Tracing Framework Guide Part 1

•

Introduction

Hello again! Some time ago I wrote a post about the Tracing Framework in SPDK. While that post provided a lot of general information, this time we’re going to go through how to add traces to SPDK in detail, with examples. We’ll also look at how to display them using spdk_trace application.

A little reminder

As I mentioned before (in the previous post), we use two types of traces in SPDK: our own SPDK tracepoints and BPF traces. The former are usually used in performance sensitive places such as the I/O path where we want as little overhead as possible. The latter we use where there are not time limitations and we can afford higher overhead due to software interrupts.

Defining SPDK Tracepoints

I will now start going over what we need to do to properly define and use a tracepoint in SPDK. We are going to look at SPDK tracepoints this time. BPF traces are going to be explained in the future post.

Definition

The tracepoint definitions are all kept in one place: in include/spdk_internal/trace_defs.h file. Here we can add definitions of new tracepoint elements. When defining a new value, we need to make sure that the name and its value are unique.

  • groups - These values are a really important set of values in SPDK tracing. They provide an organization mechanism for all of the tracepoints. Grouping tracepoints allows the SPDK to enable them in sets, instead of either enabling all of them at once. To get the list of all available groups that can be enabled on SPDK startup, we can simply build and run SPDK target application (e.g. build/bin/spdk_tgt) with -h parameter. If we were to choose thread group for example, we would enable TRACE_THREAD_IOCH_GET and TRACE_THREAD_IOCH_PUT tracepoints from section /* Thread tracepoint definitions */. Special value all enables all tracepoints.

  • objects - These names aggregate tracepoints during SPDK runtime. In include/spdk_internal/trace_defs.h we may define new types of objects to differentiate them from others. A single tracepoint object may be referenced during multiple events recorded by one or more tracepoints. We can treat them as an instance of a data structure. An example here: OBJECT_NVMF_TCP_IO inside lib/nvmf/tcp.c. This one represents IO data structure struct spdk_nvmf_tcp_req and is tied to the TCP requests inside SPDK. When we record an event for an NVMF_TCP_IO object, it will include the pointer to the spdk_nvmf_tcp_req. The trace object creation is indicated by TCP_REQ_NEW, which tells SPDK to spawn a new instance of this object (this is the only tracepoint with “1” as 5th argument; more about how this works later) and then it will be referenced many times by multiple TCP tracepoints until it finally gets destroyed during request completion event, which is marked by TCP_REQ_COMPLETED tracepoint.

  • tracepoints - each group of tracepoints has its own definition section, like /* NVMe-of TCP tracepoint definitions */. They are the most descriptive values, usually tied to a certain event or function call.

  • owners - These are mainly for associating objects with a higher order construct. For example: OBJECT_NVME_PCIE_REQ has owner OWNER_NVME_PCIE_QP, indicating on which qpair (QP) the I/O was submitted on. In this case the OWNER_NVME_PCIE_QP identifier is its NVMe queue ID. There is also special value OWNER_NONE, which is used in cases where no owner association is required (i.e. owner value must always be specified, so OWNER_NONE could be read as an “empty” value).

The way we add new definitions in include/spdk_internal/trace_defs.h is by writing #define TYPE_NAME 0xX, where 0xX is the next available ID in hexadecimal format. It is better if the IDs are in increasing order and follow already defined pattern for a section.

Now example: We want to add a new tracepoint somewhere in NVMe-of TCP related code. If the owner was not defined for our set of tracepoints, we could add it to thetrace_defs.h file:

/* Owner definitions */
...
#define OWNER_NVMF_TCP       0x30

Keep in mind that sometimes Owner for the tracepoint you want to implement is already there (like in our case), so there is no need to add a new one. Object and Trace group can be added the same way as Owner. New definitions should be placed at the bottom of a definitions section and assigned new, highest available value, following the pattern of a given section. Now we can scroll down to /* NVMe-of TCP tracepoint definitions */ section and add a new tracepoint here:

/* NVMe-of TCP tracepoint definitions */
...
#define TRACE_TCP_TEST_TPOINT       SPDK_TPOINT_ID(TRACE_GROUP_NVMF_TCP, 0x18)

SPDK_TPOINT_ID is a macro for assigning a unique ID to each tpoint, using Trace group’s own ID. The formal tracepoint registration to a Trace group happens later on.

And this is all when it comes to defining the SPDK tracepoint. In the next section I will explain the process of declaring the contents of a trace.

Declaration

We declare SPDK tracepoints in .c files where we want to record some values, using the SPDK_TRACE_REGISTER_FN macro, which defines a constructor function to register tracepoint registering functions during SPDK application start time. It also checks that trace groups’ names do not conflict with each other. The macro also lets us declare multiple tracepoints in one place, assigning them the same Owner, Object (it also formally registers both of them) and Trace group. We also need to include some header files:

#include "spdk/trace.h"
#include "spdk_internal/trace_defs.h"

There are two ways we usually go about declaring a trace:

1) For traces watching a single value:

SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP)
{
spdk_trace_register_owner(OWNER_NVMF_TCP, 'r');
spdk_trace_register_object(OBJECT_NVMF_TCP, 'r');
spdk_trace_register_description("TCP_TEST_TPOINT",
                                 TRACE_TCP_TEST_TPOINT,
                                 OWNER_NVMF_TCP, OWNER_NVMF_TCP, 1,
                                 SPDK_TRACE_ARG_TYPE_PTR, "arg1");
}

2) Where we need multiple values recorded in a tracepoint:

SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP)
{
        spdk_trace_register_owner(OWNER_NVMF_TCP, 'r');
        spdk_trace_register_object(OBJECT_NVMF_TCP, 'r');

        struct spdk_trace_tpoint_opts opts[] = {
                {
                        "TCP_TEST_TPOINT", TRACE_TCP_TEST_TPOINT,
                        OWNER_NVMF_TCP, OWNER_NVMF_TCP, 1,
                        {
                                { "arg1", SPDK_TRACE_ARG_TYPE_INT, 8 },
                                { "arg2", SPDK_TRACE_ARG_TYPE_PTR, 8 },
                                { "arg3", SPDK_TRACE_ARG_TYPE_INT, 8 },
                                { "arg4", SPDK_TRACE_ARG_TYPE_STR, 8 }
                        }
                }
        };
}

Owner registraction function call arguments should include a value matching one of the entries from /* Owner definitions */ section inside include/spdk_internal/trace_defs.h and a unique to all owners letter. Object registration should follow similar rules.

When declaring an actual tracepoint, names like arg1, arg2, should be replaced with more descriptive ones: offset, ctx or anything that a value represents - these names get emitted by spdk_trace with the corresponding value. SPDK tracepoints can take three types of arguments: integers (SPDK_TRACE_ARG_TYPE_INT), strings (SPDK_TRACE_ARG_TYPE_STR) and pointers (SPDK_TRACE_ARG_TYPE_PTR) followed by size in bytes. The difference between integers and pointers here is that an argument SPDK_TRACE_ARG_TYPE_INT will be an integer displayed in decimal format, while SPDK_TRACE_ARG_TYPE_PTR will show as a value in hexadecimal form. Letters in spdk_trace_register_owner and spdk_trace_register_object mean ID prefix, so the letter that is going to appear before the ID. A single 1 or 0 as 5th argument of spdk_trace_register_description decides whether to create a new object or not (some traces are parts of instances of a given object). For example: BDEV_IO_START uses 1, because this event marks the beginning of life of a new bdev_io object inside SPDK. When later SPDK application detects BDEV_IO_DONE event bound to the same IO, it will calculate time delta between BDEV_IO_START and BDEV_IO_DONE and display it inside a trace. Let’s look at a partial output from spdk_trace application:

...
 0: 141193.578 b00 BDEV_IO_START                         id:    i429      type:  2                ctx:   0x7ffb7d48b110   offset:12048            len:   2
...
 0: 143030.609 b00 BDEV_IO_DONE                          id:    i429      time:  1837.035         ctx:   0x7ffb7d48b110
...

In case of these two tracepoints we can see from left to right:

1) timestamp - this is the time measured from the start of SPDK application in microseconds 2) owner ID inside SPDK runtime. This value will be different from what we see in include/spdk_internal/trace_defs.h - the first character is the one passed as the second argument of spdk_trace_register_owner() function mentioned earlier, followed by poller ID from which the registration has been called. 3) tracepoint name 4) ID of the object instance from which the event was recorded 5) time, if the recorded event ends some longer process (in this case bdev IO). This value is a difference in application time between two events: 143030.609 - 141193.578 = (about) 1837.035 6) arguments provided in the tracepoint definition (in case of BDEV_IO_START: type, ctx, offset and len; in case of BDEV_IO_DONE only ctx; ctx is a pointer to the instance of tracepoint object)

about - the real value should be equal to 1837.031, but the rounding of decimal fractions was not perfect this time.

There is also a neat feature called relations. We can specify a relation between a tracepoint and a certain object, to which other tracepoints belong. Relations are defined inside SPDK_TRACE_REGISTER_FN macro by calling spdk_trace_tpoint_register_relation. Currently SPDK has only two relations defined, so lets look at them as an example. They can be found inside lib/nvmf/tcp.c file.

        spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1);
        spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0);

The first relation binds trace event called TRACE_BDEV_IO_START, which is defined in bdev library lib/bdev/bdev.c, to the new object OBJECT_NVMF_TCP_IO registered inside TCP library. We can do that by taking unique ID of a tracepoint and ID of registered object and matching them later during SPDK runtime. The last argument indicates argument index of the related object. It’s the position of the tracepoint context inside a trace entry in Trace Parser library. To understand what that means, we should look inside lib/bdev/bdev.c file and jump to the SPDK_TRACE_REGISTER_FN macro. Structure spdk_trace_tpoint_opts defines tracepoints inside bdev library and both BDEV_IO_START, and BDEV_IO_DONE are the first two definitions here. The third argument in the spdk_trace_tpoint_register_relation definitions matches the position of ctx parameter in tracepoint definitions - this is a pointer to the related object OBJECT_NVMF_TCP_IO.

Second relation works similarly, but it binds TRACE_BDEV_IO_DONE to the same object OBJECT_NVMF_TCP_IO, so it can be used later to do the time calculations.

The partial output from spdk_trace would look like this:

 0: 141184.297 t02 TCP_REQ_NEW                           id:    r520             qpair: 0x618000004080
 0: 141185.109 t02 TCP_REQ_NEED_BUFFER                   id:    r520             time:  0.810            qpair: 0x618000004080

 0: 141186.500 t02 TCP_REQ_TX_H_TO_C                     id:    r520             time:  2.199            qpair: 0x618000004080

 0: 141189.609 t02 TCP_REQ_RDY_TO_EXECUTE                id:    r520             time:  5.321            qpair: 0x618000004080
 0: 141193.578 b00 BDEV_IO_START                         id:    i429 (r520)      type:  2                ctx:   0x7ffb7d48b110   offset:12048            len:   2

 0: 143030.609 b00 BDEV_IO_DONE                          id:    i429 (r520)      time:  1837.035         ctx:   0x7ffb7d48b110
 0: 143032.734 t02 TCP_REQ_EXECUTED                      id:    r520             time:  1848.446         qpair: 0x618000004080
 0: 143033.031 t02 TCP_REQ_RDY_TO_COMPLETE               id:    r520             time:  1848.728         qpair: 0x618000004080
 0: 143034.344 t02 TCP_REQ_TRANSFER_C2H                  id:    r520             time:  1850.043         qpair: 0x618000004080

 0: 143606.969 t02 TCP_REQ_COMPLETED                     id:    r520             time:  2422.666         qpair: 0x618000004080

This time we can see that traces describing bdev IO start and finish contain additional value in parentheses, right after the object IO. In this case this is the ID of TCP request tracepoint object that was initialized during TCP request. Another object was later created for bdev IO trace events and SPDK bound them together with trace relation mechanism.

Usage of spdk_trace to record and display the traces will be explained later.

Recording a trace

Now that we have our tracepoint defined and declared, it is time to finally make use of it in our code. In the same file as we declared our tracepoint, we may call spdk_trace_record like this:

        spdk_trace_record(TRACE_TCP_TEST_TPOINT, 0, 0,
                         (uintptr_t)object_pointer, arg1);

where the arguments are:

1) tracepoint ID (the one we assigned in trace_defs.h) 2) poller ID 3) size 4) pointer to trace object instance 5) first argument

Each time the code execution hits the spdk_trace_record it will record a value of a given argument. Now let’s proceed to the next section, where we’ll be testing our change.

Getting the recorded data

Our code is now ready, we may configure and build the SPDK project. Note that I gave an example with TCP library, so we’ll be configuring SPDK with ./configure without any arguments, as TCP support is enabled by default. After that build the project and launch SPDK application (for example build/bin/nvmf_tgt) with a flag to enable the tracepoint: -e nvmf_tcp:<new tracepoint ID> or just -e nvmf_tcp, if you prefer to see all traces from TCP group.

After that run build/bin/spdk_trace with flags provided by SPDK on startup. Look for a lines that looks similar to these:

[2022-06-13 14:52:04.427428] app.c: 450:app_setup_trace: *NOTICE*: Use 'spdk_trace -s spdk_tgt -p 599051' to capture a snapshot of events at runtime.
[2022-06-13 14:52:04.427446] app.c: 455:app_setup_trace: *NOTICE*: Or copy /dev/shm/spdk_tgt_trace.pid599051 for offline analysis/debug.

If the code hit your tracepoint, you should see it printed here along with the values’ of arguments that you provided.

Final words

Summary

Traces are an important part of SPDK and a way of gathering data during application runtime. They are neatly organized into SPDK traces, which were explained in this blogpost, and bpf probes, which are going to be a topic of a future post.

Questions?

If you have any questions related to this topic, or any other in SPDK do not hesitate to ask on our Slack channels or on our GitHub! You may also prepare your own patches with traces in SPDK, if you feel like we should have them in a certain place.

Copyright © SPDK a Series of LF Projects, LLC. For web site terms of use, trademark policy, and other project policies please see https://lfprojects.org.