Storing the pointer to the pool along with the stats is quite useful as
it allows to report the name. That's what we're doing here. We could
store it in place of another field but that's not convenient as it would
require to change all functions that manipulate counters. Thus here we
store one extra field, as well as some padding because the struct turns
56 bytes long, thus better go to 64 directly. Example of output from
"show profiling memory":
2 0 48 0| 0x4bfb2c ha_quic_set_encryption_secrets+0xcc/0xb5e p_alloc(24) [pool=quic_tls_iv]
0 55252 0 10608384| 0x4bed32 main+0x2beb2 free(-192)
15 0 2760 0| 0x4be855 main+0x2b9d5 p_alloc(184) [pool=quic_frame]
1 0 1048 0| 0x4be266 ha_quic_add_handshake_data+0x2b6/0x66d p_alloc(1048) [pool=quic_crypto]
3 0 552 0| 0x4be142 ha_quic_add_handshake_data+0x192/0x66d p_alloc(184) [pool=quic_frame]
31276 0 6755616 0| 0x4bb8f9 quic_sock_fd_iocb+0x689/0x69b p_alloc(216) [pool=quic_dgram]
0 31424 0 6787584| 0x4bb7f3 quic_sock_fd_iocb+0x583/0x69b p_free(-216) [pool=quic_dgram]
152 0 32832 0| 0x4bb4d9 quic_sock_fd_iocb+0x269/0x69b p_alloc(216) [pool=quic_dgram]
Pools are being used so well that it becomes difficult to profile their
usage via the regular memory profiling. Let's add new entries for pools
there, named "p_alloc" and "p_free" that correspond to pool_alloc() and
pool_free(). Ideally it would be nice to only report those that fail
cache lookups but that's complicated, particularly on the free() path
since free lists are released in clusters to the shared pools.
It's worth noting that the alloc_tot/free_tot fields can easily be
determined by multiplying alloc_calls/free_calls by the pool's size, and
could be better used to store a pointer to the pool itself. However it
would require significant changes down the code that sorts output.
If this were to cause a measurable slowdown, an alternate approach could
consist in using a different value of USE_MEMORY_PROFILING to enable pools
profiling. Also, this profiler doesn't depend on intercepting regular malloc
functions, so we could also imagine enabling it alone or the other one alone
or both.
Tests show that the CPU overhead on QUIC (which is already an extremely
intensive user of pools) jumps from ~7% to ~10%. This is quite acceptable
in most deployments.
Right now it's not possible to feed memory profiling info from outside
activity.c, so let's export the function and move the enum and struct
to the include file.
The thread flags are touched a little bit by other threads, e.g. the STUCK
flag may be set by other ones, and they're watched a little bit. As such
we need to use atomic ops only to manipulate them. Most places were already
using them, but here we generalize the practice. Only ha_thread_dump() does
not change because it's run under isolation.
Since we don't mix tasks from different threads in the run queues
anymore, we don't need to use the eb32sc_ trees and we can switch
to the regular eb32 ones. This uses cheaper lookup and insert code,
and a 16-thread test on the queues shows a performance increase
from 570k RPS to 585k RPS.
Since we only use the shared runqueue to put tasks only assigned to
known threads, let's move that runqueue to each of these threads. The
goal will be to arrange an N*(N-1) mesh instead of a central contention
point.
The global_rqueue_ticks had to be dropped (for good) since we'll now
use the per-thread rqueue_ticks counter for both trees.
A few points to note:
- the rq_lock stlil remains the global one for now so there should not
be any gain in doing this, but should this trigger any regression, it
is important to detect whether it's related to the lock or to the tree.
- there's no more reason for using the scope-based version of the ebtree
now, we could switch back to the regular eb32_tree.
- it's worth checking if we still need TASK_GLOBAL (probably only to
delete a task in one's own shared queue maybe).
Instead of having a global mask of all the profiled threads, let's have
one flag per thread in each thread's flags. They are never accessed more
than one at a time an are better located inside the threads' contexts for
both performance and scalability.
There's no more reason for keepin the code and definitions in conn_stream,
let's move all that to stconn. The alphabetical ordering of include files
was adjusted.
This file contains all the stream-connector functions that are specific
to application layers of type stream. So let's name it accordingly so
that it's easier to figure what's located there.
The alphabetical ordering of include files was preserved.
We're starting to propagate the stream connector's new name through the
API. Most call places of these functions that retrieve the channel or its
buffer are in applets. The local variable names are not changed in order
to keep the changes small and reviewable. There were ~92 uses of cs_ic(),
~96 of cs_oc() (due to co_get*() being less factorizable than ci_put*),
and ~5 accesses to the buffer itself.
This applies the change so that the applet code stops using ci_putchk()
and friends everywhere possible, for the much saferapplet_put*() instead.
The change is mechanical but large. Two or three functions used to have no
appctx and a cs derived from the appctx instead, which was a reminiscence
of old times' stream_interface. These were simply changed to directly take
the appctx. No sensitive change was performed, and the old (more complex)
API is still usable when needed (e.g. the channel is already known).
The change touched roughly a hundred of locations, with no less than 124
lines removed.
It's worth noting that the stats applet, the oldest of the series, could
get a serious lifting, as it's still very channel-centric instead of
propagating the appctx along the chain. Given that this code doesn't
change often, there's no emergency to clean it up but it would look
better.
This renames the "struct conn_stream" to "struct stconn" and updates
the descriptions in all comments (and the rare help descriptions) to
"stream connector" or "connector". This touches a lot of files but
the change is minimal. The local variables were not even renamed, so
there's still a lot of "cs" everywhere.
This one is the pointer to the conn_stream which is always in the
endpoint that is always present in the appctx, thus it's not needed.
This patch removes it and replaces it with appctx_cs() instead. A
few occurences that were using __cs_strm(appctx->owner) were moved
directly to appctx_strm() which does the equivalent.
Remaining flags and associated functions are move in the conn-stream
scope. These flags are added on the endpoint and not the conn-stream
itself. This way it will be possible to get them from the mux or the
applet. The functions to get or set these flags are renamed accordingly with
the "cs_" prefix and updated to manipualte a conn-stream instead of a
stream-interface.
At many places, we now use the new CS functions to get a stream or a channel
from a conn-stream instead of using the stream-interface API. It is the
first step to reduce the scope of the stream-interfaces. The main change
here is about the applet I/O callback functions. Before the refactoring, the
stream-interface was the appctx owner. Thus, it was heavily used. Now, as
far as possible,the conn-stream is used. Of course, it remains many calls to
the stream-interface API.
Because appctx is now an endpoint of the conn-stream, there is no reason to
still have the stream-interface as appctx owner. Thus, the conn-stream is
now the appctx owner.
The current model causes an issue when trying to spot memory leaks,
because malloc(0) or realloc(0) do not count as allocations since we only
account for the application-usable size. This is the problem that made
issue #1406 not to appear as a leak.
What we're doing now is to account for one extra pointer (the one that
memory allocators usually place before the returned area), so that a
malloc(0) will properly account for 4 or 8 bytes. We don't need something
exact, we just need something non-zero so that a realloc(X) followed by a
realloc(0) without a free() gives a small non-zero result.
It was verified that the results are stable including in the presence
of lots of malloc/realloc/free as happens when stressing Lua.
It would make sense to backport this to 2.4 as it helps in bug reports.
realloc() calls are painful to analyse because they have two non-zero
columns and trying to spot a leaking one requires a bit of scripting.
Let's simply append the delta at the end of the line when alloc and
free are non-nul.
It would be useful to backport this to 2.4 to help with bug reports.
The scheduler contains a lot of stuff that is thread-local and not
exclusively tied to the scheduler. Other parts (namely thread_info)
contain similar thread-local context that ought to be merged with
it but that is even less related to the scheduler. However moving
more data into this structure isn't possible since task.h is high
level and cannot be included everywhere (e.g. activity) without
causing include loops.
In the end, it appears that the task_per_thread represents most of
the per-thread context defined with generic types and should simply
move to tinfo.h so that everyone can use them.
The struct was renamed to thread_ctx and the variable "sched" was
renamed to "th_ctx". "sched" used to be initialized manually from
run_thread_poll_loop(), now it's initialized by ha_set_tid() just
like ti, tid, tid_bit.
The memset() in init_task() was removed in favor of a bss initialization
of the array, so that other subsystems can put their stuff in this array.
Since the tasklet array has TL_CLASSES elements, the TL_* definitions
was moved there as well, but it's not a problem.
The vast majority of the change in this patch is caused by the
renaming of the structures.
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.
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.
This one is expensive in code size because it comes with xxhash.h at a
low level of dependency that's inherited at plenty of places, and for
a function does doesn't benefit from inlining and could possibly even
benefit from not being inline given that it's large and called from the
scheduler.
Moving it to activity.c reduces the LoC by 1.2% and the binary size by
~1kB.
This function has no reason for being inlined, it's called from non
critical places (once in pollers), is quite large and comes with
dependencies (time and freq_ctr). Let's move it to acitvity.c. That's
another 0.4% less LoC to build.
"show profiling" by default sorts by usage/counts, which is suitable for
occasional use. But when called from scripts to monitor/search variations,
this is not very convenient. Let's add a new "byaddr" option to support
sorting the output by address. It also eases matching alloc/free calls
from within a same library, or reading grouped tasks costs by library.
Already had to perform too many additions by external scripts, it's
time to add the totals and delay alloc-free as a last line in the
output of the "show memory profiling".
This was planned but missing in the previous attempt, we really need to
see what is used at each place, especially due to realloc(). Now we
print the function used in front of the caller's address, as well as
the average alloc/free size per call.
The realloc() function checks if the size grew or reduced in order to
count an allocation or a free, but it does so with the absolute (new
or old) value instead of the difference, resulting in realloc() often
being credited for allocating too much.
No backport is needed.
It was found that when viewing the help output from the CLI that
"set profiling" had 2 spaces in it, which was pushing it out from
the rest of similar commands.
i.e. it looked like this:
prepare acl <acl>
prepare map <acl>
set profiling <what> {auto|on|off}
set dynamic-cookie-key backend <bk> <k>
set map <map> [<key>|#<ref>] <value>
set maxconn frontend <frontend> <value>
This patch removes all of the double spaces within the command and
unifies them to single spacing, which is what is observed within the
rest of the commands.
There were 102 CLI commands whose help were zig-zagging all along the dump
making them unreadable. This patch realigns all these messages so that the
command now uses up to 40 characters before the delimiting colon. About a
third of the commands did not correctly list their arguments which were
added after the first version, so they were all updated. Some abuses of
the term "id" were fixed to use a more explanatory term. The
"set ssl ocsp-response" command was not listed because it lacked a help
message, this was fixed as well. The deprecated enable/disable commands
for agent/health/server were prominently written as deprecated. Whenever
possible, clearer explanations were provided.
When memory profiling is enabled, realloc() can occasionally get the area
size wrong due to the wrong pointer being used to check the new size. When
the old area gets unmapped in the operation, this may even result in a
crash. There's no impact without memory profiling though.
No backport is needed as this is exclusively 2.4-dev.
Now the memory usage stats are dumped. They are first sorted by total
alloc+free so that the first ones are always the most relevant, and
that most symmetric alloc/free pairs appear next to each other. This
way it becomes convenient to only show a small part of them such as:
show profiling memory 20
It's worth noting that the sorting is performed upon each call to the
iohandler so it is technically possible that an entry could appear
twice or be dropped if the ordering changes between two calls. In
practice it is not an issue but it's worth being mentioned.
Let's rearrange it to make it more configurable and allow to iterate
over multiple parts (header, tasks, memory etc), to restart from a
given line number (previously it didn't work, though fortunately it
didn't happen), and to support dumping only certain parts and a given
number of lines. A few entries from ctx.cli are now used to store a
restart point and the current step.
When built with USE_MEMORY_PROFILING the main memory allocation functions
are diverted to collect statistics per caller. It is a bit tricky because
the only way to call the original ones is to find their pointer, which
requires dlsym(), and which is not available everywhere.
Thus all functions are designed to call their fallback function (the
original one), which is preset to an initialization function that is
supposed to call dlsym() to resolve the missing symbols, and vanish.
This saves expensive tests in the critical path.
A second problem is that dlsym() calls calloc() to initialize some
error messages. After plenty of tests with posix_memalign(), valloc()
and friends, it turns out that returning NULL still makes it happy.
Thus we currently use a visit counter (in_memprof) to detect if we're
reentering, in which case all allocation functions return NULL.
In order to convert a return address to an entry in the stats, we
perform a cheap hash consisting in multiplying the pointer by a
balanced number (as many zeros as ones) and keeping the middle bits.
The hash is already pretty good like this, achieving to store up to
638 entries in a 2048-entry table without collision. But in order to
further refine this and improve the fill ratio of the table, in case
of collision we move up to 16 adjacent entries to find a free place.
This remains quite cheap and manages to store all of these inside a
1024-entries hash table with even less risk of collision.
Also, free(NULL) does not produce any stats. By doing so we reduce
from 638 to 208 the average number of entries needed for a basic
config using SSL. free(NULL) not only provides no information as it's
a NOP, but keeping it is pure pollution as it happens all the time.
When DEBUG_MEM_STATS is enabled, malloc/calloc/realloc are redefined as
macros, preventing the code from compiling. Thus, when this option is
detected, the macros are undefined as they are pointless there anyway.
The functions are optimized to quickly jump to the fallback and as such
become almost invisible in terms of processing time, execpt an extra
"if" on a read_mostly variable and a jump. Considering that this only
happens for pool misses and library routines, this remains acceptable.
Performance tests in SSL (the most stressful test) shows less than 1%
performance loss when profiling is enabled on 2c4t.
The code was written in a way to ease backporting to modern versions
(2.2+) if needed, so it keeps the long names for integers and doesn't
use the _INC version of the atomic ops.
We'll need to store for each call place, the pointer to the caller
(the return address to be more exact as with free() it's not uncommon
to see tail calls), the number of calls to alloc/free and the total
alloc/free bytes. realloc() will be counted either as alloc or free
depending on the balance of the size before vs after.
We store 1024+1 entries. The first ones are used as hashes and the
last one for collisions.
When profiling is enabled via the CLI, all the stats are reset.
This adds the necessary flags to permit run-time enabling/disabling of
memory profiling. For now this is disabled.
A few words were added to the management doc about it and recalling that
this is limited to certain OSes.
These ones are only read by the scheduler and occasionally written to
by the CLI parser, so let's move them to read_mostly so that they do
not risk to suffer from cache line pollution.
The default proxy was passed as a variable to all parsers instead of a
const, which is not without risk, especially when some timeout parsers used
to make some int pointers point to the default values for comparisons. We
want to be certain that none of these parsers will modify the defaults
sections by accident, so it's important to mark this proxy as const.
This patch touches all occurrences found (89).
During the call to thread_isolate(), some other threads might have
performed some task_wakeup() which will have a call date past the
one we retrieved. It could be avoided by taking the current date
once we're alone but this would significantly affect the latency
measurements by adding the isolation time. Instead we're now only
accounting positive times, so that late wakeups normally appear
with a zero latency.
No backport is needed, this is 2.4.
This finally adds the long-awaited solution to inspect the run queues
and figure what is eating the CPU or causing latencies. We can even see
the experienced latencies when profiling is enabled. Example on a
saturated process:
> show tasks
Running tasks: 14983 (4 threads)
function places % lat_tot lat_avg
process_stream 4948 33.0 5.840m 70.82ms
h1_io_cb 2535 16.9 - -
main+0x9e670 2508 16.7 2.930m 70.10ms
ssl_sock_io_cb 2499 16.6 - -
si_cs_io_cb 2493 16.6 - -
If a user enables profiling by hand, it makes sense to reset the stats
counters to provide fresh new measurements. Therefore it's worth using
this as the standard method to reset counters.