The prior blocking implementation seemed to be prone to locking up on some (especially recent) kernel versions. Since we were unable to reproduce the issue in a development environment we can only speculate as to the cause, but there is a good chance that blocking sockets were the issue or contributed to the issue.
So move to a non-blocking implementation to hopefully clear up these issues. Testing in production environments that were prone to locking shows that the approach is promising and at the very least not a regression.
The main differences from the blocking version are the non-blocking connect() implementation and handling of WANT_READ/WANT_WRITE retries for all SSL*() functions.
Timeouts in the tests needed to be increased because socket connect() and TLS SSL_connect() were not included in the timeout before. The tests don't run any slower, though. In fact, all platforms but Ubuntu 12.04 worked fine with the shorter timeouts.
select() is a bit old-fashioned and cumbersome to use. Since the select() code needed to be modified to handle write ready this seems like a good time to upgrade to poll().
poll() has been around for a long time so there doesn't seem to be any need to provide a fallback to select().
Also change the error on timeout from FileReadError to ProtocolError. This works better for read vs. write and failure to poll() is indicative of a protocol error or unexpected EOF.
The prior behavior introduced in dcddf3a5 could possibly lead to postgresql.conf or postgresql.auto.conf being truncated in the backup since they are copied via tmp files and could change size during the backup.
In general it seems safer to limit this feature to WAL-logged files which will be reconstructed during recovery.
A session looks much the same whether it is initiated from the client or the server, so use the session objects to implement the TLS, HTTP, and S3 test servers.
For TLS, at least, there are some differences between client and server sessions so add a client/server type to SocketSession to determine how the session was initiated.
Aside from reducing code duplication, the main advantage is that the test server will now timeout rather than hanging indefinitely when less input that expected is received.
Previously an error was only thrown when errno was set but in practice this is usually not the case. This may have something to do with getting errno late but attempts to get it earlier have not been successful. It appears that errno usually gets cleared and spot research seems to indicate that other users have similar issues.
An error at this point indicates unexpected EOF so it seems better to just throw an error all the time and be consistent.
To test this properly our test server needs to call SSL_shutdown() except when the client expects this error.
This abstraction allows the session code to be shared between the TLS client and (upcoming) server code.
Session management is no longer implemented in TlsClient so the HttpClient was updated to free and create sessions as needed. No test changes were required for HttpClient so the functionality should be unchanged.
Mechanical changes to the TLS tests were required to use TlsSession where appropriate rather than TlsClient. There should be no change in functionality other than how sessions are managed, i.e. using tlsClientOpen()/tlsSessionFree() rather than just tlsClientOpen().
The errorInternalThrowSys*() functions were marked as returning during coverage testing even when they had no possibility to return, i.e. the error parameter was set to constant true. This meant the compiler would treat the functions as returning even when they would not.
Instead create completely separate functions for coverage to use for THROW_ON_SYS_ERROR*() that can return and leave the regular functions marked __noreturn__.
This abstraction allows the session code to be shared between the socket client and (upcoming) server code. There should no difference in how the code works -- only the organization has changed. Note that no changes to the tests were required.
This same abstraction will be required for TlsClient but that will be done in a separate commit because it requires test changes.
The storage driver requires two list functions to be implemented, list and infoList. But the former is a subset of the latter so implementing both in every driver is wasteful. The reason both exist is that in Posix it is cheaper to get a list of names than it is to stat files to get size, time, etc. In S3 these operations are equivalent.
Introduce storageInfoLevelType to determine the amount of information required by the caller. That way Posix can work efficiently and all drivers can return only the data required which saves some bandwidth. The storageList() and storageInfoList() functions remain in the storage interface since they are useful -- the only change is simplifying the drivers with no external impact.
Note that since list() accepted an expression infoList() must now do so. Checking the expression is optional for the driver but can be used to limit results or save IO costs.
Similarly, exists() and pathExists() are just specialized forms of info() so adapt them to call info() instead.
Prior to 2.25 the individual TCP keep-alive options were not being configured due to a missing header. In 2.25 they were being configured incorrectly due to a disconnect between the timeout specified in ms and what was expected by the TCP options, i.e. seconds.
Instead make the TCP keep-alive options directly configurable, with correct units and better testing. Keep-alive is enabled by default (though it can be defaulted to the system setting instead) and the rest of the options are not set by default. This is in line with what PostgreSQL does, though PostgreSQL does not allow keep-alive to be defaulted.
Also move configuration of TCP options before connect() as PostgreSQL does.
Features:
* Add lz4 compression support. Note that setting compress-type=lz4 will make new backups and archive incompatible (unrestorable) with prior versions of pgBackRest. (Reviewed by Cynthia Shang.)
* Add --dry-run option to the expire command. Use dry-run to see which backups/archive would be removed by the expire command without actually removing anything. (Contributed by Cynthia Shang, Luca Ferrari.)
Improvements:
* Improve performance of remote manifest build. (Suggested by Jens Wilke.)
* Fix detection of keepalive options on Linux. (Contributed by Marc Cousin.)
* Add configure host detection to set standards flags correctly. (Contributed by Marc Cousin.)
* Remove compress/compress-level options from commands where unused. These commands (e.g. restore, archive-get) never used the compress options but allowed them to be passed on the command line. Now they will error when these options are passed on the command line. If these errors occur then remove the unused options. (Reviewed by Cynthia Shang.)
* Limit backup file copy size to size reported at backup start. If a file grows during the backup it will be reconstructed by WAL replay during recovery so there is no need to copy the additional data. (Reviewed by Cynthia Shang.)
Other storage*InfoList() functions do this but it was missed here.
memResize()/memFree() operations become more expensive as the mem context grows larger so freeing it periodically saves processing time.
If a file grows during the backup it will be reconstructed by WAL replay during recovery so there is no need to copy the additional data.
This also reduces the likelihood of seeing torn pages during the copy. Torn pages can still occur in the middle of the file, though, so they must be handled.
This code stanza was not being included on Linux platforms because of a missing header file.
Also update the order of operations and make the timeout calculations more sensible.
LZ4 compresses data faster than gzip but at a lower ratio. This can be a good tradeoff in certain scenarios.
Note that setting compress-type=lz4 will make new backups and archive incompatible (unrestorable) with prior versions of pgBackRest.
Add compress-type option and deprecate compress option. Since the compress option is boolean it won't work with multiple compression types. Add logic to cfgLoadUpdateOption() to update compress-type if it is not set directly. The compress option should no longer be referenced outside the cfgLoadUpdateOption() function.
Add common/compress/helper module to contain interface functions that work with multiple compression types. Code outside this module should no longer call specific compression drivers, though it may be OK to reference a specific compression type using the new interface (e.g., saving backup history files in gz format).
Unit tests only test compression using the gz format because other formats may not be available in all builds. It is the job of integration tests to exercise all compression types.
Additional compression types will be added in future commits.
These commands (e.g. restore, archive-get) never used the compress options but allowed them to be passed on the command line. Now they will error when these options are passed on the command line. If these errors occur then remove the unused options.
The prior code used TRY...CATCH blocks to cleanup mem contexts when an error occurred. This included freeing new mem contexts that were still being initialized when the error occurred and ensuring that the prior memory context was restored.
This worked fine in production but it involved a lot of setjmp()/longjmp() calls that resulted in longer compilation times and sluggish performance under valgrind, profiling, and coverage testing.
Instead maintain a stack of new contexts and context switches that can be used to do cleanup after an error. Normally, the stack is not used for this purpose and pushing/popping is a cheap operation. In the prior implementation most of the TRY...CATCH logic needed to be run even on success.
One bonus is that the binary is about 8% smaller after this change. Another benefit is that new contexts *must* be explicitly freed/discarded or an error will occur. See info/manifest.c for an example of where this is useful outside the standard macros.
Bug Fixes:
* Prevent defunct processes in asynchronous archive commands. (Reviewed by Stephen Frost. Reported by Adam Brusselback, ejberdecia.)
* Error when archive-get/archive-push/restore are not run on a PostgreSQL host. (Reviewed by Stephen Frost. Reported by Jesper St John.)
* Read HTTP content to eof when size/encoding not specified. (Reviewed by Cynthia Shang. Reported by Christian ROUX.)
* Fix resume when the resumable backup was created by Perl. In this case the resumable backup should be ignored, but the C code was not able to load the partial manifest written by Perl since the format differs slightly. Add validations to catch this case and continue gracefully. (Reported by Kacey Holston.)
Features:
* Auto-select backup set on restore when time target is specified. Auto-selection is performed only when --set is not specified. If a backup set for the given target time cannot not be found, the latest (default) backup set will be used. (Contributed by Cynthia Shang.)
Improvements:
* Skip pg_internal.init temp file during backup. (Reviewed by Cynthia Shang. Suggested by Michael Paquier.)
* Add more validations to the manifest on backup. (Reviewed by Cynthia Shang.)
Documentation Improvements:
* Prevent lock-bot from adding comments to locked issues. (Suggested by Christoph Berg.)
If PostgreSQL crashes it can leave behind a pg_internal.init temp file with the pid as the extension, as discussed in https://www.postgresql.org/message-id/flat/20200131045352.GB2631%40paquier.xyz#7700b9481ef5b0dd5f09cc410b4750f6. On restart this file is not cleaned up so it can persist for the lifetime of the cluster or until another process with the same id happens to write pg_internal.init.
This is arguably a bug in PostgreSQL, but in any case it makes sense not to backup this file.
This error was lost during the migration to C. The error that occurred instead (generally an SSH auth error) was hard to debug.
Restore the original behavior by throwing an error immediately if pg1-host is configured for any of these commands. reset-pg1-host can be used to suppress the error when required.
The main improvement is a double-fork to prevent zombie processes if the parent process exits after the (child) async process. This is a real possibility since the parent process sticks around to monitor the results of the async process.
In the first fork, ignore SIGCHLD in the very unlikely case that the async process exits before the first fork. This is probably only possible if the async process exits immediately, perhaps due to a chdir() failure. Set SIGCHLD back to default in the async process so waitpid() will work as expected.
Also update the comment on chdir() to more accurately reflect what is happening.
Finally, add a test in certain debug builds to ensure the first fork exits very quickly. This only works when valgrind is not in use because valgrind makes forking so slow that it is hard to tell if the async process performed work or not (in the case that the second fork goes missing and the async process is a direct child).
In this case the resumable backup should be ignored, but the C code was not able to load the partial manifest written by Perl since the format differs slightly. Add validations to catch this case and continue gracefully.
Generally, the content-size or content-encoding headers will be used to specify how much content should be expected.
There is a special case where the server sends 'Connection:close' without the content headers and the content may be read up until eof.
This appears to be an atypical usage but it is required by the specification.
Auto-selection is performed only when --set is not specified. If a backup set for the given target time cannot not be found, the latest (default) backup set will be used.
Currently a limited number of date formats are recognized and timezone names are not allowed, only timezone offsets.
Validate that checksums exist for zero size files. This means that the checksums for zero size files are explicitly set by backup even though they'll always be the same. Also validate that zero length files have the correct checksum.
Validate that repo size is > 0 if size is > 0. No matter what compression type is used a non-zero amount of data cannot be stored in zero bytes.
Bug Fixes:
* Fix missing files corrupting the manifest. If a file was removed by PostgreSQL during the backup (or was missing from the standby) then the next file might not be copied and updated in the manifest. If this happened then the backup would error when restored. (Reviewed by Cynthia Shang. Reported by Vitaliy Kukharik.)
Improvements:
* Use pkg-config instead of xml2-config for libxml2 build options. (Contributed by David Steele, Adrian Vondendriesch.)
* Validate checksums are set in the manifest on backup/restore. (Reviewed by Cynthia Shang.)
This is a modest start but it addresses the specific issue that was caused by the bug fixed in 45ec694a. This validation will produce an immediate error rather than erroring out partway through the restore.
More validations are planned but this is the most important one and seems safest for this release.
If a file was removed by PostgreSQL during the backup (or was missing from the standby) then the next file might not be copied and updated in the manifest. If this happened then the backup would error when restored.
The issue was that removing files from the manifest invalidated the pointers stored in the processing queues. When a file was removed, all the pointers shifted to the next file in the list, causing a file to be unprocessed. Since the unprocessed file was still in the manifest it would be saved with no checksum, causing a failure on restore.
When process-max was > 1 then the bug would often not express since the file had already been pulled from the queue and updates to the manifest are done by name rather than by pointer.
pkg-config is a generic way to get build options rather than relying on a package-specific utility.
XML2_CONFIG can be used to override this utility for systems that do not ship pkg-config.
Bug Fixes:
* Fix error in timeline conversion. The timeline is required to verify WAL segments in the archive after a backup. The conversion was performed base 10 instead of 16, which led to errors when the timeline was ≥ 0xA. (Reported by Lukas Ertl, Eric Veldhuyzen.)
The timeline is required to verify WAL segments in the archive after a backup. The conversion was performed base 10 instead of 16, which led to errors when the timeline was ≥ 0xA.
Bug Fixes:
* Fix options being ignored by asynchronous commands. The asynchronous archive-get/archive-push processes were not loading options configured in command configuration sections, e.g. [global:archive-get]. (Reviewed by Cynthia Shang. Reported by Urs Kramer.)
* Fix handling of \ in filenames. \ was not being properly escaped when calculating the manifest checksum which prevented the manifest from loading. Since instances of \ in cluster filenames should be rare to nonexistent this does not seem likely to be a serious problem in the field.
Features:
* pgBackRest is now pure C.
* Add pg-user option. Specifies the database user name when connecting to PostgreSQL. If not specified pgBackRest will connect with the local OS user or PGUSER, which was the previous behavior. (Contributed by Mike Palmiotto.)
* Allow path-style URIs in S3 driver.
Improvements:
* The backup command is implemented entirely in C. (Reviewed by Cynthia Shang.)
The local, remote, archive-get-async, and archive-push-async commands were used to run functionality that was not directly available to the user. Unfortunately that meant they would not pick up options from the command that the user expected, e.g. backup, archive-get, etc.
Remove the internal commands and add roles which allow pgBackRest to determine what functionality is required without implementing special commands. This way the options are loaded from the expected command section.
Since remote is no longer a specific command with its own options, more manipulation is required when calling remote. This might be something we can improve in the config system but it may be worth leaving as is because it is a one-off, for now at least.
