Re: [PATCH 10/12] fs/dax: Properly refcount fs dax pages

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Alistair Popple wrote:
[..]
>> I'm not really following this scenario, or at least how it relates to
> >> the comment above. If the page is pinned for DMA it will have taken a
> >> refcount on it and so the page won't be considered free/idle per
> >> dax_wait_page_idle() or any of the other mm code.
> >
> > [ tl;dr: I think we're ok, analysis below, but I did talk myself into
> > the proposed dax_busy_page() changes indeed being broken and needing to
> > remain checking for refcount > 1, not > 0 ]
> >
> > It's not the mm code I am worried about. It's the filesystem block
> > allocator staying in-sync with the allocation state of the page.
> >
> > fs/dax.c is charged with converting idle storage blocks to pfns to
> > mapped folios. Once they are mapped, DMA can pin the folio, but nothing
> > in fs/dax.c pins the mapping. In the pagecache case the page reference
> > is sufficient to keep the DMA-busy page from being reused. In the dax
> > case something needs to arrange for DMA to be idle before
> > dax_delete_mapping_entry().
> 
> Ok. How does that work today? My current mental model is that something
> has to call dax_layout_busy_page() whilst holding the correct locks to
> prevent a new mapping being established prior to calling
> dax_delete_mapping_entry(). Is that correct?

Correct. dax_delete_mapping_entry() is invoked by the filesystem with
inode locks held. See xfs_file_fallocate() where it takes the lock,
calls xfs_break_layouts() and if that succeeds performs
xfs_file_free_space() with the lock held.

xfs_file_free_space() triggers dax_delete_mapping_entry() with knowledge
that the mapping cannot be re-established until the lock is dropped.

> > However, looking at XFS it indeed makes that guarantee. First it does
> > xfs_break_dax_layouts() then it does truncate_inode_pages() =>
> > dax_delete_mapping_entry().
> >
> > It follows that that the DMA-idle condition still needs to look for the
> > case where the refcount is > 1 rather than 0 since refcount == 1 is the
> > page-mapped-but-DMA-idle condition.
> 
> Sorry, but I'm still not following this line of reasoning. If the
> refcount == 1 the page is either mapped xor DMA-busy.

No, my expectation is the refcount is 1 while the page has a mapping
entry, analagous to an idle / allocated page cache page, and the
refcount is 2 or more for DMA, get_user_pages(), or any page walker that
takes a transient page pin.

> is enough to conclude that the page cannot be reused because it is
> either being accessed from userspace via a CPU mapping or from some
> device DMA or some other in kernel user.

Userspace access is not a problem, that access can always be safely
revoked by unmapping the page, and that's what dax_layout_busy_page()
does to force a fault and re-taking the inode + mmap locks so that the
truncate path knows it has temporary exclusive access to the page, pfn,
and storage-block association.

> The current proposal is that dax_busy_page() returns true if refcount >=
> 1, and dax_wait_page_idle() will wait until the refcount ==
> 0. dax_busy_page() will try and force the refcount == 0 by unmapping it,
> but obviously can't force other pinners to release their reference hence
> the need to wait. Callers should already be holding locks to ensure new
> mappings can't be established and hence can't become DMA-busy after the
> unmap.

Am I missing a page_ref_dec() somewhere? Are you saying that
dax_layout_busy_page() will find entries with ->mapping non-NULL and
refcount == 0?

[..]
> >> >> @@ -1684,14 +1663,21 @@ static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
> >> >>  	if (dax_fault_is_synchronous(iter, vmf->vma))
> >> >>  		return dax_fault_synchronous_pfnp(pfnp, pfn);
> >> >>  
> >> >> -	/* insert PMD pfn */
> >> >> +	page = pfn_t_to_page(pfn);
> >> >
> >> > I think this is clearer if dax_insert_entry() returns folios with an
> >> > elevated refrence count that is dropped when the folio is invalidated
> >> > out of the mapping.
> >> 
> >> I presume this comment is for the next line:
> >> 
> >> +	page_ref_inc(page);
> >>  
> >> I can move that into dax_insert_entry(), but we would still need to
> >> drop it after calling vmf_insert_*() to ensure we get the 1 -> 0
> >> transition when the page is unmapped and therefore
> >> freed. Alternatively we can make it so vmf_insert_*() don't take
> >> references on the page, and instead ownership of the reference is
> >> transfered to the mapping. Personally I prefered having those
> >> functions take their own reference but let me know what you think.
> >
> > Oh, the model I was thinking was that until vmf_insert_XXX() succeeds
> > then the page was never allocated because it was never mapped. What
> > happens with the code as proposed is that put_page() triggers page-free
> > semantics on vmf_insert_XXX() failures, right?
> 
> Right. And actually that means I can't move the page_ref_inc(page) into
> what will be called dax_create_folio(), because an entry may have been
> created previously that had a failed vmf_insert_XXX() which will
> therefore have a zero refcount folio associated with it.

I would expect a full cleanup on on vmf_insert_XXX() failure, not
leaving a zero-referenced entry.

> But I think that model is wrong. I think the model needs to be the page
> gets allocated when the entry is first created (ie. when
> dax_create_folio() is called). A subsequent free (ether due to
> vmf_insert_XXX() failing or the page being unmapped or becoming
> DMA-idle) should then delete the entry.
>
> I think that makes the semantics around dax_busy_page() nicer as well -
> no need for the truncate to have a special path to call
> dax_delete_mapping_entry().

I agree it would be lovely if the final put could clean up the mapping
entry and not depend on truncate_inode_pages_range() to do that.

...but I do not immediately see how to get there when block, pfn, and
page are so tightly coupled with dax. That's a whole new project to
introduce that paradigm, no? The page cache case gets away with
it by safely disconnecting the pfn+page from the block and then letting
DMA final put_page() take its time.

> > There is no need to invoke the page-free / final-put path on
> > vmf_insert_XXX() error because the storage-block / pfn never actually
> > transitioned into a page / folio.
> 
> It's not mapping a page/folio that transitions a pfn into a page/folio
> it is the allocation of the folio that happens in dax_create_folio()
> (aka. dax_associate_new_entry()). So we need to delete the entry (as
> noted above I don't do that currently) if the insertion fails.

Yeah, deletion on insert failure makes sense.

[..]
> >> >> @@ -519,21 +529,3 @@ void zone_device_page_init(struct page *page)
> >> >>  	lock_page(page);
> >> >>  }
> >> >>  EXPORT_SYMBOL_GPL(zone_device_page_init);
> >> >> -
> >> >> -#ifdef CONFIG_FS_DAX
> >> >> -bool __put_devmap_managed_folio_refs(struct folio *folio, int refs)
> >> >> -{
> >> >> -	if (folio->pgmap->type != MEMORY_DEVICE_FS_DAX)
> >> >> -		return false;
> >> >> -
> >> >> -	/*
> >> >> -	 * fsdax page refcounts are 1-based, rather than 0-based: if
> >> >> -	 * refcount is 1, then the page is free and the refcount is
> >> >> -	 * stable because nobody holds a reference on the page.
> >> >> -	 */
> >> >> -	if (folio_ref_sub_return(folio, refs) == 1)
> >> >> -		wake_up_var(&folio->_refcount);
> >> >> -	return true;
> >> >
> >> > It follow from the refcount disvussion above that I think there is an
> >> > argument to still keep this wakeup based on the 2->1 transitition.
> >> > pagecache pages are refcount==1 when they are dma-idle but still
> >> > allocated. To keep the same semantics for dax a dax_folio would have an
> >> > elevated refcount whenever it is referenced by mapping entry.
> >> 
> >> I'm not sold on keeping it as it doesn't seem to offer any benefit
> >> IMHO. I know both Jason and Christoph were keen to see it go so it be
> >> good to get their feedback too. Also one of the primary goals of this
> >> series was to refcount the page normally so we could remove the whole
> >> "page is free with a refcount of 1" semantics.
> >
> > The page is still free at refcount 0, no argument there. But, by
> > introducing a new "page refcount is elevated while mapped" (as it
> > should), it follows that "page is DMA idle at refcount == 1", right?
> 
> No. The page is either mapped xor DMA-busy - ie. not free. If we want
> (need?) to tell the difference we can use folio_maybe_dma_pinned(),
> assuming the driver doing DMA has called pin_user_pages() as it should.
> 
> That said I'm not sure why we care about the distinction between
> DMA-idle and mapped? If the page is not free from the mm perspective the
> block can't be reallocated by the filesystem.

"can't be reallocated", what enforces that in your view? I am hoping it
is something I am overlooking.

In my view the filesystem has no idea of this page-to-block
relationship. All it knows is that when it wants to destroy the
page-to-block association, dax notices and says "uh, oh, this is my last
chance to make sure the block can go back into the fs allocation pool so
I need to wait for the mm to say that the page is exclusive to me (dax
core) before dax_delete_mapping_entry() destroys the page-to-block
association and the fs reclaims the allocation".

> > Otherwise, the current assumption that fileystems can have
> > dax_layout_busy_page_range() poll on the state of the pfn in the mapping
> > is broken because page refcount == 0 also means no page to mapping
> > association.
> 
> And also means nothing from the mm (userspace mapping, DMA-busy, etc.)
> is using the page so the page isn't busy and is free to be reallocated
> right?

Lets take the 'map => start dma => truncate => end dma' scenario.

At the 'end dma' step, how does the filesystem learn that the block that
it truncated, potentially hours ago, is now a free block? The filesystem
thought it reclaimed the block when truncate completed. I.e. dax says,
thou shalt 'end dma' => 'truncate' in all cases.

Note "dma" can be replaced with "any non dax core page_ref".




[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux