Copyright (C) 1999-2002 VMware, Inc. All Rights Reserved The code here may be used/distributed under the terms of the standard XFree86 license. VMware SVGA Device Interface and Programming Model -------------------------------------------------- Include Files ------------- svga_reg.h SVGA register definitions, SVGA capabilities, and FIFO command definitions. svga_limits.h Included by svga_reg.h, defines maximum frame buffer and memory region sizes. guest_os.h Values for the GUEST_ID register. vm_basic_types.h Common type definitions. vm_device_version.h PCI vendor ID's and related information. Programming the VMware SVGA Device ---------------------------------- 1. Reading/writing a register: The SVGA registers are addressed by an index/value pair of 32 bit registers in the IO address space. The 0710 VMware SVGA chipset (PCI device ID PCI_DEVICE_ID_VMWARE_SVGA) has its index and value ports hardcoded at: index: SVGA_LEGACY_BASE_PORT + 4 * SVGA_INDEX_PORT value: SVGA_LEGACY_BASE_PORT + 4 * SVGA_VALUE_PORT The 0405 VMware SVGA chipset (PCI device ID PCI_DEVICE_ID_VMWARE_SVGA2) determines its index and value ports as a function of the first base address register in its PCI configuration space as: index: + SVGA_INDEX_PORT value: + SVGA_VALUE_PORT To read a register: Set the index port to the index of the register, using a dword OUT Do a dword IN from the value port To write a register: Set the index port to the index of the register, using a dword OUT Do a dword OUT to the value port Example, setting the width to 1024: mov eax, SVGA_REG_WIDTH mov edx, out dx, eax mov eax, 1024 mov edx, out dx, eax 2. Initialization Check the version number loop: Write into SVGA_REG_ID the maximum SVGA_ID_* the driver supports. Read from SVGA_REG_ID. Check if it is the value you wrote. If yes, VMware SVGA device supports it If no, decrement SVGA_ID_* and goto loop This algorithm converges. Map the frame buffer and the command FIFO Read SVGA_REG_FB_START, SVGA_REG_FB_SIZE, SVGA_REG_MEM_START, SVGA_REG_MEM_SIZE. Map the frame buffer (FB) and the FIFO memory (MEM) Get the device capabilities and frame buffer dimensions Read SVGA_REG_CAPABILITIES, SVGA_REG_MAX_WIDTH, SVGA_REG_MAX_HEIGHT, and SVGA_REG_HOST_BITS_PER_PIXEL / SVGA_REG_BITS_PER_PIXEL. Note: The capabilities can and do change without the PCI device ID changing or the SVGA_REG_ID changing. A driver should always check the capabilities register when loading before expecting any capabilities-determined feature to be available. See below for a list of capabilities as of this writing. Note: If SVGA_CAP_8BIT_EMULATION is not set, then it is possible that SVGA_REG_HOST_BITS_PER_PIXEL does not exist and SVGA_REG_BITS_PER_PIXEL should be read instead. Report the Guest Operating System Write SVGA_REG_GUEST_ID with the appropriate value from . While not required in any way, this is useful information for the virtual machine to have available for reporting and sanity checking purposes. SetMode Set SVGA_REG_WIDTH, SVGA_REG_HEIGHT, SVGA_REG_BITS_PER_PIXEL Read SVGA_REG_FB_OFFSET (SVGA_REG_FB_OFFSET is the offset from SVGA_REG_FB_START of the visible portion of the frame buffer) Read SVGA_REG_BYTES_PER_LINE, SVGA_REG_DEPTH, SVGA_REG_PSEUDOCOLOR, SVGA_REG_RED_MASK, SVGA_REG_GREEN_MASK, SVGA_REG_BLUE_MASK Note: SVGA_REG_BITS_PER_PIXEL is readonly if SVGA_CAP_8BIT_EMULATION is not set in the capabilities register. Even if it is set, values other than 8 and SVGA_REG_HOST_BITS_PER_PIXEL will be ignored. Enable SVGA Set SVGA_REG_ENABLE to 1 (to disable SVGA, set SVGA_REG_ENABLE to 0. Setting SVGA_REG_ENABLE to 0 also enables VGA.) Initialize the command FIFO The FIFO is exclusively dword (32-bit) aligned. The first four dwords define the portion of the MEM area that is used for the command FIFO. These are values are all in byte offsets from the start of the MEM area. A minimum sized FIFO would have these values: mem[SVGA_FIFO_MIN] = 16; mem[SVGA_FIFO_MAX] = 16 + (10 * 1024); mem[SVGA_FIFO_NEXT_CMD] = 16; mem[SVGA_FIFO_STOP] = 16; Set SVGA_REG_CONFIG_DONE to 1 after these values have been set. Note: Setting SVGA_REG_CONFIG_DONE to 0 will stop the device from reading the FIFO until it is reinitialized and SVGA_REG_CONFIG_DONE is set to 1 again. 3. SVGA command FIFO protocol The FIFO is empty when SVGA_FIFO_NEXT_CMD == SVGA_FIFO_STOP. The driver writes commands to the FIFO starting at the offset specified by SVGA_FIFO_NEXT_CMD, and then increments SVGA_FIFO_NEXT_CMD. The FIFO is full when SVGA_FIFO_NEXT_CMD is one word before SVGA_FIFO_STOP. When the FIFO becomes full, the FIFO should be sync'd To sync the FIFO Write SVGA_REG_SYNC Read SVGA_REG_BUSY Wait for the value in SVGA_REG_BUSY to be 0 The FIFO should be sync'd before the driver touches the frame buffer, to guarantee that any outstanding BLT's are completed. 4. Cursor When SVGA_CAP_CURSOR is set, hardware cursor support is available. In practice, SVGA_CAP_CURSOR will only be set when SVGA_CAP_CURSOR_BYPASS is also set and drivers supporting a hardware cursor should only worry about SVGA_CAP_CURSOR_BYPASS and only use the FIFO to define the cursor. See below for more information. 5. Pseudocolor When the read-only register SVGA_REG_PSEUDOCOLOR is 1, the device is in a colormapped mode whose index width and color width are both SVGA_REG_DEPTH. Thus far, 8 is the only depth at which pseudocolor is ever used. In pseudocolor, the colormap is programmed by writing to the SVGA palette registers. These start at SVGA_PALETTE_BASE and are interpreted as follows: SVGA_PALETTE_BASE + 3*n - The nth red component SVGA_PALETTE_BASE + 3*n + 1 - The nth green component SVGA_PALETTE_BASE + 3*n + 2 - The nth blue component And n ranges from 0 to ((1<. What should this be set to? Likewise with SVGA_CMD_DEFINE_PIXMAP. And when should the SCANLINE macros be used? OK, I'll use pixmaps as an example. First you have to define the pixmap: #define SVGA_CMD_DEFINE_PIXMAP 6 /* FIFO layout: Pixmap ID, Width, Height, Depth, */ The ID is something you choose, which you subsequently use to refer to this pixmap. It must be an integer between 0 and SVGA_MAX_ID. The width and height and depth are the dimensions of the pixmap. For now, the depth of the pixmap has to match the depth of the screen. The scanlines are the pixels that make up the pixmap, arranged one row at a time. Each row is required to be 32-bit aligned. The macros SVGA_PIXMAP_SCANLINE_SIZE and SVGA_PIXMAP_SIZE give the size of a single scanline, and the size of the entire pixmap, respectively, in 32-bit words. The second step is to use it: #define SVGA_CMD_RECT_PIXMAP_FILL 9 /* FIFO layout: Pixmap ID, X, Y, Width, Height */ The ID here is the one you chose when defining the pixmap. X, Y, Width, and Height define a rectangle on the screen that is to be filled with the pixmap. The pixmap is screen aligned, which means that the coordinates in the pixmap are defined by the screen coordinates modulo the pixmap dimensions. If you want a different alignment between the screen and the pixmap, then you can use this command, which allows the pixmap coordinates to be defined: #define SVGA_CMD_RECT_PIXMAP_COPY 11 /* FIFO layout: Pixmap ID, Source X, Source Y, Dest X, Dest Y, Width, Height */ The Source X and Source Y are pixmap coordinates, and the Dest X and Dest Y are screen coordinates. 5. OK, now it works briefly, then stops displaying anything. Also, my log file is filled with lines like: Unknown Command 0xff in SVGA command FIFO What's happening? The most common problem at this point is that the FIFO gets out of sync. This can happen if the amount of data in the FIFO doesn't match what the VMware SVGA device expects. To track this down, try to isolate the particular command which causes the problem. Another way this can happen is if the wraparound in the FIFO isn't done correctly. Here is some example code for writing to the FIFO (mem is an array of 32-bit integers that points to the FIFO memory region): while (TRUE) { fifo_min = mem[SVGA_FIFO_MIN] / 4; fifo_max = mem[SVGA_FIFO_MAX] / 4; fifo_next = mem[SVGA_FIFO_NEXT_CMD] / 4; fifo_stop = mem[SVGA_FIFO_STOP] / 4; tmp_next = fifo_next+1; if (tmp_next == fifo_max) tmp_next = fifo_min; // Wraparound if (tmp_next == fifo_stop) { sync_fifo(); // FIFO full continue; // retry } mem[fifo_next] = item; mem[SVGA_FIFO_NEXT_CMD] = tmp_next * 4; break; } This isn't the most efficient code, but it should work. It's important to do the increment with wraparound before the FIFO full check, and to check FIFO full before updating the next command pointer. 6. My driver tries to switch modes and either nothing happens or the display becomes completely garbled. What's going on? When you change modes, make very sure you reread all of the registers listed above under SetMode. Getting the pitch (SVGA_REG_BYTES_PER_LINE) incorrect will cause a heavily garbled display. Also, if you change SVGA_REG_BITS_PER_PIXEL, make certain that SVGA_CAP_8BIT_EMULATION is present in the SVGA_REG_CAPABILITIES register. Also, even with 8 bit emulation, the driver must still use either 8 bpp or SVGA_REG_HOST_BITS_PER_PIXEL bpp, nothing else. 7. Why does my driver's hardware cursor work when my virtual machine is in window mode, but draw/erase incorrectly or in garbled locations in fullscreen mode? You need to make sure you use SVGA_CURSOR_ON_REMOVE_FROM_FB and SVGA_CURSOR_ON_RESTORE_TO_FB _every_ time your driver or the virtual machine touches a region of the framebuffer that overlaps the cursor. If you forget to remove it then it can show up when doing save-under operations or get mixed in with other drawing. If you forget to restore it then can disappear. You also need to make sure SVGA_CAP_CURSOR_BYPASS2 is available, or else you will have to use SVGA_CURSOR_ON_SHOW and SVGA_CURSOR_ON_HIDE (which will flicker, even in window mode), or else a software cursor. Newer version of the virtual SVGA hardware will never put the hardware cursor in the framebuffer while in window mode, so everything will appear to work correctly there. 8. Why do my accelerated glyphs look funny? OR Why does the fifo complain about invalid commands when I draw accelerated glyphs? The bitmap data passed to SVGA_CMD_DRAW_GLYPH_* must not have any per-scanline alignment. If there are any remaining bits left in the last byte of a scanline, the first bits of the next scanline should use them. The bitmap data as a whole must be 4 byte aligned. $XFree86: xc/programs/Xserver/hw/xfree86/drivers/vmware/README,v 1.5 2002/10/16 22:12:53 alanh Exp $