Re: [PATCH v3 lora-next 5/5] net: lora: sx125x sx1301: allow radio to register as a clk provider

Am 31.12.18 um 23:56 schrieb Andreas Färber:

Am 31.12.18 um 18:50 schrieb Mark Brown:

quoted

On Sun, Dec 30, 2018 at 11:55:46AM +0100, Andreas Färber wrote:

quoted

Given that observed symptoms were CPU stalls, workqueue hangs and RCU
problems, requiring a power-cycle to recover, I wonder whether we are
running into some atomic/locking issue with clk_enable()? Is it valid at
all to use SPI/regmap for clk_enable()? If it is, is there a known issue
specific to spi-sun6i (A64) in 4.20.0?
I already tried setting .disable_locking = true in both regmap_configs.
Any suggestions how to further debug?

You can't use SPI for clk_enable(), clk_enable() needs to be doable in
atomic context since we need to wait for the bus operations to complete
(you can start SPI transfers in atomic context but you still need to
wait for them to complete).  Any clocks that are only accessible via a
slow bus like I2C or SPI need to do the enable/disable in the
prepare/unprepare operations which aren't done in atomic context.

regmap can be used in atomic contexts, though you need to configure it
to use spinlocks instead of mutexes and ensure that no register cache
allocations happen during I/O (eg, by providing defaults for all
registers or by not using a cache).

We have .cache_type = REGCACHE_NONE on both bus and spi regmap_configs.

I moved the regmap_field_write() from .enable to .prepare and set
.fast_io = true on both regmap_configs to force using spinlocks, but
same hang as in .enable before...

And same if I set .disable_locking = true on both.

Given that it works with one SPI driver and not with the other,
independent of the locking options applied, I assume my symptoms are not
a regmap-layer issue.

Is it allowed during a .prepare operation to call the mentioned
clk_get_rate(), which ends up calling clk_prepare_lock()?

According to my debug output in spi-sun6i.c our hanging
regmap_field_write() ends up calling sun6i_transfer_one() three times,
the first two look okay, but the third one doesn't make it past the
clk_get_rate() [...].

SysRq still works in that state! Attached is SysRq-w output.
(still with .disable_locking = true in both regmap_configs)

In the very bottom you see the "ip" task, at wait_for_completion() from
__spi_sync().
I trigger this issue with `ip link set lora2 up`, so that looks okay.

Then there's a "spi1" task at clk_prepare_lock()' mutex_lock() coming
from spi_pump_messages().
The reason for that will be that clk_prepare_lock()'s mutex_trylock()
failed (because we're holding the prepare_lock from clk_prepare_enable()
in the "ip" task) and that the prepare_owner == current check fails for
this separate task_struct, too.

So, the third invocation of sun6i_transfer_one() calling clk_get_rate()
hangs at the prepare_lock instead of reference-counting, because it runs
from a separate kthread, unlike the two previous calls?

Besides, there's also an mmc_rescan workqueue task at clk_prepare_lock()
coming from sunxi_mmc_enable() due to pm_generic_runtime_resume().
My rootfs is on microSD card.

I did not find any *regmap_init_spi() based example in drivers/clk/, and
all other "spi" mentions in drivers/clk/ appeared to be clock names.
The closest was devm_regmap_init_i2c() based clk-cdce706.c, which uses
the prepare/unprepare ops, as suggested by Mark, and does
regmap_update_bits() from there.

A quick grep in drivers/i2c/ does not find any mention of "kthread", so
probably that's the breaking difference?

Regards,
Andreas

-- 
SUSE Linux GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)