[RFC,125/162] drm/i915/lmem: Limit block size to 4G

Commit Message

Matthew Auld Nov. 27, 2020, 12:06 p.m. UTC

From: Venkata Sandeep Dhanalakota <[e-mail]>

when allocating pages to lmem object of size 4G or greater
we allocate memory blocks from buddy system. In this scenario
buddy sytem can allocate blocks that can have size >= 4G and
these blocks require >32b to represent block size with these
blocks we run into an issue with sg list construction because
sg->length field is only 32b wide.

Hence limit the max allowed block size to less than 4G.

Cc: Niranjana Vishwanathapura <[e-mail]>
Cc: Matthew Auld <[e-mail]>
Cc: CQ Tang <[e-mail]>
Signed-off-by: Venkata Sandeep Dhanalakota <[e-mail]>
---
 drivers/gpu/drm/i915/intel_memory_region.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

Comments

Chris Wilson Nov. 27, 2020, 2:02 p.m. UTC | #1

Quoting Matthew Auld (2020-11-27 12:06:41)
> From: Venkata Sandeep Dhanalakota <[e-mail]>
> 
> when allocating pages to lmem object of size 4G or greater
> we allocate memory blocks from buddy system.

Any lmem object is from the buddy system.

> In this scenario
> buddy sytem can allocate blocks that can have size >= 4G and
> these blocks require >32b to represent block size with these
> blocks we run into an issue with sg list construction because
> sg->length field is only 32b wide.

Just say the when using scatterlist, the maximum segment size is 4G. In
fact, we can ask sg what the backend maximum is, and use that as our max
order.

The only question is whether this merits a flag, or we just assume that
the buddy allocator is only used for objects and so always presented via
sg?
-Chris

Patch

diff --git a/drivers/gpu/drm/i915/intel_memory_region.c b/drivers/gpu/drm/i915/intel_memory_region.c
index 554fdd7735a8..371cd88ff6d8 100644
--- a/drivers/gpu/drm/i915/intel_memory_region.c
+++ b/drivers/gpu/drm/i915/intel_memory_region.c
@@ -101,6 +101,7 @@  __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem,
 				      struct list_head *blocks)
 {
 	unsigned int min_order = 0;
+	unsigned int max_order;
 	unsigned long n_pages;
 
 	GEM_BUG_ON(!IS_ALIGNED(size, mem->mm.chunk_size));
@@ -121,6 +122,16 @@  __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem,
 
 	n_pages = size >> ilog2(mem->mm.chunk_size);
 
+	/*
+	 * When allocating pages for an lmem object of size > 4G
+	 * the memory blocks allocated from buddy system could be
+	 * from sizes greater than 4G requiring > 32b to represent
+	 * block size. But those blocks cannot be used in sg list
+	 * construction(in caller) as sg->length is only 32b wide.
+	 * Hence limiting the block size to 4G.
+	 */
+	max_order = (ilog2(SZ_4G) - 1) - ilog2(mem->mm.chunk_size);
+
 	mutex_lock(&mem->mm_lock);
 
 	do {
@@ -128,7 +139,7 @@  __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem,
 		unsigned int order;
 		bool retry = true;
 retry:
-		order = fls(n_pages) - 1;
+		order = min_t(u32, (fls(n_pages) - 1), max_order);
 		GEM_BUG_ON(order > mem->mm.max_order);
 		GEM_BUG_ON(order < min_order);