|| ||David Howells <email@example.com>|
|| ||firstname.lastname@example.org, email@example.com|
|| ||[PATCH 0/5] WorkStruct: Shrink work_struct by two thirds|
|| ||Wed, 22 Nov 2006 13:02:22 +0000|
|| ||firstname.lastname@example.org, email@example.com|
The workqueue struct is huge, and this limits it's usefulness. On a 64-bit
architecture it's nearly 100 bytes in size, of which the timer_list is half.
These patches shrink work_struct by 8 of the 12 words it ordinarily consumes.
This is done by:
(1) Splitting the timer out so that delayable work items are defined by a
separate structure which incorporates a basic work_struct and a timer.
(2) Folding the pending bit and wq_data data together
(3) Removing the private data. This can almost always be derived from the
address of the work_struct using container_of() and the selection of the
work function. For the cases where the container of the work_struct may
go away the moment the pending bit is cleared, it is made possible to
defer the release of the structure by deferring the clearing of the
These patches reduce the size of the work_struct thusly:
#WORDS 32-bit arch 64-bit arch
=============== =============== ===============
As is 12 48 bytes 96 bytes
Non-delayable 4 16 bytes 32 bytes
Delayable 10 40 bytes 80 bytes
I've looked through most of the usages of work_structs, and I think that
probably fewer than half the work_structs used actually require delayability,
and I'm not sure that it's absolutely necessary in all cases.
With these patches applied, there are four classes of work item where
previously there was one. These are made up of a combination of the following
(*) Delayable vs Non-delayable.
Delayable work items have their execution deferred for at least a certain
amount of time; non-delayable items are executed as soon as possible.
(*) Auto-release vs Non-auto-release
Ordinarily, the work queue executor would release the work_struct for
further scheduling or deallocation by clearing the pending bit prior to
jumping to the work function. This means that, unless the driver makes
some guarantee itself that the work_struct won't go away, the work
function may not access anything else in the work_struct or its container
lest they be deallocated.. This is a problem if the auxiliary data is
taken away (as done by the last patch).
However, if the pending bit is *not* cleared before jumping to the work
function, then the work function *may* access the work_struct and its
container with no problems. But then the work function must itself
release the work_struct by calling work_release().
In most cases, automatic release is fine, so this is the default. Special
initiators exist for the non-auto-release case.
Note that this is a partial conversion. If these patches are generally
acceptable, then the rest of the kernel will also need modification. I've
tested these patches on my x86_64 testbox only, though I have built
allyesconfig on x86_64, i386 and ppc64 as best I can.
Furthermore, the timer_list struct could possibly be shrunk by 1 word if it
also lost its data member.
Whilst going through all those work items, I noticed that there appears to be a
general misunderstanding as to what cancel_delayed_work() actually does. In
many places it is called on a work_struct that has never had delayed work done
on it, and so is ineffective. I think that people assume that it also removes
the work item from the work queue - which it does not.
Furthermore, it is not generally followed up by a flush of the workqueue, even
though the even may be queued. A further auditing of all cancel_delayed_work()
calls is required. Also, I think that this function should be marked as a
must-check-result function. It's almost certainly an error not to do so.
Removing some of these cancel_delayed_work() calls would also permit some of
the structures they are invoked upon to be converted to work_structs - which
lack the timer and therefore use less space.