fix: release the stream when a bitmap preview cannot be decoded - #41835
Merged
Merged
Conversation
Bitmap::getThumbnail() opened the file and closed it only on the success path. The catch around getResizedPreview() returned without closing, so every failed decode leaked one file descriptor for the lifetime of the process. A preview pre-generation run or a cron preview job over a directory of files ImageMagick has no coder for exhausts the descriptors one file at a time. The open was also unchecked. A storage that cannot open the file returns false rather than throwing, and stream_get_contents(false) raises a TypeError - an \Error, so it escapes the \Exception handler directly underneath and surfaces as a 500 instead of the missing preview every other failure here degrades to. Closing moves into a finally block, and a false return from fopen() is handled explicitly. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com>
This comment was marked as resolved.
This comment was marked as resolved.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com>
phil-davis
approved these changes
Sep 16, 2026
This was referenced Sep 16, 2026
phil-davis
pushed a commit
that referenced
this pull request
Sep 16, 2026
…6] (#41837) * fix: release the stream when a bitmap preview cannot be decoded [10.16] Bitmap::getThumbnail() opened the file and closed it only on the success path. The catch around getResizedPreview() returned without closing, so every failed decode leaked one file descriptor for the lifetime of the process. A preview pre-generation run or a cron preview job over a directory of files ImageMagick has no coder for exhausts the descriptors one file at a time. The open was also unchecked. A storage that cannot open the file returns false rather than throwing, and stream_get_contents(false) cannot report that: on the PHP 7.4 this branch runs on it warns and hands on false, so the real cause is only ever logged as "ImageMagick says: Zero size image string passed", behind an unrelated PHP warning. (On PHP 8, which master runs, the same call raises a TypeError - an \Error, so it escapes the \Exception handler directly underneath and surfaces as a 500. That difference is why the wording here and in the changelog deviates from #41835.) Closing moves into a finally block, and a false return from fopen() is handled explicitly. 10.16 backport of #41835. (cherry picked from commit 0d0a306) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: make the backported stream tests detect their defects on PHP 7.4 Both cherry-picked test cases were written against PHP 8 and did not hold on the only version this branch supports. testReturnsFalseWhenTheFileCannotBeOpened asserted the return value, which cannot distinguish anything on 7.4: stream_get_contents(false) merely warns and hands on false, the sanitizer coerces it and Imagick rejects the empty string, so the unpatched code already returns false. The case passed with the fix reverted. What the guard actually removes on 7.4 is the noise - a warning from stream_get_contents(), and an "ImageMagick says:" line blaming ImageMagick for a file it never saw - so it now asserts that no warning is emitted, and is renamed accordingly. The handler honours error_reporting(), so diagnostics the code under test silenced with @ (the sanitizer's own loadXML warning, among any future ones) cannot fail the case; the un-suppressed warning alone detects the regression. testClosesTheStreamWhenDecodingThrows fed an XML payload, which ImageMagick sniffs as SVG. It throws here only because neither owncloudci/php:7.4 nor :8.3 registers an SVG delegate; on a build with librsvg or the internal MSVG renderer the lenient parser returns a blank canvas instead, the decode succeeds and the case fails. Replaced with content no coder claims at all, verified to be reported as format '' rather than format 'SVG' on both images. Confirmed both now fail without the fix - 2 failures, the second listing the warning verbatim - with the success-path case still passing as the control. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> --------- Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com>
oc-tmueller
added a commit
that referenced
this pull request
Sep 22, 2026
* test: stub the mime type in BitmapStreamTest so it survives the coder pin BitmapStreamTest mocks OCP\Files\File without stubbing getMimeType(), so the mock returns null. That is harmless today, but #41827 has Bitmap providers read the mime type to decide which Imagick coder to pin, and getResizedPreview() declares it as string - null there is a TypeError, which being an \Error escapes getThumbnail()'s \Exception handler rather than degrading to no preview. Merging #41827 would therefore turn these cases red on master. The success case also decoded a PNG through the Photoshop provider, which only works while ImageMagick is free to sniff the format. Once Photoshop pins the PSD coder, a PNG stops decoding and the case fails for a reason that has nothing to do with the stream. It now uses the PDF provider against testimage.pdf, so the provider, the file's mime type and the content all agree and the success path stays a success either way - guarded on the PDF coder, since pinning makes that a hard requirement. Verified against both trees: on master 3 tests / 5 assertions, and on master merged with #41827 the full tests/lib/Preview/ suite is 79 tests / 215 assertions / 0 failures, where before this change it reported 2 errors. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: decode a self-written TIFF rather than gating on the PDF coder Imagick::queryFormats('PDF') reports that the coder was compiled in. It says nothing about whether a PDF can actually be decoded: it consults neither the coder rights in policy.xml nor the presence of the Ghostscript delegate. On an image that revokes the PDF coder - the ImageMagick hardening OC10-164 is itself driving - or one without the gs binary, the guard passes, readImageBlob() throws, and the case fails red over an environment difference rather than over the stream handling it exists to check. That is the same mistake as gating a test on a coder the provider never uses, which this series has been removing elsewhere. The success case now writes its own TIFF through Imagick and decodes it through the TIFF provider. TIFF needs no external delegate, and a build cannot disagree with itself about a blob it just produced, so the remaining skip fires only where TIFF is unavailable altogether - in which case no assertion here could run anyway. It also drops a fixture dependency. The comments claiming that the mime type is read and that XML is rejected before any coder is consulted described the coder-pin change on #41827, which is not in this tree. They now say what happens here and what they anticipate. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: probe the TIFF read path, not just the write path, before asserting The guard added in the previous commit wrote a TIFF and treated that as proof the build could handle TIFF. ImageMagick grants coder rights per direction, so a policy of rights="write" for TIFF lets the blob be produced, declines to skip, and then fails red on the decode - reintroducing exactly the failure the guard exists to remove. It now reads the blob back inside the guard, so what is probed is what the assertion needs. Verified by revoking TIFF read in a throwaway container: the case skips with a clear message instead of failing. The guard also caught only \ImagickException, while ImagickPixelException extends \Exception directly and is a sibling rather than a subclass, so a pixel-wand failure would have escaped as an error rather than the intended skip. It now catches \Exception, and the Imagick handles are released in finally blocks rather than only on the success path - which matters in a test about releasing handles. Finally, the claim that no coder is consulted for the XML payload was wrong: ImageMagick's SVG coder claims any blob opening with "<?xml" and then fails on a document with no <svg> root. The comment now says that, and warns that another XML payload is not automatically substitutable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: skip only when TIFF is absent, and assert the stream before the decode The round-trip probe added in the previous commit closed one hole by opening another: it turned any TIFF failure into a skip, and a skip here costs the success-path fclose() assertion - which is the OC10-164 stream-leak guard itself. A guard quietly withholding these assertions is exactly how they came to never run in CI, so a misconfiguration should be loud, not green. The guard is now the single condition that is genuinely an absent feature rather than a broken setup: no TIFF coder registered at all. Revoked coder rights, an unparsable policy.xml or a wand that cannot be constructed all fail. TIFF can be held to that standard because no stock policy revokes it, unlike PDF, which Debian and Ubuntu deny out of the box - the reason this uses a TIFF in the first place. The stream assertion also moves ahead of the decode assertion, so an environment that cannot decode the blob still exercises the handle release under test and still reports the decode as the failure. Verified by revoking TIFF read in a throwaway container: all five assertions run, and the failure names the decode. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: separate an absent TIFF delegate from a denied one by ImageMagick's message The previous commit gated on Imagick::queryFormats('TIFF'), on the assumption that registration implies support. It does not: coders/tiff.c registers TIFF, TIF and TIFF64 unconditionally and only assigns the decoder and encoder pointers when built against libtiff, while GetMagickList() behind queryFormats() matches on the coder name alone. A build without libtiff therefore reports TIFF as registered, declines to skip, and - with the catch removed by that same commit - errors instead. That is the fourth variant of one mistake in this file: checking something adjacent to what the assertion needs. There is no registration check that can tell an absent feature from a broken setup, so this stops using a proxy and reads what ImageMagick reports. A missing delegate yields "no encode delegate for this image format" (or the decode equivalent) and skips; a policy denial yields "not allowed by the security policy" and is re-thrown, along with anything else. Both directions are probed, since coder rights are granted per direction. The success-path assertion message is also outcome-neutral now. It runs before the decode assertion, so it fires when the decode failed too, and must not claim the leak was on the success path when the decode is the actual defect. Verified in throwaway containers: a normal build passes; a policy revoking TIFF is loud rather than skipped; and MagickCore's message catalogue carries both delegate strings this matches on. The missing-delegate branch is matched against that catalogue rather than executed, since this build has libtiff. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: keep both TIFF guards, since neither covers the other's case The previous commit swapped the queryFormats() check for a message check when the two are complementary. Without libtiff, a modular ImageMagick - Debian and Ubuntu configure --with-modules - never builds coders/tiff.so, so TIFF is not registered and setImageFormat() fails with php-imagick's own "Unable to set the image format" before any delegate is consulted. That matches neither delegate substring, so it was rethrown and turned a build with no TIFF feature red. queryFormats() is what catches that case; the message check catches the non-modular build, which registers TIFF regardless and fails later at the delegate. Both are back. Also records two limits instead of implying they do not exist. A module- or coder-domain policy denial can surface as MissingDelegateError, textually identical to an absent delegate, so such a build skips - the classifier only rejects messages that name a policy outright rather than guessing. And an allowlist-style policy.xml denying all but a few coders fails here, which is the accepted cost of being loud about misconfiguration; the note explaining why TIFF rather than PDF is restored alongside it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: fail the undecodable case on bytes no coder claims The payload was '<?xml version="1.0"?><notanimage>x</notanimage>', which is not environment-independent. ImageMagick's IsSVG() claims any blob opening with "<?xml", so readImageBlob() reported "no decode delegate for this image format `SVG'" - it threw only because these images register no SVG renderer. Where librsvg or the internal MSVG renderer is present, the lenient parser returns a blank canvas rather than throwing, and the case would fail for reasons unrelated to the stream. That is the same environment coupling this file has been shedding elsewhere; the failure path had it too. It now uses bytes no coder claims. ImageMagick sniffs the format as "" and fails with "no decode delegate for this image format `'" on every build regardless of which delegates are compiled in. libmagic reads them as application/octet-stream rather than text, so they also survive #41827's mime gate and still reach the decode on that branch instead of being turned away earlier. Verified in a container: the old payload sniffs as SVG, the new one as "", and both the master tree and the tree merged with #41827 stay green. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * docs: correct the recorded reasons in BitmapStreamTest's comments Three claims in these docblocks were wrong, and the payload rationale was the one that mattered: it said an XML payload "would throw only where no SVG renderer is registered" and would otherwise return a blank canvas. Measured in three builds - stock, with libmagickcore-6.q16-6-extra installed, and with the policy opened up - it throws in all of them, as "no decode delegate `SVG'", then "not allowed by the security policy `MVG'", then MVG's own "must specify image size". The coder is MVG rather than SVG too. So the reason to prefer bytes no coder claims is not that the XML payload is unusable, it is that its failure reason varies by build and that libmagic reads it as text/xml, which #41827's mime gate rejects before the decode. The comment now says that, so nobody rules out a working option on a wrong premise. The read-back rationale claimed both directions get denied; what actually happens with TIFF rights revoked is that getImageBlob() still returns a blob and only the read raises - which is the argument for probing the read, now stated as measured. The mime-type stub was described as anticipating #41827 and reading as speculative, when omitting it is precisely what turned that PR red. It is stated as a requirement instead, so it does not invite deletion once the pin lands. Also trims the libtiff explanation. It asserted ImageMagick internals no assertion here pins and which differ across major versions, and it is where the errors above were concentrated; the two-check rationale and the PDF-vs-TIFF choice stay. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: pin why the undecodable case throws, and assert the handle first Both assertions in testClosesTheStreamWhenDecodingThrows are satisfied by any early return from getThumbnail(), and nothing tied the failure to the decode. That matters on #41827, which adds a pre-decode mime gate denying text/*: a build whose libmagic read these bytes as text would refuse them before any coder, leave this test green, and quietly stop covering the path the test is named for. The detected media type is now asserted, so that drift fails instead of hiding. The two tests also disagreed on assertion order. PHPUnit stops at the first failure, so asserting the result first meant an unexpectedly decodable payload would mask a co-occurring leak - the handle being the regression guard this file exists for. testClosesTheStreamOnSuccess already ordered it the other way and said why; the failure case now matches. The payload rationale claimed the sniffed format is "", which holds here but not under the pin, where nothing is sniffed and the pinned coder rejects the header instead. Both throw without depending on the build's delegates, which is the actual property being relied on, so the comment says that rather than one tree's mechanism. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: decode a PSD, dropping the TIFF availability guard entirely Every guard in this file existed because the success case used TIFF, and TIFF can be absent: coders/tiff.so links libtiff. PSD cannot be absent for that reason - coders/psd.so links no image library at all, ImageMagick implements the format natively - and the Photoshop provider was already here for the failure case. So the success case now writes and decodes a PSD, and the whole apparatus goes: no queryFormats() check, no write-then-read-back probe, no message classifier separating an absent delegate from a denied one, and no docblock asserting ImageMagick internals that nothing pins. The test is unconditional, which is what it should have been throughout - a skip would retire the success-path fclose() assertion, and a guard quietly withholding assertions is how the OC10-164 preview tests came to never run in CI to begin with. The file loses 44 lines. This also removes a contradiction with the branch it is written to be compatible with: CoderPinningTest::requireDecodableFixture() skips on a policy denial where the classifier here rethrew, so the same suite gave two answers for the same coder. Verified: unconditional pass on master and on the tree merged with the pin, and still red when the finally that releases the handle is removed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * test: assert the gate's own condition, and stub the third mock's mime type Two narrow corrections. The pin on the payload's detected type asserted one exact classification, application/octet-stream, while the gate it protects only refuses text/*, image/svg*, application/xml and image/x-mvg. A libmagic that matched these bytes to some other binary magic entry would still reach the decode exactly as intended and fail the assertion, which is the build-dependence this file has been shedding. It now mirrors isDangerousToDecode()'s own condition. The mime type is also stubbed on the cannot-be-opened mock, so that case does not depend on where in getThumbnail() the mime type is first read. That stub does not make the file runnable on a tree without #41835's fopen guard, and the comment no longer claims it does - measured on #41827's branch, which carries neither that guard nor the finally, the suite reports 1 error and 1 failure because every case here asserts what those two added. Failing there is correct, and it is why this lands on master rather than folded into #41834: CI builds the head-into-base merge commit, which always contains both. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * docs: scope the master-only claim, and name the deny-list this mirrors Two comment corrections, no code change. "Every case here asserts behaviour the guard and the finally introduced" is wrong for testClosesTheStreamOnSuccess: fclose() on the success path predates #41835, which only moved it into the finally, so that case passes on a tree without either. The measurement already said so - one error and one failure on #41834's branch, two cases and not three - and the claim should have been scoped to those two. The pre-decode check is also now attributed to its source, OC\Preview\Bitmap:: isDangerousToDecode(), which #41834 adds and which is private and so cannot be called from a test. Mirroring it is still preferable to pinning one exact libmagic classification, but the duplication has a cost worth stating: if that deny-list gains an entry, this copy must gain it too, or the payload starts being refused at the gate while the assertion stays green and the decode goes uncovered. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * docs: name the change that actually drifts, and drop merge-strategy prose Two more comment corrections. The maintenance note pointed the wrong maintainer at the mirror. A new text/ entry in isDangerousToDecode()'s deny-list is already matched by the text/ prefix here, so mirroring it would be busywork; the change that actually drifts is an entry of the application/xml or image/x-mvg shape, which the prefix does not catch. It now says that. isDangerousToDecode()'s own comment also enumerates what it already covers rather than anticipating additions, so that clause is gone. The claim that CI building the head-into-base merge commit is why this belongs on master rather than folded into #41834 was a non-sequitur - that same fact means folding it in would have been green too, since the failures only appear on the bare branch. The real reason is that the branch tree lacks #41835 and so cannot run the file locally, which the surrounding lines already say. Merge-strategy reasoning does not belong in a test docblock in any case; it goes in the pull request. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> * docs: state the drift rule against all three of the mirrored rules The previous wording named only the text/ prefix and treated a drifting deny-list entry as necessarily an exact match. The mirror has three rules, and an entry written as a prefix - application/postscript alongside the existing image/svg, say - drifts just as badly while a reader following that wording concludes no mirroring is needed. It also over-warned in the other direction: an added image/svg+xml is not matched by text/ but is already matched by the image/svg prefix, so it does not drift. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> --------- Signed-off-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> Co-authored-by: Thomas Müller <323649642+oc-tmueller@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Bitmap::getThumbnail()opens the file and closes it only on the success path — thecatcharoundgetResizedPreview()returns before thefclose(). Every failed decode therefore leaks one file descriptor for the lifetime of the process.Flagged as out of scope in #41827 ("that's a resource leak, not a security defect, and deserves its own focused PR"), so this is that PR. It targets
masterand is independent of the #41827/#41834 chain.Why it matters more than it looks
A failed decode is not an edge case. Any content ImageMagick has no coder for lands in that
catch, and #41834 makes throwing a designed outcome — if a build does not register a provider's coder, the pin throws rather than falling back to content sniffing. Sooccpreview pre-generation, or a cron preview job, over a directory of.heic/.psdfiles leaks a handle per file untilEMFILE.Second defect, same three lines
$file->fopen('r')was unchecked. A storage that cannot open the file returnsfalse, andstream_get_contents(false)raises aTypeError— an\Error, so it escapes thecatch (\Exception)directly underneath and surfaces as a 500 instead of the missing preview every other failure here degrades to. Confirmed, not theorised:What changed
fclose()moves into afinally, so it runs on both paths.falsereturn fromfopen()is handled explicitly and logged.Tests
New
tests/lib/Preview/BitmapStreamTest.php, 3 cases. It asserts the contract directly viais_resource($stream)on the caller's own handle rather than counting descriptors, so it is portable rather than Linux-only.Confirmed RED before the fix on
owncloudci/php:8.3— 1 failure (the stream must be closed on the failure path) and 1 error (theTypeErrorabove); the success-path case passed unfixed, as the control. GREEN after: 3 tests, 5 assertions.Verification
tests/lib/Preview/onowncloudci/php:8.3: 55 tests, 152 assertions, 0 failures.make test-php-style: 0 of 2435 files need fixing.php -lclean under PHP 7.4, so this backports to the 10.x line without syntax changes.Note on sequencing
#41827 modifies these same lines. Whichever lands second needs a trivial rebase — the conflict is confined to the
try/catch/fcloseblock.🤖 Generated with Claude Code