This is the third and hopefully for now last part of my work to fix drain. The main goal of this series is to make drain robust against graph changes that happen in any callbacks of in-flight requests while we drain a block node.
The individual patches describe the details, but the rough plan is to change all three drain types (single node, subtree and all) to work like this: 1. First call all the necessary callbacks to quiesce external sources for new requests. This includes the block driver callbacks, the child node callbacks and disabling external AioContext events. This is done recursively. Much of the trouble we had with drain resulted from the fact that the graph changed while we were traversing the graph recursively. None of the callbacks called in this phase may change the graph. 2. Then do a single AIO_WAIT_WHILE() to drain the requests of all affected nodes. The aio_poll() called by it is where graph changes can happen and we need to be careful. However, while evaluating the loop condition, the graph can't change, so we can safely call all necessary callbacks, if needed recursively, to determine whether there are still pending requests in any affected nodes. We just need to make sure that we don't rely on the set of nodes being the same between any two evaluation of the condition. There are a few more smaller, mostly self-contained changes needed before we're actually safe, but this is the main mechanism that will help you understand what we're working towards during the series. Kevin Wolf (18): test-bdrv-drain: bdrv_drain() works with cross-AioContext events block: Use bdrv_do_drain_begin/end in bdrv_drain_all() block: Remove 'recursive' parameter from bdrv_drain_invoke() block: Don't manually poll in bdrv_drain_all() tests/test-bdrv-drain: bdrv_drain_all() works in coroutines now block: Avoid unnecessary aio_poll() in AIO_WAIT_WHILE() block: Really pause block jobs on drain block: Remove bdrv_drain_recurse() block: Drain recursively with a single BDRV_POLL_WHILE() test-bdrv-drain: Test node deletion in subtree recursion block: Don't poll in parent drain callbacks test-bdrv-drain: Graph change through parent callback block: Defer .bdrv_drain_begin callback to polling phase test-bdrv-drain: Test that bdrv_drain_invoke() doesn't poll block: Allow AIO_WAIT_WHILE with NULL ctx block: Move bdrv_drain_all_begin() out of coroutine context block: Allow graph changes in bdrv_drain_all_begin/end sections test-bdrv-drain: Test graph changes in drain_all section Max Reitz (1): test-bdrv-drain: Add test for node deletion include/block/aio-wait.h | 25 +- include/block/block.h | 17 + include/block/block_int.h | 8 + include/block/blockjob_int.h | 8 + block.c | 31 +- block/io.c | 280 +++++++++------- block/mirror.c | 8 + blockjob.c | 21 ++ tests/test-bdrv-drain.c | 738 ++++++++++++++++++++++++++++++++++++++++--- 9 files changed, 974 insertions(+), 162 deletions(-) -- 2.13.6
