Kutrace User Guide
Kutrace User Guide
Kutrace User Guide
kutrace_mb=20
specifies the number of MB of kernel memory to reserve for the trace buffer. The default is 20MB.
kutrace_nocheck=0
specifies whether to check for the caller of kutrace_control having the proper privilege. The default is
0. Setting this to 1 or any other nonzero value disables the checking, which might be appropriate for
a standalone machine used for a class. The required privilege is PRIV_KMEM_READ.
Packet filtering, as described below for Linux, is not yet implemented for FreeBSD.
The command
$ sudo kldload kutrace_mod.ko
will load and initialize the loadable module, and kldunload removes it.
tracemb=20
specifies the number of MB of kernel memory to reserve for the trace buffer. The default is 2MB.
The module implements an optional network packet-tracing facility. Packets are filtered in a simple way,
with those passing the filter adding one KUtrace entry each at the time that the kernel TCP or UDP code
transfers the packet to/from a NIC. The filtering is done via a hash of the first 24 payload bytes of a
packet, i.e. those immediately after the packet headers. These are the data bytes that are visible to user-
mode software sending network messages, so if user code records similar entries it becomes possible to
see the time delay between user code and NIC on both ends of a transmission.
pktmask=0x0000000f
specifies which bytes of a network packet to hash for a possible tracing match. The low 24 bits
correspond to the first 24 payload bytes and the high 8 bits are ignored. A one-bit uses that byte
while a zero-bit sets it to zero. The hash is a four-byte-wide simple XOR. In this example, the first
four data bytes are used and the other 20 zeroed.
pktmatch=0xd1c517e5
specifies the network packet hash value to match; matched packets get KUtrace entries. In this
example, a packet matches if its first four bytes are exactly 0xd1c517e5. This idea is to be able to
pick off packets that are the beginnings of possibly multi-packet messages, based on some known bit
pattern (signature) at the front of a message. The default parameters here match the RPC library used
in the book.
The command
$ sudo insmod kutrace_mod.ko tracemb=2 pktmask=0x0000000f pktmatch=0xd1c517e5
will initialize the loadable module packet filter with these parameters, which are also the defaults.
Because every single inbound and outbound packet has the filtering calculation done when tracing is on,
the calculation is kept very simple and fast, about 10 nsec per packet on a 3.9 GHz Intel i3 chip. For
messages that are typically many packets, matching on just the first one is usually sufficient to
understand networking delays. However, a mask of zero and a match of zero will trace all packets, and a
mask of zero and match of non-zero will trace no packets.
Changing these parameters requires removing (rmmod) and then re-inserting the loadable module.
Region 1, Controls
The Controls region has a number of buttons and text boxes to control what is displayed in the main
timeline region 3. By default, the timeline region shows execution timespans (termed spans hereafter)
for the CPU group and also shows non-execution (waiting) for the Process and RPC groups. Several
optional overlays can accompany the main timelines, specified by the Annot: and Option: controls.
The Search: controls specify text annotations for spans with matching names. When a KUtrace
HTML file is first loaded, the far right of the Controls area shows the build version and date of the
Linux kernel software that created the trace, the processor model, and the size in pixels of the browser
window. The size disappears after 5 seconds -- its purpose is to allow you to create consistent displays
for screen shots. The kernel version and processor model disappear as soon as you pan or zoom.
In addition to the obvious buttons, all the bright blue text items in the Controls and other regions are also
control buttons.
The Annot: group controls adding short text annotations for some spans. By default, the text for these
has accompanying vertical lines, solid below a span and dotted above. This allows you to quickly see
which row matched and see the time alignment of spans on different rows. An alternate style of
annotation just has the name and a short line below a span. Clicking the blue Annot: button toggles to
lower-case annot: and toggles the annotation style. Span names longer than 9 characters are elided,
using just the first 7 characters, tilde, and last character. System calls have the low 16 bits of the first
argument in parentheses followed by an equal sign and 16 bits of the return value.
The User button annotates the first occurrence of each user-mode process ID that is on-screen in
Region 3. This can give you a quick view of what programs and threads are in the displayed portion of a
trace.
The All button annotates every span on the screen, again using short text. It is most useful when
zoomed in.
The Mark button cycles though four states: showing all manually-inserted mark_a/b/c/d trace entries,
showing just the text mark_a/b/c ones, showing just the numeric mark_d ones, and showing none. The
button is grayed out if there are no marks in the trace.
The Arc button cycle through three states: showing wakeup arcs between threads, showing them
bolder, and not showing them. To avoid clutter, short wakeups are only visible when you zoom in to just
a few milliseconds across the display.
The Lock button cycles through three states: showing software locks being acquired or held, showing
them with the lock-held lines bolder, and not showing them. Locks are shown above associated rows in
the PID and RPC groups; locks are not associated with CPUs, so are not shown in that group. To avoid
clutter, locks are only visible when there is enough room vertically and they occupy at least a few
horizontal pixels across the display. The button is grayed out if there are no software lock events in the
trace.
The Freq button cycles through three states: showing a CPU clock-frequency overlay, showing it
bolder, or not showing it. The overlay shows very light green for clock frequencies that are in the top
12% of the frequency range recorded in the trace, dark red for the bottom 12%, and successively darker
yellows for the middle 75%. Because a slow CPU clock can sometimes completely explain slow
performance or is otherwise correlated with execution anomalies, this display is on by default. On some
machines (x86), the individual frequency of each CPU core is sampled at timer interrupts, while on
others (Raspberry Pi), the single frequency for all cores is recorded just when it changes. To avoid
clutter, frequency spans are only visible when they occupy a few pixels across the display. The button is
grayed out if there are no frequency events in the trace or if they all specify the same frequency.
The IPC button cycles through four states: showing instructions-per-cycle speedometer triangles for all
kernel- and user-mode execution spans, showing them just for kernel-mode, just for user-mode, and
showing none. IPC triangles are shown on top of each execution span in all three CPU, PID and RPC
groups. To avoid clutter, IPC triangles are only visible when they occupy at least a few horizontal pixels
across the display. The redundant IPC 0 is not shown for idle spans. When IPC triangles are selected, the
IPC legend is shown in Region 4.
The Samp button toggles on and off the showing of PC samples. Samples are shown as slightly-rising
lines above associated rows in all three CPU, PID and RPC groups. To avoid clutter, samples are only
visible when there is enough room vertically and they occupy at least a few horizontal pixels across the
display. Also to avoid clutter, PC samples of the idle loop are suppressed. PC samples are taken at every
timer interrupt. Kernel-mode samples are shown as thick dashed lines and user-mode as thinner dashed
lines. The little "v" mark at the right end of a sample line shows the time it was taken. The slight slope
of the overlay line is suggestive of the decreasing probability that the sampled code was being executed
earlier. As explained in the book, the samptoname_k and samptoname_u postprocessing programs
can be used to turn hex PC addresses into routine names in kernel- and user-code respectively.
The CB button toggles a change in timeline span colors, as a small aid to color-blind users. It simply
rotates the red/green/blue values of each color. This can sometimes be helpful is distinguishing the
spans.
The Search: group of controls select on-screen spans (those starting at or after the tic mark at the left
end of the x-axis and before the tic mark at the right end) to annotate. The text box accepts a string or
JavaScript regular expression to match, and the [!] button negates the match, much like "-v" for grep.
The other two text boxes specify minimum and maximum durations for matching spans. By default,
these are microseconds, but the blue µsec button can be clicked to rotate among nsec, µsec, and msec
values.
Each search applies short-text annotations to each matching span and then displays on the far upper right
the match count, the total duration of those matches, and the minimum and maximum matching span
durations. To avoid clutter, only one annotation is shown per horizontal pixel across the display. All of
the matches are counted, though.
The search facility is quite powerful. Searching for the name of a process shows all the on-screen places
where it is running. Searching for the name of a system call shows all of those. Searching for "ipi"
shows all interprocessor interrupts: sending one and also the resulting reschedule_ipi interrupt.
Searching for "futex" shows system calls that block on software locks and the points at which they
resume. Searching for "wait_cpu" shows all the waiting-for-CPU spans in PID and RPC groups.
Searching for "mwait" (x86) or "wfi" (Raspberry Pi) finds all the events that trigger low-power idle.
Searching for "=1" can find system calls that return one. Searching for "PC=00556f" finds PC samples
with high bits in that address range. Note the capitalization.
Searching for "rx|tx" (no spaces) shows all the network receive or transmit events on-screen, if packet
tracing was specified. Searching for "write\(3" with the default Annot: style (not annot:) finds all
the writes to stderr. Note the use of backslash escapes and the dependency on the annotation style to get
parentheses at all.
Point events in the trace (as opposed to spans that always have non-zero duration) have an artificial
duration of 10 nsec. You can find these by searching for empty search text and a maximum duration of
10 nsec.
A few synthetic-event names are spelled with a leading hyphen: "-c-exit-", "-freq-", "-idle-", "-idlelp-",
"-sched-", and "-wakeup-". To avoid clutter, these only match in searches when the leading hyphen is
part of the search string.
To the right of the search group, the operating system version and CPU model name are shown at startup
if these items are recorded in a trace.
The July 2022 update adds the ability to search for a time in seconds and the display will change to be
centered on that time. To avoid unintended triggering, the time must have at least one fraction digit. The
width of the resulting displayed timespan is a function of the number of fraction digits. It can be
adjusted by appending one or more plus or minus signs on the right. For example, searching for 50.085+
in the supplied hello_world_trace.html will center the screen on the exit_group call.
As a convenience, searching for @@ will calculate a time search string that encompasses the current
display, zooming out a little to quantize the width.
Whenever a <CR> is typed in the search box, the search string is put on the front of a list of recent
searches. Subsequent search typing will show the most recent strings that contain the typed characters.
Click on one of those to search for it.
Searching for something that does not exist, such as zzzz, and then clicking the [!] button matches
everything currently on screen. The resulting match count and duration information can be useful for
summarizing the visible activity.
Region 2, Y-axis
The y-axis has up to three groups of rows and a label for each row. Each group header line has an
expand/contract triangle, the group name, and the number of rows in that group.
The PID group labels are process-name dot PID number, sorted by PID number. The idle process, PID
number zero, is not shown. Some process names have plus signs, showing that the name changed during
execution, for example bash+time_getpid.11046 indicates that PID 11046 was first named
bash but then changed (via execve for example) to time_getpid. Linux PID numbers less than
1024 are usually kernel threads. In reality, the PID numbers are TID numbers -- Linux thread numbers
or FreeBSD thread numbers with the leading 100000 removed. The name comes from the Linux kernel's
misnamed "pid" variables.
The RPC row labels have method-name dot RPCID number, sorted by start time. Thus, reading top to
bottom, RPCs are shown in order of arrival.
When there are too many rows to show all the names readably, some names are omitted, but the tic mark
for each row is always shown, so you can tell if any names are missing.
Clicking the leading triangle in a group name initially toggles between two states: expanded and
collapsed. The active click area for this action is the leftmost 1/4 of Region 2. Mouse click-drag and
wheel turns in the rightmost 3/4 allow vertical pan and zoom. Clicking and dragging vertically pans
rows, and using the mouse wheel zooms the rows.
Shift-click on a row name highlights the row by making the text bold, and highlights all the spans in
that row by graying out all other spans. These span highlights carry across the CPU/PID/RPC groups,
but do not cause rows in other groups to be highlighted. You can manually highlight these other-group
rows if desired. Shift-clicking multiple rows ORs all the highlighted spans. When any row in a group is
highlighted, clicking the leading triangle in a group name will rotate among three states instead of two:
expanded, expanded only for highlighted rows, and collapsed.
As a convenient shortcut, shift-right-click on a row name toggles all the rows with the same prefix, up
to the first non-letter/digit/underscore/hyphen. This makes it easy to highlight multiple related threads or
multiple related RPCs.
Only highlighted items are considered on-screen for User, All, and Search annotations. This can be
useful for careful filtering of matches by specific CPU, PID, or RPC.
Region 3, Timelines
There is a lot going on with the timelines. The initial display has all the kernel- user- and idle-mode
execution spans on all CPUs for the entire trace start-to-finish time. Clicking the red dot where the axes
meet reverts to this view at any time.
To avoid clutter and to substantially speed up the display of large traces, only one span is shown per
horizontal pixel across the display and only one row is shown per vertical pixel. In the horizontal
direction, the JavaScript internally accumulates spans that are less than one pixel wide until the deferred
time reaches one pixel wide. Then a representative 1-pixel-wide single-color span is drawn. The
deferred time separately counts kernel, user, and idle time; the representative span's height and color is
based on the largest of these. For many traces, this design can give a quite choppy initial display,
possibly showing mostly timer interrupts. As you zoom in, more and more detail will be revealed.
In the vertical direction, a similar mechanism reduces very skinny rows to single colors with overlays
dropped.
If you zoom in far enough, the approximate KUtrace overhead for each event is shown as diagonal
white lines at the leading edge of a span. This is to remind you that there is overhead and to let you see
if it might be enough to distort observations. The approximate values are determined by running the
time_getpid program with no tracing, with KUtrace go, and with KUtrace goipc (which
unfortunately has higher overhead because of the terribly slow read of the instructions-retired
performance counter). For the Intel i3 at 3.9 GHz, the measured overhead is about 20 nsec for go, and
about 50 nsec for goipc. For the Raspberry Pi at 1.5 GHz, the overhead is about 100 nsec for go, and
about 150 nsec for goipc.
The User button can be initially helpful to see all the process names and then pick time areas that you
care about.
Within the timeline region, mouse click-drag pans and mouse wheel zooms in and out. Shift-click-
drag annotates the start and end points and shows the elapsed time difference at the bottom of the
screen. Shift-click on any span annotates it. If there is enough vertical room, annotation text rotates
through several lines, reducing overwriting of names. As a convenience, shift-right-click on a span
annotates it and highlights all matching spans by graying out everything else. This also shows the match
count and duration information across all the matches.
A shift-click annotation in the default style gives the long-form version: starting time of a span relative
to the base time at the far left of Region 5 (avoiding excess-digit clutter), the name of the span, its
duration, and if available its IPC. For system calls, the low 16 bits of the first argument and the low 16
bits of the return value are shown as foo(1234)=12. Alternate-style annotations give just the name.
If shift is still down when you unclick, it clears annotations. Unshifting first leaves annotations showing,
allowing you to annotate many spans. If you pan or zoom horizontally, the single annotation nearest the
mouse x-position is kept and converted to long form. This can help keep you located near something of
interest while you pan and zoom.
If the Lock or PC sample overlay is visible, shift-clicking just above a timeline will select the overlay
span instead of the CPU/wait span.
If packet tracing was specified for the loadable module, then the KUtrace packet entries will show as
little diagonal lines above CPU 0, downward-slanting for incoming rx packets, and upward-slanting for
outgoing tx packets. Their activity can then be correlated with the sending or receiving process activity.
A search for "tx" or "rx" will annotate these entries, showing in uppercase hex 16 bits of the unmasked
data hash over the first 32 data bytes of a packet. These hash values can help correlate each traced
packet across client-server connections between two machines, and across kernel-user interfaces within
one machine.
Region 5, X-axis
The x-axis shows time. To avoid clutter, most of the time is shown to the left of the x-axis: day, hours,
minutes, seconds. As you zoom in, the x-axis units change from seconds to milliseconds to
microseconds to nanoseconds and the base time adds high-order seconds, milliseconds and then
microseconds digits. When the low-order digits along the x-axis wrap around, the next second,
millisecond, etc. is shown followed by an underscore in the corresponding label.
Region 6, Save/Restore
It is useful, when you find an anomaly in a view, to record the time and control settings such that you
can get back to that view. At the lower left there are four circled-digit buttons for this purpose. If you
shift-click on one of these, the current display state is saved and the button turns from gray to blue. If
you later click (no shift) on a blue button, that view will be restored. The back arrow will go back
exactly one view when clicked.
You can also add new buttons by clicking the circle-plus sign. A popup offers you a name for the new
button and clicking OK creates the button and stores the current view state there.
Unfortunately, there is no simple way to save this information across HTML file opens, because many
security mechanisms prevent updating an HTML file on disk from within a browser. Instead, I have
implemented two kludges.
First, whenever the browser closes a KUtrace HTML file, the current view state is saved as a text string
in localStorage (a limited 2MB of browser file space holding <key, value> pairs). Each KUtrace
HTML file has a 32-bit random ID, and that is used as the key. Whenever a KUtrace HTML file is
opened, the localStorage is checked for the matching key. If found, the view state is loaded but not
immediately used. Instead, the back-arrow button is turned blue to indicate that a previous state is
available. Clicking it uses the saved state. Not using the state automatically makes the mechanism robust
against corrupted or redesigned state. This facility gives a one-view local-machine state saving.
Second, the circle-1 through circle-4 states and any added buttons and callouts can be exported in text
form to the clipboard and then hand-edited into an underlying JSON file and thence via makeself
turned into a new HTML file that carries along the saved states. This is further explained in the next
section. I hope someone can do better.
At the right side of Region 6 is a blue [more] button. Click it to expand the sketchy directions on that
line to somewhat more text. Click again (it will read [less]) to shrink.
Secondary Controls
Shift-click on the red dot to toggle on some secondary controls. These are mainly for tweaking the
display for the purpose of making presentations (or book diagrams). There are five text boxes for
numbers and four buttons.
The Aspect box accepts a viewing aspect ratio of n:m, where each is a single digit. For example, "4:3"
makes a view that is the same aspect ratio as old TV sets, and "9:4" closely matches the 16:9 aspect ratio
of new TV sets, while "0:0" turns off the restriction. When the aspect ratio is active, the height of the
display area is restricted to be a multiple of 100 pixels and the width is determined by the aspect ratio.
The upper right of Region 1 will show the display area size in pixels for five seconds after you change
the browser window size.
The Ychars and Ypx boxes specify the approximate number of characters reserved for the y-axis labels
and their font size in pixels. The Ypx value is also used for the font size of the x-axis labels.
The txt and spn boxes specify the use of vertical space for each row of the timeline display. The txt box
specifies the number of lines of text to be used for annotations, and the spn box specifies the number of
same-size lines to be used as the space for drawing timelines. So for each timeline row, txt / (txt+spn) of
the height is used for annotations and spn / (txt+spn) is used for the kernel/user/idle drawing. For
example, 2 and 4 specifies 2/6 of the row height for 2 lines of annotation text and 4/6 for timelines,
while 3 and 6 specifies the same proportional split of vertical space for text and timelines but with three
lines of annotation text. The number of lines of text and the y-axis zoom determine the font size of the
annotations.
The Legend button cycles through three states: showing the normal timeline display, showing the
KUtrace HTML Legend from the back of the book in landscape orientation and showing it in portrait
orientation.
The Fade button grays out all the main timeline diagram elements, leaving just the overlay items. This
can be occasionally useful for emphasizing some overlay data.
The Speech-bubble button allows you to insert a speech-bubble callout in the diagram. A popup asks
for the callout's text. Clicking OK displays the callout. You can then drag the spike part to the span you
want to label and then drag the bubble part to its desired position. There is an underlying grid for the
bubble positions, to make it easier to align several related bubbles. If exactly one span is annotated when
you start, the callout will be positioned at that span.
The Export button copies the set of view buttons (described above in the Save/Restore section) and the
callout bubbles to the clipboard, as JSON text. The last line of the JSON for a full trace is "]}". You can
replace this line with the clipboard text, beginning with "]," and ending with "]}". This adds two more
arrays to the JSON, the "extra" array giving callouts and the "savedview" array giving view buttons,
including any you have added. The resulting JSON can then go through makeself to create an
augmented HTML file.
At the right of these buttons, the x: value gives the time offset and width of the x-axis, both in
microseconds, and the y: value gives the vertical offset and height of the y-axis in units of "tracks"
where there are 20 tracks per timeline row. These numbers can be used to align and size screen shots
across multiple views or multiple HTML files. Finally, the version date of the show_cpu.html template
is shown at the far right (for July 2022 templates and later).
These per-row profiles thus show the aggregated kernel/user/idle time for each CPU or PID or RPC.
They also show the overlays: aggregated lock holding and waiting time, CPU frequency time, and PC
samples. The overlays will no longer line up with the execution spans or each other, but they correctly
show the overall distribution. The IPC values for each kernel and user interval are proper weighted-by-
duration sums of the individual IPC values.
These profiles show where larger amounts of execution and non-execution times go, and can reveal
significant differences between similar threads or similar RPCs, showing how slower ones differ from
normal ones. Note that just the user-mode PC-samples in this information for a single PID matches what
a simple profiler, one that covers one process and only user-mode execution, would show.
The -group command-line flag accumulates row profiles as above and then within each CPU and PID
and RPC group combines multiple rows based on their names and elapsed time.
Within the CPU group, all the time for all CPUs is accumulated and then divided by the number of
CPUs, giving the average time breakdown across all CPUs in an entire trace. Note that just the PC-
sample portion of this information matches what a full-CPU profiler, one that covers all processes and
both kernel- and user-mode execution, would show.
Within the PID group, multiple threads with the same initial name -- letters/digits/underscore/hyphen --
are accumulated into power-of-two buckets based on elapsed time from start to finish of each thread.
There are ten microsecond buckets 1, 2, 4, ... ten millisecond buckets, and four seconds buckets 1, 2, 4,
8+. In addition, for each group of like-named threads there is an average bucket. These profiles can
reveal what is different between faster-than-average and slower-than-average threads doing similar
work.
Within the RPC group, multiple RPCs with the same initial method name are accumulated into power-
of-two buckets. This can be particularly revealing about what is different about slow RPCs, be it
execution time, waiting time, low IPC, and so on.
To avoid clutter groups of just one row are dropped from the group profile unless the -all command-
line flag is used. Their group profile would just be a duplicate of their row profile.
Keep in mind that you may need to zoom in substantially near time 0.00 to see the detail in short profile
rows that initially appear to be empty.