1================================ 2I2C muxes and complex topologies 3================================ 4 5There are a couple of reasons for building more complex I2C topologies 6than a straight-forward I2C bus with one adapter and one or more devices. 7 81. A mux may be needed on the bus to prevent address collisions. 9 102. The bus may be accessible from some external bus master, and arbitration 11 may be needed to determine if it is ok to access the bus. 12 133. A device (particularly RF tuners) may want to avoid the digital noise 14 from the I2C bus, at least most of the time, and sits behind a gate 15 that has to be operated before the device can be accessed. 16 17Etc 18=== 19 20These constructs are represented as I2C adapter trees by Linux, where 21each adapter has a parent adapter (except the root adapter) and zero or 22more child adapters. The root adapter is the actual adapter that issues 23I2C transfers, and all adapters with a parent are part of an "i2c-mux" 24object (quoted, since it can also be an arbitrator or a gate). 25 26Depending of the particular mux driver, something happens when there is 27an I2C transfer on one of its child adapters. The mux driver can 28obviously operate a mux, but it can also do arbitration with an external 29bus master or open a gate. The mux driver has two operations for this, 30select and deselect. select is called before the transfer and (the 31optional) deselect is called after the transfer. 32 33 34Locking 35======= 36 37There are two variants of locking available to I2C muxes, they can be 38mux-locked or parent-locked muxes. As is evident from below, it can be 39useful to know if a mux is mux-locked or if it is parent-locked. The 40following list was correct at the time of writing: 41 42In drivers/i2c/muxes/: 43 44====================== ============================================= 45i2c-arb-gpio-challenge Parent-locked 46i2c-mux-gpio Normally parent-locked, mux-locked iff 47 all involved gpio pins are controlled by the 48 same I2C root adapter that they mux. 49i2c-mux-gpmux Normally parent-locked, mux-locked iff 50 specified in device-tree. 51i2c-mux-ltc4306 Mux-locked 52i2c-mux-mlxcpld Parent-locked 53i2c-mux-pca9541 Parent-locked 54i2c-mux-pca954x Parent-locked 55i2c-mux-pinctrl Normally parent-locked, mux-locked iff 56 all involved pinctrl devices are controlled 57 by the same I2C root adapter that they mux. 58i2c-mux-reg Parent-locked 59====================== ============================================= 60 61In drivers/iio/: 62 63====================== ============================================= 64gyro/mpu3050 Mux-locked 65imu/inv_mpu6050/ Mux-locked 66====================== ============================================= 67 68In drivers/media/: 69 70======================= ============================================= 71dvb-frontends/lgdt3306a Mux-locked 72dvb-frontends/m88ds3103 Parent-locked 73dvb-frontends/rtl2830 Parent-locked 74dvb-frontends/rtl2832 Mux-locked 75dvb-frontends/si2168 Mux-locked 76usb/cx231xx/ Parent-locked 77======================= ============================================= 78 79 80Mux-locked muxes 81---------------- 82 83Mux-locked muxes does not lock the entire parent adapter during the 84full select-transfer-deselect transaction, only the muxes on the parent 85adapter are locked. Mux-locked muxes are mostly interesting if the 86select and/or deselect operations must use I2C transfers to complete 87their tasks. Since the parent adapter is not fully locked during the 88full transaction, unrelated I2C transfers may interleave the different 89stages of the transaction. This has the benefit that the mux driver 90may be easier and cleaner to implement, but it has some caveats. 91 92==== ===================================================================== 93ML1. If you build a topology with a mux-locked mux being the parent 94 of a parent-locked mux, this might break the expectation from the 95 parent-locked mux that the root adapter is locked during the 96 transaction. 97 98ML2. It is not safe to build arbitrary topologies with two (or more) 99 mux-locked muxes that are not siblings, when there are address 100 collisions between the devices on the child adapters of these 101 non-sibling muxes. 102 103 I.e. the select-transfer-deselect transaction targeting e.g. device 104 address 0x42 behind mux-one may be interleaved with a similar 105 operation targeting device address 0x42 behind mux-two. The 106 intension with such a topology would in this hypothetical example 107 be that mux-one and mux-two should not be selected simultaneously, 108 but mux-locked muxes do not guarantee that in all topologies. 109 110ML3. A mux-locked mux cannot be used by a driver for auto-closing 111 gates/muxes, i.e. something that closes automatically after a given 112 number (one, in most cases) of I2C transfers. Unrelated I2C transfers 113 may creep in and close prematurely. 114 115ML4. If any non-I2C operation in the mux driver changes the I2C mux state, 116 the driver has to lock the root adapter during that operation. 117 Otherwise garbage may appear on the bus as seen from devices 118 behind the mux, when an unrelated I2C transfer is in flight during 119 the non-I2C mux-changing operation. 120==== ===================================================================== 121 122 123Mux-locked Example 124------------------ 125 126 127:: 128 129 .----------. .--------. 130 .--------. | mux- |-----| dev D1 | 131 | root |--+--| locked | '--------' 132 '--------' | | mux M1 |--. .--------. 133 | '----------' '--| dev D2 | 134 | .--------. '--------' 135 '--| dev D3 | 136 '--------' 137 138When there is an access to D1, this happens: 139 140 1. Someone issues an I2C transfer to D1. 141 2. M1 locks muxes on its parent (the root adapter in this case). 142 3. M1 calls ->select to ready the mux. 143 4. M1 (presumably) does some I2C transfers as part of its select. 144 These transfers are normal I2C transfers that locks the parent 145 adapter. 146 5. M1 feeds the I2C transfer from step 1 to its parent adapter as a 147 normal I2C transfer that locks the parent adapter. 148 6. M1 calls ->deselect, if it has one. 149 7. Same rules as in step 4, but for ->deselect. 150 8. M1 unlocks muxes on its parent. 151 152This means that accesses to D2 are lockout out for the full duration 153of the entire operation. But accesses to D3 are possibly interleaved 154at any point. 155 156 157Parent-locked muxes 158------------------- 159 160Parent-locked muxes lock the parent adapter during the full select- 161transfer-deselect transaction. The implication is that the mux driver 162has to ensure that any and all I2C transfers through that parent 163adapter during the transaction are unlocked I2C transfers (using e.g. 164__i2c_transfer), or a deadlock will follow. There are a couple of 165caveats. 166 167==== ==================================================================== 168PL1. If you build a topology with a parent-locked mux being the child 169 of another mux, this might break a possible assumption from the 170 child mux that the root adapter is unused between its select op 171 and the actual transfer (e.g. if the child mux is auto-closing 172 and the parent mux issues I2C transfers as part of its select). 173 This is especially the case if the parent mux is mux-locked, but 174 it may also happen if the parent mux is parent-locked. 175 176PL2. If select/deselect calls out to other subsystems such as gpio, 177 pinctrl, regmap or iio, it is essential that any I2C transfers 178 caused by these subsystems are unlocked. This can be convoluted to 179 accomplish, maybe even impossible if an acceptably clean solution 180 is sought. 181==== ==================================================================== 182 183 184Parent-locked Example 185--------------------- 186 187:: 188 189 .----------. .--------. 190 .--------. | parent- |-----| dev D1 | 191 | root |--+--| locked | '--------' 192 '--------' | | mux M1 |--. .--------. 193 | '----------' '--| dev D2 | 194 | .--------. '--------' 195 '--| dev D3 | 196 '--------' 197 198When there is an access to D1, this happens: 199 200 1. Someone issues an I2C transfer to D1. 201 2. M1 locks muxes on its parent (the root adapter in this case). 202 3. M1 locks its parent adapter. 203 4. M1 calls ->select to ready the mux. 204 5. If M1 does any I2C transfers (on this root adapter) as part of 205 its select, those transfers must be unlocked I2C transfers so 206 that they do not deadlock the root adapter. 207 6. M1 feeds the I2C transfer from step 1 to the root adapter as an 208 unlocked I2C transfer, so that it does not deadlock the parent 209 adapter. 210 7. M1 calls ->deselect, if it has one. 211 8. Same rules as in step 5, but for ->deselect. 212 9. M1 unlocks its parent adapter. 213 10. M1 unlocks muxes on its parent. 214 215 216This means that accesses to both D2 and D3 are locked out for the full 217duration of the entire operation. 218 219 220Complex Examples 221================ 222 223Parent-locked mux as parent of parent-locked mux 224------------------------------------------------ 225 226This is a useful topology, but it can be bad:: 227 228 .----------. .----------. .--------. 229 .--------. | parent- |-----| parent- |-----| dev D1 | 230 | root |--+--| locked | | locked | '--------' 231 '--------' | | mux M1 |--. | mux M2 |--. .--------. 232 | '----------' | '----------' '--| dev D2 | 233 | .--------. | .--------. '--------' 234 '--| dev D4 | '--| dev D3 | 235 '--------' '--------' 236 237When any device is accessed, all other devices are locked out for 238the full duration of the operation (both muxes lock their parent, 239and specifically when M2 requests its parent to lock, M1 passes 240the buck to the root adapter). 241 242This topology is bad if M2 is an auto-closing mux and M1->select 243issues any unlocked I2C transfers on the root adapter that may leak 244through and be seen by the M2 adapter, thus closing M2 prematurely. 245 246 247Mux-locked mux as parent of mux-locked mux 248------------------------------------------ 249 250This is a good topology:: 251 252 .----------. .----------. .--------. 253 .--------. | mux- |-----| mux- |-----| dev D1 | 254 | root |--+--| locked | | locked | '--------' 255 '--------' | | mux M1 |--. | mux M2 |--. .--------. 256 | '----------' | '----------' '--| dev D2 | 257 | .--------. | .--------. '--------' 258 '--| dev D4 | '--| dev D3 | 259 '--------' '--------' 260 261When device D1 is accessed, accesses to D2 are locked out for the 262full duration of the operation (muxes on the top child adapter of M1 263are locked). But accesses to D3 and D4 are possibly interleaved at 264any point. Accesses to D3 locks out D1 and D2, but accesses to D4 265are still possibly interleaved. 266 267 268Mux-locked mux as parent of parent-locked mux 269--------------------------------------------- 270 271This is probably a bad topology:: 272 273 .----------. .----------. .--------. 274 .--------. | mux- |-----| parent- |-----| dev D1 | 275 | root |--+--| locked | | locked | '--------' 276 '--------' | | mux M1 |--. | mux M2 |--. .--------. 277 | '----------' | '----------' '--| dev D2 | 278 | .--------. | .--------. '--------' 279 '--| dev D4 | '--| dev D3 | 280 '--------' '--------' 281 282When device D1 is accessed, accesses to D2 and D3 are locked out 283for the full duration of the operation (M1 locks child muxes on the 284root adapter). But accesses to D4 are possibly interleaved at any 285point. 286 287This kind of topology is generally not suitable and should probably 288be avoided. The reason is that M2 probably assumes that there will 289be no I2C transfers during its calls to ->select and ->deselect, and 290if there are, any such transfers might appear on the slave side of M2 291as partial I2C transfers, i.e. garbage or worse. This might cause 292device lockups and/or other problems. 293 294The topology is especially troublesome if M2 is an auto-closing 295mux. In that case, any interleaved accesses to D4 might close M2 296prematurely, as might any I2C transfers part of M1->select. 297 298But if M2 is not making the above stated assumption, and if M2 is not 299auto-closing, the topology is fine. 300 301 302Parent-locked mux as parent of mux-locked mux 303--------------------------------------------- 304 305This is a good topology:: 306 307 .----------. .----------. .--------. 308 .--------. | parent- |-----| mux- |-----| dev D1 | 309 | root |--+--| locked | | locked | '--------' 310 '--------' | | mux M1 |--. | mux M2 |--. .--------. 311 | '----------' | '----------' '--| dev D2 | 312 | .--------. | .--------. '--------' 313 '--| dev D4 | '--| dev D3 | 314 '--------' '--------' 315 316When D1 is accessed, accesses to D2 are locked out for the full 317duration of the operation (muxes on the top child adapter of M1 318are locked). Accesses to D3 and D4 are possibly interleaved at 319any point, just as is expected for mux-locked muxes. 320 321When D3 or D4 are accessed, everything else is locked out. For D3 322accesses, M1 locks the root adapter. For D4 accesses, the root 323adapter is locked directly. 324 325 326Two mux-locked sibling muxes 327---------------------------- 328 329This is a good topology:: 330 331 .--------. 332 .----------. .--| dev D1 | 333 | mux- |--' '--------' 334 .--| locked | .--------. 335 | | mux M1 |-----| dev D2 | 336 | '----------' '--------' 337 | .----------. .--------. 338 .--------. | | mux- |-----| dev D3 | 339 | root |--+--| locked | '--------' 340 '--------' | | mux M2 |--. .--------. 341 | '----------' '--| dev D4 | 342 | .--------. '--------' 343 '--| dev D5 | 344 '--------' 345 346When D1 is accessed, accesses to D2, D3 and D4 are locked out. But 347accesses to D5 may be interleaved at any time. 348 349 350Two parent-locked sibling muxes 351------------------------------- 352 353This is a good topology:: 354 355 .--------. 356 .----------. .--| dev D1 | 357 | parent- |--' '--------' 358 .--| locked | .--------. 359 | | mux M1 |-----| dev D2 | 360 | '----------' '--------' 361 | .----------. .--------. 362 .--------. | | parent- |-----| dev D3 | 363 | root |--+--| locked | '--------' 364 '--------' | | mux M2 |--. .--------. 365 | '----------' '--| dev D4 | 366 | .--------. '--------' 367 '--| dev D5 | 368 '--------' 369 370When any device is accessed, accesses to all other devices are locked 371out. 372 373 374Mux-locked and parent-locked sibling muxes 375------------------------------------------ 376 377This is a good topology:: 378 379 .--------. 380 .----------. .--| dev D1 | 381 | mux- |--' '--------' 382 .--| locked | .--------. 383 | | mux M1 |-----| dev D2 | 384 | '----------' '--------' 385 | .----------. .--------. 386 .--------. | | parent- |-----| dev D3 | 387 | root |--+--| locked | '--------' 388 '--------' | | mux M2 |--. .--------. 389 | '----------' '--| dev D4 | 390 | .--------. '--------' 391 '--| dev D5 | 392 '--------' 393 394When D1 or D2 are accessed, accesses to D3 and D4 are locked out while 395accesses to D5 may interleave. When D3 or D4 are accessed, accesses to 396all other devices are locked out. 397