libclang_rt.asan and libclang_rt.ubsan_standalone both contain
sanitizer_common symbols, causing duplicate symbol errors when linked
together. When combining -fsanitize=address,undefined, clang uses the
ASan runtime to cover UBSan instrumentation as well, so only the ASan
runtime libraries (asan_static, asan, asan_cxx) should be linked.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

The combined ASan+UBSan binary is larger than pure ASan — each
clickhouse-client process uses ~350MB RSS. Under parallel test mode
with multiple concurrent clients the 5GB cgroup limit is exceeded,
causing OOM kills. Raise it to 10GB for asan_ubsan jobs only.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

The is_coverage/JIT settings block was accidentally included from the
working tree. Only the collect_build_flags refactoring (parsing all
sanitizer values from -fsanitize=... instead of elif chain) belongs
to this branch.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

`clickhouse local` (used for post-test log collection) was aborting at
exit under ASan+LSan with 45 KB of leaks from OpenSSL provider
initialization (`ossl_provider_new`, `OSSL_PROVIDER_try_load_ex`, etc.).
OpenSSL 3.x intentionally does not free all global state at exit, so
these are false positives.

Added `__lsan_default_suppressions` alongside the existing
`__lsan_default_options` in `programs/main.cpp` to suppress the five
known OpenSSL leak patterns. This fixes the "Unknown error / Lost
connection to server" failures in:
  - Stateless tests (amd_asan_ubsan, distributed plan, parallel, 1/2)
  - Stateless tests (amd_asan_ubsan, distributed plan, parallel, 2/2)
  - Stateless tests (arm_asan_ubsan, azure, sequential)

CI report: https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=99657&sha=24b9445c3f9c8dfc420b55f1a2fb369ae17220be&name_0=PR

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

Two issues remained after the previous fix:

1. The ASan stack-size warning block (4 lines) was only partially
   filtered in `clickhouse_proc.py`. The last line:
     "For details see google/sanitizers#189"
   was not excluded, leaving `sanitizer_hits` non-empty and triggering
   a spurious "Unknown error / Lost connection to server" result for
   jobs that had no real sanitizer errors. Add the missing exclusion.

2. OpenSSL's EVP method cache (`evp_md_new`, `construct_evp_method`,
   `CRYPTO_THREAD_lock_new`) is never freed at process exit. These
   objects are allocated when the S3 client initializes HMAC via
   `AWS SDK -> HMAC_Init_ex -> EVP_MD_fetch`. LSan reports them as
   leaks, causing the server to abort with SIGABRT (`halt_on_error=1`),
   which `clickhouse-test` records as "Unknown error / Lost connection
   to server". Add suppressions to `__lsan_default_suppressions`.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…fect

The previous suppressions were ineffective:

1. `src:*/contrib/arrow/cpp/src/arrow/array/array_binary.h*` - the path
   glob did not match the absolute path used during compilation; broadened
   to `src:*/contrib/arrow/*` to cover all Arrow contrib sources.

2. `fun:*DB::BufferBase::BufferBase*` - per the file's own caveats, `fun:`
   does not work for inlined functions, and `BufferBase::BufferBase` is a
   header-only constructor that is always inlined at call sites. Replaced
   with `src:*/src/IO/BufferBase.h`.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

UBSan's compile-time `src:` suppression for inlined functions must target
the TRANSLATION UNIT (.cpp file) where the inlined code is compiled, not
the header file where it is defined. All the inlined constructors in the
call chain compile down to instructions attributed to the outer .cpp file.

- Arrow: `arrow::BaseBinaryArray::value_offset` (misaligned load in
  `array_binary.h`) is inlined into `ArrowColumnToCHColumn.cpp`.
  Suppress that translation unit.

- RocksDB: `BufferBase::BufferBase` and friends (pointer overflow) are
  inlined through `ReadBufferFromString` → `KVStorageUtils.h` into
  `StorageEmbeddedRocksDB.cpp`. Suppress that translation unit.
  Issue: #99862

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

The ubsan_ignorelist.txt approach was ineffective because sccache keys
its object-file cache on compiler arguments (including the ignorelist
PATH) but not on the ignorelist FILE CONTENT. Changing the ignorelist
without changing source files serves stale cached objects, leaving the
UBSan instrumentation in place.

Fix by adding NO_SANITIZE_UNDEFINED directly to the affected functions,
which changes the preprocessed source content and forces sccache to
recompile:

1. ArrowColumnToCHColumn.cpp: `readColumnWithStringData<ArrowArray>` —
   Arrow IPC offset arrays may be unaligned (misaligned-load). This is
   intentional in the Arrow IPC format and safe on x86.

2. StorageEmbeddedRocksDB.cpp: `EmbeddedRocksDBSource::generateFullScan`
   — inlined BufferBase pointer arithmetic overflows when a RocksDB slice
   returns a corrupted size. Pre-existing bug tracked in:
   #99862

Also revert the now-unnecessary src: entries added to ubsan_ignorelist.txt
in the previous attempts.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…geEmbeddedRocksDB

Source annotations (NO_SANITIZE_UNDEFINED) were ineffective because the
UBSan checks for inlined callees are attributed to the callee's source
location, not the annotated caller. ubsan_ignorelist.txt was also
ineffective due to sccache not including ignorelist file content in its
cache key.

Fix: use set_source_files_properties in CMakeLists.txt to pass
-fno-sanitize=<check> flags for specific translation units. Changing
compiler flags is part of sccache's cache key, guaranteeing a recompile.

- ArrowColumnToCHColumn.cpp: -fno-sanitize=alignment
  Arrow IPC offset arrays are intentionally unaligned (packed int32).
  The misaligned read is safe on x86.

- StorageEmbeddedRocksDB.cpp: -fno-sanitize=pointer-overflow
  Pre-existing bug: corrupted RocksDB slice (size = 0xFFFFFFFFFFFFFFFE)
  causes pointer arithmetic overflow in BufferBase when iterating in
  read-only mode. Tracked in:
  #99862

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

Using relative paths in `set_source_files_properties` was not reliably
targeting the correct source files. Switch to `${CMAKE_CURRENT_SOURCE_DIR}`
absolute paths to ensure the suppression flags are applied.

Also remove `NO_SANITIZE_UNDEFINED` source annotations from
`ArrowColumnToCHColumn.cpp` and `StorageEmbeddedRocksDB.cpp` now that
the CMake approach is fixed.

Remove `amd_release` and `arm_asan_ubsan` targeted integration test
job configs from `job_configs.py` as these artifact variants don't
exist yet.

CI: #99657

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…ol function

Convert coverage addresses to file offsets by subtracting the binary's load base
so that dumps are portable across processes with different ASLR bases. This matches
how `addressToSymbol` works: the fallback path in `SymbolIndex::findSymbol` treats
values as raw file offsets when they're not in any mapped object.

Add `SymbolIndex::diagnose` to explain why symbol resolution succeeds or fails:
returns `"found:<name>"`, `"no_object"`, or `"no_symbol[object=<path>:offset=0x<hex>]"`.

Add `debugAddressToSymbol` SQL function backed by `diagnose`, useful for
investigating why `coverageCurrent()` produces empty symbols.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

When targeted integration tests run with `--count 10 --random-order`,
the same test functions execute 10 times in a single cluster session.
`test_read_only` and `test_dirctory_missing_after_stop` were not
idempotent: a failed run would leave tables or RocksDB lock files behind,
causing all subsequent runs to fail:

- After a clean run of `test_read_only`, the directory
  `/var/lib/clickhouse/user_files/test_rocksdb_read_only` persists on
  disk (dropping a table does not remove its RocksDB directory). In the
  next run, `CREATE TABLE ... read_only=1` on that path succeeds instead
  of raising, because a valid database now exists in the directory.
- A mid-run failure in either test can leave a table alive, keeping the
  RocksDB LOCK file held by the server process. Subsequent runs then fail
  with "lock hold by current process".

Fix: add cleanup at the start of each test that drops any leftover tables
and removes the RocksDB directories before the test body runs.

CI: #99657

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…ssToSymbol function"

This reverts commit 75df705.

…stem commands"

This reverts commit 38c7f78.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

When targeted integration tests run with --count 10, pytest-xdist spawns
multiple workers that all reuse the same ClickHouseCluster (same Docker
container, same server process).  RocksDB keeps a per-process in-memory
lock map: once a path is open, any concurrent Open() from another thread
in the same server fails with "lock hold by current process".

Fix: derive a per-worker suffix from PYTEST_XDIST_WORKER so each worker
uses distinct RocksDB paths and table names.  Also add cleanup at the
start of each test function to handle leftover state from a failed
previous run on the same worker (sequential reruns within one worker).

CI: #99657

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

When a Buffer is default-constructed (_ptr = nullptr, _used = 0),
calling setCapacity with newCapacity > 0 and preserveContent = true
computes newSz = 0 and then calls memcpy(dst, nullptr, 0).

Passing a null pointer to memcpy is undefined behaviour even when size
is 0 (the nonnull attribute on argument 2 applies regardless of size).
UBSan correctly flags this and with abort_on_error=1 the ClickHouse
server process aborts during Poco::CompressedLogFile construction
before binding port 9000, causing integration tests to time out.

Fix: skip the memcpy when newSz == 0.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

This reverts commit 707d72b.