In practice it's all pollers except select(). It turns out that we're
keeping some legacy code only for select and enforcing it on all
pollers, let's offer the pollers the ability to declare that they
do not need that.
There's currently a big ambiguity on our use of POLLHUP because we
currently map POLLHUP and POLLRDHUP to FD_POLL_HUP. The first one
indicates a close in *both* directions while the second one indicates
a unidirectional close. Since we don't know from the resulting flag
we always have to read when reported. Furthermore kqueue only reports
unidirectional responses which are mapped to FD_POLL_HUP as well, and
their write closes are mapped to a general error.
We could add a new FD_POLL_RDHUP flag to improve the mapping, or
switch only to the POLL* flags, but that further complicates the
portability for operating systems like FreeBSD which do not have
POLLRDHUP but have its semantics.
Let's instead directly use the per-direction flag values we already
have, and it will be a first step in the direction of finer states.
Thus we introduce an ERR and a SHUT status for each direction, that
the pollers will be able to compute and pass to fd_update_events().
It's worth noting that FD_EV_STATUS already sees the two new flags,
but they are harmless since used only by fd_{recv,send}_state() which
are never called. Thus in its current state this patch must be totally
transparent.
Now that we don't have to update FD_EV_POLLED_* at the same time as
FD_EV_ACTIVE_*, we don't need to use a CAS anymore, a bit-test-and-set
operation is enough. Doing so reduces the code size by a bit more than
1 kB. One function was special, fd_done_recv(), whose comments and doc
were inaccurate for the part related to the lack of polling.
Since commit 7ac0e35f2 in 1.9-dev1 ("MAJOR: fd: compute the new fd polling
state out of the fd lock") we've started to update the FD POLLED bit a
bit more aggressively. Lately with the removal of the FD cache, this bit
is always equal to the ACTIVE bit. There's no point continuing to watch
it and update it anymore, all it does is create confusion and complicate
the code. One interesting side effect is that it now becomes visible that
all fd_*_{send,recv}() operations systematically call updt_fd_polling(),
except fd_cant_recv()/fd_cant_send() which never saw it change.
The purpose is to be able to remember that initialization was already
done for a file descriptor. This will allow to get rid of some dirty
hacks performed in the logs or fd sinks where the init state of the
fd has to be guessed.
The "cache" entry was still present in the fdtab struct and it was
reported in "show sess". Removing it broke the cache-line alignment
on 64-bit machines which is important for threads, so it was fixed
by adding an attribute(aligned()) when threads are in use. Doing it
only in this case allows 32-bit thread-less platforms to see the
struct fit into 32 bytes.
Currently both logs and event sinks may use a file descriptor to
atomically emit some output contents. The two may use the same FD though
nothing is done to make sure they use the same lock. Also there is quite
some redundancy between the two. Better make a specific function to send
a fragmented message to a file descriptor which will take care of the
locking via the fd's lock. The function is also able to truncate a
message and to enforce addition of a trailing LF when building the
output message.
While experimenting with potentially improved fairness and latency using
ticket locks on a Ryzen 16-thread/8-core, a very strange situation happened
a lot for some levels of traffic. Around 300k connections per second, no
more connections would be accepted on the multi-threaded listener but all
others would continue to work fine. All attempts to trace showed that the
threads were all in the trylock in the fd cache, or in the spinlock of
fd_update_events(), or in the one of fd_may_recv(). But as indicated this
was not a deadlock since the process continues to work fine.
After quite some investigation it appeared that the issue is caused by a
lack of fairness between the fdcache's trylock and these functions' spin
locks above. In fact, regardless of the success or failure of the fdcache's
attempt at grabbing the lock, the poller was calling fd_update_events()
which locks the FD once for something that can be done with a CAS, and
then calls fd_may_recv() with another lock for something that most often
didn't change. The high contention on these spinlocks leaves no chance to
any other thread to grab the lock using trylock(), and once this happens,
there is no thread left to process incoming connection events nor to stop
polling on the FD, leaving all threads at 100% CPU but partially operational.
This patch addresses the issue by using bit-test-and-set instead of the OR
in fd_may_recv() / fd_may_send() so that nothing is done if the FD was
already configured as expected. It does the same in fd_update_events()
using a CAS to check if the FD's events need to be changed at all or not.
With this patch applied, it became impossible to reproduce the issue, and
now there's no way to saturate all 16 CPUs with the load used for testing,
as no more than 1350-1400 were noticed at 300+kcps vs 1600.
Ideally this patch should go further and try to remove the remaining
incarnations of the fdlock as this seems possible, but it's difficult
enough to be done in a distinct patch that will not have to be backported.
It is possible that workloads involving a high connection rate may slightly
benefit from this patch and observe a slightly lower CPU usage even when
the service doesn't misbehave.
This patch must be backported to 2.0 and 1.9.
We have been abusing the do_poll()'s timeout for a while, making it zero
whenever there is some known activity. The problem this poses is that it
complicates activity diagnostic by incrementing the poll_exp field for
each known activity. It also requires extra computations that could be
avoided.
This change passes a "wake" argument to say that the poller must not
sleep. This simplifies the operations and allows one to differenciate
expirations from activity.
The polled_mask is only used in the pollers, and removing it from the
struct fdtab makes it fit in one 64B cacheline again, on a 64bits machine,
so make it a separate array.
With the old model, any fd shared by multiple threads, such as listeners
or dns sockets, would only be updated on one threads, so that could lead
to missed event, or spurious wakeups.
To avoid this, add a global list for fd that are shared, using the same
implementation as the fd cache, and only remove entries from this list
when every thread as updated its poller.
[wt: this will need to be backported to 1.8 but differently so this patch
must not be backported as-is]
An fd cache entry might be removed and added at the end of the list, while
another thread is parsing it, if that happens, we may miss fd cache entries,
to avoid that, add a new field in the struct fdtab, "added_mask", which
contains a mask for potentially affected threads, if it is set, the
corresponding thread will set its bit in fd_cache_mask, to avoid waiting in
poll while it may have more work to do.
Create a local, per-thread, fdcache, for file descriptors that only belongs
to one thread, and make the global fd cache mostly lockless, as we can get
a lot of contention on the fd cache lock.
Since only select() and poll() still make use of maxfd, let's move
its computation right there in the pollers themselves, and only
during each fd update pass. The computation doesn't need a lock
anymore, only a few atomic ops. It will be accurate, be done much
less often and will not be required anymore in the FD's fast patch.
This provides a small performance increase of about 1% in connection
rate when using epoll since we get rid of this computation which was
performed under a lock.
Some pollers like epoll() need to know if the fd is already known or
not in order to compute the operation to perform (add, mod, del). For
now this is performed based on the difference between the previous FD
state and the new state but this will not be usable anymore once threads
become responsible for their own polling.
Here we come with a different approach : a bitmask is stored with the
fd to indicate which pollers already know it, and the pollers will be
able to simply perform the add/mod/del operations based on this bit
combined with the new state.
This patch only adds the bitmask declaration and initialization, it
is it not yet used. It will be needed by the next two fixes and will
need to be backported to 1.8.
Since the fd update tables are per-thread, we need to have a bit per
thread to indicate whether an update exists, otherwise this can lead
to lost update events every time multiple threads want to update the
same FD. In practice *for now*, it only happens at start time when
listeners are enabled and ask for polling after facing their first
EAGAIN. But since the pollers are still shared, a lost event is still
recovered by a neighbor thread. This will not reliably work anymore
with per-thread pollers, where it has been observed a few times on
startup that a single-threaded listener would not always accept
incoming connections upon startup.
It's worth noting that during this code review it appeared that the
"new" flag in the fdtab isn't used anymore.
This fix should be backported to 1.8.
The struct is not cache line aligned but at least, every time the lock
will appear in the same cache line as the fd it will benefit from being
accessed first. This improves the performance by about 2% on fd-intensive
workloads with 4 threads.
This macro should be used to declare variables or struct members depending on
the USE_THREAD compile option. It avoids the encapsulation of such declarations
between #ifdef/#endif. It is used to declare all lock variables.
It was a leftover from the last cleaning session; this mask applies
to threads and calling it process_mask is a bit confusing. It's the
same in fd, task and applets.
Many changes have been made to do so. First, the fd_updt array, where all
pending FDs for polling are stored, is now a thread-local array. Then 3 locks
have been added to protect, respectively, the fdtab array, the fd_cache array
and poll information. In addition, a lock for each entry in the fdtab array has
been added to protect all accesses to a specific FD or its information.
For pollers, according to the poller, the way to manage the concurrency is
different. There is a poller loop on each thread. So the set of monitored FDs
may need to be protected. epoll and kqueue are thread-safe per-se, so there few
things to do to protect these pollers. This is not possible with select and
poll, so there is no sharing between the threads. The poller on each thread is
independant from others.
Finally, per-thread init/deinit functions are used for each pollers and for FD
part for manage thread-local ressources.
Now, you must be carefull when a FD is created during the HAProxy startup. All
update on the FD state must be made in the threads context and never before
their creation. This is mandatory because fd_updt array is thread-local and
initialized only for threads. Because there is no pollers for the main one, this
array remains uninitialized in this context. For this reason, listeners are now
enabled in run_thread_poll_loop function, just like the worker pipe.
We'll need to differenciate between pollers which can report hangup at
the same time as read (POLL_RDHUP) from the other ones, because only
these ones may benefit from the fd_done_recv() optimization. Epoll has
had support for EPOLLRDHUP since Linux 2.6.17 and has always been used
this way in haproxy, so now we only set the flag once we've observed it
once in a response. It means that some initial requests may try to
perform a second recv() call, but after the first closed connection it
will be enough to know that the second call is not needed anymore.
Later we may extend these flags to designate event-triggered pollers.
Setting an FD to -1 when closed isn't the most easily noticeable thing
to do when we're chasing accidental reuse of a stale file descriptor.
Instead set it to that large a negative value that it will overflow the
fdtab and provide an analysable core at the moment the issue happens.
Care was taken to ensure it doesn't overflow nor change sign on 32-bit
machines when multiplied by fdtab, and that it also remains negative for
the various checks that exist. The value equals 0xFDDEADFD which happens
to be easily spotted in a debugger.
There's quite some inconsistency in the internal API. listener_accept()
which is the main accept() function returns void but is declared as int
in the include file. It's assigned to proto->accept() for all stream
protocols where an int is expected but the result is never checked (nor
is it documented by the way). This proto->accept() is in turn assigned
to fd->iocb() which is supposed to return an int composed of FD_WAIT_*
flags, but which is never checked either.
So let's fix all this mess :
- nobody checks accept()'s return
- nobody checks iocb()'s return
- nobody sets a return value
=> let's mark all these functions void and keep the current ones intact.
Additionally we now include listener.h from listener.c to ensure we won't
silently hide this incoherency in the future.
Note that this patch could/should be backported to 1.6 and even 1.5 to
simplify debugging sessions.
When run in daemon mode (i.e. with at least one forked process) and using
the epoll poller, sending USR1 (graceful shutdown) to the worker processes
can cause some workers to start running at 100% CPU. Precondition is having
an established HTTP keep-alive connection when the signal is received.
The cloned (during fork) listening sockets do not get closed in the parent
process, thus they do not get removed from the epoll set automatically
(see man 7 epoll). This can lead to the process receiving epoll events
that it doesn't feel responsible for, resulting in an endless loop around
epoll_wait() delivering these events.
The solution is to explicitly remove these file descriptors from the epoll
set. To not degrade performance, care was taken to only do this when
neccessary, i.e. when the file descriptor was cloned during fork.
Signed-off-by: Conrad Hoffmann <conrad@soundcloud.com>
[wt: a backport to 1.4 could be studied though chances to catch the bug are low]
This commit heavily changes the polling system in order to definitely
fix the frequent breakage of SSL which needs to remember the last
EAGAIN before deciding whether to poll or not. Now we have a state per
direction for each FD, as opposed to a previous and current state
previously. An FD can have up to 8 different states for each direction,
each of which being the result of a 3-bit combination. These 3 bits
indicate a wish to access the FD, the readiness of the FD and the
subscription of the FD to the polling system.
This means that it will now be possible to remember the state of a
file descriptor across disable/enable sequences that generally happen
during forwarding, where enabling reading on a previously disabled FD
would result in forgetting the EAGAIN flag it met last time.
Several new state manipulation functions have been introduced or
adapted :
- fd_want_{recv,send} : enable receiving/sending on the FD regardless
of its state (sets the ACTIVE flag) ;
- fd_stop_{recv,send} : stop receiving/sending on the FD regardless
of its state (clears the ACTIVE flag) ;
- fd_cant_{recv,send} : report a failure to receive/send on the FD
corresponding to EAGAIN (clears the READY flag) ;
- fd_may_{recv,send} : report the ability to receive/send on the FD
as reported by poll() (sets the READY flag) ;
Some functions are used to report the current FD status :
- fd_{recv,send}_active
- fd_{recv,send}_ready
- fd_{recv,send}_polled
Some functions were removed :
- fd_ev_clr(), fd_ev_set(), fd_ev_rem(), fd_ev_wai()
The POLLHUP/POLLERR flags are now reported as ready so that the I/O layers
knows it can try to access the file descriptor to get this information.
In order to simplify the conditions to add/remove cache entries, a new
function fd_alloc_or_release_cache_entry() was created to be used from
pollers while scanning for updates.
The following pollers have been updated :
ev_select() : done, built, tested on Linux 3.10
ev_poll() : done, built, tested on Linux 3.10
ev_epoll() : done, built, tested on Linux 3.10 & 3.13
ev_kqueue() : done, built, tested on OpenBSD 5.2
We're completely changing the way FDs will be polled. There will be no
more speculative I/O since we'll know the exact FD state, so these will
only be cached events.
First, let's fix a few field names which become confusing. "spec_e" was
used to store a speculative I/O event state. Now we'll store the whole
R/W states for the FD there. "spec_p" was used to store a speculative
I/O cache position. Now let's clearly call it "cache".
We're completely changing the way FDs will be polled. First, let's fix
a few field names which become confusing. "spec_e" was used to store a
speculative I/O event state. Now we'll store the whole R/W states for
the FD there.
Right now we see many places doing their own setsockopt(SO_LINGER).
Better only do it just before the close() in fd_delete(). For this
we add a new flag on the file descriptor, indicating if it's safe or
not to linger. If not (eg: after a connect()), then the setsockopt()
call is automatically performed before a close().
The flag automatically turns to safe when receiving a read0.
At the moment sepoll is not 100% event-driven, because a call to fd_set()
on an event which is already being polled will not change its state.
This causes issues with OpenSSL because if some I/O processing is interrupted
after clearing the I/O event (eg: read all data from a socket, can't put it
all into the buffer), then there is no way to call the SSL_read() again once
the buffer releases some space.
The only real solution is to go 100% event-driven. The principle is to use
the spec list as an event cache and that each time an I/O event is reported
by epoll_wait(), this event is automatically scheduled for addition to the
spec list for future calls until the consumer explicitly asks for polling
or stopping.
Doing this is a bit tricky because sepoll used to provide a substantial
number of optimizations such as event merging. These optimizations have
been maintained : a dedicated update list is affected when events change,
but not the event list, so that updates may cancel themselves without any
side effect such as displacing events. A specific case was considered for
handling newly created FDs as soon as they are detected from within the
poll loop. This ensures that their read or write operation will always be
attempted as soon as possible, thus reducing the number of poll loops and
process_session wakeups. This is especially true for newly accepted fds
which immediately perform their first recv() call.
Two new flags were added to the fdtab[] struct to tag the fact that a file
descriptor already exists in the update list. One flag indicates that a
file descriptor is new and has just been created (fdtab[].new) and the other
one indicates that a file descriptor is already referenced by the update list
(fdtab[].updated). Even if the FD state changes during operations or if the
fd is closed and replaced, it's not an issue because the update flag remains
and is easily spotted during list walks. The flag must absolutely reflect the
presence of the fd in the update list in order to avoid overflowing the update
list with more events than there are distinct fds.
Note that this change also recovers the small performance loss introduced
by its connection counter-part and goes even beyond.
It appears that fd.h includes a number of unneeded files and was
included from standard.h, and as such served as an intermediary
to provide almost everything to everyone.
By removing its useless includes, a long dependency chain broke
but could easily be fixed.
The "spec" sub-struct was using 8 bytes for only 5 needed. There is no
reason to keep it as a struct, it doesn't bring any value. By flattening
it, we can merge the single byte with the next single byte, resulting in
an immediate saving of 4 bytes (20%). Interestingly, tests have shown a
steady performance gain of 0.6% after this change, which can possibly be
attributed to a more cache-line friendly struct.
These flags were added for TCP_CORK. They were only set at various places
but never checked by any user since TCP_CORK was replaced with MSG_MORE.
Simply get rid of this now.
The old EV_FD_SET() macro was confusing, as it would enable receipt but there
was no way to indicate that EAGAIN was received, hence the recently added
FD_WAIT_* flags. They're not enough as we're still facing a conflict between
EV_FD_* and FD_WAIT_*. So let's offer I/O functions what they need to explicitly
request polling.
These primitives were initially introduced so that callers were able to
conditionally set/disable polling on a file descriptor and check in return
what the state was. It's been long since we last had an "if" on this, and
all pollers' functions were the same for cond_* and their systematic
counter parts, except that this required a check and a specific return
value that are not always necessary.
So let's simplify the FD API by removing this now unused distinction and
by making all specific functions return void.
fdtab[].state was only used to know whether a connection was in progress
or an error was encountered. Instead we now use connection->flags to store
a flag for both. This way, connection management will be able to update the
connection status on I/O.
In an attempt to get rid of fdtab[].state, and to move the relevant
parts to the connection struct, we remove the FD_STCLOSE state which
can easily be deduced from the <owner> pointer as there is a 1:1 match.
The destination address is purely a connection thing and not an fd thing.
It's also likely that later the address will be stored into the connection
and linked to by the SI.
struct fdinfo only keeps the pointer to the port range and the local port
for now. All of this also needs to move to the connection but before this
the release of the port range must move from fd_delete() to a new function
dedicated to the connection.
These pointers were used to hold pointers to buffers in the past, but
since we introduced the stream interface, they're no longer used but
they were still sometimes set.
Removing them shrink the struct fdtab from 32 to 24 bytes on 32-bit machines,
and from 52 to 36 bytes on 64-bit machines, which is a significant saving. A
quick tests shows a steady 0.5% performance gain, probably due to the better
cache efficiency.
The fd_list[] used by sepoll was indexed on the fd number and was only
used to store the equivalent of an integer. Changing it to be merged
with fdtab reduces the number of pointer computations, the code size
and some initialization steps. It does not harm other pollers much
either, as only one integer was added to the fdtab array.
Some rarely information are stored in fdtab, making it larger for no
reason (source port ranges, remote address, ...). Such information
lie there because the checks can't find them anywhere else. The goal
will be to move these information to the stream interface once the
checks make use of it.
For now, we move them to an fdinfo array. This simple change might
have improved the cache hit ratio a little bit because a 0.5% of
performance increase has measured.
