Hi Greg
On Sat, 6 Sept 2025 at 20:03, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Thu, Aug 28, 2025 at 01:59:48PM +0800, Zhangfei Gao wrote:
> > Hi, Greg
> >
> > On Fri, 22 Aug 2025 at 19:46, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Aug 22, 2025 at 06:39:03PM +0800, Chenghai Huang wrote:
> > > > From: Yang Shen <shenyang39@xxxxxxxxxx>
> > > >
> > > > The current uacce_vm_ops does not support the mremap operation of
> > > > vm_operations_struct. Implement .mremap to return -EPERM to remind
> > > > users
> > >
> > > Why is this needed? If mremap is not set, what is the value returned?
> >
> > Did some debug locally.
> >
> > By default, mremap is permitted.
> >
> > With mremap, the original vma is released,
> > The vma_close is called and free resources, including q->qfr.
> >
> > However, vma->vm_private_data (q) is copied to the new vma.
> > When the new vma is closed, vma_close will get q and q->qft=0.
> >
> > So disable mremap here looks safer.
> >
> > >
> > > And why is -EPERM the correct value to return here? That's not what the
> > > man pages say is valid :(
> >
> > if disable mremap, -1 is returned as MAP_FAILED.
> > The errno is decided by the return value, -EPERM (-1) or -EINVAL (-22).
> > man mremap only lists -EINVAL.
> >
> > However, here the driver wants to disable mremap, looks -EPERM is more suitable.
>
> Disabling mremap is not a permission issue, it's more of an invalid
OK，it's fine.
> call? I don't know, what do other drivers do?
Found one example in kernel/events/uprobes.c
commit c16e2fdd746c78f5b2ce3c2ab8a26a61b6ed09e5
Author: Oleg Nesterov <oleg@xxxxxxxxxx>
Date: Sun Sep 29 16:42:58 2024 +0200
uprobes: deny mremap(xol_vma)
+static int xol_mremap(const struct vm_special_mapping *sm, struct
vm_area_struct *new_vma)
+{
+ return -EPERM;
+}
Thanks
