Instead of fiddling with before_poll and after_poll in
activity_count_runtime(), the function is now called by
clock_entering_poll() which passes it the number of microseconds
spent working. This allows to remove all calls to
activity_count_runtime() from the pollers.
The entering_poll/leaving_poll/measure_idle functions that were hard
to classify and used to move to various locations have now been placed
into clock.c since it's precisely about time-keeping. The functions
were renamed to clock_*. The samp_time and idle_time values are now
static since there is no reason for them to be read from outside.
There is currently a problem related to time keeping. We're mixing
the functions to perform calculations with the os-dependent code
needed to retrieve and adjust the local time.
This patch extracts from time.{c,h} the parts that are solely dedicated
to time keeping. These are the "now" or "before_poll" variables for
example, as well as the various now_*() functions that make use of
gettimeofday() and clock_gettime() to retrieve the current time.
The "tv_*" functions moved there were also more appropriately renamed
to "clock_*".
Other parts used to compute stolen time are in other files, they will
have to be picked next.
time.h is a horrible place to put activity calculation, it's a
historical mistake because the functions were there. We already have
most of the parts in sched.{c,h} and these ones make an exception in
the middle, forcing time.h to include some thread stuff and to access
the before/after_poll and idle_pct values.
Let's move these 3 functions to task.h with the other ones. They were
prefixed with "sched_" instead of the historical "tv_" which already
made no sense anymore.
The current principle of running under isolation was made to access
sensitive data while being certain that no other thread was using them
in parallel, without necessarily having to place locks everywhere. The
main use case are "show sess" and "show fd" which run over long chains
of pointers.
The thread_isolate() call relies on the "harmless" bit that indicates
for a given thread that it's not currently doing such sensitive things,
which is advertised using thread_harmless_now() and which ends usings
thread_harmless_end(), which also waits for possibly concurrent threads
to complete their work if they took this opportunity for starting
something tricky.
As some system calls were notoriously slow (e.g. mmap()), a bunch of
thread_harmless_now() / thread_harmless_end() were placed around them
to let waiting threads do their work while such other threads were not
able to modify memory contents.
But this is not sufficient for performing memory modifications. One such
example is the server deletion code. By modifying memory, it not only
requires that other threads are not playing with it, but are not either
in the process of touching it. The fact that a pool_alloc() or pool_free()
on some structure may call thread_harmless_now() and let another thread
start to release the same object's memory is not acceptable.
This patch introduces the concept of "idle threads". Threads entering
the polling loop are idle, as well as those that are waiting for all
others to become idle via the new function thread_isolate_full(). Once
thread_isolate_full() is granted, the thread is not idle anymore, and
it is released using thread_release() just like regular isolation. Its
users have to keep in mind that across this call nothing is granted as
another thread might have performed shared memory modifications. But
such users are extremely rare and are actually expecting this from their
peers as well.
Note that that in case of backport, this patch depends on previous patch:
MINOR: threads: make thread_release() not wait for other ones to complete
This function already performs a number of checks prior to calling the
IOCB, and detects the change of thread (FD migration). Half of the
controls are still in each poller, and these pollers also maintain
activity counters for various cases.
Note that the unreliable test on thread_mask was removed so that only
the one performed by fd_set_running() is now used, since this one is
reliable.
Let's centralize all that fd-specific logic into the function and make
it return a status among:
FD_UPDT_DONE, // update done, nothing else to be done
FD_UPDT_DEAD, // FD was already dead, ignore it
FD_UPDT_CLOSED, // FD was closed
FD_UPDT_MIGRATED, // FD was migrated, ignore it now
Some pollers already used to call it last and have nothing to do after
it, regardless of the result. epoll has to delete the FD in case a
migration is detected. Overall this removes more code than it adds.
This one only contains the list of per-thread epoll FDs, and is used
a lot during updates. Let's mark it read_mostly to avoid false sharing
of FDs placed at the extremities.
This patch replaces roughly all occurrences of an HA_ATOMIC_ADD(&foo, 1)
or HA_ATOMIC_SUB(&foo, 1) with the equivalent HA_ATOMIC_INC(&foo) and
HA_ATOMIC_DEC(&foo) respectively. These are 507 changes over 45 files.
This makes the code more readable and less prone to copy-paste errors.
In addition, it allows to place some __builtin_constant_p() predicates
to trigger a link-time error in case the compiler knows that the freed
area is constant. It will also produce compile-time error if trying to
free something that is not a regular pointer (e.g. a function).
The DEBUG_MEM_STATS macro now also defines an instance for ha_free()
so that all these calls can be checked.
178 occurrences were converted. The vast majority of them were handled
by the following Coccinelle script, some slightly refined to better deal
with "&*x" or with long lines:
@ rule @
expression E;
@@
- free(E);
- E = NULL;
+ ha_free(&E);
It was verified that the resulting code is the same, more or less a
handful of cases where the compiler optimized slightly differently
the temporary variable that holds the copy of the pointer.
A non-negligible amount of {free(str);str=NULL;str_len=0;} are still
present in the config part (mostly header names in proxies). These
ones should also be cleaned for the same reasons, and probably be
turned into ist strings.
When DEBUG_FD is set at build time, we'll keep a counter of per-FD events
in the fdtab. This counter is reported in "show fd" even for closed FDs if
not zero. The purpose is to help spot situations where an apparently closed
FD continues to be reported in loops, or where some events are dismissed.
Some of the recent optimizations around the polling to save a few
epoll_ctl() calls have shown that they could also cause some trouble.
However, over time our code base has become totally asynchronous with
I/Os always attempted from the upper layers and only retried at the
bottom, making it look like we're getting closer to EPOLLET support.
There are showstoppers there such as the listeners which cannot support
this. But given that most of the epoll_ctl() dance comes from the
connections, we can try to enable edge-triggered polling on connections.
What this patch does is to add a new global tunable "tune.fd.edge-triggered",
that makes fd_insert() automatically set an et_possible bit on the fd if
the I/O callback is conn_fd_handler. When the epoll code sees an update
for such an FD, it immediately registers it in both directions the first
time and doesn't update it anymore.
On a few tests it proved quite useful with a 14% request rate increase in
a H2->H1 scenario, reducing the epoll_ctl() calls from 2 per request to
2 per connection.
The option is obviously disabled by default as bugs are still expected,
particularly around the subscribe() code where it is possible that some
layers do not always re-attempt reading data after being woken up.
We have poll_drop, poll_dead and poll_skip which are confusingly named
like their poll_io and poll_exp counterparts except that they are not
per poll() call but per-fd. This patch renames them to poll_drop_fd(),
poll_dead_fd() and poll_skip_fd() for this reason.
The "show activity" output mentions a number of indicators to explain
wake up reasons but doesn't have the number of times poll() sees some
I/O. And given that multiple events can happen simultaneously, it's
not always possible to deduce this metric by subtracting.
This patch adds a new "poll_io" counter that allows one to see how
often poll() returns with at least one active FD. This should help
detect stuck events and measure various ratios of poll sub-metrics.
This patch fixes all the leftovers from the include cleanup campaign. There
were not that many (~400 entries in ~150 files) but it was definitely worth
doing it as it revealed a few duplicates.
Since these are used as type attributes or conditional clauses, they
are used about everywhere and should not require a dependency on
thread.h. Moving them to compiler.h along with other similar statements
like ALIGN() etc looks more logical; this way they become part of the
base API. This allowed to remove thread-t.h from ~12 files, one was
found to only require thread-t and not thread and dict.c was found to
require thread.h.
global.h was one of the messiest files, it has accumulated tons of
implicit dependencies and declares many globals that make almost all
other file include it. It managed to silence a dependency loop between
server.h and proxy.h by being well placed to pre-define the required
structs, forcing struct proxy and struct server to be forward-declared
in a significant number of files.
It was split in to, one which is the global struct definition and the
few macros and flags, and the rest containing the functions prototypes.
The UNIX_MAX_PATH definition was moved to compat.h.
A few includes were missing in each file. A definition of
struct polled_mask was moved to fd-t.h. The MAX_POLLERS macro was
moved to defaults.h
Stdio used to be silently inherited from whatever path but it's needed
for list_pollers() which takes a FILE* and which can thus not be
forward-declared.
And also rename standard.c to tools.c. The original split between
tools.h and standard.h dates from version 1.3-dev and was mostly an
accident. This patch moves the files back to what they were expected
to be, and takes care of not changing anything else. However this
time tools.h was split between functions and types, because it contains
a small number of commonly used macros and structures (e.g. name_desc)
which in turn cause the massive list of includes of tools.h to conflict
with the callers.
They remain the ugliest files of the whole project and definitely need
to be cleaned and split apart. A few types are defined there only for
functions provided there, and some parts are even OS-specific and should
move somewhere else, such as the symbol resolution code.
This moves types/activity.h to haproxy/activity-t.h and
proto/activity.h to haproxy/activity.h.
The macros defining the bit field values for the profiling variable
were moved to the type file to be more future-proof.
This one is included almost everywhere and used to rely on a few other
.h that are not needed (unistd, stdlib, standard.h). It could possibly
make sense to split it into multiple parts to distinguish operations
performed on timers and the internal time accounting, but at this point
it does not appear much important.
This splits the hathreads.h file into types+macros and functions. Given
that most users of this file used to include it only to get the definition
of THREAD_LOCAL and MAXTHREADS, the bare minimum was placed into thread-t.h
(i.e. types and macros).
All the thread management was left to haproxy/thread.h. It's worth noting
the drop of the trailing "s" in the name, to remove the permanent confusion
that arises between this one and the system implementation (no "s") and the
makefile's option (no "s").
For consistency, src/hathreads.c was also renamed thread.c.
A number of files were updated to only include thread-t which is the one
they really needed.
Some future improvements are possible like replacing empty inlined
functions with macros for the thread-less case, as building at -O0 disables
inlining and causes these ones to be emitted. But this really is cosmetic.
This one used to be stored into debug.h but the debug tools got larger
and require a lot of other includes, which can't use BUG_ON() anymore
because of this. It does not make sense and instead this macro should
be placed into the lower includes and given its omnipresence, the best
solution is to create a new bug.h with the few surrounding macros needed
to trigger bugs and place assertions anywhere.
Another benefit is that it won't be required to add include <debug.h>
anymore to use BUG_ON, it will automatically be covered by api.h. No
less than 32 occurrences were dropped.
The FSM_PRINTF macro was dropped since not used at all anymore (probably
since 1.6 or so).
All files that were including one of the following include files have
been updated to only include haproxy/api.h or haproxy/api-t.h once instead:
- common/config.h
- common/compat.h
- common/compiler.h
- common/defaults.h
- common/initcall.h
- common/tools.h
The choice is simple: if the file only requires type definitions, it includes
api-t.h, otherwise it includes the full api.h.
In addition, in these files, explicit includes for inttypes.h and limits.h
were dropped since these are now covered by api.h and api-t.h.
No other change was performed, given that this patch is large and
affects 201 files. At least one (tools.h) was already freestanding and
didn't get the new one added.
This was made to support epoll on patched 2.4 kernels, and on early 2.6
using alternative libcs thanks to the arch-specific syscall definitions.
All the features we support have been around since 2.6.2 and present in
glibc since 2.3.2, neither of which are found in field anymore. Let's
simply drop this and use epoll normally.
valgrind complains that epoll_ctl() uses an epoll_event in which we
have only set the part we use from the data field (i.e. the fd). Tests
show that pre-initializing the struct in the stack doesn't have a
measurable impact so let's do it.
This used to be a minor optimization on ix86 where registers are scarce
and the calling convention not very efficient, but this platform is not
relevant enough anymore to warrant all this dirt in the code for the sake
of saving 1 or 2% of performance. Modern platforms don't use this at all
since their calling convention already defaults to using several registers
so better get rid of this once for all.
Historically we used to have a global epoll_event for various
manipulations involving epoll_ctl() and when threads were added,
this was turned to a thread_local, which is needlessly expensive
since it's just a temporary variable. Let's move it to a local
variable wherever it's called instead.
The cost of enabling polling in one direction with epoll is very high
because it requires one syscall per FD and per direction change. In
addition we don't know about input readiness until we either try to
receive() or enable polling and watch the result. With HTTP keep-alive,
both are equally expensive as it's very uncommon to see the server
instantly respond (unless it's a second stage of the same process on
localhost, which has become much less common with threads).
But when a connection is established it's also quite usual to have to
poll for sending (except on localhost or UNIX sockets where it almost
always instantly works). So this cost of polling could be factored out
with the second step if both were enabled together.
This is the idea behind this patch. What it does is to always enable
polling for Rx if it's not ready and at least one direction is active.
This means that if it's not explicitly disabled, or if it was but in a
state that causes the loss of the information (rx ready cannot be
guessed), then let's take any opportunity for a polling change to
enable it at the same time, and learn about rx readiness for free.
In addition the FD never gets unregistered for Rx unless it's ready
and was blocked (buffer full). This avoids a lot of the flip-flop
behaviour at beginning and end of requests.
On a test with 10k requests in keep-alive, the difference is quite
noticeable:
Before:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
83.67 0.010847 0 20078 epoll_ctl
16.33 0.002117 0 2231 epoll_wait
0.00 0.000000 0 20 20 connect
------ ----------- ----------- --------- --------- ----------------
100.00 0.012964 22329 20 total
After:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
96.35 0.003351 1 2644 epoll_wait
2.36 0.000082 4 20 20 connect
1.29 0.000045 0 66 epoll_ctl
------ ----------- ----------- --------- --------- ----------------
100.00 0.003478 2730 20 total
It may also save a recvfrom() after connect() by changing the following
sequence, effectively saving one epoll_ctl() and one recvfrom() :
before | after
-----------------------------+----------------------------
- connect() | - connect()
- epoll_ctl(add,out) | - epoll_ctl(add, in|out)
- sendto() | - epoll_wait() = out
- epoll_ctl(mod,in|out) | - send()
- epoll_wait() = out | - epoll_wait() = in|out
- recvfrom() = EAGAIN | - recvfrom() = OK
- epoll_ctl(mod,in) | - recvfrom() = EAGAIN
- epoll_wait() = in | - epoll_ctl(mod, in)
- recvfrom() = OK | - epoll_wait()
- recvfrom() = EAGAIN |
- epoll_wait() |
(...)
Now on a 10M req test on 16 threads with 2k concurrent conns and 415kreq/s,
we see 190k updates total and 14k epoll_ctl() only.
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.
As mentioned in previous commit, these flags do not map well to
modern poller capabilities. Let's use the FD_EV_*_{R,W} flags instead.
This first patch only performs a 1-to-1 mapping making sure that the
previously reported flags are still reported identically while using
the closest possible semantics in the pollers.
It's worth noting that kqueue will now support improvements such as
returning distinctions between shut and errors on each direction,
though this is not exploited for now.
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.
In the poller code, instead of just remembering if we're currently polling
a fd or not, remember if we're polling it for writing and/or for reading, that
way, we can avoid to modify the polling if it's already polled as needed.
Now that the architecture was changed so that attempts to receive/send data
always come from the upper layers, instead of them only trying to do so when
the lower layer let them know they could try, we can finally get rid of the
fd cache. We don't really need it anymore, and removing it gives us a small
performance boost.
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.
In some situations, especially when dealing with low latency on processors
supporting a variable frequency or when running inside virtual machines,
each time the process waits for an I/O using the poller, the processor
goes back to sleep or is offered to another VM for a long time, and it
causes excessively high latencies.
A solution to this provided by this patch is to enable busy polling using
a global option. When busy polling is enabled, the pollers never sleep and
loop over themselves waiting for an I/O event to happen or for a timeout
to occur. On multi-processor machines it can significantly overheat the
processor but it usually results in much lower latencies.
A typical test consisting in injecting traffic over a single connection at
a time over the loopback shows a bump from 4640 to 8540 connections per
second on forwarded connections, indicating a latency reduction of 98
microseconds for each connection, and a bump from 12500 to 21250 for
locally terminated connections (redirects), indicating a reduction of
33 microseconds.
It is only usable with epoll and kqueue because select() and poll()'s
API is not convenient for such usages, and the level of performance they
are used in doesn't benefit from this anyway.
The option, which obviously remains disabled by default, can be turned
on using "busy-polling" in the global section, and turned off later
using "no busy-polling". Its status is reported in "show info" to help
troubleshooting suspicious CPU spikes.
At the moment the situation with activity measurement is quite tricky
because the struct activity is defined in global.h and declared in
haproxy.c, with operations made in time.h and relying on freq_ctr
which are defined in freq_ctr.h which itself includes time.h. It's
barely possible to touch any of these files without breaking all the
circular dependency.
Let's move all this stuff to activity.{c,h} and be done with it. The
measurement of active and stolen time is now done in a dedicated
function called just after tv_before_poll() instead of mixing the two,
which used to be a lazy (but convenient) decision.
No code was changed, stuff was just moved around.
By placing this code into time.h (tv_entering_poll() and tv_leaving_poll())
we can remove the logic from the pollers and prepare for extending this to
offer more accurate time measurements.
The 4 pollers all contain the same code used to compute the poll timeout.
This is pointless, let's centralize this into fd.h. This also gets rid of
the useless SCHEDULER_RESOLUTION macro which used to work arond a very old
linux 2.2 bug causing select() to wake up slightly before the timeout.
The current synchronization point enforces certain restrictions which
are hard to workaround in certain areas of the code. The fact that the
critical code can only be called from the sync point itself is a problem
for some callback-driven parts. The "show fd" command for example is
fragile regarding this.
Also it is expensive in terms of CPU usage because it wakes every other
thread just to be sure all of them join to the rendez-vous point. It's a
problem because the sleeping threads would not need to be woken up just
to know they're doing nothing.
Here we implement a different approach. We keep track of harmless threads,
which are defined as those either doing nothing, or doing harmless things.
The rendez-vous is used "for others" as a way for a thread to isolate itself.
A thread then requests to be alone using thread_isolate() when approaching
the dangerous area, and then waits until all other threads are either doing
the same or are doing something harmless (typically polling). The function
only returns once the thread is guaranteed to be alone, and the critical
section is terminated using thread_release().
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]
Clearing the update_mask bit in fd_insert may lead to duplicate insertion
of fd in fd_updt, that could lead to a write past the end of the array.
Instead, make sure the update_mask bit is cleared by the pollers no matter
what.
This should be backported to 1.8.
[wt: warning: 1.8 doesn't have the lockless fdcache changes and will
require some careful changes in the pollers]
