fbdev: Don't sort deferred-I/O pages by default

Fbdev's deferred I/O sorts all dirty pages by default, which incurs a
significant overhead. Make the sorting step optional and update the few
drivers that require it. Use a FIFO list by default.

Most fbdev drivers with deferred I/O build a bounding rectangle around
the dirty pages or simply flush the whole screen. The only two affected
DRM drivers, generic fbdev and vmwgfx, both use a bounding rectangle.
In those cases, the exact order of the pages doesn't matter. The other
drivers look at the page index or handle pages one-by-one. The patch
sets the sort_pagelist flag for those, even though some of them would
probably work correctly without sorting. Driver maintainers should update
their driver accordingly.

Sorting pages by memory offset for deferred I/O performs an implicit
bubble-sort step on the list of dirty pages. The algorithm goes through
the list of dirty pages and inserts each new page according to its
index field. Even worse, list traversal always starts at the first
entry. As video memory is most likely updated scanline by scanline, the
algorithm traverses through the complete list for each updated page.

For example, with 1024x768x32bpp each page covers exactly one scanline.
Writing a single screen update from top to bottom requires updating
768 pages. With an average list length of 384 entries, a screen update
creates (768 * 384 =) 294912 compare operation.

Fix this by making the sorting step opt-in and update the few drivers
that require it. All other drivers work with unsorted page lists. Pages
are appended to the list. Therefore, in the common case of writing the
framebuffer top to bottom, pages are still sorted by offset, which may
have a positive effect on performance.

Playing a video [1] in mplayer's benchmark mode shows the difference
(i7-4790, FullHD, simpledrm, kernel with debugging).

  mplayer -benchmark -nosound -vo fbdev ./big_buck_bunny_720p_stereo.ogg

With sorted page lists:

  BENCHMARKs: VC:  32.960s VO:  73.068s A:   0.000s Sys:   2.413s =  108.441s
  BENCHMARK%: VC: 30.3947% VO: 67.3802% A:  0.0000% Sys:  2.2251% = 100.0000%

With unsorted page lists:

  BENCHMARKs: VC:  31.005s VO:  42.889s A:   0.000s Sys:   2.256s =   76.150s
  BENCHMARK%: VC: 40.7156% VO: 56.3219% A:  0.0000% Sys:  2.9625% = 100.0000%

VC shows the overhead of video decoding, VO shows the overhead of the
video output. Using unsorted page lists reduces the benchmark's run time
by ~32s/~25%.

v2:
	* Make sorted pagelists the special case (Sam)
	* Comment on drivers' use of pagelist (Sam)
	* Warn about the overhead in comment

Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Link: https://download.blender.org/peach/bigbuckbunny_movies/big_buck_bunny_720p_stereo.ogg # [1]
Link: https://patchwork.freedesktop.org/patch/msgid/20220211094640.21632-3-tzimmermann@suse.de

drivers/staging/fbtft/fbtft-core.c

@@ -654,6 +654,7 @@ struct fb_info *fbtft_framebuffer_alloc(struct fbtft_display *display,
 	fbops->fb_blank     =      fbtft_fb_blank;
 
 	fbdefio->delay =           HZ / fps;
+	fbdefio->sort_pagelist =   true;
 	fbdefio->deferred_io =     fbtft_deferred_io;
 	fb_deferred_io_init(info);
 

drivers/video/fbdev/broadsheetfb.c

@@ -1059,6 +1059,7 @@ static const struct fb_ops broadsheetfb_ops = {
 
 static struct fb_deferred_io broadsheetfb_defio = {
 	.delay		= HZ/4,
+	.sort_pagelist	= true,
 	.deferred_io	= broadsheetfb_dpy_deferred_io,
 };
 

drivers/video/fbdev/core/fb_defio.c

@@ -96,7 +96,7 @@ static vm_fault_t fb_deferred_io_mkwrite(struct vm_fault *vmf)
 	struct page *page = vmf->page;
 	struct fb_info *info = vmf->vma->vm_private_data;
 	struct fb_deferred_io *fbdefio = info->fbdefio;
-	struct page *cur;
+	struct list_head *pos = &fbdefio->pagelist;
 
 	/* this is a callback we get when userspace first tries to
 	write to the page. we schedule a workqueue. that workqueue
@@ -137,14 +137,24 @@ static vm_fault_t fb_deferred_io_mkwrite(struct vm_fault *vmf)
 	if (!list_empty(&page->lru))
 		goto page_already_added;
 
-	/* we loop through the pagelist before adding in order
-	to keep the pagelist sorted */
-	list_for_each_entry(cur, &fbdefio->pagelist, lru) {
-		if (cur->index > page->index)
-			break;
+	if (unlikely(fbdefio->sort_pagelist)) {
+		/*
+		 * We loop through the pagelist before adding in order to
+		 * keep the pagelist sorted. This has significant overhead
+		 * of O(n^2) with n being the number of written pages. If
+		 * possible, drivers should try to work with unsorted page
+		 * lists instead.
+		 */
+		struct page *cur;
+
+		list_for_each_entry(cur, &fbdefio->pagelist, lru) {
+			if (cur->index > page->index)
+				break;
+		}
+		pos = &cur->lru;
 	}
 
-	list_add_tail(&page->lru, &cur->lru);
+	list_add_tail(&page->lru, pos);
 
 page_already_added:
 	mutex_unlock(&fbdefio->lock);

drivers/video/fbdev/metronomefb.c

@@ -568,6 +568,7 @@ static const struct fb_ops metronomefb_ops = {
 
 static struct fb_deferred_io metronomefb_defio = {
 	.delay		= HZ,
+	.sort_pagelist	= true,
 	.deferred_io	= metronomefb_dpy_deferred_io,
 };
 

drivers/video/fbdev/udlfb.c

@@ -980,6 +980,7 @@ static int dlfb_ops_open(struct fb_info *info, int user)
 
 		if (fbdefio) {
 			fbdefio->delay = DL_DEFIO_WRITE_DELAY;
+			fbdefio->sort_pagelist = true;
 			fbdefio->deferred_io = dlfb_dpy_deferred_io;
 		}
 

include/linux/fb.h

@@ -204,6 +204,7 @@ struct fb_pixmap {
 struct fb_deferred_io {
 	/* delay between mkwrite and deferred handler */
 	unsigned long delay;
+	bool sort_pagelist; /* sort pagelist by offset */
 	struct mutex lock; /* mutex that protects the page list */
 	struct list_head pagelist; /* list of touched pages */
 	/* callback */