mmc: msm_sdcc: Fix deadlock in mmc_suspend_host and mmc_rescan
SDCC runtime suspend can race with mmc_rescan causing deadlock.
The call flow would be this:
1) When runtime suspend is triggered:
msmsdcc_runtime_suspend()
-> mmc_suspend_host()
-> mmc_flush_scheduled_work()
2) mmc_flush_schedule_work() waits for mmc_rescan() work
to be completed if it is already in runqueue.
3) When mmc_rescan() is scheduled to run:
mmc_claim_host()
-> msmsdcc_enable()
->pm_runtime_get_sync()
4) pm_runtime_get_sync() waits for runtime_suspend to complete,
and hence cause two threads to wait upon each other.
Fix this deadlock by aborting runtime suspend when mmc_rescan
is scheduled.
CRs-Fixed: 326610
Change-Id: I4ca88651f80f5a1bfccb6e0c07e3ea83fadcdc57
Signed-off-by: Sujit Reddy Thumma <sthumma@codeaurora.org>
diff --git a/drivers/mmc/host/msm_sdcc.c b/drivers/mmc/host/msm_sdcc.c
index 4b6525c..db83404 100644
--- a/drivers/mmc/host/msm_sdcc.c
+++ b/drivers/mmc/host/msm_sdcc.c
@@ -4544,7 +4544,11 @@
* simple become pm usage counter increment operations.
*/
pm_runtime_get_noresume(dev);
- rc = mmc_suspend_host(mmc);
+ /* If there is pending detect work abort runtime suspend */
+ if (unlikely(work_busy(&mmc->detect.work)))
+ rc = -EAGAIN;
+ else
+ rc = mmc_suspend_host(mmc);
pm_runtime_put_noidle(dev);
if (!rc) {
@@ -4575,7 +4579,7 @@
if (rc && wake_lock_active(&host->sdio_suspend_wlock))
wake_unlock(&host->sdio_suspend_wlock);
}
- pr_debug("%s: %s: end\n", mmc_hostname(mmc), __func__);
+ pr_debug("%s: %s: ends with err=%d\n", mmc_hostname(mmc), __func__, rc);
return rc;
}