[PATCH 01/13] PCI/P2PDMA: Release the per-cpu ref of pgmap when vm_insert_page() fails

[Alistair Popple]

On 2026-01-12 at 11:12 +1100, Alistair Popple <apopple at nvidia.com> wrote...
> On 2026-01-12 at 10:21 +1100, Alistair Popple <apopple at nvidia.com> wrote...
> > On 2026-01-10 at 02:03 +1100, Bjorn Helgaas <helgaas at kernel.org> wrote...
> > > On Fri, Jan 09, 2026 at 11:41:51AM +1100, Alistair Popple wrote:
> > > > On 2026-01-09 at 02:55 +1100, Bjorn Helgaas <helgaas at kernel.org> wrote...
> > > > > On Thu, Jan 08, 2026 at 02:23:16PM +1100, Alistair Popple wrote:
> > > > > > On 2025-12-20 at 15:04 +1100, Hou Tao <houtao at huaweicloud.com> wrote...
> > > > > > > From: Hou Tao <houtao1 at huawei.com>
> > > > > > > 
> > > > > > > When vm_insert_page() fails in p2pmem_alloc_mmap(), p2pmem_alloc_mmap()
> > > > > > > doesn't invoke percpu_ref_put() to free the per-cpu ref of pgmap
> > > > > > > acquired after gen_pool_alloc_owner(), and memunmap_pages() will hang
> > > > > > > forever when trying to remove the PCIe device.
> > > > > > > 
> > > > > > > Fix it by adding the missed percpu_ref_put().
> > > > ...
> > > 
> > > > > Looking at this again, I'm confused about why in the normal, non-error
> > > > > case, we do the percpu_ref_tryget_live_rcu(ref), followed by another
> > > > > percpu_ref_get(ref) for each page, followed by just a single
> > > > > percpu_ref_put() at the exit.
> > > > > 
> > > > > So we do ref_get() "1 + number of pages" times but we only do a single
> > > > > ref_put().  Is there a loop of ref_put() for each page elsewhere?
> > > > 
> > > > Right, the per-page ref_put() happens when the page is freed (ie. the struct
> > > > page refcount drops to zero) - in this case free_zone_device_folio() will call
> > > > p2pdma_folio_free() which has the corresponding percpu_ref_put().
> > > 
> > > I don't see anything that looks like a loop to call ref_put() for each
> > > page in free_zone_device_folio() or in p2pdma_folio_free(), but this
> > > is all completely out of my range, so I'll take your word for it :)  
> > 
> > That's brave :-)
> > 
> > What happens is the core mm takes over managing the page life time once
> > vm_insert_page() has been (successfully) called to map the page:
> > 
> > 	VM_WARN_ON_ONCE_PAGE(!page_ref_count(page), page);
> > 	set_page_count(page, 1);
> > 	ret = vm_insert_page(vma, vaddr, page);
> > 	if (ret) {
> > 		gen_pool_free(p2pdma->pool, (uintptr_t)kaddr, len);
> > 		return ret;
> > 	}
> > 	percpu_ref_get(ref);
> > 	put_page(page);
> > 
> > In the above sequence vm_insert_page() takes a page ref for each page it maps
> > into the user page tables with folio_get(). This reference is dropped when the
> > user page table entry is removed, typically by the loop in zap_pte_range().
> > 
> > Normally the user page table mapping is the only thing holding a reference so
> > it ends up calling folio_put()->free_zone_device_folio->...->ref_put() one page
> > at a time as the PTEs are removed from the page tables. At least that's what
> > happens conceptually - the TLB batching code makes it hard to actually see where
> > the folio_put() is called in this sequence.
> > 
> > Note the extra set_page_count(1) and put_page(page) in the above sequence is
> > just to make vm_insert_page() happy - it complains it you try and insert a page
> > with a zero page ref.
> > 
> > And looking at that sequence there is another minor bug - in the failure
> > path we are exiting the loop with the failed page ref count set to
> > 1 from set_page_count(page, 1). That needs to be reset to zero with
> > set_page_count(page, 0) to avoid the VM_WARN_ON_ONCE_PAGE() if the page gets
> > reused. I will send a fix for that.
> 
> Actually the whole failure path above seems wrong to me - we
> free the entire allocation with gen_pool_free() even though
> vm_insert_page() may have succeeded in mapping some pages. AFAICT the
> generic VFS mmap code will call unmap_region() to undo any partial
> mapping (see __mmap_new_file_vma) but that should end up calling
> folio_put()->zone_free_device_range()->p2pdma_folio_free()->gen_pool_free_owner()
> for the mapped pages even though we've already freed the entire pool.

Oh nevermind, I hit send too soon. Ignore the above paragraph - I hadn't noticed
kaddr/len gets updated at the end of the loop to account for the successful
mappings.

> >  - Alistair
> > 
> > > Bjorn
> > 
>