On 25-Mar-21 8:21 AM, xiangxia.m....@gmail.com wrote:
From: Tonghao Zhang <xiangxia.m....@gmail.com>The hugepage of different size, 2MB, 1GB may be mounted on the same directory (e.g /dev/hugepages). Then dpdk primary process will be blocked. To address this issue, add the LOCK_NB flags to flock(). $ cat /proc/mounts ... none /dev/hugepages hugetlbfs rw,seclabel,relatime,pagesize=1024M 0 0 none /dev/hugepages hugetlbfs rw,seclabel,relatime,pagesize=2M 0 0 Add more details for err logs. Signed-off-by: Tonghao Zhang <xiangxia.m....@gmail.com> --- lib/librte_eal/linux/eal_hugepage_info.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/lib/librte_eal/linux/eal_hugepage_info.c b/lib/librte_eal/linux/eal_hugepage_info.c index d97792cadeb6..1ff76e539053 100644 --- a/lib/librte_eal/linux/eal_hugepage_info.c +++ b/lib/librte_eal/linux/eal_hugepage_info.c @@ -451,9 +451,12 @@ hugepage_info_init(void) hpi->lock_descriptor = open(hpi->hugedir, O_RDONLY);/* if blocking lock failed */- if (flock(hpi->lock_descriptor, LOCK_EX) == -1) { + if (flock(hpi->lock_descriptor, LOCK_EX | LOCK_NB) == -1) { RTE_LOG(CRIT, EAL, - "Failed to lock hugepage directory!\n"); + "Failed to lock hugepage directory! " + "The hugepage dir (%s) was locked by " + "other processes or self twice.\n", + hpi->hugedir); break; } /* clear out the hugepages dir from unused pages */
Use cases such as "having two hugetlbfs page sizes on the same hugetlbfs mountpoint" are user error, but i agree that deadlocking is probably not the way we want to go about it.
An alternative way would be to check if we already have a mountpoint with the same path, and this would produce a better error message (as a user, "hugepage dir is locked by self twice" doesn't tell me anything useful), at a cost of slightly more complicated code.
I'm not sure which way i want to go here. Normally, hugetlbfs shouldn't be staying locked for long, so i'm wary of adding a LOCK_NB here, so i feel slightly uneasy about this patch. Do you have any opinions?
Also, do other OS's EALs need similar fix? -- Thanks, Anatoly
