summaryrefslogtreecommitdiff
path: root/docs/devel/fuzzing.txt
blob: 96d71c94d7e7ab797325bab2eb19761806be7a30 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
= Fuzzing =

== Introduction ==

This document describes the virtual-device fuzzing infrastructure in QEMU and
how to use it to implement additional fuzzers.

== Basics ==

Fuzzing operates by passing inputs to an entry point/target function. The
fuzzer tracks the code coverage triggered by the input. Based on these
findings, the fuzzer mutates the input and repeats the fuzzing.

To fuzz QEMU, we rely on libfuzzer. Unlike other fuzzers such as AFL, libfuzzer
is an _in-process_ fuzzer. For the developer, this means that it is their
responsibility to ensure that state is reset between fuzzing-runs.

== Building the fuzzers ==

NOTE: If possible, build a 32-bit binary. When forking, the 32-bit fuzzer is
much faster, since the page-map has a smaller size. This is due to the fact that
AddressSanitizer mmaps ~20TB of memory, as part of its detection. This results
in a large page-map, and a much slower fork().

To build the fuzzers, install a recent version of clang:
Configure with (substitute the clang binaries with the version you installed).
Here, enable-sanitizers, is optional but it allows us to reliably detect bugs
such as out-of-bounds accesses, use-after-frees, double-frees etc.

    CC=clang-8 CXX=clang++-8 /path/to/configure --enable-fuzzing \
                                                --enable-sanitizers

Fuzz targets are built similarly to system/softmmu:

    make i386-softmmu/fuzz

This builds ./i386-softmmu/qemu-fuzz-i386

The first option to this command is: --fuzz-target=FUZZ_NAME
To list all of the available fuzzers run qemu-fuzz-i386 with no arguments.

For example:
    ./i386-softmmu/qemu-fuzz-i386 --fuzz-target=virtio-scsi-fuzz

Internally, libfuzzer parses all arguments that do not begin with "--".
Information about these is available by passing -help=1

Now the only thing left to do is wait for the fuzzer to trigger potential
crashes.

== Useful libFuzzer flags ==

As mentioned above, libFuzzer accepts some arguments. Passing -help=1 will list
the available arguments. In particular, these arguments might be helpful:

$CORPUS_DIR/ : Specify a directory as the last argument to libFuzzer. libFuzzer
stores each "interesting" input in this corpus directory. The next time you run
libFuzzer, it will read all of the inputs from the corpus, and continue fuzzing
from there. You can also specify multiple directories. libFuzzer loads existing
inputs from all specified directories, but will only write new ones to the
first one specified.

-max_len=4096 : specify the maximum byte-length of the inputs libFuzzer will
generate.

-close_fd_mask={1,2,3} : close, stderr, or both. Useful for targets that
trigger many debug/error messages, or create output on the serial console.

-jobs=4 -workers=4 : These arguments configure libFuzzer to run 4 fuzzers in
parallel (4 fuzzing jobs in 4 worker processes). Alternatively, with only
-jobs=N, libFuzzer automatically spawns a number of workers less than or equal
to half the available CPU cores. Replace 4 with a number appropriate for your
machine. Make sure to specify a $CORPUS_DIR, which will allow the parallel
fuzzers to share information about the interesting inputs they find.

-use_value_profile=1 : For each comparison operation, libFuzzer computes 
(caller_pc&4095) | (popcnt(Arg1 ^ Arg2) << 12) and places this in the coverage
table. Useful for targets with "magic" constants. If Arg1 came from the fuzzer's
input and Arg2 is a magic constant, then each time the Hamming distance
between Arg1 and Arg2 decreases, libFuzzer adds the input to the corpus.

-shrink=1 : Tries to make elements of the corpus "smaller". Might lead to
better coverage performance, depending on the target.

Note that libFuzzer's exact behavior will depend on the version of
clang and libFuzzer used to build the device fuzzers.

== Generating Coverage Reports ==
Code coverage is a crucial metric for evaluating a fuzzer's performance.
libFuzzer's output provides a "cov: " column that provides a total number of
unique blocks/edges covered. To examine coverage on a line-by-line basis we
can use Clang coverage:

 1. Configure libFuzzer to store a corpus of all interesting inputs (see
    CORPUS_DIR above)
 2. ./configure the QEMU build with:
    --enable-fuzzing \
    --extra-cflags="-fprofile-instr-generate -fcoverage-mapping"
 3. Re-run the fuzzer. Specify $CORPUS_DIR/* as an argument, telling libfuzzer
    to execute all of the inputs in $CORPUS_DIR and exit. Once the process
    exits, you should find a file, "default.profraw" in the working directory.
 4. Execute these commands to generate a detailed HTML coverage-report:
 llvm-profdata merge -output=default.profdata default.profraw
 llvm-cov show ./path/to/qemu-fuzz-i386 -instr-profile=default.profdata \
 --format html -output-dir=/path/to/output/report

== Adding a new fuzzer ==
Coverage over virtual devices can be improved by adding additional fuzzers.
Fuzzers are kept in tests/qtest/fuzz/ and should be added to
tests/qtest/fuzz/Makefile.include

Fuzzers can rely on both qtest and libqos to communicate with virtual devices.

1. Create a new source file. For example ``tests/qtest/fuzz/foo-device-fuzz.c``.

2. Write the fuzzing code using the libqtest/libqos API. See existing fuzzers
for reference.

3. Register the fuzzer in ``tests/fuzz/Makefile.include`` by appending the
corresponding object to fuzz-obj-y

Fuzzers can be more-or-less thought of as special qtest programs which can
modify the qtest commands and/or qtest command arguments based on inputs
provided by libfuzzer. Libfuzzer passes a byte array and length. Commonly the
fuzzer loops over the byte-array interpreting it as a list of qtest commands,
addresses, or values.

= Implementation Details =

== The Fuzzer's Lifecycle ==

The fuzzer has two entrypoints that libfuzzer calls. libfuzzer provides it's
own main(), which performs some setup, and calls the entrypoints:

LLVMFuzzerInitialize: called prior to fuzzing. Used to initialize all of the
necessary state

LLVMFuzzerTestOneInput: called for each fuzzing run. Processes the input and
resets the state at the end of each run.

In more detail:

LLVMFuzzerInitialize parses the arguments to the fuzzer (must start with two
dashes, so they are ignored by libfuzzer main()). Currently, the arguments
select the fuzz target. Then, the qtest client is initialized. If the target
requires qos, qgraph is set up and the QOM/LIBQOS modules are initialized.
Then the QGraph is walked and the QEMU cmd_line is determined and saved.

After this, the vl.c:qemu__main is called to set up the guest. There are
target-specific hooks that can be called before and after qemu_main, for
additional setup(e.g. PCI setup, or VM snapshotting).

LLVMFuzzerTestOneInput: Uses qtest/qos functions to act based on the fuzz
input. It is also responsible for manually calling the main loop/main_loop_wait
to ensure that bottom halves are executed and any cleanup required before the
next input.

Since the same process is reused for many fuzzing runs, QEMU state needs to
be reset at the end of each run. There are currently two implemented
options for resetting state:
1. Reboot the guest between runs.
   Pros: Straightforward and fast for simple fuzz targets.
   Cons: Depending on the device, does not reset all device state. If the
   device requires some initialization prior to being ready for fuzzing
   (common for QOS-based targets), this initialization needs to be done after
   each reboot.
   Example target: i440fx-qtest-reboot-fuzz
2. Run each test case in a separate forked process and copy the coverage
   information back to the parent. This is fairly similar to AFL's "deferred"
   fork-server mode [3]
   Pros: Relatively fast. Devices only need to be initialized once. No need
   to do slow reboots or vmloads.
   Cons: Not officially supported by libfuzzer. Does not work well for devices
   that rely on dedicated threads.
   Example target: virtio-net-fork-fuzz