net/tg3: resolve deadlock in tg3_reset_task() during EEH
authorDavid Christensen <drc@linux.vnet.ibm.com>
Tue, 24 Jan 2023 18:53:39 +0000 (13:53 -0500)
committerJakub Kicinski <kuba@kernel.org>
Thu, 26 Jan 2023 06:35:42 +0000 (22:35 -0800)
During EEH error injection testing, a deadlock was encountered in the tg3
driver when tg3_io_error_detected() was attempting to cancel outstanding
reset tasks:

crash> foreach UN bt
...
PID: 159    TASK: c0000000067c6000  CPU: 8   COMMAND: "eehd"
...
 #5 [c00000000681f990] __cancel_work_timer at c00000000019fd18
 #6 [c00000000681fa30] tg3_io_error_detected at c00800000295f098 [tg3]
 #7 [c00000000681faf0] eeh_report_error at c00000000004e25c
...

PID: 290    TASK: c000000036e5f800  CPU: 6   COMMAND: "kworker/6:1"
...
 #4 [c00000003721fbc0] rtnl_lock at c000000000c940d8
 #5 [c00000003721fbe0] tg3_reset_task at c008000002969358 [tg3]
 #6 [c00000003721fc60] process_one_work at c00000000019e5c4
...

PID: 296    TASK: c000000037a65800  CPU: 21  COMMAND: "kworker/21:1"
...
 #4 [c000000037247bc0] rtnl_lock at c000000000c940d8
 #5 [c000000037247be0] tg3_reset_task at c008000002969358 [tg3]
 #6 [c000000037247c60] process_one_work at c00000000019e5c4
...

PID: 655    TASK: c000000036f49000  CPU: 16  COMMAND: "kworker/16:2"
...:1

 #4 [c0000000373ebbc0] rtnl_lock at c000000000c940d8
 #5 [c0000000373ebbe0] tg3_reset_task at c008000002969358 [tg3]
 #6 [c0000000373ebc60] process_one_work at c00000000019e5c4
...

Code inspection shows that both tg3_io_error_detected() and
tg3_reset_task() attempt to acquire the RTNL lock at the beginning of
their code blocks.  If tg3_reset_task() should happen to execute between
the times when tg3_io_error_deteced() acquires the RTNL lock and
tg3_reset_task_cancel() is called, a deadlock will occur.

Moving tg3_reset_task_cancel() call earlier within the code block, prior
to acquiring RTNL, prevents this from happening, but also exposes another
deadlock issue where tg3_reset_task() may execute AFTER
tg3_io_error_detected() has executed:

crash> foreach UN bt
PID: 159    TASK: c0000000067d2000  CPU: 9   COMMAND: "eehd"
...
 #4 [c000000006867a60] rtnl_lock at c000000000c940d8
 #5 [c000000006867a80] tg3_io_slot_reset at c0080000026c2ea8 [tg3]
 #6 [c000000006867b00] eeh_report_reset at c00000000004de88
...
PID: 363    TASK: c000000037564000  CPU: 6   COMMAND: "kworker/6:1"
...
 #3 [c000000036c1bb70] msleep at c000000000259e6c
 #4 [c000000036c1bba0] napi_disable at c000000000c6b848
 #5 [c000000036c1bbe0] tg3_reset_task at c0080000026d942c [tg3]
 #6 [c000000036c1bc60] process_one_work at c00000000019e5c4
...

This issue can be avoided by aborting tg3_reset_task() if EEH error
recovery is already in progress.

Fixes: db84bf43ef23 ("tg3: tg3_reset_task() needs to use rtnl_lock to synchronize")
Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
Reviewed-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Link: https://lore.kernel.org/r/20230124185339.225806-1-drc@linux.vnet.ibm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/ethernet/broadcom/tg3.c

index 59debdc..5874729 100644 (file)
@@ -11166,7 +11166,7 @@ static void tg3_reset_task(struct work_struct *work)
        rtnl_lock();
        tg3_full_lock(tp, 0);
 
-       if (!netif_running(tp->dev)) {
+       if (tp->pcierr_recovery || !netif_running(tp->dev)) {
                tg3_flag_clear(tp, RESET_TASK_PENDING);
                tg3_full_unlock(tp);
                rtnl_unlock();
@@ -18101,6 +18101,9 @@ static pci_ers_result_t tg3_io_error_detected(struct pci_dev *pdev,
 
        netdev_info(netdev, "PCI I/O error detected\n");
 
+       /* Want to make sure that the reset task doesn't run */
+       tg3_reset_task_cancel(tp);
+
        rtnl_lock();
 
        /* Could be second call or maybe we don't have netdev yet */
@@ -18117,9 +18120,6 @@ static pci_ers_result_t tg3_io_error_detected(struct pci_dev *pdev,
 
        tg3_timer_stop(tp);
 
-       /* Want to make sure that the reset task doesn't run */
-       tg3_reset_task_cancel(tp);
-
        netif_device_detach(netdev);
 
        /* Clean up software state, even if MMIO is blocked */