drm/udl/udl_transfer.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Copyright (C) 2012 Red Hat
0004  * based in parts on udlfb.c:
0005  * Copyright (C) 2009 Roberto De Ioris <roberto@unbit.it>
0006  * Copyright (C) 2009 Jaya Kumar <jayakumar.lkml@gmail.com>
0007  * Copyright (C) 2009 Bernie Thompson <bernie@plugable.com>
0008  */
0009
0010 #include <asm/unaligned.h>
0011
0012 #include "udl_drv.h"
0013
0014 #define MAX_CMD_PIXELS      255
0015
0016 #define RLX_HEADER_BYTES    7
0017 #define MIN_RLX_PIX_BYTES       4
0018 #define MIN_RLX_CMD_BYTES   (RLX_HEADER_BYTES + MIN_RLX_PIX_BYTES)
0019
0020 #define RLE_HEADER_BYTES    6
0021 #define MIN_RLE_PIX_BYTES   3
0022 #define MIN_RLE_CMD_BYTES   (RLE_HEADER_BYTES + MIN_RLE_PIX_BYTES)
0023
0024 #define RAW_HEADER_BYTES    6
0025 #define MIN_RAW_PIX_BYTES   2
0026 #define MIN_RAW_CMD_BYTES   (RAW_HEADER_BYTES + MIN_RAW_PIX_BYTES)
0027
0028 /*
0029  * Trims identical data from front and back of line
0030  * Sets new front buffer address and width
0031  * And returns byte count of identical pixels
0032  * Assumes CPU natural alignment (unsigned long)
0033  * for back and front buffer ptrs and width
0034  */
0035 #if 0
0036 static int udl_trim_hline(const u8 *bback, const u8 **bfront, int *width_bytes)
0037 {
0038     int j, k;
0039     const unsigned long *back = (const unsigned long *) bback;
0040     const unsigned long *front = (const unsigned long *) *bfront;
0041     const int width = *width_bytes / sizeof(unsigned long);
0042     int identical = width;
0043     int start = width;
0044     int end = width;
0045
0046     for (j = 0; j < width; j++) {
0047         if (back[j] != front[j]) {
0048             start = j;
0049             break;
0050         }
0051     }
0052
0053     for (k = width - 1; k > j; k--) {
0054         if (back[k] != front[k]) {
0055             end = k+1;
0056             break;
0057         }
0058     }
0059
0060     identical = start + (width - end);
0061     *bfront = (u8 *) &front[start];
0062     *width_bytes = (end - start) * sizeof(unsigned long);
0063
0064     return identical * sizeof(unsigned long);
0065 }
0066 #endif
0067
0068 static inline u16 pixel32_to_be16(const uint32_t pixel)
0069 {
0070     return (((pixel >> 3) & 0x001f) |
0071         ((pixel >> 5) & 0x07e0) |
0072         ((pixel >> 8) & 0xf800));
0073 }
0074
0075 static inline u16 get_pixel_val16(const uint8_t *pixel, int log_bpp)
0076 {
0077     u16 pixel_val16;
0078     if (log_bpp == 1)
0079         pixel_val16 = *(const uint16_t *)pixel;
0080     else
0081         pixel_val16 = pixel32_to_be16(*(const uint32_t *)pixel);
0082     return pixel_val16;
0083 }
0084
0085 /*
0086  * Render a command stream for an encoded horizontal line segment of pixels.
0087  *
0088  * A command buffer holds several commands.
0089  * It always begins with a fresh command header
0090  * (the protocol doesn't require this, but we enforce it to allow
0091  * multiple buffers to be potentially encoded and sent in parallel).
0092  * A single command encodes one contiguous horizontal line of pixels
0093  *
0094  * The function relies on the client to do all allocation, so that
0095  * rendering can be done directly to output buffers (e.g. USB URBs).
0096  * The function fills the supplied command buffer, providing information
0097  * on where it left off, so the client may call in again with additional
0098  * buffers if the line will take several buffers to complete.
0099  *
0100  * A single command can transmit a maximum of 256 pixels,
0101  * regardless of the compression ratio (protocol design limit).
0102  * To the hardware, 0 for a size byte means 256
0103  *
0104  * Rather than 256 pixel commands which are either rl or raw encoded,
0105  * the rlx command simply assumes alternating raw and rl spans within one cmd.
0106  * This has a slightly larger header overhead, but produces more even results.
0107  * It also processes all data (read and write) in a single pass.
0108  * Performance benchmarks of common cases show it having just slightly better
0109  * compression than 256 pixel raw or rle commands, with similar CPU consumpion.
0110  * But for very rl friendly data, will compress not quite as well.
0111  */
0112 static void udl_compress_hline16(
0113     const u8 **pixel_start_ptr,
0114     const u8 *const pixel_end,
0115     uint32_t *device_address_ptr,
0116     uint8_t **command_buffer_ptr,
0117     const uint8_t *const cmd_buffer_end, int log_bpp)
0118 {
0119     const int bpp = 1 << log_bpp;
0120     const u8 *pixel = *pixel_start_ptr;
0121     uint32_t dev_addr  = *device_address_ptr;
0122     uint8_t *cmd = *command_buffer_ptr;
0123
0124     while ((pixel_end > pixel) &&
0125            (cmd_buffer_end - MIN_RLX_CMD_BYTES > cmd)) {
0126         uint8_t *raw_pixels_count_byte = NULL;
0127         uint8_t *cmd_pixels_count_byte = NULL;
0128         const u8 *raw_pixel_start = NULL;
0129         const u8 *cmd_pixel_start, *cmd_pixel_end = NULL;
0130         uint16_t pixel_val16;
0131
0132         *cmd++ = 0xaf;
0133         *cmd++ = 0x6b;
0134         *cmd++ = (uint8_t) ((dev_addr >> 16) & 0xFF);
0135         *cmd++ = (uint8_t) ((dev_addr >> 8) & 0xFF);
0136         *cmd++ = (uint8_t) ((dev_addr) & 0xFF);
0137
0138         cmd_pixels_count_byte = cmd++; /*  we'll know this later */
0139         cmd_pixel_start = pixel;
0140
0141         raw_pixels_count_byte = cmd++; /*  we'll know this later */
0142         raw_pixel_start = pixel;
0143
0144         cmd_pixel_end = pixel + (min3(MAX_CMD_PIXELS + 1UL,
0145                     (unsigned long)(pixel_end - pixel) >> log_bpp,
0146                     (unsigned long)(cmd_buffer_end - 1 - cmd) / 2) << log_bpp);
0147
0148         pixel_val16 = get_pixel_val16(pixel, log_bpp);
0149
0150         while (pixel < cmd_pixel_end) {
0151             const u8 *const start = pixel;
0152             const uint16_t repeating_pixel_val16 = pixel_val16;
0153
0154             put_unaligned_be16(pixel_val16, cmd);
0155
0156             cmd += 2;
0157             pixel += bpp;
0158
0159             while (pixel < cmd_pixel_end) {
0160                 pixel_val16 = get_pixel_val16(pixel, log_bpp);
0161                 if (pixel_val16 != repeating_pixel_val16)
0162                     break;
0163                 pixel += bpp;
0164             }
0165
0166             if (unlikely(pixel > start + bpp)) {
0167                 /* go back and fill in raw pixel count */
0168                 *raw_pixels_count_byte = (((start -
0169                         raw_pixel_start) >> log_bpp) + 1) & 0xFF;
0170
0171                 /* immediately after raw data is repeat byte */
0172                 *cmd++ = (((pixel - start) >> log_bpp) - 1) & 0xFF;
0173
0174                 /* Then start another raw pixel span */
0175                 raw_pixel_start = pixel;
0176                 raw_pixels_count_byte = cmd++;
0177             }
0178         }
0179
0180         if (pixel > raw_pixel_start) {
0181             /* finalize last RAW span */
0182             *raw_pixels_count_byte = ((pixel - raw_pixel_start) >> log_bpp) & 0xFF;
0183         } else {
0184             /* undo unused byte */
0185             cmd--;
0186         }
0187
0188         *cmd_pixels_count_byte = ((pixel - cmd_pixel_start) >> log_bpp) & 0xFF;
0189         dev_addr += ((pixel - cmd_pixel_start) >> log_bpp) * 2;
0190     }
0191
0192     if (cmd_buffer_end <= MIN_RLX_CMD_BYTES + cmd) {
0193         /* Fill leftover bytes with no-ops */
0194         if (cmd_buffer_end > cmd)
0195             memset(cmd, 0xAF, cmd_buffer_end - cmd);
0196         cmd = (uint8_t *) cmd_buffer_end;
0197     }
0198
0199     *command_buffer_ptr = cmd;
0200     *pixel_start_ptr = pixel;
0201     *device_address_ptr = dev_addr;
0202
0203     return;
0204 }
0205
0206 /*
0207  * There are 3 copies of every pixel: The front buffer that the fbdev
0208  * client renders to, the actual framebuffer across the USB bus in hardware
0209  * (that we can only write to, slowly, and can never read), and (optionally)
0210  * our shadow copy that tracks what's been sent to that hardware buffer.
0211  */
0212 int udl_render_hline(struct drm_device *dev, int log_bpp, struct urb **urb_ptr,
0213              const char *front, char **urb_buf_ptr,
0214              u32 byte_offset, u32 device_byte_offset,
0215              u32 byte_width)
0216 {
0217     const u8 *line_start, *line_end, *next_pixel;
0218     u32 base16 = 0 + (device_byte_offset >> log_bpp) * 2;
0219     struct urb *urb = *urb_ptr;
0220     u8 *cmd = *urb_buf_ptr;
0221     u8 *cmd_end = (u8 *) urb->transfer_buffer + urb->transfer_buffer_length;
0222
0223     BUG_ON(!(log_bpp == 1 || log_bpp == 2));
0224
0225     line_start = (u8 *) (front + byte_offset);
0226     next_pixel = line_start;
0227     line_end = next_pixel + byte_width;
0228
0229     while (next_pixel < line_end) {
0230
0231         udl_compress_hline16(&next_pixel,
0232                  line_end, &base16,
0233                  (u8 **) &cmd, (u8 *) cmd_end, log_bpp);
0234
0235         if (cmd >= cmd_end) {
0236             int len = cmd - (u8 *) urb->transfer_buffer;
0237             int ret = udl_submit_urb(dev, urb, len);
0238             if (ret)
0239                 return ret;
0240             urb = udl_get_urb(dev);
0241             if (!urb)
0242                 return -EAGAIN;
0243             *urb_ptr = urb;
0244             cmd = urb->transfer_buffer;
0245             cmd_end = &cmd[urb->transfer_buffer_length];
0246         }
0247     }
0248
0249     *urb_buf_ptr = cmd;
0250
0251     return 0;
0252 }