On 10/15/18 3:52 PM, Jon Hunter wrote: > > On 30/08/18 19:36, Dmitry Osipenko wrote: >> This fixes splats like the one below if CONFIG_DEBUG_ATOMIC_SLEEP=y >> and machine (Tegra30) booted with SMP=n or all secondary CPU's are put >> offline. >> >> BUG: sleeping function called from invalid context at kernel/locking/mutex.c:254 >> in_atomic(): 1, irqs_disabled(): 128, pid: 0, name: swapper/0 >> CPU: 0 PID: 0 Comm: swapper/0 Tainted: G C 4.18.0-next-20180821-00180-gc3ebb6544e44-dirty #823 >> Hardware name: NVIDIA Tegra SoC (Flattened Device Tree) >> [<c01134f4>] (unwind_backtrace) from [<c010db2c>] (show_stack+0x20/0x24) >> [<c010db2c>] (show_stack) from [<c0bd0f3c>] (dump_stack+0x94/0xa8) >> [<c0bd0f3c>] (dump_stack) from [<c0151df8>] (___might_sleep+0x13c/0x174) >> [<c0151df8>] (___might_sleep) from [<c0151ea0>] (__might_sleep+0x70/0xa8) >> [<c0151ea0>] (__might_sleep) from [<c0bec2b8>] (mutex_lock+0x2c/0x70) >> [<c0bec2b8>] (mutex_lock) from [<c0589844>] (tegra_powergate_is_powered+0x44/0xa8) >> [<c0589844>] (tegra_powergate_is_powered) from [<c0581a60>] (tegra30_cpu_rail_off_ready+0x30/0x74) >> [<c0581a60>] (tegra30_cpu_rail_off_ready) from [<c0122244>] (tegra30_idle_lp2+0xa0/0x108) >> [<c0122244>] (tegra30_idle_lp2) from [<c0853438>] (cpuidle_enter_state+0x140/0x540) >> [<c0853438>] (cpuidle_enter_state) from [<c08538a4>] (cpuidle_enter+0x40/0x4c) >> [<c08538a4>] (cpuidle_enter) from [<c01595e0>] (call_cpuidle+0x30/0x48) >> [<c01595e0>] (call_cpuidle) from [<c01599f8>] (do_idle+0x238/0x28c) >> [<c01599f8>] (do_idle) from [<c0159d28>] (cpu_startup_entry+0x28/0x2c) >> [<c0159d28>] (cpu_startup_entry) from [<c0be76c8>] (rest_init+0xd8/0xdc) >> [<c0be76c8>] (rest_init) from [<c1200f50>] (start_kernel+0x41c/0x430) > > Given the above, rather than converting to a spinlock I wonder if we are > just better off dropping the mutex completely from > tegra_powergate_is_powered()? Otherwise ... > >> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx> >> --- >> drivers/soc/tegra/pmc.c | 36 ++++++++++++++++++------------------ >> 1 file changed, 18 insertions(+), 18 deletions(-) >> >> diff --git a/drivers/soc/tegra/pmc.c b/drivers/soc/tegra/pmc.c >> index 2d6f3fcf3211..d6bc9f66f1cd 100644 >> --- a/drivers/soc/tegra/pmc.c >> +++ b/drivers/soc/tegra/pmc.c >> @@ -186,7 +186,7 @@ struct tegra_pmc_soc { >> * @lp0_vec_phys: physical base address of the LP0 warm boot code >> * @lp0_vec_size: size of the LP0 warm boot code >> * @powergates_available: Bitmap of available power gates >> - * @powergates_lock: mutex for power gate register access >> + * @powergates_lock: lock for power gate register access >> */ >> struct tegra_pmc { >> struct device *dev; >> @@ -215,7 +215,7 @@ struct tegra_pmc { >> u32 lp0_vec_size; >> DECLARE_BITMAP(powergates_available, TEGRA_POWERGATE_MAX); >> >> - struct mutex powergates_lock; >> + spinlock_t powergates_lock; >> }; >> >> static struct tegra_pmc *pmc = &(struct tegra_pmc) { >> @@ -288,10 +288,10 @@ static int tegra_powergate_set(unsigned int id, bool new_state) >> if (id == TEGRA_POWERGATE_3D && pmc->soc->has_gpu_clamps) >> return -EINVAL; >> >> - mutex_lock(&pmc->powergates_lock); >> + spin_lock(&pmc->powergates_lock); >> >> if (tegra_powergate_state(id) == new_state) { >> - mutex_unlock(&pmc->powergates_lock); >> + spin_unlock(&pmc->powergates_lock); >> return 0; >> } >> >> @@ -300,7 +300,7 @@ static int tegra_powergate_set(unsigned int id, bool new_state) >> err = readx_poll_timeout(tegra_powergate_state, id, status, >> status == new_state, 10, 100000); >> >> - mutex_unlock(&pmc->powergates_lock); >> + spin_unlock(&pmc->powergates_lock); > > ... the above readx_poll_timeout needs to be converted to the atomic > version. Furthermore, the above 100ms timeout is probably not suited to > spinlock. It's converted in the second patch. Seems mutex indeed could be dropped from tegra_powergate_is_powered, at least for now I can't recall why decided to keep the locking. Thank you for the review, I'll try to drop the mutex and come back with v2 if it will be fine.