[RFC PATCH] drm/i915/gvt: Try cma allocation first for oos page setup

[Feng Tang]

Hi Yakui,

Thanks for the review!

On Mon, Jan 28, 2019 at 11:28:28AM +0800, Zhao, Yakui wrote:
> 
> 
> >-----Original Message-----
> >From: intel-gvt-dev [mailto:intel-gvt-dev-bounces at lists.freedesktop.org] On
> >Behalf Of Feng Tang
> >Sent: Monday, January 28, 2019 10:46 AM
> >To: intel-gvt-dev at lists.freedesktop.org; Zhenyu Wang
> ><zhenyuw at linux.intel.com>; Wang, Zhi A <zhi.a.wang at intel.com>; Xiao
> >Zheng <xiao.zheng at intel.com>; He, Min <min.he at intel.com>; Bing Niu
> ><bing.niu at intel.com>
> >Cc: Tang, Feng <feng.tang at intel.com>
> >Subject: [RFC PATCH] drm/i915/gvt: Try cma allocation first for oos page setup
> >
> >When doing boottime profiling, we found the oos page setup takes about
> >25~30 ms, while using cma allocation takes less than 10 ms. And if the
> >preallocated_oos_pages increases in future, it will save more.
> >
> >This commit will try to allocate the memory from CMA first, and fall back to
> >the original kzalloc if CMA is not available.
> >
> >Signed-off-by: Feng Tang <feng.tang at intel.com>
> >---
> > drivers/gpu/drm/i915/gvt/gtt.c | 26 +++++++++++++++++++++-----
> > 1 file changed, 21 insertions(+), 5 deletions(-)
> >
> >diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
> >index c7103dd..e8d6f10 100644
> >--- a/drivers/gpu/drm/i915/gvt/gtt.c
> >+++ b/drivers/gpu/drm/i915/gvt/gtt.c
> >@@ -2496,18 +2496,34 @@ static void clean_spt_oos(struct intel_gvt *gvt)
> >static int setup_spt_oos(struct intel_gvt *gvt)  {
> > 	struct intel_gvt_gtt *gtt = &gvt->gtt;
> >-	struct intel_vgpu_oos_page *oos_page;
> >+	struct intel_vgpu_oos_page *oos_page, *oos_page_base;
> >+	struct page *cma_page;
> >+	int npages;
> > 	int i;
> > 	int ret;
> >
> > 	INIT_LIST_HEAD(&gtt->oos_page_free_list_head);
> > 	INIT_LIST_HEAD(&gtt->oos_page_use_list_head);
> >
> >+	/* Try cma allocation first to save setup time */
> >+	npages = preallocated_oos_pages * sizeof(*oos_page);
> >+	npages = DIV_ROUND_UP(npages, PAGE_SIZE);
> >+
> >+	cma_page = dma_alloc_from_contiguous(NULL, npages, 1, TRUE);
> >+	if (cma_page) {
> >+		oos_page_base = page_address(cma_page);
> >+		memset(oos_page_base, 0, npages << PAGE_SHIFT);
> >+	}
> >+
> 
> Currently it will try to call kzalloc for 8192 times.
> So can we firstly use the kcalloc to allocate the memory for all the *oos_page and
> then initialize the oos_page list?

One concern with this is the size of the oos memory is about 33M, while kmalloc
may not be able to allocate so big a block (MAX_ORDER is about 11, which is 8MB).

Another question is if the kzalloc is necessary for this setup, while kmalloc
will consume much less time.

> 
> For the CMA: we need to reserve some regions for CMA. This is not recommended.
This cma allocation is a one-time deal in boot phase, which should be fine. Also
for platform with no cma explicitly allocated, this is a nop.

Thanks,
Feng