Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 1 | Driver Model Compiled-in Device Tree / Platform Data |
| 2 | ==================================================== |
| 3 | |
| 4 | |
| 5 | Introduction |
| 6 | ------------ |
| 7 | |
| 8 | Device tree is the standard configuration method in U-Boot. It is used to |
| 9 | define what devices are in the system and provide configuration information |
| 10 | to these devices. |
| 11 | |
| 12 | The overhead of adding device tree access to U-Boot is fairly modest, |
| 13 | approximately 3KB on Thumb 2 (plus the size of the DT itself). This means |
| 14 | that in most cases it is best to use device tree for configuration. |
| 15 | |
| 16 | However there are some very constrained environments where U-Boot needs to |
| 17 | work. These include SPL with severe memory limitations. For example, some |
| 18 | SoCs require a 16KB SPL image which must include a full MMC stack. In this |
| 19 | case the overhead of device tree access may be too great. |
| 20 | |
| 21 | It is possible to create platform data manually by defining C structures |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 22 | for it, and reference that data in a U_BOOT_DEVICE() declaration. This |
| 23 | bypasses the use of device tree completely, effectively creating a parallel |
| 24 | configuration mechanism. But it is an available option for SPL. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 25 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 26 | As an alternative, a new 'of-platdata' feature is provided. This converts the |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 27 | device tree contents into C code which can be compiled into the SPL binary. |
| 28 | This saves the 3KB of code overhead and perhaps a few hundred more bytes due |
| 29 | to more efficient storage of the data. |
| 30 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 31 | Note: Quite a bit of thought has gone into the design of this feature. |
| 32 | However it still has many rough edges and comments and suggestions are |
| 33 | strongly encouraged! Quite possibly there is a much better approach. |
| 34 | |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 35 | |
| 36 | Caveats |
| 37 | ------- |
| 38 | |
| 39 | There are many problems with this features. It should only be used when |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 40 | strictly necessary. Notable problems include: |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 41 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 42 | - Device tree does not describe data types. But the C code must define a |
| 43 | type for each property. These are guessed using heuristics which |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 44 | are wrong in several fairly common cases. For example an 8-byte value |
| 45 | is considered to be a 2-item integer array, and is byte-swapped. A |
| 46 | boolean value that is not present means 'false', but cannot be |
| 47 | included in the structures since there is generally no mention of it |
| 48 | in the device tree file. |
| 49 | |
| 50 | - Naming of nodes and properties is automatic. This means that they follow |
| 51 | the naming in the device tree, which may result in C identifiers that |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 52 | look a bit strange. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 53 | |
| 54 | - It is not possible to find a value given a property name. Code must use |
| 55 | the associated C member variable directly in the code. This makes |
| 56 | the code less robust in the face of device-tree changes. It also |
| 57 | makes it very unlikely that your driver code will be useful for more |
| 58 | than one SoC. Even if the code is common, each SoC will end up with |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 59 | a different C struct name, and a likely a different format for the |
| 60 | platform data. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 61 | |
| 62 | - The platform data is provided to drivers as a C structure. The driver |
| 63 | must use the same structure to access the data. Since a driver |
| 64 | normally also supports device tree it must use #ifdef to separate |
| 65 | out this code, since the structures are only available in SPL. |
| 66 | |
| 67 | |
| 68 | How it works |
| 69 | ------------ |
| 70 | |
| 71 | The feature is enabled by CONFIG SPL_OF_PLATDATA. This is only available |
| 72 | in SPL and should be tested with: |
| 73 | |
| 74 | #if CONFIG_IS_ENABLED(SPL_OF_PLATDATA) |
| 75 | |
| 76 | A new tool called 'dtoc' converts a device tree file either into a set of |
| 77 | struct declarations, one for each compatible node, or a set of |
| 78 | U_BOOT_DEVICE() declarations along with the actual platform data for each |
| 79 | device. As an example, consider this MMC node: |
| 80 | |
| 81 | sdmmc: dwmmc@ff0c0000 { |
| 82 | compatible = "rockchip,rk3288-dw-mshc"; |
| 83 | clock-freq-min-max = <400000 150000000>; |
| 84 | clocks = <&cru HCLK_SDMMC>, <&cru SCLK_SDMMC>, |
| 85 | <&cru SCLK_SDMMC_DRV>, <&cru SCLK_SDMMC_SAMPLE>; |
| 86 | clock-names = "biu", "ciu", "ciu_drv", "ciu_sample"; |
| 87 | fifo-depth = <0x100>; |
| 88 | interrupts = <GIC_SPI 32 IRQ_TYPE_LEVEL_HIGH>; |
| 89 | reg = <0xff0c0000 0x4000>; |
| 90 | bus-width = <4>; |
| 91 | cap-mmc-highspeed; |
| 92 | cap-sd-highspeed; |
| 93 | card-detect-delay = <200>; |
| 94 | disable-wp; |
| 95 | num-slots = <1>; |
| 96 | pinctrl-names = "default"; |
| 97 | pinctrl-0 = <&sdmmc_clk>, <&sdmmc_cmd>, <&sdmmc_cd>, <&sdmmc_bus4>; |
| 98 | vmmc-supply = <&vcc_sd>; |
| 99 | status = "okay"; |
| 100 | u-boot,dm-pre-reloc; |
| 101 | }; |
| 102 | |
| 103 | |
| 104 | Some of these properties are dropped by U-Boot under control of the |
| 105 | CONFIG_OF_SPL_REMOVE_PROPS option. The rest are processed. This will produce |
| 106 | the following C struct declaration: |
| 107 | |
| 108 | struct dtd_rockchip_rk3288_dw_mshc { |
| 109 | fdt32_t bus_width; |
| 110 | bool cap_mmc_highspeed; |
| 111 | bool cap_sd_highspeed; |
| 112 | fdt32_t card_detect_delay; |
| 113 | fdt32_t clock_freq_min_max[2]; |
| 114 | struct phandle_2_cell clocks[4]; |
| 115 | bool disable_wp; |
| 116 | fdt32_t fifo_depth; |
| 117 | fdt32_t interrupts[3]; |
| 118 | fdt32_t num_slots; |
| 119 | fdt32_t reg[2]; |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 120 | fdt32_t vmmc_supply; |
| 121 | }; |
| 122 | |
| 123 | and the following device declaration: |
| 124 | |
| 125 | static struct dtd_rockchip_rk3288_dw_mshc dtv_dwmmc_at_ff0c0000 = { |
| 126 | .fifo_depth = 0x100, |
| 127 | .cap_sd_highspeed = true, |
| 128 | .interrupts = {0x0, 0x20, 0x4}, |
| 129 | .clock_freq_min_max = {0x61a80, 0x8f0d180}, |
| 130 | .vmmc_supply = 0xb, |
| 131 | .num_slots = 0x1, |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 132 | .clocks = {{&dtv_clock_controller_at_ff760000, 456}, |
| 133 | {&dtv_clock_controller_at_ff760000, 68}, |
| 134 | {&dtv_clock_controller_at_ff760000, 114}, |
| 135 | {&dtv_clock_controller_at_ff760000, 118}}, |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 136 | .cap_mmc_highspeed = true, |
| 137 | .disable_wp = true, |
| 138 | .bus_width = 0x4, |
| 139 | .u_boot_dm_pre_reloc = true, |
| 140 | .reg = {0xff0c0000, 0x4000}, |
| 141 | .card_detect_delay = 0xc8, |
| 142 | }; |
| 143 | U_BOOT_DEVICE(dwmmc_at_ff0c0000) = { |
| 144 | .name = "rockchip_rk3288_dw_mshc", |
| 145 | .platdata = &dtv_dwmmc_at_ff0c0000, |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 146 | .platdata_size = sizeof(dtv_dwmmc_at_ff0c0000), |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 147 | }; |
| 148 | |
| 149 | The device is then instantiated at run-time and the platform data can be |
| 150 | accessed using: |
| 151 | |
| 152 | struct udevice *dev; |
| 153 | struct dtd_rockchip_rk3288_dw_mshc *plat = dev_get_platdata(dev); |
| 154 | |
| 155 | This avoids the code overhead of converting the device tree data to |
| 156 | platform data in the driver. The ofdata_to_platdata() method should |
| 157 | therefore do nothing in such a driver. |
| 158 | |
Simon Glass | 2cce586 | 2017-06-13 21:10:06 -0600 | [diff] [blame] | 159 | Where a node has multiple compatible strings, a #define is used to make them |
| 160 | equivalent, e.g.: |
| 161 | |
| 162 | #define dtd_rockchip_rk3299_dw_mshc dtd_rockchip_rk3288_dw_mshc |
| 163 | |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 164 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 165 | Converting of-platdata to a useful form |
| 166 | --------------------------------------- |
| 167 | |
| 168 | Of course it would be possible use the of-platdata directly in your driver |
| 169 | whenever configuration information is required. However this meands that the |
| 170 | driver will not be able to support device tree, since the of-platdata |
| 171 | structure is not available when device tree is used. It would make no sense |
| 172 | to use this structure if device tree were available, since the structure has |
| 173 | all the limitations metioned in caveats above. |
| 174 | |
| 175 | Therefore it is recommended that the of-platdata structure should be used |
| 176 | only in the probe() method of your driver. It cannot be used in the |
| 177 | ofdata_to_platdata() method since this is not called when platform data is |
| 178 | already present. |
| 179 | |
| 180 | |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 181 | How to structure your driver |
| 182 | ---------------------------- |
| 183 | |
| 184 | Drivers should always support device tree as an option. The of-platdata |
| 185 | feature is intended as a add-on to existing drivers. |
| 186 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 187 | Your driver should convert the platdata struct in its probe() method. The |
| 188 | existing device tree decoding logic should be kept in the |
| 189 | ofdata_to_platdata() method and wrapped with #if. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 190 | |
| 191 | For example: |
| 192 | |
| 193 | #include <dt-structs.h> |
| 194 | |
| 195 | struct mmc_platdata { |
| 196 | #if CONFIG_IS_ENABLED(SPL_OF_PLATDATA) |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 197 | /* Put this first since driver model will copy the data here */ |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 198 | struct dtd_mmc dtplat; |
| 199 | #endif |
| 200 | /* |
| 201 | * Other fields can go here, to be filled in by decoding from |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 202 | * the device tree (or the C structures when of-platdata is used). |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 203 | */ |
| 204 | int fifo_depth; |
| 205 | }; |
| 206 | |
| 207 | static int mmc_ofdata_to_platdata(struct udevice *dev) |
| 208 | { |
| 209 | #if !CONFIG_IS_ENABLED(SPL_OF_PLATDATA) |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 210 | /* Decode the device tree data */ |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 211 | struct mmc_platdata *plat = dev_get_platdata(dev); |
| 212 | const void *blob = gd->fdt_blob; |
Simon Glass | e160f7d | 2017-01-17 16:52:55 -0700 | [diff] [blame] | 213 | int node = dev_of_offset(dev); |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 214 | |
| 215 | plat->fifo_depth = fdtdec_get_int(blob, node, "fifo-depth", 0); |
| 216 | #endif |
| 217 | |
| 218 | return 0; |
| 219 | } |
| 220 | |
| 221 | static int mmc_probe(struct udevice *dev) |
| 222 | { |
| 223 | struct mmc_platdata *plat = dev_get_platdata(dev); |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 224 | |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 225 | #if CONFIG_IS_ENABLED(SPL_OF_PLATDATA) |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 226 | /* Decode the of-platdata from the C structures */ |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 227 | struct dtd_mmc *dtplat = &plat->dtplat; |
| 228 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 229 | plat->fifo_depth = dtplat->fifo_depth; |
| 230 | #endif |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 231 | /* Set up the device from the plat data */ |
| 232 | writel(plat->fifo_depth, ...) |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 233 | } |
| 234 | |
| 235 | static const struct udevice_id mmc_ids[] = { |
| 236 | { .compatible = "vendor,mmc" }, |
| 237 | { } |
| 238 | }; |
| 239 | |
| 240 | U_BOOT_DRIVER(mmc_drv) = { |
| 241 | .name = "mmc", |
| 242 | .id = UCLASS_MMC, |
| 243 | .of_match = mmc_ids, |
| 244 | .ofdata_to_platdata = mmc_ofdata_to_platdata, |
| 245 | .probe = mmc_probe, |
| 246 | .priv_auto_alloc_size = sizeof(struct mmc_priv), |
| 247 | .platdata_auto_alloc_size = sizeof(struct mmc_platdata), |
| 248 | }; |
| 249 | |
| 250 | |
| 251 | In the case where SPL_OF_PLATDATA is enabled, platdata_auto_alloc_size is |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 252 | still used to allocate space for the platform data. This is different from |
| 253 | the normal behaviour and is triggered by the use of of-platdata (strictly |
| 254 | speaking it is a non-zero platdata_size which triggers this). |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 255 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 256 | The of-platdata struct contents is copied from the C structure data to the |
| 257 | start of the newly allocated area. In the case where device tree is used, |
| 258 | the platform data is allocated, and starts zeroed. In this case the |
| 259 | ofdata_to_platdata() method should still set up the platform data (and the |
| 260 | of-platdata struct will not be present). |
| 261 | |
| 262 | SPL must use either of-platdata or device tree. Drivers cannot use both at |
| 263 | the same time, but they must support device tree. Supporting of-platdata is |
| 264 | optional. |
| 265 | |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 266 | The device tree becomes in accessible when CONFIG_SPL_OF_PLATDATA is enabled, |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 267 | since the device-tree access code is not compiled in. A corollary is that |
| 268 | a board can only move to using of-platdata if all the drivers it uses support |
| 269 | it. There would be little point in having some drivers require the device |
| 270 | tree data, since then libfdt would still be needed for those drivers and |
| 271 | there would be no code-size benefit. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 272 | |
| 273 | Internals |
| 274 | --------- |
| 275 | |
| 276 | The dt-structs.h file includes the generated file |
| 277 | (include/generated//dt-structs.h) if CONFIG_SPL_OF_PLATDATA is enabled. |
| 278 | Otherwise (such as in U-Boot proper) these structs are not available. This |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 279 | prevents them being used inadvertently. All usage must be bracketed with |
| 280 | #if CONFIG_IS_ENABLED(SPL_OF_PLATDATA). |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 281 | |
| 282 | The dt-platdata.c file contains the device declarations and is is built in |
| 283 | spl/dt-platdata.c. |
| 284 | |
| 285 | Some phandles (thsoe that are recognised as such) are converted into |
| 286 | points to platform data. This pointer can potentially be used to access the |
| 287 | referenced device (by searching for the pointer value). This feature is not |
| 288 | yet implemented, however. |
| 289 | |
| 290 | The beginnings of a libfdt Python module are provided. So far this only |
| 291 | implements a subset of the features. |
| 292 | |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 293 | The 'swig' tool is needed to build the libfdt Python module. If this is not |
| 294 | found then the Python model is not used and a fallback is used instead, which |
| 295 | makes use of fdtget. |
| 296 | |
| 297 | |
| 298 | Credits |
| 299 | ------- |
| 300 | |
| 301 | This is an implementation of an idea by Tom Rini <trini@konsulko.com>. |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 302 | |
| 303 | |
| 304 | Future work |
| 305 | ----------- |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 306 | - Consider programmatically reading binding files instead of device tree |
| 307 | contents |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 308 | - Complete the phandle feature |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 309 | - Move to using a full Python libfdt module |
| 310 | |
| 311 | -- |
| 312 | Simon Glass <sjg@chromium.org> |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 313 | Google, Inc |
Simon Glass | 39782af | 2016-07-04 11:58:07 -0600 | [diff] [blame] | 314 | 6/6/16 |
Simon Glass | 1269625 | 2016-07-04 11:58:42 -0600 | [diff] [blame] | 315 | Updated Independence Day 2016 |