Bogus struct page layout on 32-bit

[Grygorii Strashko]

Hi Ilias, All,

On 10/04/2021 11:52, Ilias Apalodimas wrote:
> +CC Grygorii for the cpsw part as Ivan's email is not valid anymore
> 
> Thanks for catching this. Interesting indeed...
> 
> On Sat, 10 Apr 2021 at 09:22, Jesper Dangaard Brouer <brouer at redhat.com> wrote:
>>
>> On Sat, 10 Apr 2021 03:43:13 +0100
>> Matthew Wilcox <willy at infradead.org> wrote:
>>
>>> On Sat, Apr 10, 2021 at 06:45:35AM +0800, kernel test robot wrote:
>>>>>> include/linux/mm_types.h:274:1: error: static_assert failed due to requirement '__builtin_offsetof(struct page, lru) == __builtin_offsetof(struct folio, lru)' "offsetof(struct page, lru) == offsetof(struct folio, lru)"
>>>>     FOLIO_MATCH(lru, lru);
>>>>     include/linux/mm_types.h:272:2: note: expanded from macro 'FOLIO_MATCH'
>>>>             static_assert(offsetof(struct page, pg) == offsetof(struct folio, fl))
>>>
>>> Well, this is interesting.  pahole reports:
>>>
>>> struct page {
>>>          long unsigned int          flags;                /*     0     4 */
>>>          /* XXX 4 bytes hole, try to pack */
>>>          union {
>>>                  struct {
>>>                          struct list_head lru;            /*     8     8 */
>>> ...
>>> struct folio {
>>>          union {
>>>                  struct {
>>>                          long unsigned int flags;         /*     0     4 */
>>>                          struct list_head lru;            /*     4     8 */
>>>
>>> so this assert has absolutely done its job.
>>>
>>> But why has this assert triggered?  Why is struct page layout not what
>>> we thought it was?  Turns out it's the dma_addr added in 2019 by commit
>>> c25fff7171be ("mm: add dma_addr_t to struct page").  On this particular
>>> config, it's 64-bit, and ppc32 requires alignment to 64-bit.  So
>>> the whole union gets moved out by 4 bytes.
>>
>> Argh, good that you are catching this!
>>
>>> Unfortunately, we can't just fix this by putting an 'unsigned long pad'
>>> in front of it.  It still aligns the entire union to 8 bytes, and then
>>> it skips another 4 bytes after the pad.
>>>
>>> We can fix it like this ...
>>>
>>> +++ b/include/linux/mm_types.h
>>> @@ -96,11 +96,12 @@ struct page {
>>>                          unsigned long private;
>>>                  };
>>>                  struct {        /* page_pool used by netstack */
>>> +                       unsigned long _page_pool_pad;
>>
>> I'm fine with this pad.  Matteo is currently proposing[1] to add a 32-bit
>> value after @dma_addr, and he could use this area instead.
>>
>> [1] https://lore.kernel.org/netdev/20210409223801.104657-3-mcroce@linux.microsoft.com/
>>
>> When adding/changing this, we need to make sure that it doesn't overlap
>> member @index, because network stack use/check page_is_pfmemalloc().
>> As far as my calculations this is safe to add.  I always try to keep an
>> eye out for this, but I wonder if we could have a build check like yours.
>>
>>
>>>                          /**
>>>                           * @dma_addr: might require a 64-bit value even on
>>>                           * 32-bit architectures.
>>>                           */
>>> -                       dma_addr_t dma_addr;
>>> +                       dma_addr_t dma_addr __packed;
>>>                  };
>>>                  struct {        /* slab, slob and slub */
>>>                          union {
>>>
>>> but I don't know if GCC is smart enough to realise that dma_addr is now
>>> on an 8 byte boundary and it can use a normal instruction to access it,
>>> or whether it'll do something daft like use byte loads to access it.
>>>
>>> We could also do:
>>>
>>> +                       dma_addr_t dma_addr __packed __aligned(sizeof(void *));
>>>
>>> and I see pahole, at least sees this correctly:
>>>
>>>                  struct {
>>>                          long unsigned int _page_pool_pad; /*     4     4 */
>>>                          dma_addr_t dma_addr __attribute__((__aligned__(4))); /*     8     8 */
>>>                  } __attribute__((__packed__)) __attribute__((__aligned__(4)));
>>>
>>> This presumably affects any 32-bit architecture with a 64-bit phys_addr_t
>>> / dma_addr_t.  Advice, please?
>>
>> I'm not sure that the 32-bit behavior is with 64-bit (dma) addrs.
>>
>> I don't have any 32-bit boards with 64-bit DMA.  Cc. Ivan, wasn't your
>> board (572x ?) 32-bit with driver 'cpsw' this case (where Ivan added
>> XDP+page_pool) ?

Sry, for delayed reply.

The TI platforms am3/4/5 (cpsw) and Keystone 2 (netcp) can do only 32bit DMA even in case of LPAE (dma-ranges are used).
Originally, as I remember, CONFIG_ARCH_DMA_ADDR_T_64BIT has not been selected for the LPAE case
on TI platforms and the fact that it became set is the result of multi-paltform/allXXXconfig/DMA
optimizations and unification.
(just checked - not set in 4.14)

Probable commit 4965a68780c5 ("arch: define the ARCH_DMA_ADDR_T_64BIT config symbol in lib/Kconfig").

The TI drivers have been updated, finally to accept ARCH_DMA_ADDR_T_64BIT=y by using things like (__force u32)
for example.

Honestly, I've done sanity check of CPSW with LPAE=y (ARCH_DMA_ADDR_T_64BIT=y) very long time ago.

-- 
Best regards,
grygorii