On Tue, Sep 5, 2023 at 11:05 AM Artemy Kovalyov <artem...@nvidia.com> wrote: > > > > + /* memory_hotplug_lock is taken in rte_eal_init(), so it's > > > + * safe to call thread-unsafe version. > > > + */ > > > > Nit: the lock is really taken in rte_eal_memory_init(). > > Probably "The lock is held during initialization, so..." > > would more robust against code changes and differences between platforms. > > It was previously located differently, but in the current version, it has > been shifted to rte_eal_init(). It might be worth noting this to ensure that > if there are further code changes in the future, the locking problem becomes > more apparent. We had discussed this in the bug report.
One option to explore is lock annotations. One note thought: those annotations do not get inherited in called code. So some special care is needed to maintain/annotate all code leading to the locations where the locks do matter. Quick example with rte_memseg_list_walk: https://github.com/david-marchand/dpdk/commit/mem_annotations And clang catches a deadlock: https://github.com/david-marchand/dpdk/actions/runs/6082842080/job/16501450978#step:19:816 -- David Marchand
