Skip to contentSkip to conversation

fuzz: Disable QEMU's SIG{INT,HUP,TERM} handlers

PATCH Applied Alexander Bulekov 3 participants 3 messages Oct 14, 2020 – Oct 15, 2020

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 fuzz: Disable QEMU's SIG{INT,HUP,TERM} handlers

Prior to this patch, the only way I found to terminate the fuzzer was either to:
1. Explicitly specify the number of fuzzer runs with the -runs= flag
2. SIGKILL the process with "pkill -9 qemu-fuzz-*" or similar

In addition to being annoying to deal with, SIGKILLing the process skips over any exit handlers(e.g. registered with atexit()). This is bad, since some fuzzers might create temporary files that should ideally be removed on exit using an exit handler. The only way to achieve a clean exit now is to specify -runs=N , but the desired "N" is tricky to identify prior to fuzzing.

Why doesn't the process exit with standard SIGINT,SIGHUP,SIGTERM signals? QEMU installs its own handlers for these signals in os-posix.c:os_setup_signal_handling, which notify the main loop that an exit was requested. The fuzzer, however, does not run qemu_main_loop, which performs the main_loop_should_exit() check. This means that the fuzzer effectively ignores these signals. As we don't really care about cleanly stopping the disposable fuzzer "VM", this patch uninstalls QEMU's signal handlers. Thus, we can stop the fuzzer with SIG{INT,HUP,TERM} and the fuzzing code can optionally use atexit() to clean up temporary files/resources.

---
 tests/qtest/fuzz/fuzz.c | 8 ++++++++
 1 file changed, 8 insertions(+)
tests/qtest/fuzz/fuzz.c +8 -0
index d926c490c5..eb0070437f 100644
@@ -217,5 +217,13 @@ int LLVMFuzzerInitialize(int *argc, char ***argv, char ***envp)
     /* re-enable the rcu atfork, which was previously disabled in qemu_init */
     rcu_enable_atfork();
 
+    /*
+     * Disable QEMU's signal handlers, since we manually control the main_loop,
+     * and don't check for main_loop_should_exit
+     */
+    signal(SIGINT, SIG_DFL);
+    signal(SIGHUP, SIG_DFL);
+    signal(SIGTERM, SIG_DFL);
+
     return 0;
 }
Signature · 1 line
2.28.0
#1 lore raw
Quoted 23 lines · Alexander Bulekov

Prior to this patch, the only way I found to terminate the fuzzer was either to:
1. Explicitly specify the number of fuzzer runs with the -runs= flag
2. SIGKILL the process with "pkill -9 qemu-fuzz-*" or similar

In addition to being annoying to deal with, SIGKILLing the process skips over any exit handlers(e.g. registered with atexit()). This is bad, since some fuzzers might create temporary files that should ideally be removed on exit using an exit handler. The only way to achieve a clean exit now is to specify -runs=N , but the desired "N" is tricky to identify prior to fuzzing.

Why doesn't the process exit with standard SIGINT,SIGHUP,SIGTERM signals? QEMU installs its own handlers for these signals in os-posix.c:os_setup_signal_handling, which notify the main loop that an exit was requested. The fuzzer, however, does not run qemu_main_loop, which performs the main_loop_should_exit() check. This means that the fuzzer effectively ignores these signals. As we don't really care about cleanly stopping the disposable fuzzer "VM", this patch uninstalls QEMU's signal handlers. Thus, we can stop the fuzzer with SIG{INT,HUP,TERM} and the fuzzing code can optionally use atexit() to clean up temporary files/resources.

Quoted 2 lines

Much clearer Alex, thanks for rewording it :)

Darren.

Quoted 24 lines
---
 tests/qtest/fuzz/fuzz.c | 8 ++++++++
 1 file changed, 8 insertions(+)
tests/qtest/fuzz/fuzz.c +8 -0
index d926c490c5..eb0070437f 100644
@@ -217,5 +217,13 @@ int LLVMFuzzerInitialize(int *argc, char ***argv, char ***envp)
     /* re-enable the rcu atfork, which was previously disabled in qemu_init */
     rcu_enable_atfork();
 
+    /*
+     * Disable QEMU's signal handlers, since we manually control the main_loop,
+     * and don't check for main_loop_should_exit
+     */
+    signal(SIGINT, SIG_DFL);
+    signal(SIGHUP, SIG_DFL);
+    signal(SIGTERM, SIG_DFL);
+
     return 0;
 }
Signature · 1 line
2.28.0
#2 lore raw
In reply to Alexander Bulekov
Quoted 38 lines · Alexander Bulekov

Prior to this patch, the only way I found to terminate the fuzzer was either to:
1. Explicitly specify the number of fuzzer runs with the -runs= flag
2. SIGKILL the process with "pkill -9 qemu-fuzz-*" or similar

In addition to being annoying to deal with, SIGKILLing the process skips over any exit handlers(e.g. registered with atexit()). This is bad, since some fuzzers might create temporary files that should ideally be removed on exit using an exit handler. The only way to achieve a clean exit now is to specify -runs=N , but the desired "N" is tricky to identify prior to fuzzing.

Why doesn't the process exit with standard SIGINT,SIGHUP,SIGTERM signals? QEMU installs its own handlers for these signals in os-posix.c:os_setup_signal_handling, which notify the main loop that an exit was requested. The fuzzer, however, does not run qemu_main_loop, which performs the main_loop_should_exit() check. This means that the fuzzer effectively ignores these signals. As we don't really care about cleanly stopping the disposable fuzzer "VM", this patch uninstalls QEMU's signal handlers. Thus, we can stop the fuzzer with SIG{INT,HUP,TERM} and the fuzzing code can optionally use atexit() to clean up temporary files/resources.

---
 tests/qtest/fuzz/fuzz.c | 8 ++++++++
 1 file changed, 8 insertions(+)
tests/qtest/fuzz/fuzz.c +0 -0
index d926c490c5..eb0070437f 100644
@@ -217,5 +217,13 @@ int LLVMFuzzerInitialize(int *argc, char ***argv, char ***envp)
     /* re-enable the rcu atfork, which was previously disabled in qemu_init */
     rcu_enable_atfork();
 

tests/qtest/fuzz/fuzz.c int LLVMFuzzerInitialize(int *argc, char ***argv, char ***envp) · line 229

Quoted 10 lines
tests/qtest/fuzz/fuzz.c +8 -0
+    /*
+     * Disable QEMU's signal handlers, since we manually control the main_loop,
+     * and don't check for main_loop_should_exit
+     */
+    signal(SIGINT, SIG_DFL);
+    signal(SIGHUP, SIG_DFL);
+    signal(SIGTERM, SIG_DFL);
+
     return 0;
 }

Queued, thanks.

Paolo