There are 2 problems with the old iosf PMIC I2C bus arbritration code which
need to be addressed:
1. The lockdep code complains about a possible deadlock in the
iosf_mbi_[un]block_punit_i2c_access code:
[ 6.712662] ======================================================
[ 6.712673] WARNING: possible circular locking dependency detected
[ 6.712685] 5.3.0-rc2+ #79 Not tainted
[ 6.712692] ------------------------------------------------------
[ 6.712702] kworker/0:1/7 is trying to acquire lock:
[ 6.712712] 00000000df1c5681 (iosf_mbi_block_punit_i2c_access_count_mutex){+.+.}, at: iosf_mbi_unblock_punit_i2c_access+0x13/0x90
[ 6.712739]
but task is already holding lock:
[ 6.712749] 0000000067cb23e7 (iosf_mbi_punit_mutex){+.+.}, at: iosf_mbi_block_punit_i2c_access+0x97/0x186
[ 6.712768]
which lock already depends on the new lock.
[ 6.712780]
the existing dependency chain (in reverse order) is:
[ 6.712792]
-> #1 (iosf_mbi_punit_mutex){+.+.}:
[ 6.712808] __mutex_lock+0xa8/0x9a0
[ 6.712818] iosf_mbi_block_punit_i2c_access+0x97/0x186
[ 6.712831] i2c_dw_acquire_lock+0x20/0x30
[ 6.712841] i2c_dw_set_reg_access+0x15/0xb0
[ 6.712851] i2c_dw_probe+0x57/0x473
[ 6.712861] dw_i2c_plat_probe+0x33e/0x640
[ 6.712874] platform_drv_probe+0x38/0x80
[ 6.712884] really_probe+0xf3/0x380
[ 6.712894] driver_probe_device+0x59/0xd0
[ 6.712905] bus_for_each_drv+0x84/0xd0
[ 6.712915] __device_attach+0xe4/0x170
[ 6.712925] bus_probe_device+0x9f/0xb0
[ 6.712935] deferred_probe_work_func+0x79/0xd0
[ 6.712946] process_one_work+0x234/0x560
[ 6.712957] worker_thread+0x50/0x3b0
[ 6.712967] kthread+0x10a/0x140
[ 6.712977] ret_from_fork+0x3a/0x50
[ 6.712986]
-> #0 (iosf_mbi_block_punit_i2c_access_count_mutex){+.+.}:
[ 6.713004] __lock_acquire+0xe07/0x1930
[ 6.713015] lock_acquire+0x9d/0x1a0
[ 6.713025] __mutex_lock+0xa8/0x9a0
[ 6.713035] iosf_mbi_unblock_punit_i2c_access+0x13/0x90
[ 6.713047] i2c_dw_set_reg_access+0x4d/0xb0
[ 6.713058] i2c_dw_probe+0x57/0x473
[ 6.713068] dw_i2c_plat_probe+0x33e/0x640
[ 6.713079] platform_drv_probe+0x38/0x80
[ 6.713089] really_probe+0xf3/0x380
[ 6.713099] driver_probe_device+0x59/0xd0
[ 6.713109] bus_for_each_drv+0x84/0xd0
[ 6.713119] __device_attach+0xe4/0x170
[ 6.713129] bus_probe_device+0x9f/0xb0
[ 6.713140] deferred_probe_work_func+0x79/0xd0
[ 6.713150] process_one_work+0x234/0x560
[ 6.713160] worker_thread+0x50/0x3b0
[ 6.713170] kthread+0x10a/0x140
[ 6.713180] ret_from_fork+0x3a/0x50
[ 6.713189]
other info that might help us debug this:
[ 6.713202] Possible unsafe locking scenario:
[ 6.713212] CPU0 CPU1
[ 6.713221] ---- ----
[ 6.713229] lock(iosf_mbi_punit_mutex);
[ 6.713239] lock(iosf_mbi_block_punit_i2c_access_count_mutex);
[ 6.713253] lock(iosf_mbi_punit_mutex);
[ 6.713265] lock(iosf_mbi_block_punit_i2c_access_count_mutex);
[ 6.713276]
*** DEADLOCK ***
In practice can never happen because only the first caller which
increments iosf_mbi_block_punit_i2c_access_count will also take
iosf_mbi_punit_mutex, that is the whole purpose of the counter, which
itself is protected by iosf_mbi_block_punit_i2c_access_count_mutex.
But there is no way to tell the lockdep code about this and we really
want to be able to run a kernel with lockdep enabled without these
warnings being triggered.
2. The lockdep warning also points out another real problem, if 2 threads
both are in a block of code protected by iosf_mbi_block_punit_i2c_access
and the first thread to acquire the block exits before the second thread
then the second thread will call mutex_unlock on iosf_mbi_punit_mutex,
but it is not the thread which took the mutex and unlocking by another
thread is not allowed.
Fix this by getting rid of the notion of holding a mutex for the entire
duration of the PMIC accesses, be it either from the PUnit side, or from an
in kernel I2C driver. In general holding a mutex after exiting a function
is a bad idea and the above problems show this case is no different.
Instead 2 counters are now used, one for PMIC accesses from the PUnit
and one for accesses from in kernel I2C code. When access is requested
now the code will wait (using a waitqueue) for the counter of the other
type of access to reach 0 and on release, if the counter reaches 0 the
wakequeue is woken.
Note that the counter approach is necessary to allow nested calls.
The main reason for this is so that a series of i2c transfers can be done
with the punit blocked from accessing the bus the whole time. This is
necessary to be able to safely read/modify/write a PMIC register without
racing with the PUNIT doing the same thing.
Allowing nested iosf_mbi_block_punit_i2c_access() calls also is desirable
from a performance pov since the whole dance necessary to block the PUnit
from accessing the PMIC I2C bus is somewhat expensive.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Link: https://lkml.kernel.org/r/20190812102113.95794-1-hdegoede@redhat.com
00452ba9fd