Re: [BUG 5.14] arm64/mm: dma memory mapping fails (in some cases)

[David Hildenbrand <david@redhat.com>]

On 25.08.21 12:55, Catalin Marinas wrote:
> On Wed, Aug 25, 2021 at 12:38:31PM +0200, David Hildenbrand wrote:
>> On 25.08.21 12:20, Catalin Marinas wrote:
>>> On Tue, Aug 24, 2021 at 08:59:22PM +0200, David Hildenbrand wrote:
>>>> On 24.08.21 20:46, Robin Murphy wrote:
>>>>> On 2021-08-24 19:28, Mike Rapoport wrote:
>>>>>> On Tue, Aug 24, 2021 at 06:37:41PM +0100, Catalin Marinas wrote:
>>>>>>> On Tue, Aug 24, 2021 at 03:40:47PM +0200, Alex Bee wrote:
>>>>>>>> it seems there is a regression in arm64 memory mapping in 5.14, since it
>>>>>>>> fails on Rockchip RK3328 when the pl330 dmac tries to map with:
>>>>>>>>
>>>>>>>>    ------------[ cut here ]------------
>>>>>>>>    WARNING: CPU: 2 PID: 373 at kernel/dma/mapping.c:235 dma_map_resource+0x68/0xc0
>>>>>>>>    Modules linked in: spi_rockchip(+) fuse
>>>>>>>>    CPU: 2 PID: 373 Comm: systemd-udevd Not tainted 5.14.0-rc7 #1
>>>>>>>>    Hardware name: Pine64 Rock64 (DT)
>>>>>>>>    pstate: 80000005 (Nzcv daif -PAN -UAO -TCO BTYPE=--)
>>>>>>>>    pc : dma_map_resource+0x68/0xc0
>>>>>>>>    lr : pl330_prep_slave_fifo+0x78/0xd0
>>>>>>>>    sp : ffff800012102ae0
>>>>>>>>    x29: ffff800012102ae0 x28: ffff000005c94800 x27: 0000000000000000
>>>>>>>>    x26: ffff000000566bd0 x25: 0000000000000001 x24: 0000000000000001
>>>>>>>>    x23: 0000000000000002 x22: ffff000000628c00 x21: 0000000000000001
>>>>>>>>    x20: ffff000000566bd0 x19: 0000000000000001 x18: 0000000000000000
>>>>>>>>    x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
>>>>>>>>    x14: 0000000000000277 x13: 0000000000000001 x12: 0000000000000000
>>>>>>>>    x11: 0000000000000001 x10: 00000000000008e0 x9 : ffff800012102a80
>>>>>>>>    x8 : ffff000000d14b80 x7 : ffff0000fe7b12f0 x6 : ffff0000fe7b1100
>>>>>>>>    x5 : fffffc000000000f x4 : 0000000000000000 x3 : 0000000000000001
>>>>>>>>    x2 : 0000000000000001 x1 : 00000000ff190800 x0 : ffff000000628c00
>>>>>>>>    Call trace:
>>>>>>>>      dma_map_resource+0x68/0xc0
>>>>>>>>      pl330_prep_slave_sg+0x58/0x220
>>>>>>>>      rockchip_spi_prepare_dma+0xd8/0x2c0 [spi_rockchip]
>>>>>>>>      rockchip_spi_transfer_one+0x294/0x3d8 [spi_rockchip]
>>>>>>> [...]
>>>>>>>> Note: This does not relate to the spi driver - when disabling this device in
>>>>>>>> the device tree it fails for any other (i2s, for instance) which uses dma.
>>>>>>>> Commenting out the failing check at [1], however, helps and the mapping
>>>>>>>> works again.
>>>>>>
>>>>>>> Do you know which address dma_map_resource() is trying to map (maybe
>>>>>>> add some printk())? It's not supposed to map RAM, hence the warning.
>>>>>>> Random guess, the address is 0xff190800 (based on the x1 above but the
>>>>>>> regs might as well be mangled).
>>>>>>
>>>>>> 0xff190800 will cause this warning for sure. It has a memory map, but it is
>>>>>> not RAM so old version of pfn_valid() would return 0 and the new one
>>>>>> returns 1.
>>>>>
>>>>> How does that happen, though? It's not a memory address, and it's not
>>>>> even within the bounds of anywhere there should or could be memory. This
>>>>> SoC has a simple memory map - everything from 0 to 0xfeffffff goes to
>>>>> the DRAM controller (which may not all be populated, and may have pieces
>>>>> carved out by secure firmware), while 0xff000000-0xffffffff is MMIO. Why
>>>>> do we have pages (or at least the assumption of pages) for somewhere
>>>>> which by all rights should not have them?
>>>>
>>>> Simple: we allocate the vmemmap for whole sections (e.g., 128 MiB) to avoid
>>>> any such hacks. If there is a memory hole, it gets a memmap as well.
>>>>
>>>> Tricking pfn_valid() into returning "false" where we actually have a memmap
>>>> only makes it look like there is no memmap; but there is one, and
>>>> it's PG_reserved.
>>>
>>> I can see the documentation for pfn_valid() does not claim anything more
>>> than the presence of an memmap entry. But I wonder whether the confusion
>>> is wider-spread than just the DMA code. At a quick grep, try_ram_remap()
>>> assumes __va() can be used on pfn_valid(), though I suspect it relies on
>>> the calling function to check that the resource was RAM. The arm64
>>> kern_addr_valid() returns true based on pfn_valid() and kcore.c uses
>>> standard memcpy on it, which wouldn't work for I/O (should we change
>>> this check to pfn_is_map_memory() for arm64?).
>>
>> kern_addr_valid() checks that there is a direct map entry, and that the
>> mapped address has a valid mmap. (copied from x86-64)
> 
> It checks that there is a va->pa mapping, not necessarily in the linear
> map as it walks the page tables. So for some I/O range that happens to
> be mapped but which was in close proximity to RAM so that pfn_valid() is
> true, kern_addr_valid() would return true. I don't thin that was the
> intention.
> 
>> Would you expect to have a direct map for memory holes and similar (IOW,
>> !System RAM)?
> 
> No, but we with the generic pfn_valid(), it may return true for mapped
> MMIO (with different attributes than the direct map).

Ah, right. But can we actually run into that via kcore?

kcore builds the RAM list via walk_system_ram_range(), IOW the resource 
tree. And we end up calling kern_addr_valid() only on KCORE_RAM, 
KCORE_VMEMMAP and KCORE_TEXT.

Not saying that kern_addr_valid() shouldn't be improved.

-- 
Thanks,

David / dhildenb