| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Originally, I envisioned bthreads as replacement for pollers and
workqueues. But even without preemption, having functions you call
possibly accessing structures you are iterating over can corrupt
memory. This problem exists with pollers as well, but because of their
limited scope, it's harder to shoot your foot with them, as you
don't keep implicit state between poller activations unlike bthreads,
which maintain their stack across context switches.
Limit bthread scope instead to be a replacement for workqueues. This
still allows us to port some classes of state-machine-in-kthread kernel
code, while avoding the aforementioned pitfalls.
Cc: Jan Luebbe <j.luebbe@pengutronix.de>
Cc: Sascha Hauer <s.hauer@pengutronix.de>
Cc: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
Link: https://lore.barebox.org/20210622082617.18011-5-a.fatoum@pengutronix.de
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Workqueues are run out of poller_call, not because of a dependency, but
because when they were added, poller_call was directly called from
is_timeout.
With the addition of bthreads, there is now a general resched() function
that runs pollers and switches between bthreads. It makes sense to move
workqueue handling there as well to keep scheduling matter contained in
a single function. Do so.
Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
Link: https://lore.barebox.org/20210622082617.18011-4-a.fatoum@pengutronix.de
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
With the new setjmp/longjmp/initjmp support, we have all the
architecture support in place to have suspendable green
threads in barebox. These are expected to replace pollers and
workqueues. For now we still have a differentiation between
the main and secondary threads. The main thread is allowed
I/O access unconditionally. If it's in a delay loop, a secondary
thread running needs to be wary of not entering the same driver
and doing hardware manipulation. We already have slices as
mechanism to guard against this, but they aren't used as widely
as needed.
Preferably, in the end, threads will automatically yield until
they can claim a resource (i.e. lock a mutex). Until we are there,
take the same care when using bthreads as with pollers.
Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|
|
|
|
|
|
|
|
|
| |
Filesystem operations possibly call into arbitrary devices, so shouldn't
be used from a poller. This patch sprinkles some WARN_ONCE() when this
happens. One exception is when the file which is accessed is on ramfs
which doesn't have any dependencies to devices.
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Some code wants to run arbitrary code in the background, examples are
fastboot and ratp. Currently this is implemented in pollers. The problem
with this is that pollers are executed whenever is_timeout() is called
which may happen anywhere in the code. With this we can never tell which
resources are currently in use when the poller is executed.
This adds a work queue interface which is specifically designed to
trigger doing work in a context where it's safe to run arbitrary commands.
Code in pollers can attach work to a work queue which is at that time
only queued up. A new slice, the command slice, is added which by
default is acquired. It is only released at places where the shell waits
for input. When during this time pollers are executed the queued up
works are done before running the registered pollers.
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|
|
slices, the barebox idea of locking
barebox has pollers which execute code in the background whenever one of the
delay functions (udelay, mdelay, ...) or is_timeout() are called. This
introduces resource problems when some device triggers a poller by calling
a delay function and then the poller code calls into the same device again.
As an example consider a I2C GPIO expander which drives a LED which shall
be used as a heartbeat LED:
poller -> LED on/off -> GPIO high/low -> I2C transfer
The I2C controller has a timeout loop using is_timeout() and thus can trigger
a poller run. With this the following can happen during an unrelated I2C
transfer:
I2C transfer -> is_timeout() -> poller -> LED on/off -> GPIO high/low -> I2C transfer
We end up with issuing an I2C transfer during another I2C transfer and
things go downhill.
Due to the lack of interrupts we can't do real locking in barebox. We use
a mechanism called slices instead. A slice describes a resource to which
other slices can be attached. Whenever a slice is needed it must be acquired.
Acquiring a slice never fails, it just increases the acquired counter of
the slice and its dependent slices. when a slice shall be used inside a
poller it must first be tested if the slice is already in use. If it is,
we can't do the operation on the slice now and must return and hope that
we have more luck in the next poller call.
slices can be attached other slices as dependencies. In the example above
LED driver would add a dependency to the GPIO controller and the GPIO driver
would add a dependency to the I2C bus:
GPIO driver probe:
slice_add(&gpio->slice, i2c_device_slice(i2cdev));
LED driver probe:
slice_add(&led->slice, gpio_slice(gpio));
The GPIO code would call slice_acquire(&gpio->slice) before doing any
operation on the GPIO chip providing this GPIO, likewise the I2C core
would call slice_acquire(&i2cbus->slice) before doing an operation on
this I2C bus.
The heartbeat poller code would call slice_acquired(led_slice(led)) and
only continue when the slice is not acquired.
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
|