On Thu, 16 Sept 2021 at 17:08, Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Thu, Sep 16, 2021 at 04:55:16PM +0200, Dmitry Vyukov wrote: > > > > I noticed we also had 2 KCSAN reports that mention rdma_resolve_addr. > > > > > > On commit 1df0d896: > > > ================================================================== > > > BUG: KCSAN: data-race in addr_handler / cma_check_port > > > > > > write to 0xffff88809fa40a1c of 4 bytes by task 21 on cpu 1: > > > cma_comp_exch drivers/infiniband/core/cma.c:426 [inline] > > > addr_handler+0x9f/0x2b0 drivers/infiniband/core/cma.c:3141 > > > process_one_req+0x22f/0x300 drivers/infiniband/core/addr.c:645 > > > process_one_work+0x3e1/0x9a0 kernel/workqueue.c:2269 > > > worker_thread+0x665/0xbe0 kernel/workqueue.c:2415 > > > kthread+0x20d/0x230 kernel/kthread.c:291 > > > ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:293 > > > > > > read to 0xffff88809fa40a1c of 4 bytes by task 11997 on cpu 0: > > > cma_check_port+0xbd/0x700 drivers/infiniband/core/cma.c:3506 > > This has since been fixed, cma_check_port() no longer reads state > > > > and on commit 5863cc79: > > I can't find this commit? Current rdma_resolve_addr should not trigger > this KCSAN. > > > This does not immediately explain the use-after-free for me, but these > > races suggest that everything is not protected by a single mutex and > > that there may be some surprising interleavings. > > E.g. rdma_resolve_addr checks status, and then conditionally executes > > cma_bind_addr, but the status can change concurrently. > > It is true, they weren't, however I've fixed them all. These hits look > like they all from before it got fixed up.. Then sorry for false leads. The second commit was from https://github.com/google/ktsan.git kcsan branch. I am not sure if it's still present or was rebased. But either way it's even older than the first report on upstream (we used ktsan tree before switched to upstream).
