This patchset replaces the fcntl()-based locking used in the original DPDK memory hotplug patchset, to an flock()- and lockfile-based implementation, due to numerous (well, one, really) problems with how fcntl() locks work.
Long story short, fcntl() locks will be dropped if any fd referring to locked file, is closed - even if it's not the last fd, even if it wasn't even the fd that was used to lock the file in the first place, even if it wasn't you who closed that fd, but some other library. This patchset corrects this saddening design defect in the original implementation. One of the ways to work around this was using OFD locks, but they are only supported on kernels 3.15+, so we cannot rely on them if we want to support old kernels. Hence, we use per-segment lockfiles. The number of file descriptors we open does not end up more than in non-single file segments case - we still open the same amount of files (a file per page), plus a file per memseg list. Additionally, since flock() is not atomic, we also lock the hugepage dir to prevent multiple processes from concurrently performing operations on hugetlbfs mounts. If you know of a more enlightened way of fixing this limitation, you are certainly welcome to comment :) v3: - Made lockfile naming more consistent with hugetlbfs file naming v2: - Fixes as per review comments - Make lockfiles hidden by default Anatoly Burakov (2): mem: add memalloc init stage mem: revert to using flock() and add per-segment lockfiles lib/librte_eal/bsdapp/eal/eal_memalloc.c | 6 + lib/librte_eal/common/eal_common_memory.c | 3 + lib/librte_eal/common/eal_filesystem.h | 17 + lib/librte_eal/common/eal_memalloc.h | 3 + lib/librte_eal/linuxapp/eal/eal_hugepage_info.c | 28 +- lib/librte_eal/linuxapp/eal/eal_memalloc.c | 605 +++++++++++++++--------- lib/librte_eal/linuxapp/eal/eal_memory.c | 22 +- 7 files changed, 420 insertions(+), 264 deletions(-) -- 2.7.4
