package/kernel/mac80211/patches/316-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 @@ -4258,34 +4258,58 @@ 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         int idx = ar->wmi.num_mem_chunks;
  28 +       void *vaddr = NULL;
  29
  30 -       pool_size = num_units * round_up(unit_len, 4);
  31 +       if (ar->wmi.num_mem_chunks == ARRAY_SIZE(ar->wmi.mem_chunks))
  32 +               return -ENOMEM;
  33
  34 -       if (!pool_size)
  35 -               return -EINVAL;
  36 +       while (!vaddr && num_units) {
  37 +               pool_size = num_units * round_up(unit_len, 4);
  38 +               if (!pool_size)
  39 +                       return -EINVAL;
  40
  41 -       ar->wmi.mem_chunks[idx].vaddr = dma_alloc_coherent(ar->dev,
  42 -                                                          pool_size,
  43 -                                                          &paddr,
  44 -                                                          GFP_KERNEL);
  45 -       if (!ar->wmi.mem_chunks[idx].vaddr) {
  46 -               ath10k_warn(ar, "failed to allocate memory chunk\n");
  47 -               return -ENOMEM;
  48 +               vaddr = kzalloc(pool_size, GFP_KERNEL | __GFP_NOWARN);
  49 +               if (!vaddr)
  50 +                       num_units /= 2;
  51         }
  52
  53 -       memset(ar->wmi.mem_chunks[idx].vaddr, 0, pool_size);
  54 +       if (!num_units)
  55 +               return -ENOMEM;
  56
  57 +       paddr = dma_map_single(ar->dev, vaddr, pool_size, DMA_TO_DEVICE);
  58 +       if (dma_mapping_error(ar->dev, paddr)) {
  59 +               kfree(vaddr);
  60 +               return -ENOMEM;
  61 +       }
  62 +
  63 +       ar->wmi.mem_chunks[idx].vaddr = vaddr;
  64         ar->wmi.mem_chunks[idx].paddr = paddr;
  65         ar->wmi.mem_chunks[idx].len = pool_size;
  66         ar->wmi.mem_chunks[idx].req_id = req_id;
  67         ar->wmi.num_mem_chunks++;
  68
  69 +       return num_units;
  70 +}
  71 +
  72 +static int ath10k_wmi_alloc_host_mem(struct ath10k *ar, u32 req_id,
  73 +                                    u32 num_units, u32 unit_len)
  74 +{
  75 +       int ret;
  76 +
  77 +       while (num_units) {
  78 +               ret = ath10k_wmi_alloc_chunk(ar, req_id, num_units, unit_len);
  79 +               if (ret < 0)
  80 +                       return ret;
  81 +
  82 +               num_units -= ret;
  83 +       }
  84 +
  85         return 0;
  86  }
  87
  88 @@ -7616,10 +7640,11 @@ void ath10k_wmi_free_host_mem(struct ath
  89
  90         /* free the host memory chunks requested by firmware */
  91         for (i = 0; i < ar->wmi.num_mem_chunks; i++) {
  92 -               dma_free_coherent(ar->dev,
  93 -                                 ar->wmi.mem_chunks[i].len,
  94 -                                 ar->wmi.mem_chunks[i].vaddr,
  95 -                                 ar->wmi.mem_chunks[i].paddr);
  96 +               dma_unmap_single(ar->dev,
  97 +                                ar->wmi.mem_chunks[i].paddr,
  98 +                                ar->wmi.mem_chunks[i].len,
  99 +                                DMA_TO_DEVICE);
 100 +               kfree(ar->wmi.mem_chunks[i].vaddr);
 101         }
 102
 103         ar->wmi.num_mem_chunks = 0;