Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame^] | 1 | |
| 2 | What is udlfb? |
| 3 | =============== |
| 4 | |
| 5 | This is a driver for DisplayLink USB 2.0 era graphics chips. |
| 6 | |
| 7 | DisplayLink chips provide simple hline/blit operations with some compression, |
| 8 | pairing that with a hardware framebuffer (16MB) on the other end of the |
| 9 | USB wire. That hardware framebuffer is able to drive the VGA, DVI, or HDMI |
| 10 | monitor with no CPU involvement until a pixel has to change. |
| 11 | |
| 12 | The CPU or other local resource does all the rendering; optinally compares the |
| 13 | result with a local shadow of the remote hardware framebuffer to identify |
| 14 | the minimal set of pixels that have changed; and compresses and sends those |
| 15 | pixels line-by-line via USB bulk transfers. |
| 16 | |
| 17 | Because of the efficiency of bulk transfers and a protocol on top that |
| 18 | does not require any acks - the effect is very low latency that |
| 19 | can support surprisingly high resolutions with good performance for |
| 20 | non-gaming and non-video applications. |
| 21 | |
| 22 | Mode setting, EDID read, etc are other bulk or control transfers. Mode |
| 23 | setting is very flexible - able to set nearly arbitrary modes from any timing. |
| 24 | |
| 25 | Advantages of USB graphics in general: |
| 26 | |
| 27 | * Ability to add a nearly arbitrary number of displays to any USB 2.0 |
| 28 | capable system. On Linux, number of displays is limited by fbdev interface |
| 29 | (FB_MAX is currently 32). Of course, all USB devices on the same |
| 30 | host controller share the same 480Mbs USB 2.0 interface. |
| 31 | |
| 32 | Advantages of supporting DisplayLink chips with kernel framebuffer interface: |
| 33 | |
| 34 | * The actual hardware functionality of DisplayLink chips matches nearly |
| 35 | one-to-one with the fbdev interface, making the driver quite small and |
| 36 | tight relative to the functionality it provides. |
| 37 | * X servers and other applications can use the standard fbdev interface |
| 38 | from user mode to talk to the device, without needing to know anything |
| 39 | about USB or DisplayLink's protocol at all. A "displaylink" X driver |
| 40 | and a slightly modified "fbdev" X driver are among those that already do. |
| 41 | |
| 42 | Disadvantages: |
| 43 | |
| 44 | * Fbdev's mmap interface assumes a real hardware framebuffer is mapped. |
| 45 | In the case of USB graphics, it is just an allocated (virtual) buffer. |
| 46 | Writes need to be detected and encoded into USB bulk transfers by the CPU. |
| 47 | Accurate damage/changed area notifications work around this problem. |
| 48 | In the future, hopefully fbdev will be enhanced with an small standard |
| 49 | interface to allow mmap clients to report damage, for the benefit |
| 50 | of virtual or remote framebuffers. |
| 51 | * Fbdev does not arbitrate client ownership of the framebuffer well. |
| 52 | * Fbcon assumes the first framebuffer it finds should be consumed for console. |
| 53 | * It's not clear what the future of fbdev is, given the rise of KMS/DRM. |
| 54 | |
| 55 | How to use it? |
| 56 | ============== |
| 57 | |
| 58 | Udlfb, when loaded as a module, will match against all USB 2.0 generation |
| 59 | DisplayLink chips (Alex and Ollie family). It will then attempt to read the EDID |
| 60 | of the monitor, and set the best common mode between the DisplayLink device |
| 61 | and the monitor's capabilities. |
| 62 | |
| 63 | If the DisplayLink device is successful, it will paint a "green screen" which |
| 64 | means that from a hardware and fbdev software perspective, everything is good. |
| 65 | |
| 66 | At that point, a /dev/fb? interface will be present for user-mode applications |
| 67 | to open and begin writing to the framebuffer of the DisplayLink device using |
| 68 | standard fbdev calls. Note that if mmap() is used, by default the user mode |
| 69 | application must send down damage notifcations to trigger repaints of the |
| 70 | changed regions. Alternatively, udlfb can be recompiled with experimental |
| 71 | defio support enabled, to support a page-fault based detection mechanism |
| 72 | that can work without explicit notifcation. |
| 73 | |
| 74 | The most common client of udlfb is xf86-video-displaylink or a modified |
| 75 | xf86-video-fbdev X server. These servers have no real DisplayLink specific |
| 76 | code. They write to the standard framebuffer interface and rely on udlfb |
| 77 | to do its thing. The one extra feature they have is the ability to report |
| 78 | rectangles from the X DAMAGE protocol extension down to udlfb via udlfb's |
| 79 | damage interface (which will hopefully be standardized for all virtual |
| 80 | framebuffers that need damage info). These damage notifications allow |
| 81 | udlfb to efficiently process the changed pixels. |
| 82 | |
| 83 | Module Options |
| 84 | ============== |
| 85 | |
| 86 | Special configuration for udlfb is usually unnecessary. There are a few |
| 87 | options, however. |
| 88 | |
| 89 | From the command line, pass options to modprobe |
| 90 | modprobe udlfb fb_defio=0 console=1 shadow=1 |
| 91 | |
| 92 | Or modify options on the fly at /sys/module/udlfb/parameters directory via |
| 93 | sudo nano fb_defio |
| 94 | change the parameter in place, and save the file. |
| 95 | |
| 96 | Unplug/replug USB device to apply with new settings |
| 97 | |
| 98 | Or for permanent option, create file like /etc/modprobe.d/udlfb.conf with text |
| 99 | options udlfb fb_defio=0 console=1 shadow=1 |
| 100 | |
| 101 | Accepted boolean options: |
| 102 | |
| 103 | fb_defio Make use of the fb_defio (CONFIG_FB_DEFERRED_IO) kernel |
| 104 | module to track changed areas of the framebuffer by page faults. |
| 105 | Standard fbdev applications that use mmap but that do not |
| 106 | report damage, should be able to work with this enabled. |
| 107 | Disable when running with X server that supports reporting |
| 108 | changed regions via ioctl, as this method is simpler, |
| 109 | more stable, and higher performance. |
| 110 | default: fb_defio=1 |
| 111 | |
| 112 | console Allow fbcon to attach to udlfb provided framebuffers. |
| 113 | Can be disabled if fbcon and other clients |
| 114 | (e.g. X with --shared-vt) are in conflict. |
| 115 | default: console=1 |
| 116 | |
| 117 | shadow Allocate a 2nd framebuffer to shadow what's currently across |
| 118 | the USB bus in device memory. If any pixels are unchanged, |
| 119 | do not transmit. Spends host memory to save USB transfers. |
| 120 | Enabled by default. Only disable on very low memory systems. |
| 121 | default: shadow=1 |
| 122 | |
| 123 | Sysfs Attributes |
| 124 | ================ |
| 125 | |
| 126 | Udlfb creates several files in /sys/class/graphics/fb? |
| 127 | Where ? is the sequential framebuffer id of the particular DisplayLink device |
| 128 | |
| 129 | edid If a valid EDID blob is written to this file (typically |
| 130 | by a udev rule), then udlfb will use this EDID as a |
| 131 | backup in case reading the actual EDID of the monitor |
| 132 | attached to the DisplayLink device fails. This is |
| 133 | especially useful for fixed panels, etc. that cannot |
| 134 | communicate their capabilities via EDID. Reading |
| 135 | this file returns the current EDID of the attached |
| 136 | monitor (or last backup value written). This is |
| 137 | useful to get the EDID of the attached monitor, |
| 138 | which can be passed to utilities like parse-edid. |
| 139 | |
| 140 | metrics_bytes_rendered 32-bit count of pixel bytes rendered |
| 141 | |
| 142 | metrics_bytes_identical 32-bit count of how many of those bytes were found to be |
| 143 | unchanged, based on a shadow framebuffer check |
| 144 | |
| 145 | metrics_bytes_sent 32-bit count of how many bytes were transferred over |
| 146 | USB to communicate the resulting changed pixels to the |
| 147 | hardware. Includes compression and protocol overhead |
| 148 | |
| 149 | metrics_cpu_kcycles_used 32-bit count of CPU cycles used in processing the |
| 150 | above pixels (in thousands of cycles). |
| 151 | |
| 152 | metrics_reset Write-only. Any write to this file resets all metrics |
| 153 | above to zero. Note that the 32-bit counters above |
| 154 | roll over very quickly. To get reliable results, design |
| 155 | performance tests to start and finish in a very short |
| 156 | period of time (one minute or less is safe). |
| 157 | |
| 158 | -- |
| 159 | Bernie Thompson <bernie@plugable.com> |