On Mon, Feb 06, 2023 at 03:36:49PM +1100, Alistair Popple wrote: > >> But then I don't really know how RDS works, Santos? > >> > >> Regardless, maybe the vm_account should be stored in the > >> rds_msg_zcopy_info ? > > > > On first glance that looks like a better spot. Thanks for the > > idea. > > That works fine for RDS but not for skbuff. I would definately put the RDS stuff like that.. > We still need a vm_account in the struct sock or somewhere else for > that. For example in msg_zerocopy_realloc() we only have a struct > ubuf_info_msgzc available. We can't add a struct vm_account field to > that because ultimately it is stored in struct sk_buff->ck[] which > is not large enough to contain ubuf_info_msgzc + vm_account. Well, AFAICT this is using iov_iter to get the pages and in general iov_iter - eg as used for O_DIRECT - doesn't charge anything. If this does somehow allow a userspace to hold pin a page for a long time then it is already technically wrong because it doesn't use FOLL_LONGTERM. Arguably FOLL_LONGTERM should be the key precondition to require accounting. So I wonder if it should just be deleted? Jason
