package/kernel/mac80211/patches/303-ath10k-do-not-use-coherent-memory-for-allocated-devi.patch

   1 From: Felix Fietkau <nbd@openwrt.org>
   2 Date: Sun, 22 Nov 2015 14:03:40 +0100
   3 Subject: [PATCH] ath10k: do not use coherent memory for allocated device
   4  memory chunks
   5
   6 Coherent memory is more expensive to allocate (and constrained on some
   7 architectures where it has to be pre-allocated). It is also completely
   8 unnecessary, since the host has no reason to even access these allocated
   9 memory spaces
  10
  11 Signed-off-by: Felix Fietkau <nbd@openwrt.org>
  12 ---
  13
  14 --- a/drivers/net/wireless/ath/ath10k/wmi.c
  15 +++ b/drivers/net/wireless/ath/ath10k/wmi.c
  16 @@ -4312,34 +4312,58 @@ void ath10k_wmi_event_vdev_resume_req(st
  17         ath10k_dbg(ar, ATH10K_DBG_WMI, "WMI_VDEV_RESUME_REQ_EVENTID\n");
  18  }
  19
  20 -static int ath10k_wmi_alloc_host_mem(struct ath10k *ar, u32 req_id,
  21 -                                    u32 num_units, u32 unit_len)
  22 +static int ath10k_wmi_alloc_chunk(struct ath10k *ar, u32 req_id,
  23 +                                 u32 num_units, u32 unit_len)
  24  {
  25         dma_addr_t paddr;
  26 -       u32 pool_size;
  27 +       u32 pool_size = 0;
  28         int idx = ar->wmi.num_mem_chunks;
  29 +       void *vaddr = NULL;
  30
  31 -       pool_size = num_units * round_up(unit_len, 4);
  32 +       if (ar->wmi.num_mem_chunks == ARRAY_SIZE(ar->wmi.mem_chunks))
  33 +               return -ENOMEM;
  34
  35 -       if (!pool_size)
  36 -               return -EINVAL;
  37 +       while (!vaddr && num_units) {
  38 +               pool_size = num_units * round_up(unit_len, 4);
  39 +               if (!pool_size)
  40 +                       return -EINVAL;
  41
  42 -       ar->wmi.mem_chunks[idx].vaddr = dma_alloc_coherent(ar->dev,
  43 -                                                          pool_size,
  44 -                                                          &paddr,
  45 -                                                          GFP_KERNEL);
  46 -       if (!ar->wmi.mem_chunks[idx].vaddr) {
  47 -               ath10k_warn(ar, "failed to allocate memory chunk\n");
  48 -               return -ENOMEM;
  49 +               vaddr = kzalloc(pool_size, GFP_KERNEL | __GFP_NOWARN);
  50 +               if (!vaddr)
  51 +                       num_units /= 2;
  52         }
  53
  54 -       memset(ar->wmi.mem_chunks[idx].vaddr, 0, pool_size);
  55 +       if (!num_units)
  56 +               return -ENOMEM;
  57 +
  58 +       paddr = dma_map_single(ar->dev, vaddr, pool_size, DMA_TO_DEVICE);
  59 +       if (dma_mapping_error(ar->dev, paddr)) {
  60 +               kfree(vaddr);
  61 +               return -ENOMEM;
  62 +       }
  63
  64 +       ar->wmi.mem_chunks[idx].vaddr = vaddr;
  65         ar->wmi.mem_chunks[idx].paddr = paddr;
  66         ar->wmi.mem_chunks[idx].len = pool_size;
  67         ar->wmi.mem_chunks[idx].req_id = req_id;
  68         ar->wmi.num_mem_chunks++;
  69
  70 +       return num_units;
  71 +}
  72 +
  73 +static int ath10k_wmi_alloc_host_mem(struct ath10k *ar, u32 req_id,
  74 +                                    u32 num_units, u32 unit_len)
  75 +{
  76 +       int ret;
  77 +
  78 +       while (num_units) {
  79 +               ret = ath10k_wmi_alloc_chunk(ar, req_id, num_units, unit_len);
  80 +               if (ret < 0)
  81 +                       return ret;
  82 +
  83 +               num_units -= ret;
  84 +       }
  85 +
  86         return 0;
  87  }
  88
  89 @@ -7717,10 +7741,11 @@ void ath10k_wmi_free_host_mem(struct ath
  90
  91         /* free the host memory chunks requested by firmware */
  92         for (i = 0; i < ar->wmi.num_mem_chunks; i++) {
  93 -               dma_free_coherent(ar->dev,
  94 -                                 ar->wmi.mem_chunks[i].len,
  95 -                                 ar->wmi.mem_chunks[i].vaddr,
  96 -                                 ar->wmi.mem_chunks[i].paddr);
  97 +               dma_unmap_single(ar->dev,
  98 +                                ar->wmi.mem_chunks[i].paddr,
  99 +                                ar->wmi.mem_chunks[i].len,
 100 +                                DMA_TO_DEVICE);
 101 +               kfree(ar->wmi.mem_chunks[i].vaddr);
 102         }
 103
 104         ar->wmi.num_mem_chunks = 0;