .open = tracing_open_generic,
+       .read = event_inject_read,
+       .write = event_inject_write,
+};