This also reorganizes generic_data_done to exploit symmetry between
input and output csets. Actually, we didn't update seq_num or the
timestamp for output, but now we have the current control in sysfs,
so we'd better update it every time.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

Soft timestamping happens at zio_arm_trigger time. This removes
the various software timestamping in the individual trigger.
Moreover, arming is idempotent, so timestamp in the locked section,
only when the trigger is actually armed.

As usual, hardware can overwrite ti timestamps, so this software stamp
is only used for devices that have no better notion of time.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

zio_cset_is_self_timed() only returned true for input channels.  This
is needed, but the name is wrong: use zio_cset_early_arm() instead.

An output channel can be self-timed as well, so trigger_data_done
check the flag directly, not early_arm(). Every self-timed cset
must be rearmed when the trigger fired.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

With the new self-timed output devices, the last raw_io() when the
buffer is empty would fail, so we need to mark the trigger as
not-armed, in order for output to restart when the next push happens.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

If the trigger is enabled on creation, it may fire before the cset is
ready.  This happened to me with "modprobe zio-zero trigger=irq".
For this reason it must start disabled. Also, the create
method MUST set ti->cset before activating the trigger, because the
cset lock is needed to access the enable/disable flag.

The commit adds a WARN() if the field is not set after creation, then
it leaves it zeroed to ensure the system will crash, instead of having
a subtle race.

Also, add a pair of FIXME notes about the need to allow changing
a trigger type while leaving the new one disabled.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

This is a bug fix for overlooks in this same series of commits. I
won't fix each lock when it is introduced before proposing for master,
because it takes too much time.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

This small patch supports self-timed input csets (for example, a TDC)
by ensuring the input trigger is always armed, waiting for the device.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

The function used to be called __zio_internal_data_done, in
heklpers.c.  Being used once only, it was expanded inline by the
compiler, so nothing changes for previous code.

However, if a trigger needs to implement its own data_done method, it
will most likely need to replicate this code, so it's better made
avaiable directly to everybody.  An alternative would be running this
code every time, before or after calling t_op->data_done, but we'd
better let each trigger choose the order of its own operations.

Actually, there is one difference: the input active block is set to NULL
after being stored to the current buffer, so the later implementation
of stop_io (used in zio-irq-tdc.c) won't expose a bug.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

This is the first step of a series of changes aimed at a cleaner and
more flexible management of triggers, as well as a really transparent
"user" trigger.

This commit does the following:

- zio_fire_trigger renamed to zio_arm_trigger: the trigger is
actually armed by software, and then it fires by hardware. This
distinction is especially important for the transparent trigger: when
devices have internal timing, the software trigger must arm it
immediately (as opposed to the dumb devices, where the software trigger
really causes I/O to happen).

- ZIO_TI_BUSY renamed to ZIO_TI_ARMED. Also, ZIO_TI_COMPLETING is
removed. The new, simplified policy is like this: the trigger is armed
by software (the trigger module), but completion is driven by hardware
(the device module).  There is no need for a COMPLETING flag, because
the ARMED flag is only cleared after data_done is over for all channels
in the cset.

- the cset spinlock is used to protect all changes of the ti flags.

- change_current_trigger completely revised and fixed to match new
conventions.

- trigger->abort now takes "ti" instead of "cset" as argument; it is
more natural do do so, and no current trigger implements abort so no
harm is done.

- the trigger->abort and trigger->change_status methods are now
always called while holding the cset spinlock. ZIO core calls abort
when changing the current trigger type, to ensure no pending blocks
are there.

- zio_trigger_abort is renamed to zio_trigger_abort_disable, with
an additional argument to state whether the trigger must be atomically
disabled after the abort is over. This avoids a race condition on
trigger removal. It returns the previous "disabled" bit, to be used
when changing a buffer type while preserving trigger status.

- minor unrelated improvements in error management in objects.c.

- documentation update to match the new locking, and a few typos fixed.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>
Acked-by: Federico Vaga <federico.vaga@gmail.com>

Federico Vaga authored 12 years ago and Alessandro Rubini committed 12 years ago

Signed-off-by: Federico Vaga <federico.vaga@gmail.com>
Acked-by: Alessandro Rubini <rubini@gnudd.com>