Skip to contentSkip to conversation

Bluetooth: Collect kcov coverage from hci_rx_work

PATCH Superseded Tamas Koczka 4 participants 7 messages Jun 7, 2022 – Jul 4, 2022

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 Bluetooth: Collect kcov coverage from hci_rx_work

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +9 -1
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
 	BT_DBG("%s", hdev->name);
 
-	while ((skb = skb_dequeue(&hdev->rx_q))) {
+	/* The kcov_remote functions used for collecting packet parsing
+	 * coverage information from this background thread and associate
+	 * the coverage with the syscall's thread which originally injected
+	 * the packet. This helps fuzzing the kernel.
+	 */
+	for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+		kcov_remote_start_common(skb_get_kcov_handle(skb));
+
 		/* Send copy to monitor */
 		hci_send_to_monitor(hdev, skb);
 
Signature · 1 line
2.36.1.255.ge46751e96f-goog
#1 lore raw

This is automated email and please do not reply to this email!

Dear submitter,

Thank you for submitting the patches to the linux bluetooth mailing list. This is a CI test results with your patch series:
PW Link:https://patchwork.kernel.org/project/bluetooth/list/?series=647954

---Test result---

Test Summary:

CheckPatch                    PASS      1.24 seconds
GitLint                       PASS      0.84 seconds
SubjectPrefix                 PASS      0.62 seconds
BuildKernel                   PASS      30.66 seconds
BuildKernel32                 PASS      27.33 seconds
Incremental Build with patchesPASS      37.37 seconds
TestRunner: Setup             PASS      467.41 seconds
TestRunner: l2cap-tester      PASS      17.70 seconds
TestRunner: bnep-tester       PASS      5.98 seconds
TestRunner: mgmt-tester       PASS      100.14 seconds
TestRunner: rfcomm-tester     PASS      9.56 seconds
TestRunner: sco-tester        PASS      9.27 seconds
TestRunner: smp-tester        PASS      9.32 seconds
TestRunner: userchan-tester   PASS      6.25 seconds

---
Regards,
Linux Bluetooth

#2 lore raw
In reply to Tamas Koczka

Hello Marcel,

I added some comments into the code about what the kcov_remote calls do and why they were implemented and I also added some reasoning to the commit message.

I did not mention in the commit but these functions only run if the kernel is compiled with CONFIG_KCOV.

Thank you again for reviewing the patch!

--
Tamas

Quoted 51 lines · Tamas Koczka

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +9 -2
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
        BT_DBG("%s", hdev->name);
 
-       while ((skb = skb_dequeue(&hdev->rx_q))) {
+       /* The kcov_remote functions used for collecting packet parsing
+        * coverage information from this background thread and associate
+        * the coverage with the syscall's thread which originally injected
+        * the packet. This helps fuzzing the kernel.
+        */
+       for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+               kcov_remote_start_common(skb_get_kcov_handle(skb));
+
                /* Send copy to monitor */
                hci_send_to_monitor(hdev, skb);
 
--

2.36.1.255.ge46751e96f-goog

#3 lore raw

Hello Marcel,

I hope this was the change you originally requested, and I did not misunderstand anything, but if you need any additional modification to the code or the commit, please feel free to let me know!

Thank you,
Tamas

Quoted 67 lines · Tamás Koczka

Hello Marcel,

I added some comments into the code about what the kcov_remote calls do and why they were implemented and I also added some reasoning to the commit message.

I did not mention in the commit but these functions only run if the kernel is compiled with CONFIG_KCOV.

Thank you again for reviewing the patch!

--
Tamas

Quoted 51 lines · Tamas Koczka

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +9 -2
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
        BT_DBG("%s", hdev->name);
 
-       while ((skb = skb_dequeue(&hdev->rx_q))) {
+       /* The kcov_remote functions used for collecting packet parsing
+        * coverage information from this background thread and associate
+        * the coverage with the syscall's thread which originally injected
+        * the packet. This helps fuzzing the kernel.
+        */
+       for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+               kcov_remote_start_common(skb_get_kcov_handle(skb));
+
                /* Send copy to monitor */
                hci_send_to_monitor(hdev, skb);
 
--

2.36.1.255.ge46751e96f-goog

#4 lore raw

(Resending the reply I sent to the v1 of the patch. I sent it by mistake with HTML content, so it did not reach lore.)

I checked out v5.18.1, applied this patch and fuzzed it with syzkaller for a day. The fuzzer was indeed able to find and report more coverage of the BT subsystem than without the patch.

Quoted 78 lines · Tamás Koczka

Hello Marcel,

I hope this was the change you originally requested, and I did not misunderstand anything, but if you need any additional modification to the code or the commit, please feel free to let me know!

Thank you,
Tamas

Quoted 67 lines · Tamás Koczka

Hello Marcel,

I added some comments into the code about what the kcov_remote calls do and why they were implemented and I also added some reasoning to the commit message.

I did not mention in the commit but these functions only run if the kernel is compiled with CONFIG_KCOV.

Thank you again for reviewing the patch!

--
Tamas

Quoted 51 lines · Tamas Koczka

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +9 -2
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
        BT_DBG("%s", hdev->name);
 
-       while ((skb = skb_dequeue(&hdev->rx_q))) {
+       /* The kcov_remote functions used for collecting packet parsing
+        * coverage information from this background thread and associate
+        * the coverage with the syscall's thread which originally injected
+        * the packet. This helps fuzzing the kernel.
+        */
+       for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+               kcov_remote_start_common(skb_get_kcov_handle(skb));
+
                /* Send copy to monitor */
                hci_send_to_monitor(hdev, skb);
 
--

2.36.1.255.ge46751e96f-goog

Reviewed-by
#5 lore raw
Quoted 75 lines · Aleksandr Nogikh

(Resending the reply I sent to the v1 of the patch. I sent it by mistake with HTML content, so it did not reach lore.)

I checked out v5.18.1, applied this patch and fuzzed it with syzkaller for a day. The fuzzer was indeed able to find and report more coverage of the BT subsystem than without the patch.

Quoted 63 lines · Tamás Koczka

Hello Marcel,

I hope this was the change you originally requested, and I did not misunderstand anything, but if you need any additional modification to the code or the commit, please feel free to let me know!

Thank you,
Tamas

Quoted 52 lines · Tamás Koczka

Hello Marcel,

I added some comments into the code about what the kcov_remote calls do and why they were implemented and I also added some reasoning to the commit message.

I did not mention in the commit but these functions only run if the kernel is compiled with CONFIG_KCOV.

Thank you again for reviewing the patch!

--
Tamas

Quoted 36 lines · Tamas Koczka

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +1 -0
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
        BT_DBG("%s", hdev->name);
 
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines

net/bluetooth/hci_core.c static void hci_rx_work(struct work_struct *work) · line 3793

Quoted 11 lines
net/bluetooth/hci_core.c +8 -1
-       while ((skb = skb_dequeue(&hdev->rx_q))) {
+       /* The kcov_remote functions used for collecting packet parsing
+        * coverage information from this background thread and associate
+        * the coverage with the syscall's thread which originally injected
+        * the packet. This helps fuzzing the kernel.
+        */
+       for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+               kcov_remote_start_common(skb_get_kcov_handle(skb));
+
                /* Send copy to monitor */
                hci_send_to_monitor(hdev, skb);

Looks good to me.
Anything else needed to merge this patch?

#6 lore raw

Hello,

If you need any clarification about the patch or if you have questions or if the patch needs to be modified, please feel free to tell me.

Basically the patch should not have any effect on a kernel which is not compiled with CONFIG_KCOV and we'd like to use the patch to make the coverage of the hci_rx_work background thread visible to Syzkaller, because the BT packet parsing / handling logic happens there and this way Syzkaller will be able to more effectively mutate the packets used for fuzzing, hopefully reaching new code paths, maybe discovering and reporting new vulnerabilities before they reach the mainline.

Thank you,
Tamas

Quoted 93 lines · Dmitry Vyukov
Quoted 75 lines · Aleksandr Nogikh

(Resending the reply I sent to the v1 of the patch. I sent it by mistake with HTML content, so it did not reach lore.)

I checked out v5.18.1, applied this patch and fuzzed it with syzkaller for a day. The fuzzer was indeed able to find and report more coverage of the BT subsystem than without the patch.

Quoted 63 lines · Tamás Koczka

Hello Marcel,

I hope this was the change you originally requested, and I did not misunderstand anything, but if you need any additional modification to the code or the commit, please feel free to let me know!

Thank you,
Tamas

Quoted 52 lines · Tamás Koczka

Hello Marcel,

I added some comments into the code about what the kcov_remote calls do and why they were implemented and I also added some reasoning to the commit message.

I did not mention in the commit but these functions only run if the kernel is compiled with CONFIG_KCOV.

Thank you again for reviewing the patch!

--
Tamas

Quoted 36 lines · Tamas Koczka

Annotate hci_rx_work() with kcov_remote_start() and kcov_remote_stop() calls, so remote KCOV coverage is collected while processing the rx_q queue which is the main incoming Bluetooth packet queue.

Coverage is associated with the thread which created the packet skb.

The collected extra coverage helps kernel fuzzing efforts in finding vulnerabilities.

---
Changelog since v1:
- add comment about why kcov_remote functions are called

v1: https://lore.kernel.org/all/[email protected]/

 net/bluetooth/hci_core.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
net/bluetooth/hci_core.c +1 -0
index 45c2dd2e1590..0af43844c55a 100644
@@ -29,6 +29,7 @@
 #include <linux/rfkill.h>
 #include <linux/debugfs.h>
 #include <linux/crypto.h>
+#include <linux/kcov.h>
 #include <linux/property.h>
 #include <linux/suspend.h>
 #include <linux/wait.h>
@@ -3780,7 +3781,14 @@ static void hci_rx_work(struct work_struct *work)
 
        BT_DBG("%s", hdev->name);
 
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines

net/bluetooth/hci_core.c static void hci_rx_work(struct work_struct *work) · line 3793

Quoted 11 lines
net/bluetooth/hci_core.c +8 -1
-       while ((skb = skb_dequeue(&hdev->rx_q))) {
+       /* The kcov_remote functions used for collecting packet parsing
+        * coverage information from this background thread and associate
+        * the coverage with the syscall's thread which originally injected
+        * the packet. This helps fuzzing the kernel.
+        */
+       for (; (skb = skb_dequeue(&hdev->rx_q)); kcov_remote_stop()) {
+               kcov_remote_start_common(skb_get_kcov_handle(skb));
+
                /* Send copy to monitor */
                hci_send_to_monitor(hdev, skb);

Looks good to me.
Anything else needed to merge this patch?