<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">Am 19.09.2017 um 13:20 schrieb Tom St
      Denis:<br>
    </div>
    <blockquote type="cite"
      cite="mid:20170919112057.24377-1-tom.stdenis@amd.com">
      <pre wrap="">Signed-off-by: Tom St Denis <a class="moz-txt-link-rfc2396E" href="mailto:tom.stdenis@amd.com"><tom.stdenis@amd.com></a>

(v2): Add domain to iova debugfs
(v3): Add true read/write methods to access system memory of pages
      mapped to the device
(v4): Move get_domain call out of loop and return on error
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 114 ++++++++++++++++++++++++++++++++
 1 file changed, 114 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 50d20903de4f..71a2fdc91a85 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -43,6 +43,7 @@
 #include <linux/swap.h>
 #include <linux/pagemap.h>
 #include <linux/debugfs.h>
+#include <linux/iommu.h>
 #include "amdgpu.h"
 #include "amdgpu_trace.h"
 #include "bif/bif_4_1_d.h"
@@ -1810,6 +1811,118 @@ static const struct file_operations amdgpu_ttm_gtt_fops = {
 #endif
 
 
+static void *transform_page(uint64_t phys)
+{
+       if (PageHighMem(pfn_to_page(PFN_DOWN(phys))))
+               return kmap(pfn_to_page(PFN_DOWN(phys)));
+       else
+               return __va(phys);
+}
+
+static void untransform_page(uint64_t phys)
+{
+       if (PageHighMem(pfn_to_page(PFN_DOWN(phys))))
+               return kunmap(pfn_to_page(PFN_DOWN(phys)));
+}</pre>
    </blockquote>
    <br>
    I mentioned that in the previous patch as well, why not using
    kmap()/kunmap() here directly?<br>
    <br>
    See kmap() for x86 for example:<br>
    <blockquote type="cite">
      <pre><span class="kt">void</span> <span class="o">*</span><span class="nf"><a href="http://elixir.free-electrons.com/linux/v4.2/ident/kmap">kmap</a></span><span class="p">(</span><span class="k">struct</span> <span class="n">page</span> <span class="o">*</span><span class="n">page</span><span class="p">)</span>
<span class="p">{</span>
        <span class="n"><a href="http://elixir.free-electrons.com/linux/v4.2/ident/might_sleep">might_sleep</a></span><span class="p">();</span>
        <span class="k">if</span> <span class="p">(</span><span class="o">!</span><span class="n"><a href="http://elixir.free-electrons.com/linux/v4.2/ident/PageHighMem">PageHighMem</a></span><span class="p">(</span><span class="n">page</span><span class="p">))</span>
                <span class="k">return</span> <span class="n"><a href="http://elixir.free-electrons.com/linux/v4.2/ident/page_address">page_address</a></span><span class="p">(</span><span class="n">page</span><span class="p">);</span>
        <span class="k">return</span> <span class="n"><a href="http://elixir.free-electrons.com/linux/v4.2/ident/kmap_high">kmap_high</a></span><span class="p">(</span><span class="n">page</span><span class="p">);</span>
<span class="p">}</span></pre>
    </blockquote>
    <br>
    Both normal as well as highmem are correctly handled here.<br>
    <br>
    Regards,<br>
    Christian.<br>
    <br>
    <blockquote type="cite"
      cite="mid:20170919112057.24377-1-tom.stdenis@amd.com">
      <pre wrap="">
+
+static ssize_t amdgpu_iova_to_phys_read(struct file *f, char __user *buf,
+                                  size_t size, loff_t *pos)
+{
+       struct amdgpu_device *adev = file_inode(f)->i_private;
+       ssize_t result, n;
+       int r;
+       uint64_t phys;
+       void *ptr;
+       struct iommu_domain *dom;
+
+       dom = iommu_get_domain_for_dev(adev->dev);
+       if (!dom)
+               return -EFAULT;
+
+       result = 0;
+       while (size) {
+               // get physical address and map
+               phys = iommu_iova_to_phys(dom, *pos);
+
+               // copy upto one page
+               if (size > PAGE_SIZE)
+                       n = PAGE_SIZE;
+               else
+                       n = size;
+
+               // to end of the page
+               if (((*pos & (PAGE_SIZE - 1)) + n) >= PAGE_SIZE)
+                       n = PAGE_SIZE - (*pos & (PAGE_SIZE - 1));
+
+               ptr = transform_page(phys);
+               if (!ptr)
+                       return -EFAULT;
+
+               r = copy_to_user(buf, ptr, n);
+               untransform_page(phys);
+               if (r)
+                       return -EFAULT;
+
+               *pos += n;
+               size -= n;
+               result += n;
+       }
+
+       return result;
+}
+
+static ssize_t amdgpu_iova_to_phys_write(struct file *f, const char __user *buf,
+                                  size_t size, loff_t *pos)
+{
+       struct amdgpu_device *adev = file_inode(f)->i_private;
+       ssize_t result, n;
+       int r;
+       uint64_t phys;
+       void *ptr;
+       struct iommu_domain *dom;
+
+       dom = iommu_get_domain_for_dev(adev->dev);
+       if (!dom)
+               return -EFAULT;
+
+       result = 0;
+       while (size) {
+               // get physical address and map
+               phys = iommu_iova_to_phys(dom, *pos);
+
+               // copy upto one page
+               if (size > PAGE_SIZE)
+                       n = PAGE_SIZE;
+               else
+                       n = size;
+
+               // to end of the page
+               if (((*pos & (PAGE_SIZE - 1)) + n) >= PAGE_SIZE)
+                       n = PAGE_SIZE - (*pos & (PAGE_SIZE - 1));
+
+               ptr = transform_page(phys);
+               if (!ptr)
+                       return -EFAULT;
+
+               r = copy_from_user(ptr, buf, n);
+               untransform_page(phys);
+               if (r)
+                       return -EFAULT;
+
+               *pos += n;
+               size -= n;
+               result += n;
+       }
+
+       return result;
+}
+
+static const struct file_operations amdgpu_ttm_iova_fops = {
+       .owner = THIS_MODULE,
+       .read = amdgpu_iova_to_phys_read,
+       .write = amdgpu_iova_to_phys_write,
+       .llseek = default_llseek
+};
 
 static const struct {
        char *name;
@@ -1820,6 +1933,7 @@ static const struct {
 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
        { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
 #endif
+       { "amdgpu_iova", &amdgpu_ttm_iova_fops, TTM_PL_SYSTEM },
 };
 
 #endif
</pre>
    </blockquote>
    <p><br>
    </p>
  </body>
</html>