From: Mathieu Desnoyers Date: Fri, 10 Jul 2020 14:51:26 +0000 (-0400) Subject: Fix: coherent state not changed atomically with metadata written X-Git-Tag: v2.13.0-rc1~189 X-Git-Url: http://git.liburcu.org/?a=commitdiff_plain;h=3e75e2a7f7693107faa58606538f80a4faf73fe6;hp=3e75e2a7f7693107faa58606538f80a4faf73fe6;p=lttng-modules.git Fix: coherent state not changed atomically with metadata written commit 122c63cb4310 ("Fix: Implement RING_BUFFER_GET_NEXT_SUBBUF_METADATA_CHECK") introduces a new ioctl which returns a flag indicating whether the metadata is in consistent state at the end of the sub-buffer. That commit is meant to address metadata consistency issues observable in live sessions. However, the "consistent" state is false as soon as a producer is active (between an outermost metadata_begin/end pair). Unfortunately, if the last "RING_BUFFER_GET_NEXT_SUBBUF_METADATA_CHECK" operation is done between the last metadata printf and "end" of the transaction, the last consistency state will be false, and the consumer daemon will never send metadata to the relay daemon. This in turn causes a live viewer to wait for metadata endlessly. This issue can be reproduced by running lttng-tools: tests/regression/tools/live/test_kernel as root in a loop. We observe two things: 1) the poll operation blocks when there is no more metadata to send, which means there is no mean to unblock when the consistency state changes back to "true" without producing additional metadata, 2) Even if (1) was fixed, the expectation from an ABI perspective is that the "coherent" state is only populated when RING_BUFFER_GET_NEXT_SUBBUF_METADATA_CHECK succeeds. Therefore, there is no way to let user-space know about conherency transition unless additional metadata is generated. Fixing this requires to hold the metadata cache lock across the entire production of a coherent metadata transaction. This simpler scheme is possible because the metadata is generated in a reallocated memory area and not directly into a ring buffer anymore. This was not the case in earlier lttng-modules versions, when the metadata was generated directly into a ring buffer, which explains why this simpler scheme was not implemented. Signed-off-by: Mathieu Desnoyers ---