On Tue, Jul 13, 2021 at 2:25 AM Yunsheng Lin [off-list ref] wrote:

As suggested by Alexander, "A DMA mapping should be page
aligned anyway so the lower 12 bits would be reserved 0",
so it might make more sense to repurpose the lower 12 bits
of the dma address to store the bias for frag page support
in page pool for 32 bit systems with 64 bit dma, which
should be rare those days.

For normal system, the dma_addr[1] in 'struct page' is not
used, so we can reuse the dma_addr[1] for storing bias.

The PAGE_POOP_USE_DMA_ADDR_1 macro is used to decide where
to store the bias, as the "sizeof(dma_addr_t) > sizeof(
unsigned long)" is false for normal system, so hopefully the
compiler will optimize out the unused code for those system.

I assume the name is a typo and you meant PAGE_POOL_USE_DMA_ADDR_1?

Alternatively we could change things a bit and rename things so we
have the MSB of dma_addr where dma_addr[1] is and we rename
dma_addr[0] to pp_frag_count we could have it also contain the lower
bits and handle it like so:
    ret = page->dma_addr;
    if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT) {
        ret <<= 32;
        ret |= atomic_long_read(&page->pp_frag_count) & PAGE_MASK;
    }

+       } else {
+               ret = page->dma_addr[0];
+       }
+
        return ret;
 }

 static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
 {
        page->dma_addr[0] = addr;
-       if (sizeof(dma_addr_t) > sizeof(unsigned long))
+       if (PAGE_POOP_USE_DMA_ADDR_1)
                page->dma_addr[1] = upper_32_bits(addr);

So assuming similar logic to above we could do something like:
    if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT) {
        atomic_long_set(&page->pp_frag_count, addr & PAGE_MASK);
        addr >>= 32;
    }
    pp->dma_addr = addr;

 }

+static inline int page_pool_atomic_sub_bias_return(struct page *page, int nr)
+{
+       int bias;
+
+       if (PAGE_POOP_USE_DMA_ADDR_1) {
+               unsigned long *bias_ptr = &page->dma_addr[0];
+               unsigned long old_bias = READ_ONCE(*bias_ptr);
+               unsigned long new_bias;
+
+               do {
+                       bias = (int)(old_bias & ~PAGE_MASK);
+
+                       /* Warn when page_pool_dev_alloc_pages() is called
+                        * with PP_FLAG_PAGE_FRAG flag in driver.
+                        */
+                       WARN_ON(!bias);
+
+                       /* already the last user */
+                       if (!(bias - nr))
+                               return 0;
+
+                       new_bias = old_bias - nr;
+               } while (!try_cmpxchg(bias_ptr, &old_bias, new_bias));
+
+               WARN_ON((new_bias & PAGE_MASK) != (old_bias & PAGE_MASK));
+
+               bias = new_bias & ~PAGE_MASK;
+       } else {
+               atomic_t *v = (atomic_t *)&page->dma_addr[1];

The problem with casting like this is that it makes assumptions about
byte ordering in the case that atomic_t is a 32b value and dma_addr is
a long value.

+
+               if (atomic_read(v) == nr)
+                       return 0;
+
+               bias = atomic_sub_return(nr, v);
+               WARN_ON(bias < 0);
+       }

Rather than have 2 versions of this function it might work better to
just use the atomic_long version of these functions instead. Then you
shouldn't need to have two versions of the code.

You could just modify the block on the end to check for new_frag_count
vs old_frag_count if PAGE_POOL_USE_PP_FRAG_COUNT is true, or
new_frag_count < 0 if false.

+
+       return bias;
+}
+
+static inline void page_pool_set_bias(struct page *page, int bias)
+{
+       if (PAGE_POOP_USE_DMA_ADDR_1) {
+               unsigned long dma_addr_0 = READ_ONCE(page->dma_addr[0]);
+
+               dma_addr_0 &= PAGE_MASK;
+               dma_addr_0 |= bias;
+
+               WRITE_ONCE(page->dma_addr[0], dma_addr_0);
+       } else {
+               atomic_t *v = (atomic_t *)&page->dma_addr[1];
+
+               atomic_set(v, bias);
+       }

Similarly here you could just update bias to include the dma_addr in
the if case, and then use atomic_long_set for both cases.