On Fri, Mar 9, 2018 at 8:15 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> On Fri, Mar 9, 2018 at 4:56 AM, Jan Kara <jack@xxxxxxx> wrote:
>> On Thu 08-03-18 09:02:30, Dan Williams wrote:
>>> On Mon, Jan 8, 2018 at 5:50 AM, Jan Kara <jack@xxxxxxx> wrote:
>>> > On Sun 07-01-18 13:58:42, Dan Williams wrote:
>>> >> On Thu, Jan 4, 2018 at 3:12 AM, Jan Kara <jack@xxxxxxx> wrote:
>>> >> > On Sat 23-12-17 16:57:31, Dan Williams wrote:
>>> >> >
>>> >> >> + /*
>>> >> >> + * Flush dax_dma_lock() sections to ensure all possible page
>>> >> >> + * references have been taken, or will block on the fs
>>> >> >> + * 'mmap_lock'.
>>> >> >> + */
>>> >> >> + synchronize_rcu();
>>> >> >
>>> >> > Frankly, I don't like synchronize_rcu() in a relatively hot path like this.
>>> >> > Cannot we just abuse get_dev_pagemap() to fail if truncation is in progress
>>> >> > for the pfn? We could indicate that by some bit in struct page or something
>>> >> > like that.
>>> >>
>>> >> We would need a lockless way to take a reference conditionally if the
>>> >> page is not subject to truncation.
>>> >>
>>> >> I recall the raid5 code did something similar where it split a
>>> >> reference count into 2 fields. I.e. take page->_refcount and use the
>>> >> upper bits as a truncation count. Something like:
>>> >>
>>> >> do {
>>> >> old = atomic_read(&page->_refcount);
>>> >> if (old & trunc_mask) /* upper bits of _refcount */
>>> >> return false;
>>> >> new = cnt + 1;
>>> >> } while (atomic_cmpxchg(&page->_refcount, old, new) != old);
>>> >> return true; /* we incremented the _refcount while the truncation
>>> >> count was zero */
>>> >>
>>> >> ...the only concern is teaching the put_page() path to consider that
>>> >> 'trunc_mask' when determining that the page is idle.
>>> >>
>>> >> Other ideas?
>>> >
>>> > What I rather thought about was an update to GUP paths (like
>>> > follow_page_pte()):
>>> >
>>> > if (flags & FOLL_GET) {
>>> > get_page(page);
>>> > if (pte_devmap(pte)) {
>>> > /*
>>> > * Pairs with the barrier in the truncate path.
>>> > * Could be possibly _after_atomic version of the
>>> > * barrier.
>>> > */
>>> > smp_mb();
>>> > if (PageTruncateInProgress(page)) {
>>> > put_page(page);
>>> > ..bail...
>>> > }
>>> > }
>>> > }
>>> >
>>> > and in the truncate path:
>>> >
>>> > down_write(inode->i_mmap_sem);
>>> > walk all pages in the mapping and mark them PageTruncateInProgress().
>>> > unmap_mapping_range(...);
>>> > /*
>>> > * Pairs with the barrier in GUP path. In fact not necessary since
>>> > * unmap_mapping_range() provides us with the barrier already.
>>> > */
>>> > smp_mb();
>>> > /*
>>> > * By now we are either guaranteed to see grabbed page reference or
>>> > * GUP is guaranteed to see PageTruncateInProgress().
>>> > */
>>> > while ((page = dax_find_referenced_page(mapping))) {
>>> > ...
>>> > }
>>> >
>>> > The barriers need some verification, I've opted for the conservative option
>>> > but I guess you get the idea.
>>>
>>> [ Reviving this thread for the next rev of this patch set for 4.17
>>> consideration ]
>>>
>>> I don't think this barrier scheme can work in the presence of
>>> get_user_pages_fast(). The get_user_pages_fast() path can race
>>> unmap_mapping_range() to take out an elevated reference count on a
>>> page.
>>
>> Why the scheme cannot work? Sure you'd need to patch also gup_pte_range()
>> and a similar thing for PMDs to recheck PageTruncateInProgress() after
>> grabbing the page reference. But in principle I don't see anything
>> fundamentally different between gup_fast() and plain gup().
>
> Ah, yes I didn't grok the abort on PageTruncateInProgress() until I
> read this again (and again), I'll try that.
Ok, so the problem is that PageTruncateInProgress() for a given page
is hard to detect without trapping deeper into the filesystem, at
least in the XFS case. The usage of xfs_break_layouts() happens well
before we know that a given file offset is going to be truncated. By
the time we're at a point in the call stack where we are committed to
truncating a given page it is then awkward to drop locks and wait on
the next page collision.
In order to support an early 'break' of dax layouts before touching
the extent map we can't rely on being able to positively determine the
pages that collide with a given truncate/hole-punch range. Instead the
approach I've taken drains all pinned / referenced pages for the inode
before attempting an operation that *might* lead to an extent unmap
event. This mirrors the pNFS lease case where all leases are broken
regardless of whether they actually collide with an extent that is
under active access from a remote client.
