| Message ID | 20220210141111.5231-3-tzimmermann@suse.de |
|---|---|
| State | New |
| Headers | show |
| Series | fbdev: Significantly improve performance of fbdefio mmap | expand |
Hi Thomas, On Thu, Feb 10, 2022 at 4:24 PM Thomas Zimmermann <tzimmermann@suse.de> wrote: > 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. > > 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 a 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. What about using folios? If consecutive pages are merged into a single entry, there's much less (or nothing in the example above) to sort. Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
Hi Thomas, On Mon, Feb 14, 2022 at 9:28 AM Thomas Zimmermann <tzimmermann@suse.de> wrote: > Am 14.02.22 um 09:05 schrieb Geert Uytterhoeven: > > On Thu, Feb 10, 2022 at 4:24 PM Thomas Zimmermann <tzimmermann@suse.de> wrote: > >> 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. > >> > >> 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 a 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. > > > > What about using folios? > > If consecutive pages are merged into a single entry, there's much less > > (or nothing in the example above) to sort. > > How would the code know that? Calls to page_mkwrite happen > pagefault-by-pagefault in any order AFAICT. fb_deferred_io_mkwrite() would still be called for a page, but an adjacent page can be merged with an existing entry while adding it to the list. Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
Hi Am 14.02.22 um 10:05 schrieb Geert Uytterhoeven: > Hi Thomas, > > On Mon, Feb 14, 2022 at 9:28 AM Thomas Zimmermann <tzimmermann@suse.de> wrote: >> Am 14.02.22 um 09:05 schrieb Geert Uytterhoeven: >>> On Thu, Feb 10, 2022 at 4:24 PM Thomas Zimmermann <tzimmermann@suse.de> wrote: >>>> 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. >>>> >>>> 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 a 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. >>> >>> What about using folios? >>> If consecutive pages are merged into a single entry, there's much less >>> (or nothing in the example above) to sort. >> >> How would the code know that? Calls to page_mkwrite happen >> pagefault-by-pagefault in any order AFAICT. > > fb_deferred_io_mkwrite() would still be called for a page, but an > adjacent page can be merged with an existing entry while adding it > to the list. I still don't understand how we'd use it to our advantage. Most drivers don't need sorted pages at all. A folio has strong alignment requirements for size and offset AFAICT. We might end up flushing way too much of the display memory. Best regards Thomas > > Gr{oetje,eeting}s, > > Geert > > -- > Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org > > In personal conversations with technical people, I call myself a hacker. But > when I'm talking to journalists I just say "programmer" or something like that. > -- Linus Torvalds
diff --git a/drivers/staging/fbtft/fbtft-core.c b/drivers/staging/fbtft/fbtft-core.c index f2684d2d6851..4a35347b3020 100644 --- a/drivers/staging/fbtft/fbtft-core.c +++ b/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); diff --git a/drivers/video/fbdev/broadsheetfb.c b/drivers/video/fbdev/broadsheetfb.c index fd66f4d4a621..b9054f658838 100644 --- a/drivers/video/fbdev/broadsheetfb.c +++ b/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, }; diff --git a/drivers/video/fbdev/core/fb_defio.c b/drivers/video/fbdev/core/fb_defio.c index 3727b1ca87b1..1f672cf253b2 100644 --- a/drivers/video/fbdev/core/fb_defio.c +++ b/drivers/video/fbdev/core/fb_defio.c @@ -132,15 +132,20 @@ 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 (fbdefio->sort_pagelist) { + /* + * 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; + } + list_add_tail(&page->lru, &cur->lru); + } else { + list_add_tail(&page->lru, &fbdefio->pagelist); } - list_add_tail(&page->lru, &cur->lru); - page_already_added: mutex_unlock(&fbdefio->lock); diff --git a/drivers/video/fbdev/metronomefb.c b/drivers/video/fbdev/metronomefb.c index 952826557a0c..af858dd23ea6 100644 --- a/drivers/video/fbdev/metronomefb.c +++ b/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, }; diff --git a/drivers/video/fbdev/udlfb.c b/drivers/video/fbdev/udlfb.c index b9cdd02c1000..184bb8433b78 100644 --- a/drivers/video/fbdev/udlfb.c +++ b/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; } diff --git a/include/linux/fb.h b/include/linux/fb.h index 3d7306c9a706..9a77ab615c36 100644 --- a/include/linux/fb.h +++ b/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 */
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. 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 a 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%. Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de> Link: https://download.blender.org/peach/bigbuckbunny_movies/big_buck_bunny_720p_stereo.ogg # [1] --- drivers/staging/fbtft/fbtft-core.c | 1 + drivers/video/fbdev/broadsheetfb.c | 1 + drivers/video/fbdev/core/fb_defio.c | 19 ++++++++++++------- drivers/video/fbdev/metronomefb.c | 1 + drivers/video/fbdev/udlfb.c | 1 + include/linux/fb.h | 1 + 6 files changed, 17 insertions(+), 7 deletions(-)