This is the second attempt at importing the updated mt_list code (commit
59459ea3). The previous one was attempted with commit c618ed5ff4 ("MAJOR:
import: update mt_list to support exponential back-off") but revealed
problems with QUIC connections and was reverted.
The problem that was faced was that elements deleted inside an iterator
were no longer reset, and that if they were to be recycled in this form,
they could appear as busy to the next user. This was trivially reproduced
with this:
$ cat quic-repro.cfg
global
stats socket /tmp/sock1 level admin
stats timeout 1h
limited-quic
frontend stats
mode http
bind quic4@:8443 ssl crt rsa+dh2048.pem alpn h3
timeout client 5s
stats uri /
$ ./haproxy -db -f quic-repro.cfg &
$ h2load -c 10 -n 100000 --npn h3 https://127.0.0.1:8443/
=> hang
This was purely an API issue caused by the simplified usage of the macros
for the iterator. The original version had two backups (one full element
and one pointer) that the user had to take care of, while the new one only
uses one that is transparent for the user. But during removal, the element
still has to be unlocked if it's going to be reused.
All of this sparked discussions with Fred and Aurélien regarding the still
unclear state of locking. It was found that the lock API does too much at
once and is lacking granularity. The new version offers a much more fine-
grained control allowing to selectively lock/unlock an element, a link,
the rest of the list etc.
It was also found that plenty of places just want to free the current
element, or delete it to do anything with it, hence don't need to reset
its pointers (e.g. event_hdl). Finally it appeared obvious that the
root cause of the problem was the unclear usage of the list iterators
themselves because one does not necessarily expect the element to be
presented locked when not needed, which makes the unlock easy to overlook
during reviews.
The updated version of the list presents explicit lock status in the
macro name (_LOCKED or _UNLOCKED suffixes). When using the _LOCKED
suffix, the caller is expected to unlock the element if it intends to
reuse it. At least the status is advertised. The _UNLOCKED variant,
instead, always unlocks it before starting the loop block. This means
it's not necessary to think about unlocking it, though it's obviously
not usable with everything. A few _UNLOCKED were used at obvious places
(i.e. where the element is deleted and freed without any prior check).
Interestingly, the tests performed last year on QUIC forwarding, that
resulted in limited traffic for the original version and higher bit
rate for the new one couldn't be reproduced because since then the QUIC
stack has gaind in efficiency, and the 100 Gbps barrier is now reached
with or without the mt_list update. However the unit tests definitely
show a huge difference, particularly on EPYC platforms where the EBO
provides tremendous CPU savings.
Overall, the following changes are visible from the application code:
- mt_list_for_each_entry_safe() + 1 back elem + 1 back ptr
=> MT_LIST_FOR_EACH_ENTRY_LOCKED() or MT_LIST_FOR_EACH_ENTRY_UNLOCKED()
+ 1 back elem
- MT_LIST_DELETE_SAFE() no longer needed in MT_LIST_FOR_EACH_ENTRY_UNLOCKED()
=> just manually set iterator to NULL however.
For MT_LIST_FOR_EACH_ENTRY_LOCKED()
=> mt_list_unlock_self() (if element going to be reused) + NULL
- MT_LIST_LOCK_ELT => mt_list_lock_full()
- MT_LIST_UNLOCK_ELT => mt_list_unlock_full()
- l = MT_LIST_APPEND_LOCKED(h, e); MT_LIST_UNLOCK_ELT();
=> l=mt_list_lock_prev(h); mt_list_lock_elem(e); mt_list_unlock_full(e, l)
The test scans the whole number space in 32 bits and compares the different
functions with the reference that does one byte at a time. In 64-bit mode,
it picks 2^32 64-bit random numbers and tests that they the 64-bit functions
all produce the expected results when submitted such numbers.
It optionally takes an initial offset and step so that it can run on
multiple cores (or even machines), though the test is reasonably fast
on modern machines, around 10s per core.
This reverts commit c618ed5ff4.
The list iterator is broken. As found by Fred, running QUIC single-
threaded shows that only the first connection is accepted because the
accepter relies on the element being initialized once detached (which
is expected and matches what MT_LIST_DELETE_SAFE() used to do before).
However while doing this in the quic_sock code seems to work, doing it
inside the macro show total breakage and the unit test doesn't work
anymore (random crashes). Thus it looks like the fix is not trivial,
let's roll this back for the time it will take to fix the loop.
The new mt_list code supports exponential back-off on conflict, which
is important for use cases where there is contention on a large number
of threads. The API evolved a little bit and required some updates:
- mt_list_for_each_entry_safe() is now in upper case to explicitly
show that it is a macro, and only uses the back element, doesn't
require a secondary pointer for deletes anymore.
- MT_LIST_DELETE_SAFE() doesn't exist anymore, instead one just has
to set the list iterator to NULL so that it is not re-inserted
into the list and the list is spliced there. One must be careful
because it was usually performed before freeing the element. Now
instead the element must be nulled before the continue/break.
- MT_LIST_LOCK_ELT() and MT_LIST_UNLOCK_ELT() have always been
unclear. They were replaced by mt_list_cut_around() and
mt_list_connect_elem() which more explicitly detach the element
and reconnect it into the list.
- MT_LIST_APPEND_LOCKED() was only in haproxy so it was left as-is
in list.h. It may however possibly benefit from being upstreamed.
This required tiny adaptations to event_hdl.c and quic_sock.c. The
test case was updated and the API doc added. Note that in order to
keep include files small, the struct mt_list definition remains in
list-t.h (par of the internal API) and was ifdef'd out in mt_list.h.
A test on QUIC with both quictls 1.1.1 and wolfssl 5.6.3 on ARM64 with
80 threads shows a drastic reduction of CPU usage thanks to this and
the refined memory barriers. Please note that the CPU usage on OpenSSL
3.0.9 is significantly higher due to the excessive use of atomic ops
by openssl, but 3.1 is only slightly above 1.1.1 though:
- before: 35 Gbps, 3.5 Mpps, 7800% CPU
- after: 41 Gbps, 4.2 Mpps, 2900% CPU
The current "ADD" vs "ADDQ" is confusing because when thinking in terms
of appending at the end of a list, "ADD" naturally comes to mind, but
here it does the opposite, it inserts. Several times already it's been
incorrectly used where ADDQ was expected, the latest of which was a
fortunate accident explained in 6fa922562 ("CLEANUP: stream: explain
why we queue the stream at the head of the server list").
Let's use more explicit (but slightly longer) names now:
LIST_ADD -> LIST_INSERT
LIST_ADDQ -> LIST_APPEND
LIST_ADDED -> LIST_INLIST
LIST_DEL -> LIST_DELETE
The same is true for MT_LISTs, including their "TRY" variant.
LIST_DEL_INIT keeps its short name to encourage to use it instead of the
lazier LIST_DELETE which is often less safe.
The change is large (~674 non-comment entries) but is mechanical enough
to remain safe. No permutation was performed, so any out-of-tree code
can easily map older names to new ones.
The list doc was updated.
The code that is there to run some unit tests on some internal features
was moved to tests/unit. Ideally it should be buildable from the main
makefile though this is not yet the case.
The code that is kept for experimentation purposes (hashes, syscall
optimization etc) as well as some captures of the results was moved
to tests/exp.
A few totally obsolete files which couldn't build anymore and were
not relevant to current versions were removed.
Willy Tarreau
Amaury Denoyelle