Tarantool development patches archive
 help / color / mirror / Atom feed
From: "Ilya Kosarev" <i.kosarev@tarantool.org>
To: "Vladislav Shpilevoy" <v.shpilevoy@tarantool.org>
Cc: tarantool-patches@dev.tarantool.org
Subject: Re: [Tarantool-patches] [PATCH v3] iproto: make iproto thread more independent from tx
Date: Mon, 19 Oct 2020 12:49:54 +0300	[thread overview]
Message-ID: <1603100994.191990135@f429.i.mail.ru> (raw)
In-Reply-To: <903df397-8107-1619-e589-2e41c52d0478@tarantool.org>

[-- Attachment #1: Type: text/plain, Size: 16538 bytes --]


Hi,
 
Thanks for your review!
 
Sent v4 of the patch considering the comments and privat discussion.
 
Some answers are below.
  
>Пятница, 2 октября 2020, 2:45 +03:00 от Vladislav Shpilevoy <v.shpilevoy@tarantool.org>:
> 
>Hi! Thanks for the patch! Very nice! Extreme niceness! Optimistic connection handling!
>
>See 17 small minor comments below.
>
>1. The branch fails quite hard all the iproto-related tests on my machine,
>some with crashes. Also when I tried to run it in the console, I get
>'Peer closed' right after netbox.connect.
>
>However I see CI green, and I am not sure why does it pass there. On my
>machine all networking it totally broken on your branch.
>
>I fixed it by adding this:
>
>@@ -578,6 +578,7 @@ iproto_msg_new(struct iproto_connection *con)
>return NULL;
>}
>msg->connection = con;
>+ msg->close_connection = false;
>rmean_collect(rmean_net, IPROTO_REQUESTS, 1);
>return msg;
>}
>
>It seems it is your bug, because previously it was used only on
>connection, and initialized only in iproto_on_accept(). Now you
>use it on each request, but not initialize on each request. Also this
>should have been a huge luck it didn't fail on any CI job.
Right, my fault here. Fixed in v4.
>
>On 26.09.2020 00:53, Ilya Kosarev wrote:
>> On connection, an evio service callback is invoked to accept it. The
>> next step after acception was to process connection to tx thread
>> through cbus. This meant that any connection interaction involves
>> tx thread even before we get to decode what does the client want
>> from us. Consequently, a number of problems appears. The main one
>> is that we might get descriptor leak in case of unresponsive tx thread
>> (for example, when building secondary index).
>
>2. It is not a leak really. Just too many clients waiting for being
>processed. Descriptors were not lost completely.
>
>As for the issue - what was the problem not to start listening until
>recovery is complete?
>
>And how does this patch solve it, if the connections are still kept,
>and so their descriptors? How did you validate the problem was gone?
>It seems the descriptors still are created, and still are not closed.
>
>If the problem is not gone, that why does this patch does what it does?
This topic was discussed with voice. Short answer here:
Previously iproto had only accepted connections without any
further actions, leading to the very fast descriptors exhaustion.
We can’t not start listening before recovery: it is needed for replication.
Patches solves is, cause the connections are now greeted and
they are not trying reconnect, wasting sockets. There is a reproducer
in the ticket, which i can't see how to include in test-run. But it shows
the difference.
>> There are some other
>> cases where we might not want to spend precious tx time to process
>> the connection in case iproto can do it all alone.
>
>3. What are these other cases? I can think only of SWIM in iproto thread,
>but it is not related at all.
The cases are that we can now add more logic to request
processing in iproto, if needed, when tx is not really needed to answer.
For example, see #4646.
This can also solve some problems with improper tx answers. 
>
>> This patch allows iproto to accept connection and send greeting by
>> itself. The connection is initialized in tx thread when the real
>> request comes through iproto_msg_decode(). In case request type was not
>> recognized we can also send reply with an error without using tx. It is
>> planned to add more iproto logic to prevent extra interaction with
>> tx thread. This patch already to some extent solves descriptors leakage
>> problem as far as connection establishes and stays in fetch_schema
>> state while tx thread is unresponsive.
>> The other user visible change is that on_connect triggers won't run on
>> connections that don't provide any input, as reflected in
>> bad_trigger.test.py.
>>
>> Part of #3776
>>
>> @TarantoolBot document
>> Title: iproto: on_connect triggers execution
>> Update the documentation for on_connect triggers to reflect that they
>> are now being executed only with the first request. Though the triggers
>> are still the first thing to be executed on a new connection. While it
>> is quite a synthetic case to establish a connection without making
>> any requests it is technically possible and now your triggers won't be
>> executed in this case. Some request is required to start their
>> execution.
>
>4. I am not sure I understand. According to 3776 there was a problem, that
>the connections were made. So they didn't do any requests? And the
>case is not that syntetic as it seems?
They didn’t make any requests just because they weren’t even greeted.
Instead they were reconnecting all the time. As soon as they are greeted
they are sending requests.
>
>What was the problem to preserve the original behaviour? For example,
>you could make the on_connect execution delayed. TX thread could tell
>when recovery is done, and iproto would send the pending connections to it.
I think we don’t really need to deal with the connections that don’t send requests.
>
>> src/box/iproto.cc | 317 +++++++++++++++++++++-----------
>> test/box-py/bad_trigger.result | 1 +
>> test/box-py/bad_trigger.test.py | 22 ++-
>> test/sql/prepared.result | 4 +
>> test/sql/prepared.test.lua | 1 +
>> 5 files changed, 239 insertions(+), 106 deletions(-)
>>
>> diff --git a/src/box/iproto.cc b/src/box/iproto.cc
>> index b8f65e5ec..9f98fce86 100644
>> --- a/src/box/iproto.cc
>> +++ b/src/box/iproto.cc
>> @@ -463,6 +474,7 @@ struct iproto_connection
>> struct ev_io output;
>> /** Logical session. */
>> struct session *session;
>> + bool init_failed;
>
>5. Flags need 'is_' prefix.
Right, ok. Done in v4.
>
>6. What is this? Why is it needed? Usually we leave comments
>to each struct member. Otherwise it is hard to understand why
>some of them are needed.
Right, added the comment in v4.
>
>7. The field is accessed from tx thread only, and is close to
>fields mutated from iproto thread. That leads to false-sharing
>problem. Please, move it to iproto_connection.tx sub-struct.
Right! Fixed in v4.
>
>> ev_loop *loop;
>> /**
>> * Pre-allocated disconnect msg. Is sent right after
>> @@ -677,9 +700,19 @@ iproto_connection_close(struct iproto_connection *con)
>> * is done only once.
>> */
>> con->p_ibuf->wpos -= con->parse_size;
>> - cpipe_push(&tx_pipe, &con->disconnect_msg);
>> assert(con->state == IPROTO_CONNECTION_ALIVE);
>> con->state = IPROTO_CONNECTION_CLOSED;
>> + rlist_del(&con->in_stop_list);
>> + if (con->session != NULL) {
>> + cpipe_push(&tx_pipe, &con->disconnect_msg);
>> + } else {
>> + /*
>> + * In case session was not created we can safely
>> + * try to start destroy not involving tx thread.
>> + */
>> + iproto_connection_try_to_start_destroy(con);
>> + }
>> + return;
>
>8. You wouldn't need the 'return' and wouldn't need to duplicate
>'rlist_del(&con->in_stop_list);' call, if you would move
>'rlist_del(&con->in_stop_list);' to the beginning of this function
>from its end.
Right, done in v4.
>
>> } else if (con->state == IPROTO_CONNECTION_PENDING_DESTROY) {
>> iproto_connection_try_to_start_destroy(con);
>> } else {
>> @@ -809,6 +842,7 @@ iproto_enqueue_batch(struct iproto_connection *con, struct ibuf *in)
>> assert(rlist_empty(&con->in_stop_list));
>> int n_requests = 0;
>> bool stop_input = false;
>> + bool obuf_in_iproto = (con->session == NULL);
>
>9. We always name flags starting from 'is_' prefix or a similar one like 'has_',
>'does_'. Please, follow that agreement.
All right, added in v4.
>
>> const char *errmsg;
>> while (con->parse_size != 0 && !stop_input) {
>> if (iproto_check_msg_max()) {
>> @@ -1314,13 +1367,24 @@ iproto_msg_decode(struct iproto_msg *msg, const char **pos, const char *reqend,
>> (uint32_t) type);
>> goto error;
>> }
>> - return;
>> + return 0;
>> error:
>> /** Log and send the error. */
>> diag_log();
>> - diag_create(&msg->diag);
>> - diag_move(&fiber()->diag, &msg->diag);
>> - cmsg_init(&msg->base, error_route);
>> + if (msg->connection->session != NULL) {
>> + diag_create(&msg->diag);
>> + diag_move(&fiber()->diag, &msg->diag);
>> + cmsg_init(&msg->base, error_route);
>> + return 1;
>
>10. Why do you need 3 return values? You only check for -1. So you
>could return 0 here like it was before, and -1 below. In the
>calling code you would check != 0 like we do everywhere.
Well, the idea was that we have 2 different error cases. But actually
it should be: 0 if the route was inited. -1 otherwise. Now it is like that.
Fixed in v4.
>
>> + }
>> + /*
>> + * In case session was not created we can process error path
>> + * without tx thread.
>> + */
>> + tx_accept_wpos(msg->connection, &msg->wpos);
>> + tx_reply_error(msg);
>
>11. There is a naming convention, that all tx_ functions are called in
>TX thread always. All net_ and iproto_ functions are called in IProto
>thread. Lets not violate the rule and obfuscate the iproto.cc code
>even more that it is. Please, come up with new names reflecting where
>the functions are used. Or at least not specifying it as tx_.
Well, this convention is already broken to some extent. For example,
tx_reply_error() itself calls  iproto_reply_error() and
iproto_wpos_create(). tx_process_sql() calls  iproto_reply_sql().
Though i agree that it is a good idea to follow the convention. I decided to
introduce reply_error() and obuf_accept_wpos() in v4.
>
>> + net_send_msg(&(msg->base));
>
>12. The function is called 'iproto_msg_decode'. Lets leave it do decoding,
>and not send anything from it. It just violates its purpose and name.
Right, fixed in v4. 
>
>> + return -1;
>> }
>>
>> static void
>> @@ -1478,13 +1538,27 @@ tx_accept_wpos(struct iproto_connection *con, const struct iproto_wpos *wpos)
>> }
>> }
>>
>> -static inline struct iproto_msg *
>> -tx_accept_msg(struct cmsg *m)
>> +static inline int
>> +tx_accept_msg(struct cmsg *m, struct iproto_msg **msg)
>> {
>> - struct iproto_msg *msg = (struct iproto_msg *) m;
>> - tx_accept_wpos(msg->connection, &msg->wpos);
>> - tx_fiber_init(msg->connection->session, msg->header.sync);
>> - return msg;
>> + *msg = (struct iproto_msg *) m;
>
>13. Why do you change the return type and add an out parameter? You
>could just return NULL in case of an error.
No, i can’t, cause i will need this msg to process the error. I can’t get NULL
instead.
>
>> + /*
>> + * In case connection init failed we don't need to try anymore.
>> + */
>> + if ((*msg)->connection->init_failed)
>> + return -1;
>> + /*
>> + * In case session was not created we need to init connection in tx and
>> + * create it here.
>> + */
>> + if ((*msg)->connection->session == NULL && tx_init_connect(*msg) != 0) {
>> + (*msg)->connection->init_failed = true;
>> + (*msg)->close_connection = true;
>> + return -1;
>> + }
>> + tx_accept_wpos((*msg)->connection, &(*msg)->wpos);
>> + tx_fiber_init((*msg)->connection->session, (*msg)->header.sync);
>> + return 0;
>> }
>>
>> /**
>> @@ -1507,7 +1581,14 @@ tx_reply_error(struct iproto_msg *msg)
>> static void
>> tx_reply_iproto_error(struct cmsg *m)
>> {
>> - struct iproto_msg *msg = tx_accept_msg(m);
>> + struct iproto_msg *msg;
>> + /*
>> + * We don't need to check tx_accept_msg() return value here
>> + * as far as if we might only process iproto error in tx
>> + * in case connection session is already created and
>> + * thus tx_accept_msg() can't fail.
>> + */
>> + tx_accept_msg(m, &msg);
>
>14. Well, if it fails, you have msg == NULL, and the code below will
>crash. Otherwise it should be an assertion, not just a comment.
Right, introduced the assertion in v4.
>
>> struct obuf *out = msg->connection->tx.p_obuf;
>> iproto_reply_error(out, diag_last_error(&msg->diag),
>> msg->header.sync, ::schema_version);
>> @@ -1865,6 +1961,29 @@ tx_process_replication(struct cmsg *m)
>> }
>> }
>>
>> +/**
>> + * Check connection health and try to send an error to the client
>> + * in case of internal connection init or on_connect trigger failure.
>> + */
>> +static bool
>> +iproto_connection_fail(struct iproto_msg *msg)
>
>15. The function is called 'connection_fail' and literally does nothing
>when the connection is fine. Does not this name look wrong to you?
Ok, right, it will be iproto_connection_check_valid() in v4.
>
>> +{
>> + if (!msg->close_connection)
>> + return false;
>> + struct iproto_connection *con = msg->connection;
>> + int64_t nwr = sio_writev(con->output.fd, msg->wpos.obuf->iov,
>> + obuf_iovcnt(msg->wpos.obuf));
>
>16. You can't just write to a socket whenever you want. It may be
>not writable. Please, use libev events and callbacks to send anything.
>You can do a write only when you get EV_WRITE event from libev, who
>in turn gets it from select/poll/epoll/kqueue/whatever-else-is-provided
>by the OS.
Right. As discussed in voice, it is special error case. In case we will
do the normal way, we won’t be able to send the error. I checked the
case with bad_trigger.test and it clearly shows that the client won’t
receive  the error message in case we will try to wait for the event here,
while sio_writev() does its job.
>
>> + if (nwr > 0) {
>> + /* Count statistics. */
>> + rmean_collect(rmean_net, IPROTO_SENT, nwr);
>> + } else if (nwr < 0 && ! sio_wouldblock(errno)) {
>> + diag_log();
>> + }
>> + iproto_connection_close(con);
>> + iproto_msg_delete(msg);
>> + return true;
>> +}
>> +
>> static void
>> net_send_msg(struct cmsg *m)
>> {
>> @@ -1936,81 +2066,60 @@ net_end_subscribe(struct cmsg *m)
>> }
>>
>> /**
>> - * Handshake a connection: invoke the on-connect trigger
>> - * and possibly authenticate. Try to send the client an error
>> - * upon a failure.
>> + * Handshake a connection: prepare greeting for it.
>> */
>> static void
>> -tx_process_connect(struct cmsg *m)
>> +iproto_process_connect(struct iproto_msg *msg)
>> {
>> - struct iproto_msg *msg = (struct iproto_msg *) m;
>> struct iproto_connection *con = msg->connection;
>> struct obuf *out = msg->connection->tx.p_obuf;
>> - try { /* connect. */
>> - con->session = session_create(SESSION_TYPE_BINARY);
>> - if (con->session == NULL)
>> - diag_raise();
>> - con->session->meta.connection = con;
>> - tx_fiber_init(con->session, 0);
>> - char *greeting = (char *) static_alloc(IPROTO_GREETING_SIZE);
>> - /* TODO: dirty read from tx thread */
>> - struct tt_uuid uuid = INSTANCE_UUID;
>> - random_bytes(con->salt, IPROTO_SALT_SIZE);
>> - greeting_encode(greeting, tarantool_version_id(), &uuid,
>> - con->salt, IPROTO_SALT_SIZE);
>> - obuf_dup_xc(out, greeting, IPROTO_GREETING_SIZE);
>> - if (! rlist_empty(&session_on_connect)) {
>> - if (session_run_on_connect_triggers(con->session) != 0)
>> - diag_raise();
>> - }
>> + char *greeting = (char *) static_alloc(IPROTO_GREETING_SIZE);
>> + /* TODO: dirty read from tx thread */
>> + struct tt_uuid uuid = INSTANCE_UUID;
>> + random_bytes(con->salt, IPROTO_SALT_SIZE);
>
>17. Sorry, but wtf?! This data right now may be being written by TX thread,
>so you literally may read and return *garbage* here. It is not acceptable.
>It is not even a short term solution, it is just a bug, if you can't prove
>the value is already initialized and constant by that moment.
>
>If listen starts before recovery, and UUID is stored in the xlogs, then it
>is definitely a bug.
As discussed in voice, actually, everything is fine. My fault to leave the
wrong comment here (I was not totally sure everything is fine, but the
comment was also wrong before my patch). In v4 it is replaced with
the correct one, that explains the situation. Here it is:
/*
 * INSTANCE_UUID is guaranteed to be inited before this moment.
 * We start listening either in local_recovery() or bootstrap().
 * The INSTANCE_UUID is ensured to be inited in the beginning of
 * both methods. In case of local_recovery() it is verified that
 * INSTANCE_UUID was read from the snapshot in memtx_engine_new().
 * In bootstrap() INSTANCE_UUID is either taken from the
 * instance_uuid box.cfg{} param or created on the spot.
 */
>
>> + greeting_encode(greeting, tarantool_version_id(), &uuid,
>> + con->salt, IPROTO_SALT_SIZE);
>> + if (obuf_dup(out, greeting, IPROTO_GREETING_SIZE)
>> + != IPROTO_GREETING_SIZE) {
>> + diag_set(OutOfMemory, IPROTO_GREETING_SIZE,
>> + "greeting obuf", "dup");
>> + iproto_reply_error(out, diag_last_error(&fiber()->diag),
>> + msg->header.sync, ::schema_version);
>> iproto_wpos_create(&msg->wpos, out);
>> - } catch (Exception *e) {
>> - tx_reply_error(msg);
>> msg->close_connection = true;
>> - }
>> -}
--
Ilya Kosarev

[-- Attachment #2: Type: text/html, Size: 23534 bytes --]

  reply	other threads:[~2020-10-19  9:49 UTC|newest]

Thread overview: 4+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-09-25 22:53 Ilya Kosarev
2020-10-01 23:45 ` Vladislav Shpilevoy
2020-10-19  9:49   ` Ilya Kosarev [this message]
2020-10-27 22:06     ` Vladislav Shpilevoy

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1603100994.191990135@f429.i.mail.ru \
    --to=i.kosarev@tarantool.org \
    --cc=tarantool-patches@dev.tarantool.org \
    --cc=v.shpilevoy@tarantool.org \
    --subject='Re: [Tarantool-patches] [PATCH v3] iproto: make iproto thread more independent from tx' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox