Date:   Thu, 15 Jun 2023 16:52:40 +0200
From:   Michał Mirosław <emmir@...gle.com>
To:     Muhammad Usama Anjum <usama.anjum@...labora.com>
Cc:     Peter Xu <peterx@...hat.com>, David Hildenbrand <david@...hat.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Andrei Vagin <avagin@...il.com>,
        Danylo Mocherniuk <mdanylo@...gle.com>,
        Paul Gofman <pgofman@...eweavers.com>,
        Cyrill Gorcunov <gorcunov@...il.com>,
        Mike Rapoport <rppt@...nel.org>, Nadav Amit <namit@...are.com>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Shuah Khan <shuah@...nel.org>,
        Christian Brauner <brauner@...nel.org>,
        Yang Shi <shy828301@...il.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        "Liam R . Howlett" <Liam.Howlett@...cle.com>,
        Yun Zhou <yun.zhou@...driver.com>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Alex Sierra <alex.sierra@....com>,
        Matthew Wilcox <willy@...radead.org>,
        Pasha Tatashin <pasha.tatashin@...een.com>,
        Axel Rasmussen <axelrasmussen@...gle.com>,
        "Gustavo A . R . Silva" <gustavoars@...nel.org>,
        Dan Williams <dan.j.williams@...el.com>,
        linux-kernel@...r.kernel.org, linux-fsdevel@...r.kernel.org,
        linux-mm@...ck.org, linux-kselftest@...r.kernel.org,
        Greg KH <gregkh@...uxfoundation.org>, kernel@...labora.com
Subject: Re: [PATCH v18 2/5] fs/proc/task_mmu: Implement IOCTL to get and
 optionally clear info about PTEs

On Thu, 15 Jun 2023 at 15:58, Muhammad Usama Anjum
<usama.anjum@...labora.com> wrote:
> I'll send next revision now.
> On 6/14/23 11:00 PM, Michał Mirosław wrote:
> > (A quick reply to answer open questions in case they help the next version.)
> >
> > On Wed, 14 Jun 2023 at 19:10, Muhammad Usama Anjum
> > <usama.anjum@...labora.com> wrote:
> >> On 6/14/23 8:14 PM, Michał Mirosław wrote:
> >>> On Wed, 14 Jun 2023 at 15:46, Muhammad Usama Anjum
> >>> <usama.anjum@...labora.com> wrote:
> >>>>
> >>>> On 6/14/23 3:36 AM, Michał Mirosław wrote:
> >>>>> On Tue, 13 Jun 2023 at 12:29, Muhammad Usama Anjum
> >>>>> <usama.anjum@...labora.com> wrote:
> > [...]
> >>>>>> +       if (cur_buf->bitmap == bitmap &&
> >>>>>> +           cur_buf->start + cur_buf->len * PAGE_SIZE == addr) {
> >>>>>> +               cur_buf->len += n_pages;
> >>>>>> +               p->found_pages += n_pages;
> >>>>>> +       } else {
> >>>>>> +               if (cur_buf->len && p->vec_buf_index >= p->vec_buf_len)
> >>>>>> +                       return -ENOMEM;
> >>>>>
> >>>>> Shouldn't this be -ENOSPC? -ENOMEM usually signifies that the kernel
> >>>>> ran out of memory when allocating, not that there is no space in a
> >>>>> user-provided buffer.
> >>>> There are 3 kinds of return values here:
> >>>> * PM_SCAN_FOUND_MAX_PAGES (1) ---> max_pages have been found. Abort the
> >>>> page walk from next entry
> >>>> * 0 ---> continue the page walk
> >>>> * -ENOMEM --> Abort the page walk from current entry, user buffer is full
> >>>> which is not error, but only a stop signal. This -ENOMEM is just
> >>>> differentiater from (1). This -ENOMEM is for internal use and isn't
> >>>> returned to user.
> >>>
> >>> But why ENOSPC is not good here? I was used before, I think.
> >> -ENOSPC is being returned in form of true error from
> >> pagemap_scan_hugetlb_entry(). So I'd to remove -ENOSPC from here as it
> >> wasn't true error here, it was only a way to abort the walk immediately.
> >> I'm liking the following erturn code from here now:
> >>
> >> #define PM_SCAN_BUFFER_FULL     (-256)
> >
> > I guess this will be reworked anyway, but I'd prefer this didn't need
> > custom errors etc. If we agree to decoupling the selection and GET
> > output, it could be:
> >
> > bool is_interesting_page(p, flags); // this one does the
> > required/anyof/excluded match
> > size_t output_range(p, start, len, flags); // this one fills the
> > output vector and returns how many pages were fit
> >
> > In this setup, `is_interesting_page() && (n_out = output_range()) <
> > n_pages` means this is the final range, no more will fit. And if
> > `n_out == 0` then no pages fit and no WP is needed (no other special
> > cases).
> Right now, pagemap_scan_output() performs the work of both of these two
> functions. The part can be broken into is_interesting_pages() and we can
> leave the remaining part as it is.
>
> Saying that n_out < n_pages tells us the buffer is full covers one case.
> But there is case of maximum pages have been found and walk needs to be
> aborted.

This case is exactly what `n_out < n_pages` will cover (if scan_output
uses max_pages properly to limit n_out).
Isn't it that when the buffer is full we want to abort the scan always
(with WP if `n_out > 0`)?

> >>>>> For flags name: PM_REQUIRE_WRITE_ACCESS?
> >>>>> Or Is it intended to be checked only if doing WP (as the current name
> >>>>> suggests) and so it would be redundant as WP currently requires
> >>>>> `p->required_mask = PAGE_IS_WRITTEN`?
> >>>> This is intended to indicate that if userfaultfd is needed. If
> >>>> PAGE_IS_WRITTEN is mentioned in any of mask, we need to check if
> >>>> userfaultfd has been initialized for this memory. I'll rename to
> >>>> PM_SCAN_REQUIRE_UFFD.
> >>>
> >>> Why do we need that check? Wouldn't `is_written = false` work for vmas
> >>> not registered via uffd?
> >> UFFD_FEATURE_WP_ASYNC and UNPOPULATED needs to be set on the memory region
> >> for it to report correct written values on the memory region. Without UFFD
> >> WP ASYNC and UNPOUPULATED defined on the memory, we consider UFFD_WP state
> >> undefined. If user hasn't initialized memory with UFFD, he has no right to
> >> set is_written = false.
> >
> > How about calculating `is_written = is_uffd_registered() &&
> > is_uffd_wp()`? This would enable a user to apply GET+WP for the whole
> > address space of a process regardless of whether all of it is
> > registered.
> I wouldn't want to check if uffd is registered again and again. This is why
> we are doing it only once every walk in pagemap_scan_test_walk().

There is no need to do the checks repeatedly. If I understand the code
correctly, uffd registration is per-vma, so it can be communicated
from test_walk to entry/hole callbacks via a field in
pagemap_scan_private.

> >>> While here, I wonder if we really need to fail the call if there are
> >>> unknown bits in those masks set: if this bit set is expanded with
> >>> another category flags, a newer userspace run on older kernel would
> >>> get EINVAL even if the "treat unknown as 0" be what it requires.
> >>> There is no simple way in the API to discover what bits the kernel
> >>> supports. We could allow a no-op (no WP nor GET) call to help with
> >>> that and then rejecting unknown bits would make sense.
> >> I've not seen any examples of this. But I've seen examples of returning
> >> error if kernel doesn't support a feature. Each new feature comes with a
> >> kernel version, greater than this version support this feature. If user is
> >> trying to use advanced feature which isn't present in a kernel, we should
> >> return error and not proceed to confuse the user/kernel. In fact if we look
> >> at userfaultfd_api(), we return error immediately if feature has some bit
> >> set which kernel doesn't support.
> >
> > I think we should have a way of detecting the supported flags if we
> > don't want a forward compatibility policy for flags here. Maybe it
> > would be enough to allow all the no-op combinations for this purpose?
> Again I don't think UFFD is doing anything like this.

If it's cheap and easy to provide a user with a way to detect the
supported features - why not do it?

Best Regards
Michał Mirosław

Powered by blists - more mailing lists