mmc: msm_sdcc: Fix race in disabling sdcc core irq
o Assume SDCC driver has SDIO clients attached and
runtime clock gating has turned off the SDCC clocks.
o Now, if the SDIO client driver has initiated a request
to SDCC and since the clocks are turned off, mmc framework
tries to turn them on calling msmsdcc_set_ios().
o msmsdcc_set_ios() tries to disable the core irq first and
reenable after the clocks are turned on. While in process
of disabling the irq, if there is an asynchronous notification
from SDIO client h/w, it will preempt ongoing disable and run
msmsdcc_irq() which internally does disable_irq_nosync() since
the clocks are still off.
o While the control returns to msmsdcc_set_ios() from preemption
it will continue disabling the core irq (even though
host->sdcc_irq_disabled = 1). In effect, calls disable_irq()
twice increasing the irq desc->depth to 2.
o The enable_irq() API enables the irq in h/w if and only if
desc->depth = 1, and hence become a noop if it is called with
desc->depth = 2 causing sdcc failure because the interrupts
are not enabled in h/w.
Fix the above race condition by making sure disable irq is done
atomically.
CRs-Fixed: 357406
Change-Id: Id8a4e5cc9b0ac035a83c070ede57b2c39d96e6e9
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 e84d87b..eba80c5 100644
--- a/drivers/mmc/host/msm_sdcc.c
+++ b/drivers/mmc/host/msm_sdcc.c
@@ -2877,13 +2877,14 @@
DBG(host, "ios->clock = %u\n", ios->clock);
spin_lock_irqsave(&host->lock, flags);
if (!host->sdcc_irq_disabled) {
- spin_unlock_irqrestore(&host->lock, flags);
- disable_irq(host->core_irqres->start);
- spin_lock_irqsave(&host->lock, flags);
+ disable_irq_nosync(host->core_irqres->start);
host->sdcc_irq_disabled = 1;
}
spin_unlock_irqrestore(&host->lock, flags);
+ /* Make sure sdcc core irq is synchronized */
+ synchronize_irq(host->core_irqres->start);
+
pwr = msmsdcc_setup_pwr(host, ios);
spin_lock_irqsave(&host->lock, flags);