On 11/29, NeilBrown wrote: > > On Wed, 29 Nov 2023, Oleg Nesterov wrote: > > On 11/28, NeilBrown wrote: > > > > > > I have evidence from a customer site of 256 nfsd threads adding files to > > > delayed_fput_lists nearly twice as fast they are retired by a single > > > work-queue thread running delayed_fput(). As you might imagine this > > > does not end well (20 million files in the queue at the time a snapshot > > > was taken for analysis). > > > > On a related note... Neil, Al, et al, can you look at > > > > [PATCH 1/3] fput: don't abuse task_work_add() when possible > > https://lore.kernel.org/all/20150908171446.GA14589@xxxxxxxxxx/ > > > > Would it make sense to create a separate task_struct->delayed_fput > llist? Sure, I too thought about this, > fput() adds the file to this llist and if it was the first item on the > list, it then adds the task_work. That would probably request adding a > callback_head to struct task_struct as well. Even simpler, but perhaps I missed something... We can add a "struct file *fput_xxx" into task_struct and f_fput_xxx into the f_llist/f_rcuhead union in the struct file. fput: if (task->fput_xxx) { file->f_fput_xxx = task->fput_xxx; task->fput_xxx = file; } else { task_work_add(...); // XXX: file->f_fput_xxx != NULL task->fput_xxx = file; } ____fput: struct file *file = task->fput_xxx; struct file *tail = container_of(work, ...); // see XXX in fput() tail->f_fput_xxx = NULL; current->fput_xxx = NULL; do { next = READ_ONCE(file->f_fput_xxx); __fput(file); file = next; } while (file); Again, quite possibly I missed something, but something like this should work. But I am still trying to find a simpler solution which doesn't need another member in task_struct... Oleg.
