summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorThomas Hellström <thomas.hellstrom@linux.intel.com>2021-06-02 10:38:11 +0200
committerThomas Hellström <thomas.hellstrom@linux.intel.com>2021-06-07 16:07:08 +0200
commitb7e32bef4ae5f9149276203564b7911fac466588 (patch)
treeb8752329b4d43041fe4837a322cf684513bd697d
parent3bf3710e3718a5aebdf465343bc1125b6e8cca96 (diff)
drm: Add a prefetching memcpy_from_wc
Reading out of write-combining mapped memory is typically very slow since the CPU doesn't prefetch. However some archs have special instructions to do this. So add a best-effort memcpy_from_wc taking dma-buf-map pointer arguments that attempts to use a fast prefetching memcpy and otherwise falls back to ordinary memcopies, taking the iomem tagging into account. The code is largely copied from i915_memcpy_from_wc. Cc: Daniel Vetter <daniel@ffwll.ch> Cc: Christian König <christian.koenig@amd.com> Suggested-by: Daniel Vetter <daniel@ffwll.ch> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com> Acked-by: Christian König <christian.koenig@amd.com> Acked-by: Daniel Vetter <daniel@ffwll.ch> Link: https://lore.kernel.org/r/20210602083818.241793-5-thomas.hellstrom@linux.intel.com Link: https://patchwork.freedesktop.org/patch/msgid/20210602083818.241793-5-thomas.hellstrom@linux.intel.com
-rw-r--r--Documentation/gpu/drm-mm.rst4
-rw-r--r--drivers/gpu/drm/drm_cache.c148
-rw-r--r--drivers/gpu/drm/drm_drv.c2
-rw-r--r--include/drm/drm_cache.h7
4 files changed, 159 insertions, 2 deletions
diff --git a/Documentation/gpu/drm-mm.rst b/Documentation/gpu/drm-mm.rst
index 21be6deadc12..d5a73fa2c9ef 100644
--- a/Documentation/gpu/drm-mm.rst
+++ b/Documentation/gpu/drm-mm.rst
@@ -469,8 +469,8 @@ DRM MM Range Allocator Function References
.. kernel-doc:: drivers/gpu/drm/drm_mm.c
:export:
-DRM Cache Handling
-==================
+DRM Cache Handling and Fast WC memcpy()
+=======================================
.. kernel-doc:: drivers/gpu/drm/drm_cache.c
:export:
diff --git a/drivers/gpu/drm/drm_cache.c b/drivers/gpu/drm/drm_cache.c
index 79a50ef1250f..546599f19a93 100644
--- a/drivers/gpu/drm/drm_cache.c
+++ b/drivers/gpu/drm/drm_cache.c
@@ -28,6 +28,7 @@
* Authors: Thomas Hellström <thomas-at-tungstengraphics-dot-com>
*/
+#include <linux/dma-buf-map.h>
#include <linux/export.h>
#include <linux/highmem.h>
#include <linux/mem_encrypt.h>
@@ -35,6 +36,9 @@
#include <drm/drm_cache.h>
+/* A small bounce buffer that fits on the stack. */
+#define MEMCPY_BOUNCE_SIZE 128
+
#if defined(CONFIG_X86)
#include <asm/smp.h>
@@ -209,3 +213,147 @@ bool drm_need_swiotlb(int dma_bits)
return max_iomem > ((u64)1 << dma_bits);
}
EXPORT_SYMBOL(drm_need_swiotlb);
+
+static void memcpy_fallback(struct dma_buf_map *dst,
+ const struct dma_buf_map *src,
+ unsigned long len)
+{
+ if (!dst->is_iomem && !src->is_iomem) {
+ memcpy(dst->vaddr, src->vaddr, len);
+ } else if (!src->is_iomem) {
+ dma_buf_map_memcpy_to(dst, src->vaddr, len);
+ } else if (!dst->is_iomem) {
+ memcpy_fromio(dst->vaddr, src->vaddr_iomem, len);
+ } else {
+ /*
+ * Bounce size is not performance tuned, but using a
+ * bounce buffer like this is significantly faster than
+ * resorting to ioreadxx() + iowritexx().
+ */
+ char bounce[MEMCPY_BOUNCE_SIZE];
+ void __iomem *_src = src->vaddr_iomem;
+ void __iomem *_dst = dst->vaddr_iomem;
+
+ while (len >= MEMCPY_BOUNCE_SIZE) {
+ memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
+ memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
+ _src += MEMCPY_BOUNCE_SIZE;
+ _dst += MEMCPY_BOUNCE_SIZE;
+ len -= MEMCPY_BOUNCE_SIZE;
+ }
+ if (len) {
+ memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
+ memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
+ }
+ }
+}
+
+#ifdef CONFIG_X86
+
+static DEFINE_STATIC_KEY_FALSE(has_movntdqa);
+
+static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
+{
+ kernel_fpu_begin();
+
+ while (len >= 4) {
+ asm("movntdqa (%0), %%xmm0\n"
+ "movntdqa 16(%0), %%xmm1\n"
+ "movntdqa 32(%0), %%xmm2\n"
+ "movntdqa 48(%0), %%xmm3\n"
+ "movaps %%xmm0, (%1)\n"
+ "movaps %%xmm1, 16(%1)\n"
+ "movaps %%xmm2, 32(%1)\n"
+ "movaps %%xmm3, 48(%1)\n"
+ :: "r" (src), "r" (dst) : "memory");
+ src += 64;
+ dst += 64;
+ len -= 4;
+ }
+ while (len--) {
+ asm("movntdqa (%0), %%xmm0\n"
+ "movaps %%xmm0, (%1)\n"
+ :: "r" (src), "r" (dst) : "memory");
+ src += 16;
+ dst += 16;
+ }
+
+ kernel_fpu_end();
+}
+
+/*
+ * __drm_memcpy_from_wc copies @len bytes from @src to @dst using
+ * non-temporal instructions where available. Note that all arguments
+ * (@src, @dst) must be aligned to 16 bytes and @len must be a multiple
+ * of 16.
+ */
+static void __drm_memcpy_from_wc(void *dst, const void *src, unsigned long len)
+{
+ if (unlikely(((unsigned long)dst | (unsigned long)src | len) & 15))
+ memcpy(dst, src, len);
+ else if (likely(len))
+ __memcpy_ntdqa(dst, src, len >> 4);
+}
+
+/**
+ * drm_memcpy_from_wc - Perform the fastest available memcpy from a source
+ * that may be WC.
+ * @dst: The destination pointer
+ * @src: The source pointer
+ * @len: The size of the area o transfer in bytes
+ *
+ * Tries an arch optimized memcpy for prefetching reading out of a WC region,
+ * and if no such beast is available, falls back to a normal memcpy.
+ */
+void drm_memcpy_from_wc(struct dma_buf_map *dst,
+ const struct dma_buf_map *src,
+ unsigned long len)
+{
+ if (WARN_ON(in_interrupt())) {
+ memcpy_fallback(dst, src, len);
+ return;
+ }
+
+ if (static_branch_likely(&has_movntdqa)) {
+ __drm_memcpy_from_wc(dst->is_iomem ?
+ (void __force *)dst->vaddr_iomem :
+ dst->vaddr,
+ src->is_iomem ?
+ (void const __force *)src->vaddr_iomem :
+ src->vaddr,
+ len);
+ return;
+ }
+
+ memcpy_fallback(dst, src, len);
+}
+EXPORT_SYMBOL(drm_memcpy_from_wc);
+
+/*
+ * drm_memcpy_init_early - One time initialization of the WC memcpy code
+ */
+void drm_memcpy_init_early(void)
+{
+ /*
+ * Some hypervisors (e.g. KVM) don't support VEX-prefix instructions
+ * emulation. So don't enable movntdqa in hypervisor guest.
+ */
+ if (static_cpu_has(X86_FEATURE_XMM4_1) &&
+ !boot_cpu_has(X86_FEATURE_HYPERVISOR))
+ static_branch_enable(&has_movntdqa);
+}
+#else
+void drm_memcpy_from_wc(struct dma_buf_map *dst,
+ const struct dma_buf_map *src,
+ unsigned long len)
+{
+ WARN_ON(in_interrupt());
+
+ memcpy_fallback(dst, src, len);
+}
+EXPORT_SYMBOL(drm_memcpy_from_wc);
+
+void drm_memcpy_init_early(void)
+{
+}
+#endif /* CONFIG_X86 */
diff --git a/drivers/gpu/drm/drm_drv.c b/drivers/gpu/drm/drm_drv.c
index 3d8d68a98b95..8804ec7d3215 100644
--- a/drivers/gpu/drm/drm_drv.c
+++ b/drivers/gpu/drm/drm_drv.c
@@ -35,6 +35,7 @@
#include <linux/slab.h>
#include <linux/srcu.h>
+#include <drm/drm_cache.h>
#include <drm/drm_client.h>
#include <drm/drm_color_mgmt.h>
#include <drm/drm_drv.h>
@@ -1041,6 +1042,7 @@ static int __init drm_core_init(void)
drm_connector_ida_init();
idr_init(&drm_minors_idr);
+ drm_memcpy_init_early();
ret = drm_sysfs_init();
if (ret < 0) {
diff --git a/include/drm/drm_cache.h b/include/drm/drm_cache.h
index e9ad4863d915..cc9de1632dd3 100644
--- a/include/drm/drm_cache.h
+++ b/include/drm/drm_cache.h
@@ -35,6 +35,8 @@
#include <linux/scatterlist.h>
+struct dma_buf_map;
+
void drm_clflush_pages(struct page *pages[], unsigned long num_pages);
void drm_clflush_sg(struct sg_table *st);
void drm_clflush_virt_range(void *addr, unsigned long length);
@@ -70,4 +72,9 @@ static inline bool drm_arch_can_wc_memory(void)
#endif
}
+void drm_memcpy_init_early(void);
+
+void drm_memcpy_from_wc(struct dma_buf_map *dst,
+ const struct dma_buf_map *src,
+ unsigned long len);
#endif