The tempfile API keeps to-be-cleaned tempfiles in a
singly-linked list and never removes items from the list. A
future patch would like to start removing items, but removal
from a singly linked list is O(n), as we have to walk the
list to find the predecessor element. This means that a
process which takes "n" simultaneous lockfiles (for example,
an atomic transaction on "n" refs) may end up quadratic in
"n".
Before we start allowing items to be removed, it would be
nice to have a way to cover this case in linear time.
The simplest solution is to make an assumption about the
order in which tempfiles are added and removed from the
list. If both operations iterate over the tempfiles in the
same order, then by putting new items at the end of the list
our removal search will always find its items at the
beginning of the list. And indeed, that would work for the
case of refs. But it creates a hidden dependency between
unrelated parts of the code. If anybody changes the ref code
(or if we add a new caller that opens multiple simultaneous
tempfiles) they may unknowingly introduce a performance
regression.
Another solution is to use a better data structure. A
doubly-linked list works fine, and we already have an
implementation in list.h. But there's one snag: the elements
of "struct tempfile" are all marked as "volatile", since a
signal handler may interrupt us and iterate over the list at
any moment (even if we were in the middle of adding a new
entry).
We can declare a "volatile struct list_head", but we can't
actually use it with the normal list functions. The compiler
complains about passing a pointer-to-volatile via a regular
pointer argument. And rightfully so, as the sub-function
would potentially need different code to deal with the
volatile case.
That leaves us with a few options:
1. Drop the "volatile" modifier for the list items.
This is probably a bad idea. I checked the assembly
output from "gcc -O2", and the "volatile" really does
impact the order in which it updates memory.
2. Use macros instead of inline functions. The irony here
is that list.h is entirely implemented as trivial
inline functions. So we basically are already
generating custom code for each call. But sadly there's no
way in C to declare the inline function to take a more
generic type.
We could do so by switching the inline functions to
macros, but it does make the end result harder to read.
And it doesn't fully solve the problem (for instance,
the declaration of list_head needs to change so that
its "prev" and "next" pointers point to other volatile
structs).
3. Don't use list.h, and just make our own ad-hoc
doubly-linked list. It's not that much code to
implement the basics that we need here. But if we're
going to do so, why not add the few extra lines
required to model it after the actual list.h interface?
We can even reuse a few of the macro helpers.
So this patch takes option 3, but actually implements a
parallel "volatile list" interface in list.h, where it could
potentially be reused by other code. This implements just
enough for tempfile.c's use, though we could easily port
other functions later if need be.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The OS X build pulls in sys/queue.h, which pollutes the preprocessor
namespace with a macro generically named LIST_HEAD, and clashes with
the name we use here.
ref: http://mid.gmane.org/FB76544F-16F7-45CA-9649-FD62EE44B0DE@gmail.com
Reported-by: Lars Schneider <larsxschneider@gmail.com>
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Using the a Linux-kernel-derived doubly-linked list
implementation from the Userspace RCU library allows us to
enqueue and delete items from the object request queue in
constant time.
This change reduces enqueue times in the prefetch() function
where object request queue could grow to several thousand
objects.
I left out the list_for_each_entry* family macros from list.h
which relied on the __typeof__ operator as we support platforms
without it. Thus, list_entry (aka "container_of") needs to be
called explicitly inside macro-wrapped for loops.
The downside is this costs us an additional pointer per object
request, but this is offset by reduced overhead on queue
operations leading to improved performance and shorter queue
depths.
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
