From d21375bd0e3b5c9fb1bfd341abb294349ca2507f Mon Sep 17 00:00:00 2001
From: Mitchel Humpherys <mitchelh@codeaurora.org>
Date: Thu, 31 Jan 2013 10:30:40 -0800
Subject: [PATCH] gpu: ion: replace __GFP_ZERO with manual zero'ing
As a performance optimization, omit the __GFP_ZERO flag when
allocating individual pages and, instead, zero out all of the pages in
one fell swoop.
CRs-Fixed: 449035
Change-Id: Ieb9a895d8792727a8a40b1e27cb1bbeae098f581
Signed-off-by: Mitchel Humpherys <mitchelh@codeaurora.org>
---
drivers/gpu/ion/ion_iommu_heap.c | 42 +++++++++++++++++++++++++++++---
1 file changed, 39 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/ion/ion_iommu_heap.c b/drivers/gpu/ion/ion_iommu_heap.c
index 0e47d545b9d2..cb8fbed15a97 100644
--- a/drivers/gpu/ion/ion_iommu_heap.c
+++ b/drivers/gpu/ion/ion_iommu_heap.c
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2011-2012, The Linux Foundation. All rights reserved.
+ * Copyright (c) 2011-2013, The Linux Foundation. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 and
@@ -38,6 +38,8 @@ struct ion_iommu_priv_data {
unsigned long size;
};
+#define MAX_VMAP_RETRIES 10
+
static int ion_iommu_heap_allocate(struct ion_heap *heap,
struct ion_buffer *buffer,
unsigned long size, unsigned long align,
@@ -49,7 +51,9 @@ static int ion_iommu_heap_allocate(struct ion_heap *heap,
if (msm_use_iommu()) {
struct scatterlist *sg;
struct sg_table *table;
- unsigned int i;
+ unsigned int i, j, k;
+ void *ptr = NULL;
+ unsigned int npages_to_vmap, total_pages;
data = kmalloc(sizeof(*data), GFP_KERNEL);
if (!data)
@@ -77,7 +81,7 @@ static int ion_iommu_heap_allocate(struct ion_heap *heap,
for_each_sg(table->sgl, sg, table->nents, i) {
data->pages[i] = alloc_page(
- GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
+ GFP_KERNEL | __GFP_HIGHMEM);
if (!data->pages[i])
goto err3;
@@ -85,6 +89,38 @@ static int ion_iommu_heap_allocate(struct ion_heap *heap,
sg_dma_address(sg) = sg_phys(sg);
}
+ /*
+ * As an optimization, we omit __GFP_ZERO from
+ * alloc_page above and manually zero out all of the
+ * pages in one fell swoop here. To safeguard against
+ * insufficient vmalloc space, we only vmap
+ * `npages_to_vmap' at a time, starting with a
+ * conservative estimate of 1/8 of the total number of
+ * vmalloc pages available.
+ */
+ npages_to_vmap = ((VMALLOC_END - VMALLOC_START)/8)
+ >> PAGE_SHIFT;
+ total_pages = data->nrpages;
+ for (j = 0; j < total_pages; j += npages_to_vmap) {
+ npages_to_vmap = min(npages_to_vmap, total_pages - j);
+ for (k = 0; k < MAX_VMAP_RETRIES && npages_to_vmap;
+ ++k) {
+ ptr = vmap(&data->pages[j], npages_to_vmap,
+ VM_IOREMAP, pgprot_kernel);
+ if (ptr)
+ break;
+ else
+ npages_to_vmap >>= 1;
+ }
+ if (!ptr) {
+ pr_err("Couldn't vmap the pages for zeroing\n");
+ ret = -ENOMEM;
+ goto err3;
+ }
+ memset(ptr, 0, npages_to_vmap * PAGE_SIZE);
+ vunmap(ptr);
+ }
+
if (!ION_IS_CACHED(flags))
dma_sync_sg_for_device(NULL, table->sgl, table->nents,
DMA_BIDIRECTIONAL);
--
GitLab