That's the same principle as for ->allocated and ->used. Here we return
the summ of the raw values, so the result still needs to be fed to
swrate_avg(). It also means that we now use the local ->used instead
of the global one for the calculations and do not need to call pool_used()
anymore on fast paths. The number of samples should likely be divided by
the number of buckets, but that's not done yet (better observe first).
A function pool_needed_avg() was added to report aggregated values for
the "show pools" command.
With this change, an h2load made of 5 * 160 conn * 40 streams on 80
threads raised from 1.5M RPS to 6.7M RPS.
That's the same principle as for ->allocated. The small difference here
is that it's no longer possible to decrement ->used in batches when
releasing clusters from the cache to the shared cache, so the counter
has to be decremented for each of them. But as it provides less
contention and it's done only during forced eviction, it shouldn't be
a problem.
A function "pool_used()" was added to return the sum of the entries.
It's used by pool_alloc_nocache() and pool_free_nocache() which need
to count the number of used entries. It's not a problem since such
operations are done when picking/releasing objects to/from the OS,
but it is a reminder that the number of buckets should remain small.
With this change, an h2load test made of 5 * 160 conn * 40 streams on
80 threads raised from 812k RPS to 1.5M RPS.
The ->used counter is one of the most stressed, and it heavily
depends on the ->allocated one, so let's first move ->allocated
to a few buckets.
A function "pool_allocated()" was added to return the sum of the entries.
It's important not to abuse it as it does iterate, so everywhere it's
possible to avoid it by keeping a local counter, it's better. Currently
it's used for limited pools which need to make sure they do not allocate
too many objects. That's an acceptable tradeoff to save CPU on large
machines at the expense of spending a little bit more on small ones which
normally are not under load.
On many threads and without the shared cache, there can be extreme
contention on the ->allocated counter, the ->free_list pointer, and
the ->used counter. It's possible to limit this contention by spreading
the counters a little bit over multiple entries, that are summed up when
a consultation is needed. The criterion used to spread the values cannot
be related to the thread ID due to migrations, since we need to keep
consistent stats (allocated vs used).
Instead we'll just hash the pointer, it provides an index that does the
job and that is consistent for the object. When having just a few entries
(16 here as it showed almost identical performance between global and
non-global pools) even iterations should be short enough during
measurements to not be a problem.
A pair of functions designed to ease pointer hash bucket calculation were
added, with one of them doing it for thread IDs because allocation failures
will be associated with a thread and not a pointer.
For now this patch only brings in the relevant parts of the infrastructure,
the CONFIG_HAP_POOL_BUCKETS_BITS macro that defaults to 6 bits when 512
threads or more are supported, 5 bits when 128 or more are supported, 4
bits when 16 or more are supported, otherwise 3 bits for small setups.
The array in the pool_head and the two utility functions are already
added. It should have no measurable impact beyond inflating the pool_head
structure.
The pool's allocation counter doesn't strictly require to be updated
from these functions, it may more efficiently be done in the caller
(even out of a loop for pool_flush() and pool_gc()), and doing so will
also help us spread the counters over an array later. The functions
were renamed _noinc and _nodec to make sure we catch any possible
user in an external patch. If needed, the original functions may easily
be reimplemented in an inline function.
During tests it was noticed that the current hash is not that good
on 4- and 5- bit hashes. About 7.5% of all the 32-bit primes were tested
as candidates for the hash function, by submitting them 128 arrangements
of N pointers among 40k extracted from haproxy's pools, and the average
fill rates for 1- to 12- bit hashes were measured and compared. It was
clear that some values do not provide great hashes and other ones are
way more resistant.
The current value is not bad at all but delivers 42.6% unique 2-bit
outputs, 41.6% 3-bit, 38.0% 4-bit, 38.2% 5-bit and 37.1% 10-bit. Some
values did perform significantly better, among which 0xacd1be85 which
does 43.2% 2-bit, 42.5% 3-bit, 42.2% 4-bit, 39.2% 5-bit and 37.3% 10-bit.
The reverse value used in the ptr2_hash() was really underperforming and
was replaced with 0x9d28e4e9 which does 49.6%, 40.4%, 42.6%, 39.1%, and
37.2% respectvely.
This should slightly improve the accuracy of the task and memory
profiling, and will be useful for pools.
When testing the pointer hash on 64-bit real pointers (map entries),
it appeared that the shift by 33 bits that hoped to compensate for the
3 nul LSB degrades the hash, and the centering is more optimal on
31-(bits+1)/2. This makes sense since the topmost bit of the
multiplicator is 31, so for an input of 1 bit and 1 bit of output we
would always get zero. With the formula adjusted this way, we can get
up to ~15% more unique entries at 10 bits and ~24% more at 11 bits.
When dealing with macro-based size definitions, it is useful to be able
to hash pointers on zero bits so that the macro automatically returns a
constant 0. For now it only supports 1-32. Let's just add this special
case. It's automatically optimized out by the compiler since the function
is inlined.
LONGBITS was defined long ago with old compilers that didn't provide the
word size. It's still present as being referenced in various places in the
code, but we must not use it to define other macros that may be evaluated
at pre-processing time since it contains sizeof() and casts that are not
compatible with preprocessor conditions. Let's switch MAX_THREADS_PER_GROUP
to __WORDSIZE so that we can condition blocks of code on it if needed.
LONGBITS should really be removed by now, given that we don't support
compilers not providing __WORDSIZE anymore (gcc < 4.2).
By moving the config-time stuff after the updt_lock, we can plug some
holes without interfering with it. This allows us to get back to the
768-bytes struct. The performance was not affected at all.
The read-lock contention observed on the update lock while turning it
into an upgradable lock were due to false sharing with the nearby
updates. Simply moving the lock alone into its own cache line is
sufficient to almost double the performance again, raising from 2355
to 4480k RPS with very low contention:
Samples: 1M of event 'cycles', 4000 Hz, Event count (approx.): 743422995452 lost
Overhead Shared Object Symbol
15.88% haproxy [.] stktable_lookup_key
5.94% haproxy [.] ebmb_lookup
5.69% haproxy [.] http_wait_for_request
3.66% haproxy [.] stktable_touch_with_exp
2.62% [kernel] [k] _raw_spin_unlock_irqrestore
1.86% haproxy [.] http_action_return
1.79% haproxy [.] stream_process_counters
1.78% [kernel] [k] skb_release_data
1.77% haproxy [.] process_stream
Unfortunately, trying to move the line anywhere else didn't work,
despite the remaining holes, because this structure is not quite
clean. This adds 64 bytes to a struct that was already 768 long,
so it's now 832. It's possible to repack it a little bit and regain
these bytes by removing the THREAD_ALIGN before "keys" because we
rarely use the config stuff, but that's a bit unsafe.
Updating an entry in the updates tree is currently performed under the
table's write lock, which causes huge contention with other accesses
such as lookups and free. Aside the updates tree, the update,
localupdate and commitupdate variables, nothing is manipulated, so
let's create a distinct lock (updt_lock) to protect these together
to remove this contention. It required to add an extra lock in the
few places where we delete the update (though only if we're really
going to delete it) to protect the tree. This is very convenient
because now peer_send_teachmsgs() only needs to take this read lock,
and there is very little contention left on the stick-table.
With this alone, the performance jumped from 614k to 1140k/s on a
80-thread machine with a peers section! Stick-table updates with
no peers however now has to stand two locks and slightly regressed
from 4.0-4.1M/s to 3.9-4.0. This is fairly minimal compared to the
significant unlocking of the peers updates and considered totally
acceptable.
The structure currently mixes R/O and R/W fields, let's organize them
by access type, focusing mainly on splitting the updates from the rest
so that peers activity does not affect the rest. For now it doesn't
bring any benefit but it paves the way for splitting the lock.
Due to the ts->ref_cnt being manipulated and checked inside wrlocks,
we continue to have it updated under plenty of read locks, which have
an important cost on many-thread machines.
This patch turns them all to atomic ops and carefully moves them outside
of locks every time this is possible:
- the ref_cnt is incremented before write-unlocking on creation otherwise
the element could vanish before we can do it
- the ref_cnt is decremented after write-locking on release
- for all other cases it's updated out of locks since it's guaranteed by
the sequence that it cannot vanish
- checks are done before locking every time it's used to decide
whether we're going to release the element (saves several write locks)
- expiration tests are just done using atomic loads, since there's no
particular ordering constraint there, we just want consistent values.
For Lua, the loop that is used to dump stick-tables could switch to read
locks only, but this was not done.
For peers, the loop that builds updates in peer_send_teachmsgs is extremely
expensive in write locks and it doesn't seem this is really needed since
the only updated variables are last_pushed and commitupdate, the first
one being on the shared table (thus not used by other threads) and the
commitupdate could likely be changed using a CAS. Thus all of this could
theoretically move under a read lock, but that was not done here.
On a 80-thread machine with a peers section enabled, the request rate
increased from 415 to 520k rps.
This updates the local copy of the plock library to benefit from finer
memory ordering, EBO on more operations such as when take_w() and stow()
wait for readers to leave and refined EBO, especially on common operation
such as attempts to upgade R to S, and avoids a counter-productive prior
read in rtos() and take_r().
These changes have shown a 5% increase on regular operations on ARM,
a 33% performance increase on ARM on stick-tables and 2% on x86, and
a 14% and 4% improvements on peers updates respectively on ARM and x86.
The availability of relaxed operations will probably be useful for stats
counters which are still extremely expensive to update.
The following plock commits were included in this update:
9db830b plock: support inlining exponential backoff code
008d3c2 plock: make the rtos upgrade faster
2f76dde atomic: clean up the generic xchg()
3c6919b atomic: make sure that the no-return macros do not return a value
97c2bb7 atomic: make the fallback bts use the pointed type for the shift
f4c1880 atomic: also implement the missing pl_btr()
8329b82 atomic: guard all generic definitions to make it easier to provide specific ones
7c5cb62 atomic: use C11 atomics when available
96afaf9 atomic: prefer the C11 definitions in general
f3ec7a6 atomic: implement load/store/atomic barriers
8bdbd1e atomic: add atomic load/stores
0f604c0 atomic: add more _noret operations
3fe35db atomic: remove the (void) cast from the C11 operations
3b08a7c atomic: allow to define the fallback _noret variants
28deb22 atomic: make x86 arithmetic operations the _noret variants
8061fe2 atomic: handle modern compilers that support returning flags
b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>)
59817ca atomic: add memory order variants for most operations
a40774f plock: explicitly make use of the pl_*_noret operations
6f1861b plock: switch to pl_sub_noret_lax() for cancellation
c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd()
382eea3 plock: use a release ordering when dropping the lock
60d750d plock: use EBO when waiting for readers to leave in take_w() and stow()
fc01c4f plock: improve EBO a little bit
1ef6390 plock: switch to CAS + XADD for pl_take_r()
->xprt_ctx (struct ssl_sock_ctx) and ->conn (struct connection) must be kept
by the remaining QUIC connection object (struct quic_cc_conn) after having
release the previous one (struct quic_conn) to allow "show fd/sess" commands
to be functional without causing haproxy crashes.
No need to backport.
We're adding a new argument "relaxed" to h2_make_htx_request() so that
we can control its level of acceptance of certain invalid requests at
the proxy level with "option accept-invalid-http-request". The goal
will be to add deactivable checks that are still desirable to have by
default. For now no test is subject to it.
As its name implies, this function checks if a path component has any
forbidden headers starting at the designated location. The goal is to
seek from the result of a successful ist_find_range() for more precise
chars. Here we're focusing on 0x00-0x1F, 0x20 and 0x23 to make sure
we're not too strict at this point.
This looks up the character range <min>..<max> in the input string and
returns a pointer to the first one found. It's essentially the equivalent
of ist_find_ctl() in that it searches by 32 or 64 bits at once, but deals
with a range.
This function is not H2 specific but rather generic to HTTP. We'll
need it in H3 soon, so let's move it to HTTP and rename it to
http_header_has_forbidden_char().
When the connection enters the "connection closing" state after having sent
a datagram with CONNECTION_CLOSE frames inside its packets, a lot of memory
may be freed from quic_conn objects (QUIC connection). This is done allocating
a reduced sized object which keeps enough information to handle the remaining
incoming packets for the connection in "connection closing" state, and to
continue to send again the previous datagram with CONNECTION_CLOSE frames inside
which has already been sent.
Define a new quic_cc_conn struct which represents the connection object after
entering the "connection close" state and after having release the quic_conn
connection object.
Define <pool_head_quic_cc_conn> new pool for these quic_cc_conn struct objects.
Define QUIC_CONN_COMMON structure which is shared between quic_conn struct object
(the connection before entering "connection close" state), and new quic_cc_conn
struct object (the connection after entering "connection close"). So, all the
members inside QUIC_CONN_COMMON may be indifferently dereferenced from a
quic_conn struct or a quic_cc_conn struct pointer.
Implement qc_new_cc_conn() function to allocate such connections in
"connection close" state. This function is responsible of copying the
required information from the original connection (quic_conn) to the remaining
connection (quic_cc_conn). Among others initialization, it redefined the
QUIC packet handler task to quic_cc_conn_io_cb() and the idle timer task
to qc_cc_idle_timer_task(). quic_cc_conn_io_cb() drains the received and
resend the datagram which CONNECTION_CLOSE frame which has already been sent
when entering "connection close" state. qc_cc_idle_timer_task() only releases
the remaining quic_cc_conn struct object.
Modify quic_conn_release() to allocate quic_cc_conn struct objects from the
original connection passed as argument. It does nothing if this original
connection is not in closing state, or if the idle timer has already expired.
Implement quic_release_cc_conn() to release a "connection close" connection.
It is called when its timer expires or if an error occured when sending
a packet from this connection when the peer is no more reachable.
Add "quic_cids" new pool to allocate the ->cids trees of quic_conn objects.
Replace ->cids member of quic_conn objects by pointer to "quic_cids" and
adapt the code consequently. Nothing special.
Add a new pool <pool_head_quic_cc_buf> for buffer used when building datagram
wich CONNECTION_CLOSE frames inside with QUIC_MIN_CC_PKTSIZE(128) as minimum
size.
Add ->cc_buf_area to quic_conn struct to store such buffers.
Add ->cc_dgram_len to store the size of the "connection close" datagrams
and ->cc_buf a buffer struct to be used with ->cc_buf_area as ->area member
value.
Implement qc_get_txb() to be called in place of qc_txb_alloc() to allocate
a struct "quic_cc_buf" buffer when the connection needs an immediate close
or a buffer struct if not.
Modify qc_prep_hptks() and qc_prep_app_pkts() to allow them to use such
"quic_cc_buf" buffer when an immediate close is required.
Move rx.bytes, tx.bytes and tx.prep_bytes quic_conn struct member to
bytes anonymous struct (bytes.rx, bytes.tx and bytes.prep member respectively).
They are moved before being defined into a bytes anonoymous struct common to
a future struct to be defined.
Consequently adapt the code.
Add a BUG_ON() to quic_peer_validated_addr() to check the amplification limit
is respected when it return false(0), i.e. when the connection is not validated.
Implement quic_may_send_bytes() which returns the number of bytes which may be
sent when the connection has not already been validated and call this functions
at several places when this is the case (after having called
quic_peer_validated_addr()).
Furthermore, this patch improves the code maintainability. Some patches to
come will have to rename ->[rt]x.bytes quic_conn struct members.
A HTTP server may provide a complete response even prior receiving the
full request. In this case, RFC 9114 allows the server to abort read
with a STOP_SENDING with error code H3_NO_ERROR.
This scenario was notably reproduced with haproxy and an inactive
server. If the client send a POST request, haproxy may provide a full
HTTP 503 response before the end of the full request.
Fields of sedesc structure were documented in the comment about the
structure itself. It was not really convenient, hard to read, hard to
update. So comments about the fields are moved on the corresponding field
line, as usual.
There is a mechanisme in the H1 and H2 multiplexer to skip the payload when
a response is returned to the client when it must not contain any payload
(response to a HEAD request or a 204/304 response). However, this does not
work when the splicing is used. The H2 multiplexer does not support the
splicing, so there is no issue. But with the mux-h1, when data are sent
using the kernel splicing, the mux on the server side is not aware the
client side should skip the payload. And once the data are put in a pipe,
there is no way to stop the sending.
It is a defect of the current design. This will be easier to deal with this
case when the mux-to-mux forwarding will be implemented. But for now, to fix
the issue, we should add an HTX flag on the start-line to pass the info from
the client side to the server side and be able to disable the splicing in
necessary.
The associated reg-test was improved to be sure it does not fail when the
splicing is configured.
This patch should be backported as far as 2.4..
Move the TX part of the code to quic_tx.c.
Add quic_tx-t.h and quic_tx.h headers for this TX part code.
The definition of quic_tx_packet struct has been move from quic_conn-t.h to
quic_tx-t.h.
Same thing for the TX part:
Move the RX part of the code to quic_rx.c.
Add quic_rx-t.h and quic_rx.h headers for this TX part code.
The definition of quic_rx_packet struct has been move from quic_conn-t.h to
quic_rx-t.h.
Move the code which directly calls the functions of the OpenSSL QUIC API into
quic_ssl.c new C file.
Some code have been extracted from qc_conn_finalize() to implement only
the QUIC TLS part (see quic_tls_finalize()) into quic_tls.c.
qc_conn_finalize() has also been exported to be used from this new quic_ssl.c
C module.
To accelerate the compilation of quic_conn.c file, export the code in relation
with the traces from quic_conn.c to quic_trace.c.
Also add some headers (quic_trace-t.h and quic_trace.h).
This setting which may be used into a "global" section, enables the QUIC listener
bindings when haproxy is compiled with the OpenSSL wrapper. It has no effect
when haproxy is compiled against a TLS stack with QUIC support, typically quictls.
This wrapper needs to have an access to an encoded version of the local transport
parameter (to be sent to the peer). They are provided to the TLS stack thanks to
qc_ssl_compat_add_tps_cb() callback.
These encoded transport parameters were attached to the QUIC connection but
removed by this commit to save memory:
MINOR: quic: Stop storing the TX encoded transport parameters
This patch restores these transport parameters and attaches them again
to the QUIC connection (quic_conn struct), but only when the QUIC OpenSSL wrapper
is compiled.
Implement qc_set_quic_transport_params() to encode the transport parameters
for a connection and to set them into the stack and make this function work
for both the OpenSSL wrapper or any other TLS stack with QUIC support. Its uses
the encoded version of the transport parameters attached to the connection
when compiled for the OpenSSL wrapper, or local parameters when compiled
with TLS stack with QUIC support. These parameters are passed to
quic_transport_params_encode() and SSL_set_quic_transport_params() as before
this patch.
Include haproxy/quic_openssl_compat.h from haproxy/openssl-compat.h when the
compilation of the QUIC openssl wrapper for TLS stacks is enabled with
USE_QUIC_OPENSSLCOMPAT.
Highly inspired from nginx openssl wrapper code.
This wrapper implement this list of functions:
SSL_set_quic_method(),
SSL_quic_read_level(),
SSL_quic_write_level(),
SSL_set_quic_transport_params(),
SSL_provide_quic_data(),
SSL_process_quic_post_handshake()
and SSL_QUIC_METHOD QUIC specific bio method which are also implemented by quictls
to support QUIC from OpenSSL. So, its aims is to support QUIC from a standard OpenSSL
stack without QUIC support. It relies on the OpenSSL keylog feature to retreive
the secrets derived by the OpenSSL stack during a handshake and to pass them to
the ->set_encryption_secrets() callback as this is done by quictls. It makes
usage of a callback (quic_tls_compat_msg_callback()) to handle some TLS messages
only on the receipt path. Some of them must be passed to the ->add_handshake_data()
callback as this is done with quictls to be sent to the peer as CRYPTO data.
quic_tls_compat_msg_callback() callback also sends the received TLS alert with
->send_alert() callback.
AES 128-bits with CCM mode is not supported at this time. It is often disabled by
the OpenSSL stack, but as it can be enabled by "ssl-default-bind-ciphersuites",
the wrapper will send a TLS alerts (Handhshake failure) if this algorithm is
negotiated between the client and the server.
0rtt is also not supported by this wrapper.
The aim of this patch is to allow the building of QUIC datagrams with
as much as packets with different encryption levels inside during handshake.
At this time, this is possible only for at most two encryption levels.
That said, most of the time, a server only needs to use two encryption levels
by datagram, except during retransmissions.
Modify qc_prep_pkts(), the function responsible of building datagrams, to pass
a list of encryption levels as parameter in place of two encryption levels. This
function is also used when retransmitting datagrams. In this case this is a
customized/flexible list of encryption level which is passed to this function.
Add ->retrans new member to quic_enc_level struct, to be used as attach point
to list of encryption level used only during retransmission, and ->retrans_frms
new member which is a pointer to a list of frames to be retransmitted.
