git.openwrt.org Git - openwrt/staging/stintel.git/blob

   1 From: Yunsheng Lin <linyunsheng@huawei.com>
   2 Date: Fri, 13 Oct 2023 14:48:21 +0800
   3 Subject: [PATCH] page_pool: fragment API support for 32-bit arch with 64-bit
   4  DMA
   5
   6 Currently page_pool_alloc_frag() is not supported in 32-bit
   7 arch with 64-bit DMA because of the overlap issue between
   8 pp_frag_count and dma_addr_upper in 'struct page' for those
   9 arches, which seems to be quite common, see [1], which means
  10 driver may need to handle it when using fragment API.
  11
  12 It is assumed that the combination of the above arch with an
  13 address space >16TB does not exist, as all those arches have
  14 64b equivalent, it seems logical to use the 64b version for a
  15 system with a large address space. It is also assumed that dma
  16 address is page aligned when we are dma mapping a page aligned
  17 buffer, see [2].
  18
  19 That means we're storing 12 bits of 0 at the lower end for a
  20 dma address, we can reuse those bits for the above arches to
  21 support 32b+12b, which is 16TB of memory.
  22
  23 If we make a wrong assumption, a warning is emitted so that
  24 user can report to us.
  25
  26 1. https://lore.kernel.org/all/20211117075652.58299-1-linyunsheng@huawei.com/
  27 2. https://lore.kernel.org/all/20230818145145.4b357c89@kernel.org/
  28
  29 Tested-by: Alexander Lobakin <aleksander.lobakin@intel.com>
  30 Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
  31 CC: Lorenzo Bianconi <lorenzo@kernel.org>
  32 CC: Alexander Duyck <alexander.duyck@gmail.com>
  33 CC: Liang Chen <liangchen.linux@gmail.com>
  34 CC: Guillaume Tucker <guillaume.tucker@collabora.com>
  35 CC: Matthew Wilcox <willy@infradead.org>
  36 CC: Linux-MM <linux-mm@kvack.org>
  37 Link: https://lore.kernel.org/r/20231013064827.61135-2-linyunsheng@huawei.com
  38 Signed-off-by: Jakub Kicinski <kuba@kernel.org>
  39 ---
  40
  41 --- a/include/linux/mm_types.h
  42 +++ b/include/linux/mm_types.h
  43 @@ -125,18 +125,7 @@ struct page {
  44                         struct page_pool *pp;
  45                         unsigned long _pp_mapping_pad;
  46                         unsigned long dma_addr;
  47 -                       union {
  48 -                               /**
  49 -                                * dma_addr_upper: might require a 64-bit
  50 -                                * value on 32-bit architectures.
  51 -                                */
  52 -                               unsigned long dma_addr_upper;
  53 -                               /**
  54 -                                * For frag page support, not supported in
  55 -                                * 32-bit architectures with 64-bit DMA.
  56 -                                */
  57 -                               atomic_long_t pp_frag_count;
  58 -                       };
  59 +                       atomic_long_t pp_frag_count;
  60                 };
  61                 struct {        /* Tail pages of compound page */
  62                         unsigned long compound_head;    /* Bit zero is set */
  63 --- a/include/net/page_pool/helpers.h
  64 +++ b/include/net/page_pool/helpers.h
  65 @@ -197,7 +197,7 @@ static inline void page_pool_recycle_dir
  66         page_pool_put_full_page(pool, page, true);
  67  }
  68
  69 -#define PAGE_POOL_DMA_USE_PP_FRAG_COUNT        \
  70 +#define PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA    \
  71                 (sizeof(dma_addr_t) > sizeof(unsigned long))
  72
  73  /**
  74 @@ -211,17 +211,25 @@ static inline dma_addr_t page_pool_get_d
  75  {
  76         dma_addr_t ret = page->dma_addr;
  77
  78 -       if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
  79 -               ret |= (dma_addr_t)page->dma_addr_upper << 16 << 16;
  80 +       if (PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA)
  81 +               ret <<= PAGE_SHIFT;
  82
  83         return ret;
  84  }
  85
  86 -static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
  87 +static inline bool page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
  88  {
  89 +       if (PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA) {
  90 +               page->dma_addr = addr >> PAGE_SHIFT;
  91 +
  92 +               /* We assume page alignment to shave off bottom bits,
  93 +                * if this "compression" doesn't work we need to drop.
  94 +                */
  95 +               return addr != (dma_addr_t)page->dma_addr << PAGE_SHIFT;
  96 +       }
  97 +
  98         page->dma_addr = addr;
  99 -       if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
 100 -               page->dma_addr_upper = upper_32_bits(addr);
 101 +       return false;
 102  }
 103
 104  static inline bool page_pool_put(struct page_pool *pool)
 105 --- a/net/core/page_pool.c
 106 +++ b/net/core/page_pool.c
 107 @@ -211,10 +211,6 @@ static int page_pool_init(struct page_po
 108                  */
 109         }
 110
 111 -       if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 112 -           pool->p.flags & PP_FLAG_PAGE_FRAG)
 113 -               return -EINVAL;
 114 -
 115  #ifdef CONFIG_PAGE_POOL_STATS
 116         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 117         if (!pool->recycle_stats)
 118 @@ -363,12 +359,20 @@ static bool page_pool_dma_map(struct pag
 119         if (dma_mapping_error(pool->p.dev, dma))
 120                 return false;
 121
 122 -       page_pool_set_dma_addr(page, dma);
 123 +       if (page_pool_set_dma_addr(page, dma))
 124 +               goto unmap_failed;
 125
 126         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 127                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 128
 129         return true;
 130 +
 131 +unmap_failed:
 132 +       WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
 133 +       dma_unmap_page_attrs(pool->p.dev, dma,
 134 +                            PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 135 +                            DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 136 +       return false;
 137  }
 138
 139  static void page_pool_set_pp_info(struct page_pool *pool,