Thread
Commits
-
Add missing deflateEnd() for server-side gzip base backups
- ded9754804bc 19 (unreleased) landed
-
basebackup: add missing deflateEnd() in gzip compression sink
Jianghua Yang <yjhjstz@gmail.com> — 2026-03-20T17:02:44Z
Hi hackers, While reviewing the backup compression backends, I noticed that the gzip sink (basebackup_gzip.c) never calls deflateEnd(), unlike the lz4 and zstd sinks which properly free their compression contexts. Specifically: - lz4 calls LZ4F_freeCompressionContext() in both end_archive and a dedicated bbsink_lz4_cleanup() callback. - zstd calls ZSTD_freeCCtx() in both end_archive and a dedicated bbsink_zstd_cleanup() callback. - gzip calls deflateInit2() in begin_archive but never calls deflateEnd(). Its cleanup callback is bbsink_forward_cleanup, which just forwards to the next sink without releasing the zlib context. This means each archive (one per tablespace) leaks ~256KB of zlib internal state until the backup's memory context is destroyed. On the ERROR path, the zlib context is not released at all since there is no gzip-specific cleanup callback. The attached patch adds deflateEnd() at the end of bbsink_gzip_end_archive() for the normal path, and introduces a new bbsink_gzip_cleanup() for the error path. The new cleanup callback follows the exact same call chain as bbsink_lz4_cleanup and bbsink_zstd_cleanup: PG_FINALLY (basebackup.c:1063) -> bbsink_cleanup(sink) (basebackup.c:1065) -> sink->bbs_ops->cleanup(sink) (basebackup_sink.h:269) -> bbsink_gzip_cleanup() -- now calls deflateEnd() Previously the gzip slot in this chain pointed to bbsink_forward_cleanup, which just forwarded to the next sink without doing any gzip-specific resource release. Since z_stream is an embedded struct (not a pointer like LZ4F_cctx or ZSTD_CCtx), a bool flag "zstream_initialized" is used to track whether deflateEnd() needs to be called. Tested with pg_basebackup using gzip compression at levels 1, 5 (default), and 9, including server-side compression. All produced valid .tar.gz files that pass gzip -t integrity checks and restore correctly. Regards, Jianghua Yang -
Re: basebackup: add missing deflateEnd() in gzip compression sink
Michael Paquier <michael@paquier.xyz> — 2026-03-21T06:09:31Z
On Fri, Mar 20, 2026 at 10:02:44AM -0700, Jianghua Yang wrote: > While reviewing the backup compression backends, I noticed that the gzip > sink (basebackup_gzip.c) never calls deflateEnd(), unlike the lz4 and > zstd sinks which properly free their compression contexts. > > This means each archive (one per tablespace) leaks ~256KB of zlib > internal state until the backup's memory context is destroyed. On the > ERROR path, the zlib context is not released at all since there is no > gzip-specific cleanup callback. Yes, you are right on this one. This code could be better in cleaning up the resources it is working on (and note that all the other gzip-related code paths call the end() part if doing an init() or an init2()). Leaking that once per archive may not look like a big deal, but this is backend-side code we are talking about, and on a persistent replication connection it is not really cool to lose the context of this data with garbage coming from zlib piling up, so based on the argument of the permanent connection my take is that this cleanup warrants a backpatch. -- Michael
-
Re: basebackup: add missing deflateEnd() in gzip compression sink
Jianghua Yang <yjhjstz@gmail.com> — 2026-03-21T13:09:46Z
Thanks for reviewing, Michael. > based on the argument of the permanent connection my take is that > this cleanup warrants a backpatch. Agreed. The current code leaks zlib internal state on every archive, and on a long-lived replication connection those allocations just pile up with no cleanup until the connection ends. Michael Paquier <michael@paquier.xyz> 于2026年3月20日周五 23:09写道: > On Fri, Mar 20, 2026 at 10:02:44AM -0700, Jianghua Yang wrote: > > While reviewing the backup compression backends, I noticed that the gzip > > sink (basebackup_gzip.c) never calls deflateEnd(), unlike the lz4 and > > zstd sinks which properly free their compression contexts. > > > > This means each archive (one per tablespace) leaks ~256KB of zlib > > internal state until the backup's memory context is destroyed. On the > > ERROR path, the zlib context is not released at all since there is no > > gzip-specific cleanup callback. > > Yes, you are right on this one. This code could be better in cleaning > up the resources it is working on (and note that all the other > gzip-related code paths call the end() part if doing an init() or an > init2()). Leaking that once per archive may not look like a big deal, > but this is backend-side code we are talking about, and on a > persistent replication connection it is not really cool to lose the > context of this data with garbage coming from zlib piling up, so based > on the argument of the permanent connection my take is that this > cleanup warrants a backpatch. > -- > Michael >
-
Re: basebackup: add missing deflateEnd() in gzip compression sink
Jianghua Yang <yjhjstz@gmail.com> — 2026-03-21T21:22:25Z
Hi Michael, Thanks for the review and backpatch suggestion. Here is v1 as a two-patch series. The first patch is unchanged from the original submission. I've added a second patch that fixes another resource leak I noticed nearby in pg_basebackup: v1-0001: basebackup: add missing deflateEnd() calls in gzip compression sink v1-0002: pg_basebackup: add missing close() for incremental manifest file In the incremental backup code path, the file descriptor opened for reading the manifest is never closed after the upload completes. Since the backup can run for a long time after this point, the leaked descriptor remains open for the entire duration. Regards, Jianghua Yang Jianghua Yang <yjhjstz@gmail.com> 于2026年3月21日周六 06:09写道: > Thanks for reviewing, Michael. > > > based on the argument of the permanent connection my take is that > > this cleanup warrants a backpatch. > > Agreed. The current code leaks zlib internal state on every archive, > and on a long-lived replication connection those allocations just > pile up with no cleanup until the connection ends. > > Michael Paquier <michael@paquier.xyz> 于2026年3月20日周五 23:09写道: > >> On Fri, Mar 20, 2026 at 10:02:44AM -0700, Jianghua Yang wrote: >> > While reviewing the backup compression backends, I noticed that the gzip >> > sink (basebackup_gzip.c) never calls deflateEnd(), unlike the lz4 and >> > zstd sinks which properly free their compression contexts. >> > >> > This means each archive (one per tablespace) leaks ~256KB of zlib >> > internal state until the backup's memory context is destroyed. On the >> > ERROR path, the zlib context is not released at all since there is no >> > gzip-specific cleanup callback. >> >> Yes, you are right on this one. This code could be better in cleaning >> up the resources it is working on (and note that all the other >> gzip-related code paths call the end() part if doing an init() or an >> init2()). Leaking that once per archive may not look like a big deal, >> but this is backend-side code we are talking about, and on a >> persistent replication connection it is not really cool to lose the >> context of this data with garbage coming from zlib piling up, so based >> on the argument of the permanent connection my take is that this >> cleanup warrants a backpatch. >> -- >> Michael >> > -
Re: basebackup: add missing deflateEnd() in gzip compression sink
Michael Paquier <michael@paquier.xyz> — 2026-03-23T00:31:36Z
On Sat, Mar 21, 2026 at 02:22:25PM -0700, Jianghua Yang wrote: > v1-0001: basebackup: add missing deflateEnd() calls in gzip compression > sink After double-checking the whole code, I agree that this is a good practice to have in the tree. However, the issue is not worth bothering in back-branches as the server-side base backup gzip code relies on allocation and free callbacks, with zlib internals doing nothing with fds or more persistent states as far as I have read its code. For the current use, we'd bloat this data once per tablespace in a single base backup, safe even if the connection is persistent (missed that in my first message). What I am more worried about are future callers of this code, though, and we care about having a end() call for each matching init[2]() call in the tree in all the places that rely on gzip internals. So that's a good practice on consistency ground, at least. For these reasons, applied that on HEAD. > v1-0002: pg_basebackup: add missing close() for incremental manifest > file This one does not matter. This resource is for a backup manifest and we are talking about a single one for a single invocation of the binary. -- Michael
-
Re: basebackup: add missing deflateEnd() in gzip compression sink
Jianghua Yang <yjhjstz@gmail.com> — 2026-03-23T01:36:22Z
Hi Michael, Thank you for reviewing and committing the patch! noted on v1-0002 — understood that the incremental manifest file close isn't an issue there. Thanks again. Regards, Jianghua Yang Michael Paquier <michael@paquier.xyz> 于2026年3月22日周日 17:32写道: > On Sat, Mar 21, 2026 at 02:22:25PM -0700, Jianghua Yang wrote: > > v1-0001: basebackup: add missing deflateEnd() calls in gzip compression > > sink > > After double-checking the whole code, I agree that this is a good > practice to have in the tree. However, the issue is not worth > bothering in back-branches as the server-side base backup gzip code > relies on allocation and free callbacks, with zlib internals doing > nothing with fds or more persistent states as far as I have read its > code. For the current use, we'd bloat this data once per tablespace > in a single base backup, safe even if the connection is persistent > (missed that in my first message). > > What I am more worried about are future callers of this code, though, > and we care about having a end() call for each matching init[2]() call > in the tree in all the places that rely on gzip internals. So that's > a good practice on consistency ground, at least. For these reasons, > applied that on HEAD. > > > v1-0002: pg_basebackup: add missing close() for incremental manifest > > file > > This one does not matter. This resource is for a backup manifest and > we are talking about a single one for a single invocation of the > binary. > -- > Michael >