On Mon, Dec 02, 2019 at 10:21:03AM -0800, Sean Christopherson wrote:

On Mon, Dec 02, 2019 at 09:48:43AM -0600, Haitao Huang wrote:

quoted

On Fri, 29 Nov 2019 17:13:14 -0600, Jarkko Sakkinen
[off-list ref] wrote:

quoted

+
+	for (c = 0 ; c < addp.length; c += PAGE_SIZE) {
+		if (signal_pending(current)) {
+			ret = -ERESTARTSYS;
+			break;
+		}

This IOC is not idempotent as pages EADDed at this point can not be
re-EADDed again. So we can't return ERESTARTSYS

Ah, and now I remember why I opted for modifying the parameters directly
instead of including a "number processed" field.  Andy pointed out the
ERESTARTSYS thing in the original multi-page add RFC[*], so presumably
updating the params and returning ERESTARTSYS is legal/acceptable.

[*] https://lkml.kernel.org/r/CALCETrUb4X9_L9RXKhmyNpfSCsbNodP=BfbfO8Fz_efq24jp8w@mail.gmail.com

There are exactly two primary options to pick from given the
non-idempotent nature:

A. Return zero (since we support partial operation with the changes I
   did in v24).
B. Return -EINTR.

If we wanted to follow common syscall semantics in IO operations, then
the semantics would be a mix of these:

1. Return -EINTR if signals are pending before any pages got added.
2. Return 0 if at least one page got added.

This is how write() works for example according to the documentation
[1]. As far as the user code goes [2] I think for that it is best
idea to rely on semantics that developers are used to instead of
being clever here.

[1] http://man7.org/linux/man-pages/man2/write.2.html
[2] https://lkml.kernel.org/r/CALCETrUb4X9_L9RXKhmyNpfSCsbNodP=BfbfO8Fz_efq24jp8w@mail.gmail.com

/Jarkko