harmony 鸿蒙Trace

2022-08-09
浏览 (614)

Trace

Basic Concepts

Trace helps you learn about the kernel running process and the execution sequence of modules and tasks. With the information, you can better understand the code running process of the kernel and locate time sequence problems.

Working Principles

The kernel provides a hook framework to embed hooks in the main process of each module. In the initial startup phase of the kernel, the trace function is initialized and the trace handlers are registered with the hooks.

When a hook is triggered, the trace module encapsulates the input information and adds the trace frame header information, including the event type, ID of the running CPU, ID of the running task, and relative timestamp.

The trace module provides two working modes: offline mode and online mode.

In offline mode, trace frames are stored in a circular buffer. If too many frames are stored in the circular buffer, earlier frames will be overwritten to ensure that the information in the buffer is always the latest. Data in the circular buffer can be exported by running the shell command for further analysis. The exported information is sorted by timestamp.

The online mode must be used with the integrated development environment (IDE). Trace frames are sent to the IDE in real time. The IDE parses the records and displays them in a visualized manner.

Available APIs

Kernel Mode

The trace module of the OpenHarmony LiteOS-A kernel provides the following APIs. For more details, see API reference.

Table 1 APIs of the trace module

Category	Description
Starting/Stopping trace	LOS_TraceStart: starts trace. LOS_TraceStop: stops trace.
Managing trace records	LOS_TraceRecordDump: dumps data from the trace buffer. LOS_TraceRecordGet: obtains the start address of the trace buffer. LOS_TraceReset: clears events in the trace buffer.
Filtering trace records	LOS_TraceEventMaskSet: sets the event mask to trace only events of the specified modules.
Masking events of specified interrupt IDs	LOS_TraceHwiFilterHookReg: registers a hook to filter out events of specified interrupt IDs.
Performing function instrumentation	LOS_TRACE_EASY: performs simple instrumentation. LOS_TRACE: performs standard instrumentation.

You can perform function instrumentation in the source code to trace specific events. The system provides the following APIs for instrumentation:

LOS_TRACE_EASY(TYPE, IDENTITY, params…) for simple instrumentation

You only need to insert this API into the source code.
TYPE specifies the event type. The value range is 0 to 0xF. The meaning of each value is user-defined.
IDENTITY specifies the object of the event operation. The value is of the UIntPtr type.
Params specifies the event parameters. The value is of the UIntPtr type. Example:

Perform simple instrumentation for reading and writing files fd1 and fd2.
Set TYPE to 1 for read operations and 2 for write operations.
Insert the following to the position where the fd1 file is read:
LOS_TRACE_EASY(1, fd1, flag, size);
Insert the following to the position where the fd2 file is read:
LOS_TRACE_EASY(1, fd2, flag, size);
Insert the following to the position where the fd1 file is written:
LOS_TRACE_EASY(2, fd1, flag, size);
Insert the following in the position where the fd2 file is written:
LOS_TRACE_EASY(2, fd2, flag, size);

LOS_TRACE(TYPE, IDENTITY, params…) for standard instrumentation.

Compared with simple instrumentation, standard instrumentation supports dynamic event filtering and parameter tailoring. However, you need to extend the functions based on rules.
TYPE specifies the event type. You can define the event type in enum LOS_TRACE_TYPE in the header file los_trace.h. For details about methods and rules for defining events, see other event types.
The IDENTITY and Params are the same as those of simple instrumentation. Example:

1. Set the event mask (module-level event type) in enum LOS_TRACE_MASK.
  Format: TRACE_#MOD#_FLAG (MOD indicates the module name)
  Example:
  TRACE_FS_FLAG = 0x4000
2. Define the event type in **enum LOS_TRACE_TYPE**.
  Format: #TYPE# = TRACE_#MOD#_FLAG|NUMBER
  Example:
  FS_READ  = TRACE_FS_FLAG|0; // Read files.
  FS_WRITE = TRACE_FS_FLAG|1; // Write files.
3. Set event parameters in the #TYPE#_PARAMS(IDENTITY, parma1...) IDENTITY, ... format.
  #TYPE# is the #TYPE# defined in step 2.
  Example:
  #define FS_READ_PARAMS(fp, fd, flag, size)    fp, fd, flag, size
  The parameters defined by the macro correspond to the event parameters recorded in the trace buffer. You can modify the parameters as required.
  If no parameter is specified, events of this type are not traced.
  #define FS_READ_PARAMS(fp, fd, flag, size) // File reading events are not traced.
4. Insert a code stub in a proper position.
  Format: LOS_TRACE(#TYPE#, #TYPE#_PARAMS(IDENTITY, parma1...))
  LOS_TRACE(FS_READ, fp, fd, flag, size); // Code stub for reading files.
  #The parameters following #TYPE# are the input parameter of the **FS_READ_PARAMS** function in step 3.

NOTE
The preset trace events and parameters can be tailored in the same way. For details about the parameters, see kernel\include\los_trace.h.

For LOS_TraceEventMaskSet(UINT32 mask), only the most significant 28 bits (corresponding to the enable bit of the module in LOS_TRACE_MASK) of the mask take effect and are used only for module-based tracing. Currently, fine-grained event-based tracing is not supported. For example, in LOS_TraceEventMaskSet(0x202), the effective mask is 0x200 (TRACE_QUE_FLAG) and all events of the QUE module are collected. The recommended method is LOS_TraceEventMaskSet(TRACE_EVENT_FLAG|TRACE_MUX_FLAG|TRACE_SEM_FLAG|TRACE_QUE_FLAG);.
To enable trace of only simple instrumentation events, set Trace Mask to TRACE_MAX_FLAG.
The trace buffer has limited capacity. When the trace buffer is full, events will be overwritten. You can use LOS_TraceRecordDump to export data from the trace buffer and locate the latest records by CurEvtIndex.
The typical trace operation process includes LOS_TraceStart, LOS_TraceStop, and LOS_TraceRecordDump.
You can filter out interrupt events by interrupt ID to prevent other events from being overwritten due to frequent triggering of a specific interrupt in some scenarios. You can customize interrupt filtering rules. Example:

  BOOL Example_HwiNumFilter(UINT32 hwiNum)
  {
      if ((hwiNum == TIMER_INT)||(hwiNum == DMA_INT)) {
          return TRUE;
      }
      return FALSE;
  }
  LOS_TraceHwiFilterHookReg(Example_HwiNumFilter);

The interrupt events with interrupt ID of TIMER_INT or DMA_INT are not traced.

User Mode

The trace character device is added in /dev/trace. You can use read(), write(), and ioctl() on the device node to read, write, and control trace in user mode.

read(): reads the trace data in user mode.
write(): writes an event in user mode.
ioctl(): performs user-mode trace operations, including:

#define TRACE_IOC_MAGIC   'T'
#define TRACE_START      _IO(TRACE_IOC_MAGIC, 1)
#define TRACE_STOP       _IO(TRACE_IOC_MAGIC, 2)
#define TRACE_RESET      _IO(TRACE_IOC_MAGIC, 3)
#define TRACE_DUMP		 _IO(TRACE_IOC_MAGIC, 4)
#define TRACE_SET_MASK	 _IO(TRACE_IOC_MAGIC, 5)

The operations specified by the input parameter of ioctl() correspond to LOS_TraceStart, LOS_TraceStop, LOS_TraceReset, LOS_TraceRecordDump, and LOS_TraceEventMaskSet, respectively.

For details, see User-Mode Development Example.

Development Guidelines

Kernel-Mode Development Process

The typical trace process is as follows:

Configure the macro related to the trace module. Configure the macro LOSCFG_KERNEL_TRACE, which is disabled by default. Run the make update_config command in the kernel/liteos_a directory, choose Kernel > Enable Hook Feature, and set Enable Trace Feature to YES.

(Optional) Preset event parameters and stubs (or use the default event parameter settings and event stubs).
(Optional) Call LOS_TraceStop to stop trace and call LOS_TraceReset to clear the trace buffer. (Trace is started by default.)
(Optional) Call LOS_TraceEventMaskSet to set the mask of the events to be traced. The default event mask enables only trace of interrupts and task events. For details about the event masks, see LOS_TRACE_MASK in los_trace.h.
Call LOS_TraceStart at the start of the code where the event needs to be traced.
Call LOS_TraceStop at the end of the code where the event needs to be traced.
Call LOS_TraceRecordDump to output the data in the buffer. (The input parameter of the function is of the Boolean type. The value FALSE means to output data in the specified format, and the value TRUE means to output data to Trace IDE.)

The methods in steps 3 to 7 are encapsulated with shell commands. You can run these commands on shell. The mappings between the methods and commands are as follows:

LOS_TraceReset —— trace_reset
LOS_TraceEventMaskSet —— trace_mask
LOS_TraceStart —— trace_start
LOS_TraceStop —— trace_stop
LOS_TraceRecordDump —— trace_dump

Kernel-Mode Development Example

This example implements the following:

Create a trace task.
Set the event mask.
Start trace.
Stop trace.
Output trace data in the specified format.

Kernel-Mode Sample Code

You can add the test function of the sample code to TestTaskEntry in kernel/liteos_a/testsuites /kernel /src/osTest.c for testing.

The sample code is as follows:

#include "los_trace.h"
UINT32 g_traceTestTaskId;
VOID Example_Trace(VOID)
{
    UINT32 ret;
    LOS_TaskDelay(10);
    /* Start trace. */
    ret = LOS_TraceStart();
    if (ret != LOS_OK) {
        dprintf("trace start error\n");
        return;
    }
    /* Trigger a task switching event. */
    LOS_TaskDelay(1);
    LOS_TaskDelay(1);
    LOS_TaskDelay(1);
    /* Stop trace. */
    LOS_TraceStop();
    LOS_TraceRecordDump(FALSE);
}
UINT32 Example_Trace_test(VOID)
{
    UINT32 ret;
    TSK_INIT_PARAM_S traceTestTask;
    /* Create a trace task. */
    memset(&traceTestTask, 0, sizeof(TSK_INIT_PARAM_S));
    traceTestTask.pfnTaskEntry = (TSK_ENTRY_FUNC)Example_Trace;
    traceTestTask.pcName       = "TestTraceTsk";    /* Test task name. */
    traceTestTask.uwStackSize  = 0x800; // 0x800: trace test task stack size
    traceTestTask.usTaskPrio   = 5; // 5: trace test task priority
    traceTestTask.uwResved   = LOS_TASK_STATUS_DETACHED;
    ret = LOS_TaskCreate(&g_traceTestTaskId, &traceTestTask);
    if (ret != LOS_OK) {
        dprintf("TraceTestTask create failed .\n");
        return LOS_NOK;
    }
    /* Trace is started by default. Therefore, you can stop trace, clear the buffer, and then start trace. */                   	  
    LOS_TraceStop();
    LOS_TraceReset();
    /* Enable trace of the Task module events. */
    LOS_TraceEventMaskSet(TRACE_TASK_FLAG);
    return LOS_OK;
}
LOS_MODULE_INIT(Example_Trace_test, LOS_INIT_LEVEL_KMOD_EXTENDED);

Verification

The output is as follows:

***TraceInfo begin***
clockFreq = 50000000
CurEvtIndex = 7
Index   Time(cycles)      EventType      CurTask   Identity      params
0       0x366d5e88        0x45           0x1       0x0           0x1f         0x4       0x0
1       0x366d74ae        0x45           0x0       0x1           0x0          0x8       0x1f
2       0x36940da6        0x45           0x1       0xc           0x1f         0x4       0x9
3       0x3694337c        0x45           0xc       0x1           0x9          0x8       0x1f
4       0x36eea56e        0x45           0x1       0xc           0x1f         0x4       0x9
5       0x36eec810        0x45           0xc       0x1           0x9          0x8       0x1f
6       0x3706f804        0x45           0x1       0x0           0x1f         0x4       0x0
7       0x37070e59        0x45           0x0       0x1           0x0          0x8       0x1f
***TraceInfo end***

The output event information includes the occurrence time, event type, task in which the event occurs, object of the event operation, and other parameters of the event.

EventType: type of the event. For details, see enum LOS_TRACE_TYPE in los_trace.h.
CurrentTask: ID of the running task.
Identity: object of the event operation. For details, see #TYPE#_PARAMS in los_trace.h.
params: event parameters. For details, see #TYPE#_PARAMS in los_trace.h.

The following uses output No. 0 as an example.

Index   Time(cycles)      EventType      CurTask   Identity      params
0       0x366d5e88        0x45           0x1       0x0           0x1f         0x4

Time (cycles) can be converted into time (in seconds) by dividing the cycles by clockFreq.
0x45 indicates the task switching event. 0x1 is the ID of the task in running.
For details about the meanings of Identity and params, see the TASK_SWITCH_PARAMS macro.

#define TASK_SWITCH_PARAMS(taskId, oldPriority, oldTaskStatus, newPriority, newTaskStatus) \
taskId, oldPriority, oldTaskStatus, newPriority, newTaskStatus

Because of #TYPE#_PARAMS(IDENTITY, parma1…) IDENTITY, …, Identity is taskId (0x0) and the first parameter is oldPriority (0x1f).

NOTE

The number of parameters in params is specified by LOSCFG_TRACE_FRAME_MAX_PARAMS. The default value is 3. Excess parameters are not recorded. You need to set LOSCFG_TRACE_FRAME_MAX_PARAMS based on service requirements.

Task 0x1 is switched to Task 0x0. The priority of task 0x1 is 0x1f, and the state is 0x4. The priority of the task 0x0 is 0x0.