A "return NULL" statement was placed for error handling in the
posix_memalign() handler instead of an int errno value, by recent
commit 5ddc8b3ad4 ("MINOR: activity/memprofile: monitor non-portable
calls as well"). Surprisingly the warning only triggered on gcc-4.8.
Let's use ENOMEM instead. No backport needed.
The calls to resolv_sym_name() can be a bit expensive. Forcing to
yield more often is better for the latency and will avoid the
watchdog reporting warnings.
Note that it's still called in the sort at the end, but that one
cannot be avoided. At best we could try to rely on the list of libs
but that's not trivial and not always present.
Some dependencies might very well rely on posix_memalign(), strndup()
or other less portable callsn making us miss them when chasing memory
leaks, resulting in negative global allocation counters. Let's provide
the handlers for the following functions:
strndup() // _POSIX_C_SOURCE >= 200809L || glibc >= 2.10
valloc() // _BSD_SOURCE || _XOPEN_SOURCE>=500 || glibc >= 2.12
aligned_alloc() // _ISOC11_SOURCE
posix_memalign() // _POSIX_C_SOURCE >= 200112L
memalign() // obsolete
pvalloc() // obsolete
This time we don't fail if they're not found, we just silently forward
the calls.
Some memory profiling outputs have showed negative counters, very likely
due to some libs calling strdup(). Let's add it to the list of monitored
activities.
Actually even haproxy itself uses some. Having "profiling.memory on" in
the config reveals 35 call places.
In "show profiling memory", we need to distinguish methods which really
free memory from those which do not so that we don't account for the
free value twice. However for now it's done using multiple tests, which
are going to complicate the addition of new methods. Let's switch to a
bit field defined as a mask in a single place instead, as we don't
intend to use more than 32/64 methods!
For unknown callers we try to get the lowest known address and we
purposely ignore NULL during calculation of the min. But the side
effect is that we also report ffff in the per-DSO address. Better
catch this case and finally accept to report nil. Before it would
report this:
$ socat - /tmp/sock1 <<< "show profiling memory" |grep nil
50000 10 9600000 9440| (nil) [other] unknown(192) [delta=9590560] [pool=http_txn]
50000 10 9600000 9440| (nil) DSO:other; delta_calls=49990; delta_bytes=9590560
now it reports this:
$ socat - /tmp/sock1 <<< "show profiling memory" |grep nil
50000 11 9600000 9656| (nil) [other] unknown(192) [delta=9590344] [pool=connection]
50000 11 9600000 9656| (nil) DSO:other; delta_calls=49989; delta_bytes=9590344
There's actually a problem with memprofiles: the pool pointer is stored
in ->info but some pools are replaced during startup, such as the trash
pool, leaving a dangling pointer there.
Let's complete the API with a new function memprof_remove_stale_info()
that will remove all stale references to this info pointer. It's also
present when USE_MEMORY_PROFILING is not set so as to ease the job on
callers.
In commit 401fb0e87a ("MINOR: activity/memprofile: show per-DSO stats")
we added a summary per DSO. However the free calls/tot were not initialized
when creating a new entry because initially they were applied to any entry,
but since we don't update free calls for non-free capable callers, we still
need to reinitialize these entries when reassigning one. Because of this
bug, a "show profiling memory" output can randomly show highly negative
values on the DSO lines if it turns out that the DSO entry was created on
an alloc instead of a realloc/free.
Since the commit above was backported to 2.9, this one must go there as
well.
On systems where many libs are loaded, it's hard to track suspected
leaks. Having a per-DSO summary makes it more convenient. That's what
we're doing here by summarizing all calls per DSO before showing the
total.
It was found in a large "show profiling memory" output that a few entries
have a NULL return address, which causes confusion because this address
will be reused by the next new allocation caller, possibly resulting in
inconsistencies such as "free() ... pool=trash" which makes no sense. The
cause is in fact that the first caller had an entry->info pointing to the
trash pool from a p_alloc/p_free with a NULL return address, and the second
had a different type and reused that entry.
Let's make sure undecodable stacks causing an apparent NULL return address
all lead to the "other" bin.
While this is not exactly a bug, it would make sense to backport it to the
recent branches where the feature is used (probably at least as far as 2.8).
Thanks to the commit 5714aff4a6bf
"DEBUG: pool: store the memprof bin on alloc() and update it on free()", the
amount of memory allocations and memory "frees" is shown now on the same line,
corresponded to the caller name. This is very convenient to debug memory leaks
(haproxy should run with -dMcaller option).
The implicit drawback of this solution is that we count twice same free_calls
and same free_tot (bytes) values in cli_io_handler_show_profiling(), when
we've calculed tot_free_calls and tot_free_bytes, by adding them to the these
totalizators for p_alloc, malloc and calloc allocator types. See the details
about why this happens in a such way in __pool_free() implementation and
also in the commit message for 5714aff4a6bf.
This double addition of free counters falses 'Delta_calls' and 'Delta_bytes',
sometimes we even noticed that they show negative values.
Same problem was with the calculation of average allocated buffer size for
lines, where we show simultaneously the number of allocated and freed bytes.
The main CLI I/O handle is responsible to interrupt the processing on
shutdown/abort. It is not the responsibility of the I/O handler of CLI
commands to take care of it.
Seeing counters in "show profiling" is not always very helpful without
an indication of how long the analysis lasted nor if it's still active
or not. Let's add a pair of start/stop timers for tasks and memory so
that we can now indicate how long the measurements lasted and when they
ended (or 0 if still running).
Note that for tasks profiling set to "auto", the measurement is considered
enabled since it can automatically switch on and off on a per-thread
basis.
Let's keep two check counters per thread:
- one for "active" checks, i.e. checks that are no more sleeping
and are assigned to the thread. These include sleeping and
running checks ;
- one for "running" checks, i.e. those which are currently
executing on the thread.
By doing so, we'll be able to spread the health checks load a bit better
and refrain from sending too many at once per thread. The counters are
atomic since a migration increments the target thread's active counter.
These numbers are reported in "show activity", which allows to check
per thread and globally how many checks are currently pending and running
on the system.
Ideally, we should only consider checks in the process of establishing
a connection since that's really the expensive part (particularly with
OpenSSL 3.0). But the inner layers are really not suitable to doing
this. However knowing the number of active checks is already a good
enough hint.
We now count the number of times a check was started on each thread
and the number of times a check was adopted. This helps understand
better what is observed regarding checks.
16kB buffers are not enough to dump 4096 threads with up to 10 bytes value
on each line. By storing the column number in the applet's context, we can
now restart from the last attempted column. This requires to dump all values
as they are produced, but it doesn't cost that much: a 4096-thread output
from a fesh process produces 300kB of output in ~8ms, or ~400us per call
(19*16kB), most of which are spent in vfprintf(). Given that we don't print
more than needed, it doesn't really change anything.
The main caveat is that when interrupted on such large lines, there's a
great possibility that the total or average on the first column doesn't
match anymore the sum or average of all dumped values. In order to avoid
this whenever possible (typically less than ~1500 threads), we first try
to dump entire lines and only proceed one column at a time when we have
to retry a failed dump. This is already the same for other stats that are
dumped in an interruptible way anyway and there's little that can be done
about it at this point (and not much immediately perceived benefit in
doing this with extreme accuracy for >1500 threads).
When using many threads, it's difficult to see the end of "show activity"
due to the numerous columns which fill the buffer. For example a dump of
a 256-thread, freshly booted process yields around 15kB.
Here by arranging the dump in a loop around a switch/case block where
each case checks the code line number against the current dump position,
we have a restartable counter for free with a granularity of the line of
code, without having to maintain a matching between states and specific
lines. It just requires to reset the trash buffer for each line and to
try to dump it after each line.
Now dumping 256 threads after a few seconds of traffic happily emits 20kB.
Now each line of "show activity" will iterate over n+2 fields, one for
the line header, one for the total, and one per thread. This will soon
allow us to save the current state in a restartable way.
Now that "now" is no more a timeval, there's no point keeping a copy
of it as a timeval, let's also switch start_time to nanoseconds, it
simplifies operations.
This puts an end to the occasional confusion between the "now" date
that is internal, monotonic and not synchronized with the system's
date, and "date" which is the system's date and not necessarily
monotonic. Variable "now" was removed and replaced with a 64-bit
integer "now_ns" which is a counter of nanoseconds. It wraps every
585 years, so if all goes well (i.e. if humanity does not need
haproxy anymore in 500 years), it will just never wrap. This implies
that now_ns is never nul and that the zero value can reliably be used
as "not set yet" for a timestamp if needed. This will also simplify
date checks where it becomes possible again to do "date1<date2".
All occurrences of "tv_to_ns(&now)" were simply replaced by "now_ns".
Due to the intricacies between now, global_now and now_offset, all 3
had to be turned to nanoseconds at once. It's not a problem since all
of them were solely used in 3 functions in clock.c, but they make the
patch look bigger than it really is.
The clock_update_local_date() and clock_update_global_date() functions
are now much simpler as there's no need anymore to perform conversions
nor to round the timeval up or down.
The wrapping continues to happen by presetting the internal offset in
the short future so that the 32-bit now_ms continues to wrap 20 seconds
after boot.
The start_time used to calculate uptime can still be turned to
nanoseconds now. One interrogation concerns global_now_ms which is used
only for the freq counters. It's unclear whether there's more value in
using two variables that need to be synchronized sequentially like today
or to just use global_now_ns divided by 1 million. Both approaches will
work equally well on modern systems, the difference might come from
smaller ones. Better not change anyhting for now.
One benefit of the new approach is that we now have an internal date
with a resolution of the nanosecond and the precision of the microsecond,
which can be useful to extend some measurements given that timestamps
also have this resolution.
Another case where "now" was used instead of "date" for a publicly visible
date that was already incorrect and became worse after commit 28360dc
("MEDIUM: clock: force internal time to wrap early after boot"). No
backport is needed.
There has always been a race when checking the length of an accept queue
to determine which one is more loaded that another, because the head and
tail are read at two different moments. This is not required, we can merge
them as two 16 bit numbers inside a single 32-bit index that is always
accessed atomically. This way we read both values at once and always have
a consistent measurement.
It was missing from the output but is sometimes convenient to observe
and understand how incoming connections are distributed. The CPU usage
is reported as the instant measurement of 100-idle_pct for each thread,
and the average value is shown for the aggregated value.
This could be backported as it's helpful in certain troublehsooting
sessions.
The purpose of this patch is only a one-to-one replacement, as far as
possible.
CF_SHUTR(_NOW) and CF_SHUTW(_NOW) flags are now carried by the
stream-connecter. CF_ prefix is replaced by SC_FL_ one. Of course, it is not
so simple because at many places, we were testing if a channel was shut for
reads and writes in same time. To do the same, shut for reads must be tested
on one side on the SC and shut for writes on the other side on the opposite
SC. A special care was taken with process_stream(). flags of SCs must be
saved to be able to detect changes, just like for the channels.
We've had a start date even before the internal monotonic clock existed,
but once the monotonic clock was added, the start date was not updated
to distinguish the wall clock time units and the internal monotonic time
units. The distinction is important because both clocks do not necessarily
progress at the same speed. The very rare occurrences of the wall-clock
date are essentially for human consumption and communication with third
parties (e.g. report the start date in "show info" for monitoring
purposes). However currently this one is also used to measure the distance
to "now" as being the process' uptime. This is actually not correct. It
only works because for now the two dates are initialized at the exact
same instant at boot but could still be wrong if the system's date shows
a big jump backwards during startup for example. In addition the current
situation prevents us from enforcing an abritrary offset at boot to reveal
some heisenbugs.
This patch adds a new "start_time" at boot that is set from "now" and is
used in uptime calculations. "start_date" instead is now set from "date"
and will always reflect the system date for human consumption (e.g. in
"show info"). This way we're now sure that any drift of the internal
clock relative to the system date will not impact the reported uptime.
This could possibly be backported though it's unlikely that anyone has
ever noticed the problem.
In applets, we stop processing when a write error (CF_WRITE_ERROR) or a shutdown
for writes (CF_SHUTW) is detected. However, any write error leads to an
immediate shutdown for writes. Thus, it is enough to only test if CF_SHUTW is
set.
Initially the code was placed into cli.c to keep activity.c small and
independent of the cli stuff, but that's no longer the case anyway and
keeping that code over there makes it harder to find. Let's move it to
its more natural place now.
By default we now dump stats between caller and callee, but by
specifying "aggr" on the command line, stats get aggregated by
callee again as it used to be before the feature was available.
It may sometimes be helpful when comparing total call counts,
though that's about all.
This removes all the hard-coded 8-bit and 256 entries to use a pair of
macros instead so that we can more easily experiment with larger table
sizes if needed.
This field is misnamed because its real and important content is the
date the task was woken up, not the date it was called. It temporarily
holds the call date during execution but this remains confusing. In
fact before the latency measurements were possible it was indeed a call
date. Thus is will now be called wake_date.
This change is necessary because a subsequent fix will require the
introduction of the real call date in the thread ctx.
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.
