Hello Florian Weimer,
I'm the original author of this patch, so I'll try to answer what I can.
For some overall perspective, this patch replaces the original libdw
allocator with a thread-safe variant. The original acts both as a
suballocator (to keep from paying the malloc tax on frequent small
allocations) and a garbage collection list (to free internal structures
on dwarf_end). The patch attempts to replicate the same overall
behavior in the more volatile parallel case.
On Sat, Oct 26, 2019 at 18:14, Florian Weimer <f...@deneb.enyo.de> wrote:
* Mark Wielaard:
I'll see if I can create a case where that is a problem. Then we can
see how to adjust things to use less pthread_keys. Is there a
different
pattern we can use?
It's unclear what purpose thread-local storage serves in this context.
The thread-local storage provides the suballocator side: for each
Dwarf, each thread has its own "top block" to perform allocations from.
To make this simple, each Dwarf has a key to give threads local storage
specific to that Dwarf. Or at least that was the intent, I didn't think
to consider the limit, we didn't run into it in our use cases.
There may be other ways to handle this, I'm high-level considering
potential alternatives (with more atomics, of course). The difficulty
is mostly in providing the same performance in the single-threaded case.
You already have a Dwarf *. I would consider adding some sort of
clone function which creates a shallow Dwarf * with its own embedded
allocator or something like that.
The downside with this is that its an API addition, which we (the
Dyninst + HPCToolkit projects) would need to enforce. Which isn't a
huge deal for us, but I will need to make a case to those teams to make
the shift.
On the upside, it does provide a very understandable semantic in the
case of parallelism. For an API without synchronization clauses, this
would put our work back into the realm of "reasonably correct" (from
"technically incorrect but works.")
This assumes that memory allocation
is actually a performance problem, otherwise you could let malloc
handle the details.
In our (Dyninst + HPCToolkit) work, we have found that malloc tends to
be slow in the multithreaded case, in particular with many small
allocations. The glibc implementation (which most of our clients use)
uses a full mutex to provide thread-safety. While we could do a lot
better in our own projects with regards to memory management, the fact
remains that malloc alone is a notable facet to the performance of
libdw.
Hopefully this helps give a little light on the issue.
-Jonathon