* [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail
@ 2020-07-20 20:43 Vladislav Shpilevoy
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors Vladislav Shpilevoy
` (3 more replies)
0 siblings, 4 replies; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-20 20:43 UTC (permalink / raw)
To: tarantool-patches, gorcunov
The patchset fixes the crash when a synchronous transaction
couldn't to local master's WAL, getting an error from inside of
the WAL thread.
Note, the test is added to the existing file instead of creating
a new one intentionally. It is not so cheap to create new instance
for each new test, and also would complicate running all qsync
tests by prefix like this: `python test-run.py replication/qsync`.
Besides, the test is simple. Copy-pastable to console not counting
test_run:switch() calls.
Branch: http://github.com/tarantool/tarantool/tree/gerold103/gh-5146-qsync-crash
Issue: https://github.com/tarantool/tarantool/issues/5146
Vladislav Shpilevoy (2):
txn: single failure point for WAL and TX commit errors
txn: single failure point for WAL and TX async commit errors
src/box/txn.c | 88 +++++++++++---------------
test/replication/qsync_errinj.result | 69 ++++++++++++++++++++
test/replication/qsync_errinj.test.lua | 28 ++++++++
3 files changed, 134 insertions(+), 51 deletions(-)
--
2.21.1 (Apple Git-122.3)
^ permalink raw reply [flat|nested] 9+ messages in thread
* [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors
2020-07-20 20:43 [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
@ 2020-07-20 20:43 ` Vladislav Shpilevoy
2020-07-21 8:42 ` Cyrill Gorcunov
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async " Vladislav Shpilevoy
` (2 subsequent siblings)
3 siblings, 1 reply; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-20 20:43 UTC (permalink / raw)
To: tarantool-patches, gorcunov
Txn_commit() uses journal_write() function to send requests to WAL
thread. journal_write() can return 0/-1, but these codes have
nothing to do with the actual write result. journal_write() only
signals whether the interaction with WAL thread finished
successfully. It tells nothing about what happened in this
interaction.
To check WAL write result need to look at journal_entry.res field.
As a result, there were 2 failure points to handle. One of them
wasn't handled for synchronous transactions. Not counting 4
failure points in other places:
- When can't prepare the transaction before commit;
- When can't allocate a journal entry;
- When can't append a new entry to the qsync limbo;
- When synchronous transaction completion wait fails.
This patch merges all the failure points into one place for
txn_commit().
Closes #5146
---
src/box/txn.c | 55 +++++++++-----------
test/replication/qsync_errinj.result | 69 ++++++++++++++++++++++++++
test/replication/qsync_errinj.test.lua | 28 +++++++++++
3 files changed, 120 insertions(+), 32 deletions(-)
diff --git a/src/box/txn.c b/src/box/txn.c
index 62b91d6bb..375c01f85 100644
--- a/src/box/txn.c
+++ b/src/box/txn.c
@@ -797,11 +797,8 @@ txn_commit(struct txn *txn)
txn->fiber = fiber();
- if (txn_prepare(txn) != 0) {
- txn_rollback(txn);
- txn_free(txn);
- return -1;
- }
+ if (txn_prepare(txn) != 0)
+ goto rollback;
if (txn_commit_nop(txn)) {
txn_free(txn);
@@ -809,11 +806,8 @@ txn_commit(struct txn *txn)
}
req = txn_journal_entry_new(txn);
- if (req == NULL) {
- txn_rollback(txn);
- txn_free(txn);
- return -1;
- }
+ if (req == NULL)
+ goto rollback;
bool is_sync = txn_has_flag(txn, TXN_WAIT_SYNC);
if (is_sync) {
@@ -829,24 +823,17 @@ txn_commit(struct txn *txn)
* wouldn't be acceptable.
*/
limbo_entry = txn_limbo_append(&txn_limbo, origin_id, txn);
- if (limbo_entry == NULL) {
- txn_rollback(txn);
- txn_free(txn);
- return -1;
- }
+ if (limbo_entry == NULL)
+ goto rollback;
}
fiber_set_txn(fiber(), NULL);
- if (journal_write(req) != 0) {
+ if (journal_write(req) != 0 || req->res < 0) {
if (is_sync)
txn_limbo_abort(&txn_limbo, limbo_entry);
- fiber_set_txn(fiber(), txn);
- txn_rollback(txn);
- txn_free(txn);
-
diag_set(ClientError, ER_WAL_IO);
diag_log();
- return -1;
+ goto rollback;
}
if (is_sync) {
if (txn_has_flag(txn, TXN_WAIT_ACK)) {
@@ -861,25 +848,29 @@ txn_commit(struct txn *txn)
/* Local WAL write is a first 'ACK'. */
txn_limbo_ack(&txn_limbo, txn_limbo.instance_id, lsn);
}
- if (txn_limbo_wait_complete(&txn_limbo, limbo_entry) < 0) {
- txn_free(txn);
- return -1;
- }
+ if (txn_limbo_wait_complete(&txn_limbo, limbo_entry) < 0)
+ goto rollback;
}
if (!txn_has_flag(txn, TXN_IS_DONE)) {
txn->signature = req->res;
txn_complete(txn);
}
-
- int res = txn->signature >= 0 ? 0 : -1;
- if (res != 0) {
- diag_set(ClientError, ER_WAL_IO);
- diag_log();
- }
+ assert(txn->signature >= 0);
/* Synchronous transactions are freed by the calling fiber. */
txn_free(txn);
- return res;
+ return 0;
+
+rollback:
+ assert(txn->fiber != NULL);
+ if (!txn_has_flag(txn, TXN_IS_DONE)) {
+ fiber_set_txn(fiber(), txn);
+ txn_rollback(txn);
+ } else {
+ assert(in_txn() == NULL);
+ }
+ txn_free(txn);
+ return -1;
}
void
diff --git a/test/replication/qsync_errinj.result b/test/replication/qsync_errinj.result
index 49429cc80..1c6e357f7 100644
--- a/test/replication/qsync_errinj.result
+++ b/test/replication/qsync_errinj.result
@@ -214,6 +214,75 @@ box.space.sync:select{13}, box.space.sync:select{14}, box.space.sync:select{15}
| - - [15]
| ...
+--
+-- gh-5146: txn module should correctly handle errors from WAL
+-- thread, not only TX thread errors.
+--
+test_run:switch('default')
+ | ---
+ | - true
+ | ...
+box.cfg{replication_synchro_quorum = 3, replication_synchro_timeout = 1000}
+ | ---
+ | ...
+box.space.sync:truncate()
+ | ---
+ | ...
+-- Make a next row stuck in WAL thread.
+box.error.injection.set("ERRINJ_WAL_DELAY", true)
+ | ---
+ | - ok
+ | ...
+ok1, err1 = nil
+ | ---
+ | ...
+f1 = fiber.create(function() \
+ ok1, err1 = pcall(box.space.sync.replace, box.space.sync, {1}) \
+end)
+ | ---
+ | ...
+-- When the suspended write will be set free, it will stumble into
+-- an error.
+box.error.injection.set("ERRINJ_WAL_ROTATE", true)
+ | ---
+ | - ok
+ | ...
+-- Set the suspected write free. TX thread will receive success
+-- in terms of talking to WAL, but will get error inside the WAL
+-- response.
+box.error.injection.set("ERRINJ_WAL_DELAY", false)
+ | ---
+ | - ok
+ | ...
+test_run:wait_cond(function() return f1:status() == 'dead' end)
+ | ---
+ | - true
+ | ...
+ok1, err1
+ | ---
+ | - false
+ | - Failed to write to disk
+ | ...
+box.error.injection.set("ERRINJ_WAL_ROTATE", false)
+ | ---
+ | - ok
+ | ...
+box.cfg{replication_synchro_quorum = 2}
+ | ---
+ | ...
+box.space.sync:replace{2}
+ | ---
+ | - [2]
+ | ...
+test_run:switch('replica')
+ | ---
+ | - true
+ | ...
+box.space.sync:select{}
+ | ---
+ | - - [2]
+ | ...
+
test_run:cmd('switch default')
| ---
| - true
diff --git a/test/replication/qsync_errinj.test.lua b/test/replication/qsync_errinj.test.lua
index fe8bb4387..da554c775 100644
--- a/test/replication/qsync_errinj.test.lua
+++ b/test/replication/qsync_errinj.test.lua
@@ -77,6 +77,34 @@ box.space.sync:select{13}, box.space.sync:select{14}, box.space.sync:select{15}
test_run:switch('replica')
box.space.sync:select{13}, box.space.sync:select{14}, box.space.sync:select{15}
+--
+-- gh-5146: txn module should correctly handle errors from WAL
+-- thread, not only TX thread errors.
+--
+test_run:switch('default')
+box.cfg{replication_synchro_quorum = 3, replication_synchro_timeout = 1000}
+box.space.sync:truncate()
+-- Make a next row stuck in WAL thread.
+box.error.injection.set("ERRINJ_WAL_DELAY", true)
+ok1, err1 = nil
+f1 = fiber.create(function() \
+ ok1, err1 = pcall(box.space.sync.replace, box.space.sync, {1}) \
+end)
+-- When the suspended write will be set free, it will stumble into
+-- an error.
+box.error.injection.set("ERRINJ_WAL_ROTATE", true)
+-- Set the suspected write free. TX thread will receive success
+-- in terms of talking to WAL, but will get error inside the WAL
+-- response.
+box.error.injection.set("ERRINJ_WAL_DELAY", false)
+test_run:wait_cond(function() return f1:status() == 'dead' end)
+ok1, err1
+box.error.injection.set("ERRINJ_WAL_ROTATE", false)
+box.cfg{replication_synchro_quorum = 2}
+box.space.sync:replace{2}
+test_run:switch('replica')
+box.space.sync:select{}
+
test_run:cmd('switch default')
box.cfg{ \
--
2.21.1 (Apple Git-122.3)
^ permalink raw reply [flat|nested] 9+ messages in thread
* [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async commit errors
2020-07-20 20:43 [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors Vladislav Shpilevoy
@ 2020-07-20 20:43 ` Vladislav Shpilevoy
2020-07-21 8:43 ` Cyrill Gorcunov
2020-07-20 20:44 ` [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
2020-07-21 20:25 ` Vladislav Shpilevoy
3 siblings, 1 reply; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-20 20:43 UTC (permalink / raw)
To: tarantool-patches, gorcunov
The same as the previous commit, but applied to the async
transaction commit (txn_commit_async()). There were 6 failure
points. After the patch the failures are handled in one place.
Follow up #5146
---
src/box/txn.c | 33 ++++++++++++++-------------------
1 file changed, 14 insertions(+), 19 deletions(-)
diff --git a/src/box/txn.c b/src/box/txn.c
index 375c01f85..bd9fcdbe6 100644
--- a/src/box/txn.c
+++ b/src/box/txn.c
@@ -711,23 +711,18 @@ txn_commit_async(struct txn *txn)
* output to mark this event.
*/
diag_log();
- txn_rollback(txn);
- return -1;
+ goto rollback;
});
- if (txn_prepare(txn) != 0) {
- txn_rollback(txn);
- return -1;
- }
+ if (txn_prepare(txn) != 0)
+ goto rollback;
if (txn_commit_nop(txn))
return 0;
req = txn_journal_entry_new(txn);
- if (req == NULL) {
- txn_rollback(txn);
- return -1;
- }
+ if (req == NULL)
+ goto rollback;
bool is_sync = txn_has_flag(txn, TXN_WAIT_SYNC);
struct txn_limbo_entry *limbo_entry;
@@ -741,19 +736,16 @@ txn_commit_async(struct txn *txn)
struct trigger *trig =
region_alloc_object(&txn->region, typeof(*trig), &size);
if (trig == NULL) {
- txn_rollback(txn);
diag_set(OutOfMemory, size, "region_alloc_object",
"trig");
- return -1;
+ goto rollback;
}
/* See txn_commit(). */
uint32_t origin_id = req->rows[0]->replica_id;
limbo_entry = txn_limbo_append(&txn_limbo, origin_id, txn);
- if (limbo_entry == NULL) {
- txn_rollback(txn);
- return -1;
- }
+ if (limbo_entry == NULL)
+ goto rollback;
if (txn_has_flag(txn, TXN_WAIT_ACK)) {
int64_t lsn = req->rows[txn->n_applier_rows - 1]->lsn;
@@ -779,14 +771,17 @@ txn_commit_async(struct txn *txn)
fiber_set_txn(fiber(), NULL);
if (journal_write_async(req) != 0) {
fiber_set_txn(fiber(), txn);
- txn_rollback(txn);
-
diag_set(ClientError, ER_WAL_IO);
diag_log();
- return -1;
+ goto rollback;
}
return 0;
+
+rollback:
+ assert(txn->fiber == NULL);
+ txn_rollback(txn);
+ return -1;
}
int
--
2.21.1 (Apple Git-122.3)
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail
2020-07-20 20:43 [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors Vladislav Shpilevoy
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async " Vladislav Shpilevoy
@ 2020-07-20 20:44 ` Vladislav Shpilevoy
2020-07-21 20:25 ` Vladislav Shpilevoy
3 siblings, 0 replies; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-20 20:44 UTC (permalink / raw)
To: tarantool-patches, gorcunov
@ChangeLog:
* Fixed a crash on master WAL write fail at attempt to commit a synchronous transaction (gh-5146).
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors Vladislav Shpilevoy
@ 2020-07-21 8:42 ` Cyrill Gorcunov
0 siblings, 0 replies; 9+ messages in thread
From: Cyrill Gorcunov @ 2020-07-21 8:42 UTC (permalink / raw)
To: Vladislav Shpilevoy; +Cc: tarantool-patches
On Mon, Jul 20, 2020 at 10:43:05PM +0200, Vladislav Shpilevoy wrote:
> Txn_commit() uses journal_write() function to send requests to WAL
> thread. journal_write() can return 0/-1, but these codes have
> nothing to do with the actual write result. journal_write() only
> signals whether the interaction with WAL thread finished
> successfully. It tells nothing about what happened in this
> interaction.
>
> To check WAL write result need to look at journal_entry.res field.
> As a result, there were 2 failure points to handle. One of them
> wasn't handled for synchronous transactions. Not counting 4
> failure points in other places:
>
> - When can't prepare the transaction before commit;
> - When can't allocate a journal entry;
> - When can't append a new entry to the qsync limbo;
> - When synchronous transaction completion wait fails.
>
> This patch merges all the failure points into one place for
> txn_commit().
>
> Closes #5146
Acked-by: Cyrill Gorcunov <gorcunov@gmail.com>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async commit errors
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async " Vladislav Shpilevoy
@ 2020-07-21 8:43 ` Cyrill Gorcunov
0 siblings, 0 replies; 9+ messages in thread
From: Cyrill Gorcunov @ 2020-07-21 8:43 UTC (permalink / raw)
To: Vladislav Shpilevoy; +Cc: tarantool-patches
On Mon, Jul 20, 2020 at 10:43:06PM +0200, Vladislav Shpilevoy wrote:
> The same as the previous commit, but applied to the async
> transaction commit (txn_commit_async()). There were 6 failure
> points. After the patch the failures are handled in one place.
>
> Follow up #5146
Acked-by: Cyrill Gorcunov <gorcunov@gmail.com>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail
2020-07-20 20:43 [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
` (2 preceding siblings ...)
2020-07-20 20:44 ` [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
@ 2020-07-21 20:25 ` Vladislav Shpilevoy
2020-07-22 10:41 ` Kirill Yukhin
3 siblings, 1 reply; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-21 20:25 UTC (permalink / raw)
To: tarantool-patches, gorcunov
Pushed to master.
(Sorry, I forgot to add Acked-By, realized this after push.)
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail
2020-07-21 20:25 ` Vladislav Shpilevoy
@ 2020-07-22 10:41 ` Kirill Yukhin
2020-07-22 22:27 ` Vladislav Shpilevoy
0 siblings, 1 reply; 9+ messages in thread
From: Kirill Yukhin @ 2020-07-22 10:41 UTC (permalink / raw)
To: Vladislav Shpilevoy; +Cc: tarantool-patches
Hello,
On 21 июл 22:25, Vladislav Shpilevoy wrote:
> Pushed to master.
Why you didn't commit it to 2.5?
--
Regards, Kirill Yukhin
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail
2020-07-22 10:41 ` Kirill Yukhin
@ 2020-07-22 22:27 ` Vladislav Shpilevoy
0 siblings, 0 replies; 9+ messages in thread
From: Vladislav Shpilevoy @ 2020-07-22 22:27 UTC (permalink / raw)
To: Kirill Yukhin; +Cc: tarantool-patches
I was thinking that we continue doing 2.5 on the master branch,
was wrong.
I cherry-picked all the last qsync commits to 2.5, including this
patchset.
^ permalink raw reply [flat|nested] 9+ messages in thread
end of thread, other threads:[~2020-07-22 22:27 UTC | newest]
Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-07-20 20:43 [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 1/2] txn: single failure point for WAL and TX commit errors Vladislav Shpilevoy
2020-07-21 8:42 ` Cyrill Gorcunov
2020-07-20 20:43 ` [Tarantool-patches] [PATCH 2/2] txn: single failure point for WAL and TX async " Vladislav Shpilevoy
2020-07-21 8:43 ` Cyrill Gorcunov
2020-07-20 20:44 ` [Tarantool-patches] [PATCH 0/2] Qsync crash on WAL write fail Vladislav Shpilevoy
2020-07-21 20:25 ` Vladislav Shpilevoy
2020-07-22 10:41 ` Kirill Yukhin
2020-07-22 22:27 ` Vladislav Shpilevoy
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox