drm/i915: Stop the machine whilst capturing the GPU crash dump
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 12 Oct 2016 09:05:19 +0000 (10:05 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 12 Oct 2016 11:00:32 +0000 (12:00 +0100)
The error state is purposefully racy as we expect it to be called at any
time and so have avoided any locking whilst capturing the crash dump.
However, with multi-engine GPUs and multiple CPUs, those races can
manifest into OOPSes as we attempt to chase dangling pointers freed on
other CPUs. Under discussion are lots of ways to slow down normal
operation in order to protect the post-mortem error capture, but what it
we take the opposite approach and freeze the machine whilst the error
capture runs (note the GPU may still running, but as long as we don't
process any of the results the driver's bookkeeping will be static).

Note that by of itself, this is not a complete fix. It also depends on
the compiler barriers in list_add/list_del to prevent traversing the
lists into the void. We also depend that we only require state from
carefully controlled sources - i.e. all the state we require for
post-mortem debugging should be reachable from the request itself so
that we only have to worry about retrieving the request carefully. Once
we have the request, we know that all pointers from it are intact.

v2: Avoid drm_clflush_pages() inside stop_machine() as it may use
stop_machine() itself for its wbinvd fallback.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161012090522.367-3-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/Kconfig
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gpu_error.c

index 8844b99bd760a5c8a0317b9e4a1c06119b185c60..3eff42e4a4419c0df0c485f4ba73d61e9b64db05 100644 (file)
@@ -4,6 +4,7 @@ config DRM_I915
        depends on X86 && PCI
        select INTEL_GTT
        select INTERVAL_TREE
+       select STOP_MACHINE
        # we need shmfs for the swappable backing store, and in particular
        # the shmem_readpage() which depends upon tmpfs
        select SHMEM
index 380590b30bbf8008a6db5e326713eb6d41c0597a..4199e8aa436a701fbfac378cf4d43376f3abc435 100644 (file)
@@ -746,6 +746,8 @@ struct drm_i915_error_state {
        struct kref ref;
        struct timeval time;
 
+       struct drm_i915_private *i915;
+
        char error_msg[128];
        bool simulated;
        int iommu;
index c88c0d192a601e6e0004f6b173673a6cec0df496..159d6d7e0ceede9335c3d934f2545d483eb5f651 100644 (file)
@@ -28,6 +28,7 @@
  */
 
 #include <generated/utsrelease.h>
+#include <linux/stop_machine.h>
 #include "i915_drv.h"
 
 static const char *engine_str(int engine)
@@ -744,14 +745,12 @@ i915_error_object_create(struct drm_i915_private *dev_priv,
 
        dst->page_count = num_pages;
        while (num_pages--) {
-               unsigned long flags;
                void *d;
 
                d = kmalloc(PAGE_SIZE, GFP_ATOMIC);
                if (d == NULL)
                        goto unwind;
 
-               local_irq_save(flags);
                if (use_ggtt) {
                        void __iomem *s;
 
@@ -770,15 +769,10 @@ i915_error_object_create(struct drm_i915_private *dev_priv,
 
                        page = i915_gem_object_get_page(src, i);
 
-                       drm_clflush_pages(&page, 1);
-
                        s = kmap_atomic(page);
                        memcpy(d, s, PAGE_SIZE);
                        kunmap_atomic(s);
-
-                       drm_clflush_pages(&page, 1);
                }
-               local_irq_restore(flags);
 
                dst->pages[i++] = d;
                reloc_offset += PAGE_SIZE;
@@ -1447,6 +1441,31 @@ static void i915_capture_gen_state(struct drm_i915_private *dev_priv,
               sizeof(error->device_info));
 }
 
+static int capture(void *data)
+{
+       struct drm_i915_error_state *error = data;
+
+       /* Ensure that what we readback from memory matches what the GPU sees */
+       wbinvd();
+
+       i915_capture_gen_state(error->i915, error);
+       i915_capture_reg_state(error->i915, error);
+       i915_gem_record_fences(error->i915, error);
+       i915_gem_record_rings(error->i915, error);
+       i915_capture_active_buffers(error->i915, error);
+       i915_capture_pinned_buffers(error->i915, error);
+
+       do_gettimeofday(&error->time);
+
+       error->overlay = intel_overlay_capture_error_state(error->i915);
+       error->display = intel_display_capture_error_state(error->i915);
+
+       /* And make sure we don't leave trash in the CPU cache */
+       wbinvd();
+
+       return 0;
+}
+
 /**
  * i915_capture_error_state - capture an error record for later analysis
  * @dev: drm device
@@ -1478,18 +1497,9 @@ void i915_capture_error_state(struct drm_i915_private *dev_priv,
        }
 
        kref_init(&error->ref);
+       error->i915 = dev_priv;
 
-       i915_capture_gen_state(dev_priv, error);
-       i915_capture_reg_state(dev_priv, error);
-       i915_gem_record_fences(dev_priv, error);
-       i915_gem_record_rings(dev_priv, error);
-       i915_capture_active_buffers(dev_priv, error);
-       i915_capture_pinned_buffers(dev_priv, error);
-
-       do_gettimeofday(&error->time);
-
-       error->overlay = intel_overlay_capture_error_state(dev_priv);
-       error->display = intel_display_capture_error_state(dev_priv);
+       stop_machine(capture, error, NULL);
 
        i915_error_capture_msg(dev_priv, error, engine_mask, error_msg);
        DRM_INFO("%s\n", error->error_msg);