On Sat, Oct 31, 2026 at 02:15:40PM -0400, Michael S. Tsirkin wrote: > On Sat, Oct 31, 2026 at 11:10:04AM -0700, Shakeel Butt wrote: > > A vhost task is a user worker: a thread of a user process that runs only > > kernel code. It never returns to userspace, so it never runs the work > > that is queued for the way back. resume_user_mode_work() does that work > > when TIF_NOTIFY_RESUME is set. For a vhost task this means: > > > > - memory.high is handled at the wrong time. When a charge goes over > > memory.high, the task counts the pages, and should reclaim, and sleep > > if needed, on its way back to userspace. A vhost task only reclaims > > when more than a batch of pages has built up and the charge can block. > > Then it reclaims and sleeps in the middle of the charge, maybe while > > it holds a lock. vhost-scsi, for example, can fault in guest memory > > while it holds vq->mutex. If the charges can't block, nothing is done > > at all, and the count keeps growing. > > > > - Block cgroup throttling never runs. blkcg_schedule_throttle() asks for > > it, but only resume_user_mode_work() acts on it. > > > > - Task work queued with TWA_RESUME, like a deferred fput() or a NUMA > > balancing scan, waits until the task gets a signal or exits. > > > > io_uring workers had the same problem, and io_run_task_work() calls > > resume_user_mode_work() for them. Do the same in vhost_task_fn(): if > > TIF_NOTIFY_RESUME is set, call resume_user_mode_work() at the top of the > > loop. That is between work items, no locks are held there, and the task > > is running. Passing a NULL regs is fine, io_uring and KVM already do it. > > > > This covers every user of vhost_task_create(): the vhost workers and > > KVM's NX huge page recovery thread. > > > > Tested in a VM with a version of tools/virtio/vhost_net_test that puts > > each RX buffer on a new page, so the vhost worker faults in, and > > charges, 256M while its cgroup has memory.high=64M. Before, all 285 > > calls to __mem_cgroup_handle_over_high() came from try_charge_memcg(), > > inside handle_rx() with vq->mutex held. After, all 589 came from > > vhost_task_fn(). memory.peak stayed at 64M, and the run took about the > > same time (0.66s before, 0.67s after). A lockdep kernel showed no > > warnings. > > > > Signed-off-by: Shakeel Butt <[email protected]> > > Is this different from: > > Message-ID: > <as8p251mb00012469e6ecbea2b7d87bb0c8...@as8p251mb0001.eurp251.prod.outlook.com>
Oh wow, it is same but somewhat for different reasons. I am fine that patch, I will comment on it.
