[Intel-gfx] [PATCH 2/2] drm/i915: Disable page-faults around the fast pwrite/pread paths

[Chris Wilson]

These paths hold onto the struct mutex whilst accessing pages. In
order, to prevent a recursive dead-lock should we fault-in a GTT mapped
page we need to return -EFAULT and fallback to the slow path.

Lockdep has complained before about the potential dead-lock, but rvis is
the first application found to sufficiently abuse the API to trigger it.

Cursory performance regression testing on a 1GiB PineView system using
x11perf, cairo-perf-trace, glxgears and a few game benchmarks suggested
no large regressions with just a 2% slowdown for firefox. The caveat is
that this was an otherwise idle system and that for 32-bit systems
io_mapping_map_atomic_wc() already disabled page-faults.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38115
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem.c |   22 ++++++++++++++++++++--
 1 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 2fce620..ecb27fd 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -365,9 +365,15 @@ i915_gem_shmem_pread_fast(struct drm_device *dev,
 			return PTR_ERR(page);
 
 		vaddr = kmap_atomic(page);
+		/* We have to disable faulting here in case the user address
+		 * is really a GTT mapping and so we can not enter
+		 * i915_gem_fault() whilst already holding struct_mutex.
+		 */
+		pagefault_disable();
 		ret = __copy_to_user_inatomic(user_data,
 					      vaddr + page_offset,
 					      page_length);
+		pagefault_enable();
 		kunmap_atomic(vaddr);
 
 		mark_page_accessed(page);
@@ -593,8 +599,14 @@ fast_user_write(struct io_mapping *mapping,
 	unsigned long unwritten;
 
 	vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
+	/* We have to disable faulting here in case the user address
+	 * is really a GTT mapping and so we can not enter
+	 * i915_gem_fault() whilst already holding struct_mutex.
+	 */
+	pagefault_disable();
 	unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
 						      user_data, length);
+	pagefault_enable();
 	io_mapping_unmap_atomic(vaddr_atomic);
 	return unwritten;
 }
@@ -812,11 +824,17 @@ i915_gem_shmem_pwrite_fast(struct drm_device *dev,
 		if (IS_ERR(page))
 			return PTR_ERR(page);
 
-		vaddr = kmap_atomic(page, KM_USER0);
+		vaddr = kmap_atomic(page);
+		/* We have to disable faulting here in case the user address
+		 * is really a GTT mapping and so we can not enter
+		 * i915_gem_fault() whilst already holding struct_mutex.
+		 */
+		pagefault_disable();
 		ret = __copy_from_user_inatomic(vaddr + page_offset,
 						user_data,
 						page_length);
-		kunmap_atomic(vaddr, KM_USER0);
+		pagefault_enable();
+		kunmap_atomic(vaddr);
 
 		set_page_dirty(page);
 		mark_page_accessed(page);
-- 
1.7.5.4