Re: [PATCH v3 1/3] s390/pci: Fix s390_mmio_read/write syscall page fault handling

[David Hildenbrand <david@redhat.com>]

On 11.06.24 15:23, Niklas Schnelle wrote:
> On Tue, 2024-06-11 at 14:08 +0200, Niklas Schnelle wrote:
>> On Tue, 2024-06-11 at 13:21 +0200, Niklas Schnelle wrote:
>>> On Wed, 2024-05-29 at 13:36 +0200, Niklas Schnelle wrote:
>>>> The s390 MMIO syscalls when using the classic PCI instructions do not
>>>> cause a page fault when follow_pte() fails due to the page not being
>>>> present. Besides being a general deficiency this breaks vfio-pci's mmap()
>>>> handling once VFIO_PCI_MMAP gets enabled as this lazily maps on first
>>>> access. Fix this by following a failed follow_pte() with
>>>> fixup_user_page() and retrying the follow_pte().
>>>>
>>>> Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
>>>> Reviewed-by: Matthew Rosato <mjrosato@linux.ibm.com>
>>>> Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
>>>> ---
>>>>   arch/s390/pci/pci_mmio.c | 18 +++++++++++++-----
>>>>   1 file changed, 13 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/arch/s390/pci/pci_mmio.c b/arch/s390/pci/pci_mmio.c
>>>> index 5398729bfe1b..80c21b1a101c 100644
>>>> --- a/arch/s390/pci/pci_mmio.c
>>>> +++ b/arch/s390/pci/pci_mmio.c
>>>> @@ -170,8 +170,12 @@ SYSCALL_DEFINE3(s390_pci_mmio_write, unsigned long, mmio_addr,
>>>>   		goto out_unlock_mmap;
>>>>   
>>>>   	ret = follow_pte(vma, mmio_addr, &ptep, &ptl);
>>>> -	if (ret)
>>>> -		goto out_unlock_mmap;
>>>> +	if (ret) {
>>>> +		fixup_user_fault(current->mm, mmio_addr, FAULT_FLAG_WRITE, NULL);
>>>> +		ret = follow_pte(vma, mmio_addr, &ptep, &ptl);
>>>> +		if (ret)
>>>> +			goto out_unlock_mmap;
>>>> +	}
>>>>   
>>>>   	io_addr = (void __iomem *)((pte_pfn(*ptep) << PAGE_SHIFT) |
>>>>   			(mmio_addr & ~PAGE_MASK));
>>>> @@ -305,12 +309,16 @@ SYSCALL_DEFINE3(s390_pci_mmio_read, unsigned long, mmio_addr,
>>>>   	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
>>>>   		goto out_unlock_mmap;
>>>>   	ret = -EACCES;
>>>> -	if (!(vma->vm_flags & VM_WRITE))
>>>> +	if (!(vma->vm_flags & VM_READ))
>>>>   		goto out_unlock_mmap;
>>>>   
>>>>   	ret = follow_pte(vma, mmio_addr, &ptep, &ptl);
>>>> -	if (ret)
>>>> -		goto out_unlock_mmap;
>>>> +	if (ret) {
>>>> +		fixup_user_fault(current->mm, mmio_addr, 0, NULL);
>>>> +		ret = follow_pte(vma, mmio_addr, &ptep, &ptl);
>>>> +		if (ret)
>>>> +			goto out_unlock_mmap;
>>>> +	}
>>>>   
>>>>   	io_addr = (void __iomem *)((pte_pfn(*ptep) << PAGE_SHIFT) |
>>>>   			(mmio_addr & ~PAGE_MASK));
>>>>
>>>
>>> Ughh, I think I just stumbled over a problem with this. This is a
>>> failing lock held assertion via __is_vma_write_locked() in
>>> remap_pfn_range_notrack() but I'm not sure yet what exactly causes this
>>>
>>> [   67.338855] ------------[ cut here ]------------
>>> [   67.338865] WARNING: CPU: 15 PID: 2056 at include/linux/rwsem.h:85 remap_pfn_range_notrack+0x596/0x5b0
>>> [   67.338874] Modules linked in: <--- 8< --->
>>> [   67.338931] CPU: 15 PID: 2056 Comm: vfio-test Not tainted 6.10.0-rc1-pci-pfault-00004-g193e3a513cee #5
>>> [   67.338934] Hardware name: IBM 3931 A01 701 (LPAR)
>>> [   67.338935] Krnl PSW : 0704c00180000000 000003e54c9730ea (remap_pfn_range_notrack+0x59a/0x5b0)
>>> [   67.338940]            R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
>>> [   67.338944] Krnl GPRS: 0000000000000100 000003655915fb78 000002d80b9a5928 000003ff7fa00000
>>> [   67.338946]            0004008000000000 0000000000004000 0000000000000711 000003ff7fa04000
>>> [   67.338948]            000002d80c533f00 000002d800000100 000002d81bbe6c28 000002d80b9a5928
>>> [   67.338950]            000003ff7fa00000 000002d80c533f00 000003e54c973120 000003655915fab0
>>> [   67.338956] Krnl Code: 000003e54c9730de: a708ffea            lhi     %r0,-22
>>>                            000003e54c9730e2: a7f4fff6            brc     15,000003e54c9730ce
>>>                           #000003e54c9730e6: af000000            mc      0,0
>>>                           >000003e54c9730ea: a7f4fd6e            brc     15,000003e54c972bc6
>>>                            000003e54c9730ee: af000000            mc      0,0
>>>                            000003e54c9730f2: af000000            mc      0,0
>>>                            000003e54c9730f6: 0707                bcr     0,%r7
>>>                            000003e54c9730f8: 0707                bcr     0,%r7
>>> [   67.339025] Call Trace:
>>> [   67.339027]  [<000003e54c9730ea>] remap_pfn_range_notrack+0x59a/0x5b0
>>> [   67.339032]  [<000003e54c973120>] remap_pfn_range+0x20/0x30
>>> [   67.339035]  [<000003e4cce5396c>] vfio_pci_mmap_fault+0xec/0x1d0 [vfio_pci_core]
>>> [   67.339043]  [<000003e54c977240>] handle_mm_fault+0x6b0/0x25a0
>>> [   67.339046]  [<000003e54c966328>] fixup_user_fault+0x138/0x310
>>> [   67.339048]  [<000003e54c63a91c>] __s390x_sys_s390_pci_mmio_read+0x28c/0x3a0
>>> [   67.339051]  [<000003e54c5e200a>] do_syscall+0xea/0x120
>>> [   67.339055]  [<000003e54d5f9954>] __do_syscall+0x94/0x140
>>> [   67.339059]  [<000003e54d611020>] system_call+0x70/0xa0
>>> [   67.339063] Last Breaking-Event-Address:
>>> [   67.339065]  [<000003e54c972bc2>] remap_pfn_range_notrack+0x72/0x5b0
>>> [   67.339067] ---[ end trace 0000000000000000 ]---
>>>
>>
>> This has me a bit confused so far as __is_vma_write_locked() checks
>> mmap_assert_write_locked(vma->vm_mm) but most other users of
>> fixup_user_fault() hold mmap_read_lock() just like this code and
>> clearly in the non page fault case we only need the read lock.

This is likely the 
vm_flags_set()->vma_start_write(vma)->__is_vma_write_locked()

which checks mmap_assert_write_locked().

Setting VMA flags would be racy with the mmap lock in read mode.


remap_pfn_range() documents: "this is only safe if the mm semaphore is 
held when called." which doesn't spell out if it needs to be held in 
write mode (which I think it does) :)


My best guess is: if you are using remap_pfn_range() from a fault 
handler (not during mmap time) you are doing something wrong, that's why 
you get that report.

vmf_insert_pfn() and friends might be better alternatives, that make 
sure that the VMA already received the proper VMA flags at mmap time.

>>
> 
> And it gets weirder, as I could have sworn that I properly tested this
> on v1, I retested with v1 (tags/sent/vfio_pci_mmap-v1 on my
> git.kernel.org/niks and based on v6.9) and there I don't get the above
> warning. I also made sure that it's not caused by my change to
> "current->mm" for v2. But I'm also not hitting the checks David moved
> into follow_pte() so yeah not sure what's going on here.


You mean the mmap_assert_locked()? Yeah, that only checks if you have it 
in read mode, but not in write mode.

-- 
Cheers,

David / dhildenb