doc/howto_libipt.md

Decoding Intel(R) Processor Trace Using libipt {#libipt}
========================================================

<!---
 ! Copyright (c) 2013-2021, Intel Corporation
 !
 ! Redistribution and use in source and binary forms, with or without
 ! modification, are permitted provided that the following conditions are met:
 !
 !  * Redistributions of source code must retain the above copyright notice,
 !    this list of conditions and the following disclaimer.
 !  * Redistributions in binary form must reproduce the above copyright notice,
 !    this list of conditions and the following disclaimer in the documentation
 !    and/or other materials provided with the distribution.
 !  * Neither the name of Intel Corporation nor the names of its contributors
 !    may be used to endorse or promote products derived from this software
 !    without specific prior written permission.
 !
 ! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 ! AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 ! IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 ! ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
 ! LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 ! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 ! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 ! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 ! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 ! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 ! POSSIBILITY OF SUCH DAMAGE.
 !-->

This chapter describes how to use libipt for various tasks around Intel
Processor Trace (Intel PT).  For code examples, refer to the sample tools that
are contained in the source tree:

  * *ptdump*    A packet dumper example.
  * *ptxed*     A control-flow reconstruction example.
  * *pttc*      A packet encoder example.


For detailed information about Intel PT, please refer to the respective chapter
in Volume 3 of the Intel Software Developer's Manual at
http://www.intel.com/sdm.


## Introduction

The libipt decoder library provides multiple layers of abstraction ranging from
packet encoding and decoding to full execution flow reconstruction.  The layers
are organized as follows:

  * *packets*               This layer deals with raw Intel PT packets.

  * *events*                This layer deals with packet combinations that
                            encode higher-level events.

  * *instruction flow*      This layer deals with the execution flow on the
                            instruction level.

  * *block*                 This layer deals with the execution flow on the
                            instruction level.

                            It is faster than the instruction flow decoder but
                            requires a small amount of post-processing.


Each layer provides its own encoder or decoder struct plus a set of functions
for allocating and freeing encoder or decoder objects and for synchronizing
decoders onto the Intel PT packet stream.  Function names are prefixed with
`pt_<lyr>_` where `<lyr>` is an abbreviation of the layer name.  The following
abbreviations are used:

  * *enc*     Packet encoding (packet layer).
  * *pkt*     Packet decoding (packet layer).
  * *qry*     Event (or query) layer.
  * *insn*    Instruction flow layer.
  * *blk*     Block layer.


Here is some generic example code for working with decoders:

~~~{.c}
    struct pt_<layer>_decoder *decoder;
    struct pt_config config;
    int errcode;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;
    config...

    decoder = pt_<lyr>_alloc_decoder(&config);
    if (!decoder)
        <handle error>(errcode);

    errcode = pt_<lyr>_sync_<where>(decoder);
    if (errcode < 0)
        <handle error>(errcode);

    <use decoder>(decoder);

    pt_<lyr>_free_decoder(decoder);
~~~

First, configure the decoder.  As a minimum, the size of the config struct and
the `begin` and `end` of the buffer containing the Intel PT data need to be set.
Configuration options details will be discussed later in this chapter.  In the
case of packet encoding, this is the begin and end address of the pre-allocated
buffer, into which Intel PT packets shall be written.

Next, allocate a decoder object for the layer you are interested in.  A return
value of NULL indicates an error.  There is no further information available on
the exact error condition.  Most of the time, however, the error is the result
of an incomplete or inconsistent configuration.

Before the decoder can be used, it needs to be synchronized onto the Intel PT
packet stream specified in the configuration.  The only exception to this is the
packet encoder, which is implicitly synchronized onto the beginning of the Intel
PT buffer.

Depending on the type of decoder, one or more synchronization options are
available.

  * `pt_<lyr>_sync_forward()`     Synchronize onto the next PSB in forward
                                  direction (or the first PSB if not yet
                                  synchronized).

  * `pt_<lyr>_sync_backward()`    Synchronize onto the next PSB in backward
                                  direction (or the last PSB if not yet
                                  synchronized).

  * `pt_<lyr>_sync_set()`         Set the synchronization position to a
                                  user-defined location in the Intel PT packet
                                  stream.
                                  There is no check whether the specified
                                  location makes sense or is valid.


After synchronizing, the decoder can be used.  While decoding, the decoder
stores the location of the last PSB it encountered during normal decode.
Subsequent calls to pt_<lyr>_sync_forward() will start searching from that
location.  This is useful for re-synchronizing onto the Intel PT packet stream
in case of errors.  An example of a typical decode loop is given below:

~~~{.c}
    for (;;) {
        int errcode;

        errcode = <use decoder>(decoder);
        if (errcode >= 0)
            continue;

        if (errcode == -pte_eos)
            return;

        <report error>(errcode);

        do {
            errcode = pt_<lyr>_sync_forward(decoder);

            if (errcode == -pte_eos)
                return;
        } while (errcode < 0);
    }
~~~

You can get the current decoder position as offset into the Intel PT buffer via:

    pt_<lyr>_get_offset()


You can get the position of the last synchronization point as offset into the
Intel PT buffer via:

    pt_<lyr>_get_sync_offset()


Each layer will be discussed in detail below.  In the remainder of this section,
general functionality will be considered.


### Version

You can query the library version using:

  * `pt_library_version()`


This function returns a version structure that can be used for compatibility
checks or simply for reporting the version of the decoder library.


### Errors

The library uses a single error enum for all layers.

  * `enum pt_error_code`      An enumeration of encode and decode errors.


Errors are typically represented as negative pt_error_code enumeration constants
and returned as an int.  The library provides two functions for dealing with
errors:

  * `pt_errcode()`            Translate an int return value into a pt_error_code
                              enumeration constant.

  * `pt_errstr()`             Returns a human-readable error string.


Not all errors may occur on every layer.  Every API function specifies the
errors it may return.


### Configuration

Every encoder or decoder allocation function requires a configuration argument.
Some of its fields have already been discussed in the example above.  Refer to
the `intel-pt.h` header for detailed and up-to-date documentation of each field.

As a minimum, the `size` field needs to be set to `sizeof(struct pt_config)` and
`begin` and `end` need to be set to the Intel PT buffer to use.

The size is used for detecting library version mismatches and to provide
backwards compatibility.  Without the proper `size`, decoder allocation will
fail.

Although not strictly required, it is recommended to also set the `cpu` field to
the processor, on which Intel PT has been collected (for decoders), or for which
Intel PT shall be generated (for encoders).  This allows implementing
processor-specific behavior such as erratum workarounds.


## The Packet Layer

This layer deals with Intel PT packet encoding and decoding.  It can further be
split into three sub-layers: opcodes, encoding, and decoding.


### Opcodes

The opcodes layer provides enumerations for all the bits necessary for Intel PT
encoding and decoding.  The enumeration constants can be used without linking to
the decoder library.  There is no encoder or decoder struct associated with this
layer.  See the intel-pt.h header file for details.


### Packet Encoding

The packet encoding layer provides support for encoding Intel PT
packet-by-packet.  Start by configuring and allocating a `pt_packet_encoder` as
shown below:

~~~{.c}
    struct pt_encoder *encoder;
    struct pt_config config;
    int errcode;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;

    encoder = pt_alloc_encoder(&config);
    if (!encoder)
        <handle error>(errcode);
~~~

For packet encoding, only the mandatory config fields need to be filled in.

The allocated encoder object will be implicitly synchronized onto the beginning
of the Intel PT buffer.  You may change the encoder's position at any time by
calling `pt_enc_sync_set()` with the desired buffer offset.

Next, fill in a `pt_packet` object with details about the packet to be encoded.
You do not need to fill in the `size` field.  The needed size is computed by the
encoder.  There is no consistency check with the size specified in the packet
object.  The following example encodes a TIP packet:

~~~{.c}
    struct pt_packet_encoder *encoder = ...;
    struct pt_packet packet;
    int errcode;

    packet.type = ppt_tip;
    packet.payload.ip.ipc = pt_ipc_update_16;
    packet.payload.ip.ip = <ip>;
~~~

For IP packets, for example FUP or TIP.PGE, there is no need to mask out bits in
the `ip` field that will not be encoded in the packet due to the specified IP
compression in the `ipc` field.  The encoder will ignore them.

There are no consistency checks whether the specified IP compression in the
`ipc` field is allowed in the current context or whether decode will result in
the full IP specified in the `ip` field.

Once the packet object has been filled, it can be handed over to the encoder as
shown here:

~~~{.c}
    errcode = pt_enc_next(encoder, &packet);
    if (errcode < 0)
        <handle error>(errcode);
~~~

The encoder will encode the packet, write it into the Intel PT buffer, and
advance its position to the next byte after the packet.  On a successful encode,
it will return the number of bytes that have been written.  In case of errors,
nothing will be written and the encoder returns a negative error code.


### Packet Decoding

The packet decoding layer provides support for decoding Intel PT
packet-by-packet.  Start by configuring and allocating a `pt_packet_decoder` as
shown here:

~~~{.c}
    struct pt_packet_decoder *decoder;
    struct pt_config config;
    int errcode;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;
    config.decode.callback = <decode function>;
    config.decode.context = <decode context>;

    decoder = pt_pkt_alloc_decoder(&config);
    if (!decoder)
        <handle error>(errcode);
~~~

For packet decoding, an optional decode callback function may be specified in
addition to the mandatory config fields.  If specified, the callback function
will be called for packets the decoder does not know about.  If there is no
decode callback specified, the decoder will return `-pte_bad_opc`.  In addition
to the callback function pointer, an optional pointer to user-defined context
information can be specified.  This context will be passed to the decode
callback function.

Before the decoder can be used, it needs to be synchronized onto the Intel PT
packet stream.  Packet decoders offer three synchronization functions.  To
iterate over synchronization points in the Intel PT packet stream in forward or
backward direction, use one of the following two functions respectively:

    pt_pkt_sync_forward()
    pt_pkt_sync_backward()


To manually synchronize the decoder at a particular offset into the Intel PT
packet stream, use the following function:

    pt_pkt_sync_set()


There are no checks to ensure that the specified offset is at the beginning of a
packet.  The example below shows synchronization to the first synchronization
point:

~~~{.c}
    struct pt_packet_decoder *decoder;
    int errcode;

    errcode = pt_pkt_sync_forward(decoder);
    if (errcode < 0)
        <handle error>(errcode);
~~~

The decoder will remember the last synchronization packet it decoded.
Subsequent calls to `pt_pkt_sync_forward` and `pt_pkt_sync_backward` will use
this as their starting point.

You can get the current decoder position as offset into the Intel PT buffer via:

    pt_pkt_get_offset()


You can get the position of the last synchronization point as offset into the
Intel PT buffer via:

    pt_pkt_get_sync_offset()


Once the decoder is synchronized, you can iterate over packets by repeated calls
to `pt_pkt_next()` as shown in the following example:

~~~{.c}
    struct pt_packet_decoder *decoder;
    int errcode;

    for (;;) {
        struct pt_packet packet;

        errcode = pt_pkt_next(decoder, &packet, sizeof(packet));
        if (errcode < 0)
            break;

        <process packet>(&packet);
    }
~~~


## The Event Layer

The event layer deals with packet combinations that encode higher-level events.
It is used for reconstructing execution flow for users who need finer-grain
control not available via the instruction flow layer or for users who want to
integrate execution flow reconstruction with other functionality more tightly
than it would be possible otherwise.

This section describes how to use the query decoder for reconstructing execution
flow.  See the instruction flow decoder as an example.  Start by configuring and
allocating a `pt_query_decoder` as shown below:

~~~{.c}
    struct pt_query_decoder *decoder;
    struct pt_config config;
    int errcode;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;
    config.decode.callback = <decode function>;
    config.decode.context = <decode context>;

    decoder = pt_qry_alloc_decoder(&config);
    if (!decoder)
        <handle error>(errcode);
~~~

An optional packet decode callback function may be specified in addition to the
mandatory config fields.  If specified, the callback function will be called for
packets the decoder does not know about.  The query decoder will ignore the
unknown packet except for its size in order to skip it.  If there is no decode
callback specified, the decoder will abort with `-pte_bad_opc`.  In addition to
the callback function pointer, an optional pointer to user-defined context
information can be specified.  This context will be passed to the decode
callback function.

Before the decoder can be used, it needs to be synchronized onto the Intel PT
packet stream.  To iterate over synchronization points in the Intel PT packet
stream in forward or backward direction, the query decoders offer the following
two synchronization functions respectively:

    pt_qry_sync_forward()
    pt_qry_sync_backward()


To manually synchronize the decoder at a synchronization point (i.e. PSB packet)
in the Intel PT packet stream, use the following function:

    pt_qry_sync_set()


After successfully synchronizing, the query decoder will start reading the PSB+
header to initialize its internal state.  If tracing is enabled at this
synchronization point, the IP of the instruction, at which decoding should be
started, is returned.  If tracing is disabled at this synchronization point, it
will be indicated in the returned status bits (see below).  In this example,
synchronization to the first synchronization point is shown:

~~~{.c}
    struct pt_query_decoder *decoder;
    uint64_t ip;
    int status;

    status = pt_qry_sync_forward(decoder, &ip);
    if (status < 0)
        <handle error>(status);
~~~

In addition to a query decoder, you will need an instruction decoder for
decoding and classifying instructions.


#### In A Nutshell

After synchronizing, you begin decoding instructions starting at the returned
IP.  As long as you can determine the next instruction in execution order, you
continue on your own.  Only when the next instruction cannot be determined by
examining the current instruction, you would ask the query decoder for guidance:

  * If the current instruction is a conditional branch, the
    `pt_qry_cond_branch()` function will tell whether it was taken.

  * If the current instruction is an indirect branch, the
    `pt_qry_indirect_branch()` function will provide the IP of its destination.


~~~{.c}
    struct pt_query_decoder *decoder;
    uint64_t ip;

    for (;;) {
        struct <instruction> insn;

        insn = <decode instruction>(ip);

        ip += <instruction size>(insn);

        if (<is cond branch>(insn)) {
            int status, taken;

            status = pt_qry_cond_branch(decoder, &taken);
            if (status < 0)
                <handle error>(status);

            if (taken)
                ip += <branch displacement>(insn);
        } else if (<is indirect branch>(insn)) {
            int status;

            status = pt_qry_indirect_branch(decoder, &ip);
            if (status < 0)
                <handle error>(status);
        }
    }
~~~


Certain aspects such as, for example, asynchronous events or synchronizing at a
location where tracing is disabled, have been ignored so far.  Let us consider
them now.


#### Queries

The query decoder provides four query functions:

  * `pt_qry_cond_branch()`      Query whether the next conditional branch was
                                taken.

  * `pt_qry_indirect_branch()`  Query for the destination IP of the next
                                indirect branch.

  * `pt_qry_event()`            Query for the next event.

  * `pt_qry_time()`             Query for the current time.


Each function returns either a positive vector of status bits or a negative
error code.  For details on status bits and error conditions, please refer to
the `pt_status_flag` and `pt_error_code` enumerations in the intel-pt.h header.

The `pts_ip_suppressed` status bit is used to indicate that no IP is available
at functions that are supposed to return an IP.  Examples are the indirect
branch query function and both synchronization functions.

The `pts_event_pending` status bit is used to indicate that there is an event
pending.  You should query for this event before continuing execution flow
reconstruction.

The `pts_eos` status bit is used to indicate the end of the trace.  Any
subsequent query will return -pte_eos.


#### Events

Events are signaled ahead of time.  When you query for pending events as soon as
they are indicated, you will be aware of asynchronous events before you reach
the instruction associated with the event.

For example, if tracing is disabled at the synchronization point, the IP will be
suppressed.  In this case, it is very likely that a tracing enabled event is
signaled.  You will also get events for initializing the decoder state after
synchronizing onto the Intel PT packet stream.  For example, paging or execution
mode events.

See the `enum pt_event_type` and `struct pt_event` in the intel-pt.h header for
details on possible events.  This document does not give an example of event
processing.  Refer to the implementation of the instruction flow decoder in
pt_insn.c for details.


#### Timing

To be able to signal events, the decoder reads ahead until it arrives at a query
relevant packet.  Errors encountered during that time will be postponed until
the respective query call.  This reading ahead affects timing.  The decoder will
always be a few packets ahead.  When querying for the current time, the query
will return the time at the decoder's current packet.  This corresponds to the
time at our next query.


#### Return Compression

If Intel PT has been configured to compress returns, a successfully compressed
return is represented as a conditional branch instead of an indirect branch.
For a RET instruction, you first query for a conditional branch.  If the query
succeeds, it should indicate that the branch was taken.  In that case, the
return has been compressed.  A not taken branch indicates an error.  If the
query fails, the return has not been compressed and you query for an indirect
branch.

There is no guarantee that returns will be compressed.  Even though return
compression has been enabled, returns may still be represented as indirect
branches.

To reconstruct the execution flow for compressed returns, you would maintain a
stack of return addresses.  For each call instruction, push the IP of the
instruction following the call onto the stack.  For compressed returns, pop the
topmost IP from the stack.  See pt_retstack.h and pt_retstack.c for a sample
implementation.


## The Instruction Flow Layer

The instruction flow layer provides a simple API for iterating over instructions
in execution order.  Start by configuring and allocating a `pt_insn_decoder` as
shown below:

~~~{.c}
    struct pt_insn_decoder *decoder;
    struct pt_config config;
    int errcode;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;
    config.decode.callback = <decode function>;
    config.decode.context = <decode context>;

    decoder = pt_insn_alloc_decoder(&config);
    if (!decoder)
        <handle error>(errcode);
~~~

An optional packet decode callback function may be specified in addition to the
mandatory config fields.  If specified, the callback function will be called for
packets the decoder does not know about.  The decoder will ignore the unknown
packet except for its size in order to skip it.  If there is no decode callback
specified, the decoder will abort with `-pte_bad_opc`.  In addition to the
callback function pointer, an optional pointer to user-defined context
information can be specified.  This context will be passed to the decode
callback function.

The image argument is optional.  If no image is given, the decoder will use an
empty default image that can be populated later on and that is implicitly
destroyed when the decoder is freed.  See below for more information on this.


#### The Traced Image

In addition to the Intel PT configuration, the instruction flow decoder needs to
know the memory image for which Intel PT has been recorded.  This memory image
is represented by a `pt_image` object.  If decoding failed due to an IP lying
outside of the traced memory image, `pt_insn_next()` will return `-pte_nomap`.

Use `pt_image_alloc()` to allocate and `pt_image_free()` to free an image.
Images may not be shared.  Every decoder must use a different image.  Use this
to prepare the image in advance or if you want to switch between images.

Every decoder provides an empty default image that is used if no image is
specified during allocation.  The default image is implicitly destroyed when the
decoder is freed.  It can be obtained by calling `pt_insn_get_image()`.  Use
this if you only use one decoder and one image.

An image is a collection of contiguous, non-overlapping memory regions called
`sections`.  Starting with an empty image, it may be populated with repeated
calls to `pt_image_add_file()` or `pt_image_add_cached()`, one for each section,
or with a call to `pt_image_copy()` to add all sections from another image.  If
a newly added section overlaps with an existing section, the existing section
will be truncated or split to make room for the new section.

In some cases, the memory image may change during the execution.  You can use
the `pt_image_remove_by_filename()` function to remove previously added sections
by their file name and `pt_image_remove_by_asid()` to remove all sections for an
address-space.

In addition to adding sections, you can register a callback function for reading
memory using `pt_image_set_callback()`.  The `context` parameter you pass
together with the callback function pointer will be passed to your callback
function every time it is called.  There can only be one callback at any time.
Adding a new callback will remove any previously added callback.  To remove the
callback function, pass `NULL` to `pt_image_set_callback()`.

Callback and files may be combined.  The callback function is used whenever
the memory cannot be found in any of the image's sections.

If more than one process is traced, the memory image may change when the process
context is switched.  To simplify handling this case, an address-space
identifier may be passed to each of the above functions to define separate
images for different processes at the same time.  The decoder will select the
correct image based on context switch information in the Intel PT trace.  If
you want to manage this on your own, you can use `pt_insn_set_image()` to
replace the image a decoder uses.


#### The Traced Image Section Cache

When using multiple decoders that work on related memory images it is desirable
to share image sections between decoders.  The underlying file sections will be
mapped only once per image section cache.

Use `pt_iscache_alloc()` to allocate and `pt_iscache_free()` to free an image
section cache.  Freeing the cache does not destroy sections added to the cache.
They remain valid until they are no longer used.

Use `pt_iscache_add_file()` to add a file section to an image section cache.
The function returns an image section identifier (ISID) that uniquely identifies
the section in this cache.  Use `pt_image_add_cached()` to add a file section
from an image section cache to an image.

Multiple image section caches may be used at the same time but it is recommended
not to mix sections from different image section caches in one image.

A traced image section cache can also be used for reading an instruction's
memory via its IP and ISID as provided in `struct pt_insn`.

The image section cache provides a cache of recently mapped sections and keeps
them mapped when they are unmapped by the images that used them.  This avoid
repeated unmapping and re-mapping of image sections in some parallel debug
scenarios or when reading memory from the image section cache.

Use `pt_iscache_set_limit()` to set the limit of this cache in bytes.  This
accounts for the extra memory that will be used for keeping image sections
mapped including any block caches associated with image sections.  To disable
caching, set the limit to zero.


#### Synchronizing

Before the decoder can be used, it needs to be synchronized onto the Intel PT
packet stream.  To iterate over synchronization points in the Intel PT packet
stream in forward or backward directions, the instruction flow decoders offer
the following two synchronization functions respectively:

    pt_insn_sync_forward()
    pt_insn_sync_backward()


To manually synchronize the decoder at a synchronization point (i.e. PSB packet)
in the Intel PT packet stream, use the following function:

    pt_insn_sync_set()


The example below shows synchronization to the first synchronization point:

~~~{.c}
    struct pt_insn_decoder *decoder;
    int errcode;

    errcode = pt_insn_sync_forward(decoder);
    if (errcode < 0)
        <handle error>(errcode);
~~~

The decoder will remember the last synchronization packet it decoded.
Subsequent calls to `pt_insn_sync_forward` and `pt_insn_sync_backward` will use
this as their starting point.

You can get the current decoder position as offset into the Intel PT buffer via:

    pt_insn_get_offset()


You can get the position of the last synchronization point as offset into the
Intel PT buffer via:

    pt_insn_get_sync_offset()


#### Iterating

Once the decoder is synchronized, you can iterate over instructions in execution
flow order by repeated calls to `pt_insn_next()` as shown in the following
example:

~~~{.c}
    struct pt_insn_decoder *decoder;
    int status;

    for (;;) {
        struct pt_insn insn;

        status = pt_insn_next(decoder, &insn, sizeof(insn));

        if (insn.iclass != ptic_unknown)
            <process instruction>(&insn);

        if (status < 0)
            break;

        ...
    }
~~~

Note that the example ignores non-error status returns.

For each instruction, you get its IP, its size in bytes, the raw memory, an
identifier for the image section that contained it, the current execution mode,
and the speculation state, that is whether the instruction has been executed
speculatively.  In addition, you get a coarse classification that can be used
for further processing without the need for a full instruction decode.

If a traced image section cache is used the image section identifier can be used
to trace an instruction back to the binary file that contained it.  This allows
mapping the instruction back to source code using the debug information
contained in or reachable via the binary file.

Beware that `pt_insn_next()` may indicate errors that occur after the returned
instruction.  The returned instruction is valid if its `iclass` field is set.


#### Events

The instruction flow decoder uses an event system similar to the query
decoder's.  Pending events are indicated by the `pts_event_pending` flag in the
status flag bit-vector returned from `pt_insn_sync_<where>()`, `pt_insn_next()`
and `pt_insn_event()`.

When the `pts_event_pending` flag is set on return from `pt_insn_next()`, use
repeated calls to `pt_insn_event()` to drain all queued events.  Then switch
back to calling `pt_insn_next()` to resume with instruction flow decode as
shown in the following example:

~~~{.c}
    struct pt_insn_decoder *decoder;
    int status;

    for (;;) {
        struct pt_insn insn;

        status = pt_insn_next(decoder, &insn, sizeof(insn));
        if (status < 0)
            break;

        <process instruction>(&insn);

        while (status & pts_event_pending) {
            struct pt_event event;

            status = pt_insn_event(decoder, &event, sizeof(event));
            if (status < 0)
                <handle error>(status);

            <process event>(&event);
        }
    }
~~~


#### The Instruction Flow Decode Loop

If we put all of the above examples together, we end up with a decode loop as
shown below:

~~~{.c}
    int handle_events(struct pt_insn_decoder *decoder, int status)
    {
        while (status & pts_event_pending) {
            struct pt_event event;

            status = pt_insn_event(decoder, &event, sizeof(event));
            if (status < 0)
                break;

            <process event>(&event);
        }

        return status;
    }

    int decode(struct pt_insn_decoder *decoder)
    {
        int status;

        for (;;) {
            status = pt_insn_sync_forward(decoder);
            if (status < 0)
                break;

            for (;;) {
                struct pt_insn insn;

                status = handle_events(decoder, status);
                if (status < 0)
                    break;

                status = pt_insn_next(decoder, &insn, sizeof(insn));

                if (insn.iclass != ptic_unknown)
                    <process instruction>(&insn);

                if (status < 0)
                    break;
            }

            <handle error>(status);
        }

        <handle error>(status);

        return status;
    }
~~~


## The Block Layer

The block layer provides a simple API for iterating over blocks of sequential
instructions in execution order.  The instructions in a block are sequential in
the sense that no trace is required for reconstructing the instructions.  The IP
of the first instruction is given in `struct pt_block` and the IP of other
instructions in the block can be determined by decoding and examining the
previous instruction.

Start by configuring and allocating a `pt_block_decoder` as shown below:

~~~{.c}
    struct pt_block_decoder *decoder;
    struct pt_config config;

    memset(&config, 0, sizeof(config));
    config.size = sizeof(config);
    config.begin = <pt buffer begin>;
    config.end = <pt buffer end>;
    config.cpu = <cpu identifier>;
    config.decode.callback = <decode function>;
    config.decode.context = <decode context>;

    decoder = pt_blk_alloc_decoder(&config);
~~~

An optional packet decode callback function may be specified in addition to the
mandatory config fields.  If specified, the callback function will be called for
packets the decoder does not know about.  The decoder will ignore the unknown
packet except for its size in order to skip it.  If there is no decode callback
specified, the decoder will abort with `-pte_bad_opc`.  In addition to the
callback function pointer, an optional pointer to user-defined context
information can be specified.  This context will be passed to the decode
callback function.


#### Synchronizing

Before the decoder can be used, it needs to be synchronized onto the Intel PT
packet stream.  To iterate over synchronization points in the Intel PT packet
stream in forward or backward directions, the block decoder offers the following
two synchronization functions respectively:

    pt_blk_sync_forward()
    pt_blk_sync_backward()


To manually synchronize the decoder at a synchronization point (i.e. PSB packet)
in the Intel PT packet stream, use the following function:

    pt_blk_sync_set()


The example below shows synchronization to the first synchronization point:

~~~{.c}
    struct pt_block_decoder *decoder;
    int errcode;

    errcode = pt_blk_sync_forward(decoder);
    if (errcode < 0)
        <handle error>(errcode);
~~~

The decoder will remember the last synchronization packet it decoded.
Subsequent calls to `pt_blk_sync_forward` and `pt_blk_sync_backward` will use
this as their starting point.

You can get the current decoder position as offset into the Intel PT buffer via:

    pt_blk_get_offset()


You can get the position of the last synchronization point as offset into the
Intel PT buffer via:

    pt_blk_get_sync_offset()


#### Iterating

Once the decoder is synchronized, it can be used to iterate over blocks of
instructions in execution flow order by repeated calls to `pt_blk_next()` as
shown in the following example:

~~~{.c}
    struct pt_block_decoder *decoder;
    int status;

    for (;;) {
        struct pt_block block;

        status = pt_blk_next(decoder, &block, sizeof(block));

        if (block.ninsn > 0)
            <process block>(&block);

        if (status < 0)
            break;

        ...
    }
~~~

Note that the example ignores non-error status returns.

A block contains enough information to reconstruct the instructions.  See
`struct pt_block` in `intel-pt.h` for details.  Note that errors returned by
`pt_blk_next()` apply after the last instruction in the provided block.

It is recommended to use a traced image section cache so the image section
identifier contained in a block can be used for reading the memory containing
the instructions in the block.  This also allows mapping the instructions back
to source code using the debug information contained in or reachable via the
binary file.

In some cases, the last instruction in a block may cross image section
boundaries.  This can happen when a code segment is split into more than one
image section.  The block is marked truncated in this case and provides the raw
bytes of the last instruction.

The following example shows how instructions can be reconstructed from a block:

~~~{.c}
    struct pt_image_section_cache *iscache;
    struct pt_block *block;
    uint16_t ninsn;
    uint64_t ip;

    ip = block->ip;
    for (ninsn = 0; ninsn < block->ninsn; ++ninsn) {
        uint8_t raw[pt_max_insn_size];
        <struct insn> insn;
        int size;

        if (block->truncated && ((ninsn +1) == block->ninsn)) {
            memcpy(raw, block->raw, block->size);
            size = block->size;
        } else {
            size = pt_iscache_read(iscache, raw, sizeof(raw), block->isid, ip);
            if (size < 0)
                break;
        }

        errcode = <decode instruction>(&insn, raw, size, block->mode);
        if (errcode < 0)
            break;

        <process instruction>(&insn);

        ip = <determine next ip>(&insn);
    }
~~~


#### Events

The block decoder uses an event system similar to the query decoder's.  Pending
events are indicated by the `pts_event_pending` flag in the status flag
bit-vector returned from `pt_blk_sync_<where>()`, `pt_blk_next()` and
`pt_blk_event()`.

When the `pts_event_pending` flag is set on return from `pt_blk_sync_<where>()`
or `pt_blk_next()`, use repeated calls to `pt_blk_event()` to drain all queued
events.  Then switch back to calling `pt_blk_next()` to resume with block decode
as shown in the following example:

~~~{.c}
    struct pt_block_decoder *decoder;
    int status;

    for (;;) {
        struct pt_block block;

        status = pt_blk_next(decoder, &block, sizeof(block));
        if (status < 0)
            break;

        <process block>(&block);

        while (status & pts_event_pending) {
            struct pt_event event;

            status = pt_blk_event(decoder, &event, sizeof(event));
            if (status < 0)
                <handle error>(status);

            <process event>(&event);
        }
    }
~~~


#### The Block Decode Loop

If we put all of the above examples together, we end up with a decode loop as
shown below:

~~~{.c}
    int process_block(struct pt_block *block,
                      struct pt_image_section_cache *iscache)
    {
        uint16_t ninsn;
        uint64_t ip;

        ip = block->ip;
        for (ninsn = 0; ninsn < block->ninsn; ++ninsn) {
            struct pt_insn insn;

            memset(&insn, 0, sizeof(insn));
            insn->speculative = block->speculative;
            insn->isid = block->isid;
            insn->mode = block->mode;
            insn->ip = ip;

            if (block->truncated && ((ninsn +1) == block->ninsn)) {
                insn.truncated = 1;
                insn.size = block->size;

                memcpy(insn.raw, block->raw, insn.size);
            } else {
                int size;

                size = pt_iscache_read(iscache, insn.raw, sizeof(insn.raw),
                                       insn.isid, insn.ip);
                if (size < 0)
                    return size;

                insn.size = (uint8_t) size;
            }

            <decode instruction>(&insn);
            <process instruction>(&insn);

            ip = <determine next ip>(&insn);
        }

        return 0;
    }

    int handle_events(struct pt_blk_decoder *decoder, int status)
    {
        while (status & pts_event_pending) {
            struct pt_event event;

            status = pt_blk_event(decoder, &event, sizeof(event));
            if (status < 0)
                break;

            <process event>(&event);
        }

        return status;
    }

    int decode(struct pt_blk_decoder *decoder,
               struct pt_image_section_cache *iscache)
    {
        int status;

        for (;;) {
            status = pt_blk_sync_forward(decoder);
            if (status < 0)
                break;

            for (;;) {
                struct pt_block block;
                int errcode;

                status = handle_events(decoder, status);
                if (status < 0)
                    break;

                status = pt_blk_next(decoder, &block, sizeof(block));

                errcode = process_block(&block, iscache);
                if (errcode < 0)
                    status = errcode;

                if (status < 0)
                    break;
            }

            <handle error>(status);
        }

        <handle error>(status);

        return status;
    }
~~~


## Parallel Decode

Intel PT splits naturally into self-contained PSB segments that can be decoded
independently.  Use the packet or query decoder to search for PSB's using
repeated calls to `pt_pkt_sync_forward()` and `pt_pkt_get_sync_offset()` (or
`pt_qry_sync_forward()` and `pt_qry_get_sync_offset()`).  The following example
shows this using the query decoder, which will already give the IP needed in
the next step.

~~~{.c}
    struct pt_query_decoder *decoder;
    uint64_t offset, ip;
    int status, errcode;

    for (;;) {
        status = pt_qry_sync_forward(decoder, &ip);
        if (status < 0)
            break;

        errcode = pt_qry_get_sync_offset(decoder, &offset);
        if (errcode < 0)
            <handle error>(errcode);

        <split trace>(offset, ip, status);
    }
~~~

The individual trace segments can then be decoded using the query, instruction
flow, or block decoder as shown above in the previous examples.

When stitching decoded trace segments together, a sequence of linear (in the
sense that it can be decoded without Intel PT) code has to be filled in.  Use
the `pts_eos` status indication to stop decoding early enough.  Then proceed
until the IP at the start of the succeeding trace segment is reached.  When
using the instruction flow decoder, `pt_insn_next()` may be used for that as
shown in the following example:

~~~{.c}
    struct pt_insn_decoder *decoder;
    struct pt_insn insn;
    int status;

    for (;;) {
        status = pt_insn_next(decoder, &insn, sizeof(insn));
        if (status < 0)
            <handle error>(status);

        if (status & pts_eos)
            break;

        <process instruction>(&insn);
    }

    while (insn.ip != <next segment's start IP>) {
        <process instruction>(&insn);

        status = pt_insn_next(decoder, &insn, sizeof(insn));
        if (status < 0)
            <handle error>(status);
    }
~~~


## Threading

The decoder library API is *partly* thread-safe.  Specifically:

 * It's OK for different threads to allocate and use *different* decoder
   objects at the same time, but no single decoder object can be used
   concurrently by different threads.  A decoder allocated in one thread may be
   passed to, and used from, another thread, as long as the decoder is not used
   concurrently.

 * Different decoder objects must not use the same image object, even if the
   decoders are owned by the same thread.  You can, however, use
   `pt_image_copy()` to give each decoder its own copy of a shared master
   image.  Copying an image is lightweight and does not copy large amounts of
   data.

 * Image section cache objects *can* be shared across threads and be used
   concurrently, but only if you build libipt with `FEATURE_THREADS=ON`.