FreeCalypso > hg > freecalypso-tools
comparison doc/TIFFS-old-description @ 0:e7502631a0f9
initial import from freecalypso-sw rev 1033:5ab737ac3ad7
| author | Mychaela Falconia <falcon@freecalypso.org> |
|---|---|
| date | Sat, 11 Jun 2016 00:13:35 +0000 |
| parents | |
| children |
comparison
equal
deleted
inserted
replaced
| -1:000000000000 | 0:e7502631a0f9 |
|---|---|
| 1 The description of TIFFS that follows was originally written in the summer of | |
| 2 SE52 (A.D. 2013), before the major TI source discoveries which happened later | |
| 3 that year. The text following the dividing line below has not been edited in | |
| 4 content since it was written; for a newer write-up based on the current source- | |
| 5 enabled understanding and reflecting the current FreeCalypso plans with respect | |
| 6 to this FFS, see TIFFS-Overview. | |
| 7 | |
| 8 ------------------------------------------------------------------------------- | |
| 9 | |
| 10 This is a description, based on reverse engineering, of the flash file system | |
| 11 (FFS) implemented in Pirelli's original firmware for the DP-L10 GSM/WiFi dual | |
| 12 mode mobile phone, and in the Closedmoko GTA0x modem firmware. Not knowing the | |
| 13 "proper" name for this FFS, and needing _some_ identifier to refer to it, I | |
| 14 have named it Mokopir-FFS, from "Moko" and "Pirelli" - sometimes abbreviated | |
| 15 further to MPFFS. | |
| 16 | |
| 17 (I have previously called the FFS in question MysteryFFS; but now that I've | |
| 18 successfully reverse-engineered it, it isn't as much of a mystery any more :-) | |
| 19 | |
| 20 At a high functional level, Mokopir-FFS presents the following features: | |
| 21 | |
| 22 * Has a directory tree structure like UNIX file systems; | |
| 23 | |
| 24 * The file system API that must be implemented inside the proprietary firmware | |
| 25 appears to use UNIX-style pathnames; doing strings on firmware images reveals | |
| 26 pathname strings like these: | |
| 27 | |
| 28 /var/dbg/dar | |
| 29 /gsm/l3/rr_white_list | |
| 30 /gsm/l3/rr_medium_rxlev_thr | |
| 31 /gsm/l3/rr_upper_rxlev_thr | |
| 32 /gsm/l3/shield | |
| 33 | |
| 34 Parsing the corresponding FFS image with tools included in the present | |
| 35 package has confirmed that the directory structure implied by these pathnames | |
| 36 does indeed exist in the FFS. | |
| 37 | |
| 38 * Absolutely no DOS-ish semantics seen anywhere: no 8.3 filenames and no | |
| 39 colon-separated device names (seen in the TSM30 file system source, for | |
| 40 example) are visible in the Closedmoko/Pirelli FFS. | |
| 41 | |
| 42 * File contents are stored uncompressed, but not necessarily contiguous: one | |
| 43 could probably store a file in FFS which is bigger than the flash sector | |
| 44 size, it which case it can never be contiguous in a writable FFS (see below), | |
| 45 and the firmware implementation seems to limit chunk sizes to a fairly small | |
| 46 number: on the Pirelli phones all largish files are divided into chunks of | |
| 47 8 KiB each, and on my GTA02 the largest observed chunk size is only 2 KiB. | |
| 48 | |
| 49 The smaller files, like the IMEI and the firmware ID strings in my GTA02 FFS, | |
| 50 are contiguous. | |
| 51 | |
| 52 * The FFS structure is such that the length of "user" payload data stored in | |
| 53 each chunk (and consequently, in each file) can be known exactly in bytes, | |
| 54 with the files/chunks able to contain arbitrary binary data. (This property | |
| 55 may seem obvious or trivial, as all familiar UNIX and DOS file systems have | |
| 56 it, but contrast with RT-11 for example.) | |
| 57 | |
| 58 * The flash file system is a writable one: the running firmware can create, | |
| 59 delete and overwrite files (and possibly directories too) in the live FFS; | |
| 60 thus the FFS design is such that allows these operations to be performed | |
| 61 within the physical constraints of NOR flash write operations. | |
| 62 | |
| 63 I have reverse-engineered this Mokopir-FFS on a read-only level. What it means | |
| 64 is that I, or anyone else who can read this document and the accompanying | |
| 65 source for the listing/extraction utilities, can take a Mokopir-FFS image read | |
| 66 out of a device and see/extract its full content: the complete directory tree | |
| 67 and the exact binary byte content of all files contained therein. | |
| 68 | |
| 69 However, the knowledge possessed by the present hacker (and conveyed in this | |
| 70 document and the accompanying source code) is NOT sufficient for constructing a | |
| 71 valid Mokopir-FFS image "in vitro" given a tree of directories and files, or | |
| 72 for making modifications to the file or directory content of an existing image | |
| 73 and producing a content-modified image that is also valid; valid as in suitable | |
| 74 for the original proprietary firmware to make its normal read and write | |
| 75 operations without noticing anything amiss. | |
| 76 | |
| 77 Constructing "de novo" Mokopir-FFS images or modifying existing images in such | |
| 78 a way that they remain 100% valid for all read and write operations of the | |
| 79 original proprietary firmware would, at the very minimum, require an | |
| 80 understanding of the meaning of *all* fields of the on-media FFS format. Some | |
| 81 of these fields are still left as "non-understood" for now though: a read-only | |
| 82 implementation can get away with simply ignoring them, but a writer/generator | |
| 83 would have to put *something* in those fields. | |
| 84 | |
| 85 As you read the "read-only" description of the Mokopir-FFS on-media format in | |
| 86 the remainder of this document, it should become fairly obvious which pieces | |
| 87 are missing before our understanding of this FFS can be elevated to a | |
| 88 "writable" level. | |
| 89 | |
| 90 However, when it comes to writing new code to run on the two Calypso phones in | |
| 91 question (Closedmoko and Pirelli), it seems, at least to the present hacker, | |
| 92 that a read-only understanding of Mokopir-FFS should be sufficient: | |
| 93 | |
| 94 * In the case of Closedmoko GTA0x modems, the FFS is seen to contain the IMEI | |
| 95 and the RF calibration data. The format of the former is obvious; the latter | |
| 96 not so much - but in any case, the information of interest is clearly of a | |
| 97 read-only nature. It's difficult to tell (or rather, I haven't bothered to | |
| 98 experiment enough) whether the Closedmoko firmware does any writes to FFS or | |
| 99 if the FFS is treated as read-only outside of the production line environment, | |
| 100 but in any case, it seems to me that for any 3rd party replacement firmware, | |
| 101 the best strategy would be to treat the FFS as a read-only source of IMEI and | |
| 102 RF calibration data, and nothing more. | |
| 103 | |
| 104 * In the case of Pirelli phones, the FFS is used to store user data: sent and | |
| 105 received SMS (and MMS/email/whatever), call history, UI settings, pictures | |
| 106 taken with the camera, and whatever else. It also stores a ton of files | |
| 107 which I can only presume were meant to be immutable except at the time of | |
| 108 firmware updates: graphics for the UI, ringtones, i18n UI strings, and even | |
| 109 "helper" firmware images for the WiFi and VoIP processors. However, no IMEI | |
| 110 or RF calibration data are anywhere to be found in the FFS - instead this | |
| 111 information appears to be stored in the "factory block" at the end of the | |
| 112 flash (in its own sector) outside of the FFS. | |
| 113 | |
| 114 Being able to parse FFS images extracted out of Pirelli phones "in vitro" | |
| 115 allows us to steal some of these helper files (UI artwork, ringtones, | |
| 116 WiFi/VoIP helpers), and some of these might even come useful to firmware | |
| 117 replacement projects, but it seems to me that a replacement firmware would | |
| 118 be better off using its own FFS design for storing user data, and as to | |
| 119 retrieving the original IMEI and RF calibration data, the original FFS isn't | |
| 120 of any use for that anyway. | |
| 121 | |
| 122 ======================= | |
| 123 Moko/Pirelli FFS format | |
| 124 ======================= | |
| 125 | |
| 126 OK, now that I'm done with the introduction, we can get to the actual | |
| 127 Mokopir-FFS format. | |
| 128 | |
| 129 * On the GTA0x modem (or at least on my GTA02; my sample size is 1) the FFS | |
| 130 occupies 7 flash sectors of 64 KiB each at offsets 0x380000 through 0x3E0000, | |
| 131 inclusive. | |
| 132 | |
| 133 (The 4 MiB NOR flash chip used by Closedmoko has an independent R/W bank | |
| 134 division between the first 3 MiB and the last 1 MiB. The first 3 MiB are used | |
| 135 to hold the field-flashable closed firmware images distributed as *.m0 files; | |
| 136 the independent last megabyte holds the FFS, and thus the FW could be | |
| 137 implemented to do FFS writes while running from flash in the main bank. | |
| 138 Less than half of that last megabyte appears to be used for the FFS though; | |
| 139 the rest appears to be unused - blank flash observed.) | |
| 140 | |
| 141 * On the Pirelli the FFS occupies 18 sectors of 256 KiB each at offsets 0 | |
| 142 through 0x440000 (inclusive) of the 2nd flash chip select, the one wired to | |
| 143 nCS3 on the Calypso. | |
| 144 | |
| 145 Each flash sector allocated to FFS begins with the following signature: | |
| 146 | |
| 147 00000000: 46 66 73 23 10 02 xx yy zz FF FF FF FF FF FF FF Ffs#............ | |
| 148 | |
| 149 The bytes shown as xx and yy above serve a non-understood purpose; as a guess, | |
| 150 they may hold some info for the flash wear leveling algorithm: in a "virgin" | |
| 151 FFS image like that found in my GTA02 (which never had a SIM card in it and | |
| 152 never made or received a call) or read out of a "virgin" Pirelli phone that | |
| 153 hasn't seen any active use yet, both of these bytes are FFs, but when I look at | |
| 154 FFS images read out of the Pirelli which I currently use as my everyday-use | |
| 155 cellphone, I see other values in sectors which must have been erased and | |
| 156 rewritten. A read-only implementation can ignore these bytes, as mine does. | |
| 157 | |
| 158 The byte shown as zz is more important though, even to a read-only | |
| 159 implementation. The 3 values I've encountered in this byte so far are AB, BD | |
| 160 and BF. Per my current understanding, in a "healthy" FFS exactly one sector | |
| 161 will have AB in its header, exactly one will have BF, and the rest will have | |
| 162 BD. The meanings are (or appear to be): | |
| 163 | |
| 164 AB: the sector holds a vital data structure which I have called the active | |
| 165 index block; | |
| 166 BD: the sector holds regular data; | |
| 167 BF: the sector is blank except for the header, can be turned into a new AB or | |
| 168 BD. | |
| 169 | |
| 170 (Note that a flash program operation, which can turn 1s into 0s but not the | |
| 171 other way around, can turn BF into either AB or BD - but neither AB nor BD can | |
| 172 be turned into any other valid value.) | |
| 173 | |
| 174 In a "virgin" FFS image (as explained above) the first FFS sector is AB, the | |
| 175 last one is BF, and the ones in between are BDs. | |
| 176 | |
| 177 An FFS read operation (a search for a given pathname, or a listing of all | |
| 178 present directories and files) needs to start with locating the active index | |
| 179 block - the FFS sector with AB in the header. Following this header, which is | |
| 180 treated as being 16 bytes long (almost everything in Mokopir-FFS is aligned on | |
| 181 16-byte boundaries), the active index block contains a linear array of 16-byte | |
| 182 records, each record describing an FFS object: directory, file or file | |
| 183 continuation chunk. | |
| 184 | |
| 185 Here is my current understanding of the 16-byte index block record structure: | |
| 186 | |
| 187 2 bytes: Length of the described chunk in bytes | |
| 188 1 byte: Purpose/meaning not understood, ignored by my current code | |
| 189 1 byte: Object type | |
| 190 2 bytes: Descendant pointer | |
| 191 2 bytes: Sibling pointer | |
| 192 4 bytes: Data pointer | |
| 193 4 bytes: Purpose/meaning not understood, ignored by my current code | |
| 194 | |
| 195 (On the Calypso phones of interest, all multibyte fields are in the native | |
| 196 little-endian byte order of the ARM7TDMI processor.) | |
| 197 | |
| 198 The active index block gets filled with these records as objects are created; | |
| 199 the first record goes right after the 'Ffs#'...AB header (padded to 16 bytes); | |
| 200 the last record (at any given moment) is followed by blank flash for the | |
| 201 remainder of the sector. Records thus appear in the order in which they are | |
| 202 created, which bears no direct relation to the directory tree structure. | |
| 203 | |
| 204 The objects, each described by a record in the index block, are organized into | |
| 205 a tree structure by the descendant and sibling pointers, plus the object type | |
| 206 indicator byte. Let's start with the latter; the following objtype byte values | |
| 207 have been observed: | |
| 208 | |
| 209 00: deleted object - a read-only implementation should ignore everything except | |
| 210 the descendant and sibling pointers. (A write-capable implementation would | |
| 211 need more care - it would need a way of reclaiming dirty flash space taken | |
| 212 up by deleted/overwritten files.) | |
| 213 | |
| 214 E1: a special file - see the description of the /.journal file further down | |
| 215 F1: a regular file (head chunk thereof) | |
| 216 F2: a directory | |
| 217 F4: file continuation chunk (explained below) | |
| 218 | |
| 219 Each record in the index block has an associated chunk in one of the data | |
| 220 sectors; the index record contains fields giving the address and length of this | |
| 221 chunk. The length of a chunk is always a nonzero multiple of 16 bytes, and is | |
| 222 stored (as a number in bytes) in the first 16-bit field of the 16-byte index | |
| 223 entry. The address of each chunk is given by the data pointer field of the | |
| 224 index record, and it is reckoned in 16-byte units (thereby 16-byte alignment is | |
| 225 required) from the beginning of the FFS sector group in the flash address space. | |
| 226 | |
| 227 For objects of type F1 and F2 (regular files and directories) the just-described | |
| 228 chunk begins with the name of the file or subdirectory as a NUL-terminated ASCII | |
| 229 string. This name is just for the current level of the directory tree, just | |
| 230 like in UNIX directories, thus one will have chunk names like gsm, l3, eplmn | |
| 231 etc, rather than /gsm/l3/eplmn. One practical effect is that one can't readily | |
| 232 see pathnames or any of the directory structure by looking at an FFS image as a | |
| 233 raw hex dump; the structure is only revealed when one uses a parsing program | |
| 234 like those which accompany this document. | |
| 235 | |
| 236 In the case of directories, the "chunk" part of the object contains only the | |
| 237 name of the directory itself, padded with FFs to a 16-byte boundary. For | |
| 238 example, an FFS directory named /gsm would be represented by an object | |
| 239 consisting of two flash writes: a 16-byte entry in the active index block, with | |
| 240 the object type byte set to F2, and a corresponding 16-byte chunk in one of the | |
| 241 data sectors, with the 16 bytes containing "gsm", a terminating NUL byte, and | |
| 242 12 FF bytes to pad up to 16. In the case of files, this name may be followed | |
| 243 by the first chunk of file data content, as explained further down. | |
| 244 | |
| 245 In order to parse the FFS directory tree (whether the objective is to dump the | |
| 246 whole thing recursively or to find a specific file given a pathname), one needs | |
| 247 to first (well, after finding the active AB block) find the root directory node. | |
| 248 The root directory object is similar to other directory objects: it has a type | |
| 249 of F2, and an associated chunk of 16 bytes in one of the data sectors. The | |
| 250 latter contains the name of the root node: on the Pirelli it is "/", whereas on | |
| 251 my GTA02 it is "/ffs-root". | |
| 252 | |
| 253 The astute reader should notice that it really makes no sense to store a name | |
| 254 for the root node, and indeed, this name plays no part in the traversal of the | |
| 255 directory tree given an absolute pathname. But instead this name, or rather | |
| 256 its first character, appears to be used for the purpose of locating the root | |
| 257 node itself. At first I had assumed that the index record for the root node is | |
| 258 always the first record in the active index block right after the signature | |
| 259 header - that is how it is in "virgin" FFS images, and also in some quite non- | |
| 260 virgin ones I have pulled from my daily-use Pirelli. Naturally my first version | |
| 261 of the Mokopir-FFS (then called MysteryFFS) extraction utility expected the root | |
| 262 node to always be at index #1. But then I got some additional Pirelli phones, | |
| 263 and discovered that in certain cases, index record #1 is a deleted object (the | |
| 264 original root node which has been deleted), and the new active root node is | |
| 265 somewhere in the middle of the index! | |
| 266 | |
| 267 Thus it appears that in order to find the active root node, one needs to scan | |
| 268 the active index block linearly from the beginning (disregarding the tree | |
| 269 structure pointers in this initial pass), looking for a non-deleted object of | |
| 270 type F2 (a directory) whose corresponding name chunk sports a name beginning | |
| 271 with the '/' character. (Anyone who's been raised in UNIX will immediately | |
| 272 know that the path separator character '/' is the only character other than NUL | |
| 273 that's absolutely forbidden in the individual filenames - so this special | |
| 274 "root node name" is the only case of a '/' character appearing in what would | |
| 275 otherwise be a regular filename.) | |
| 276 | |
| 277 [What causes the root node to be somewhere other than at index #1? I assume it | |
| 278 has to do with the dirty space reclamation / data movement algorithm. In a | |
| 279 "virgin" FFS image the very first sector is the active index block, and the | |
| 280 following sector is the first to hold chunks, beginning with the name chunk of | |
| 281 the root node. Now what happens if all data in that sector aside from the | |
| 282 root node name and some other mostly-static directory names becomes dirty, | |
| 283 i.e., belonging to deleted or overwritten files? How would that flash space | |
| 284 get reclaimed? I assume that the FFS firmware algorithm moves all still-active | |
| 285 chunks to a new flash sector, invalidating the old copies - turning the latter | |
| 286 into deleted objects. The root node will be among them. Then at some point | |
| 287 the active index block is going to fill up too, and will need to be rewritten | |
| 288 into a new sector - at which point the previously-deleted index entries are | |
| 289 omitted and the root node becomes #1 again...] | |
| 290 | |
| 291 Tree structure | |
| 292 | |
| 293 Once the root node has been found, the descendant and sibling pointers are used | |
| 294 to traverse the tree structure. For each directory object, including the root | |
| 295 node, the descendant pointer points to the first child object of this directory: | |
| 296 the first file or subdirectory contained therein. (Descendant and sibling | |
| 297 pointers take the form of index numbers in the active index block. A "nil" | |
| 298 pointer is indicated by all 1s (FFFF) - the usual all-0s NULL pointer convention | |
| 299 couldn't be used because it's flash, where the blank state is all 1s.) If the | |
| 300 descendant pointer of a directory object is nil, that means an empty directory. | |
| 301 The sibling pointer of each file or directory points to its next sibling, i.e., | |
| 302 the next member of the same parent directory. The sibling pointer of the root | |
| 303 node is nil. | |
| 304 | |
| 305 Data content of files | |
| 306 | |
| 307 Objects of type F1 are the head chunks of files. Each file has a head chunk, | |
| 308 and may or may not have continuation chunks. More precisely, the head chunk | |
| 309 may contain only the name (or viewed alternatively, 0 bytes of data), or it may | |
| 310 contain a nonzero number of payload bytes; orthogonally to this variability, | |
| 311 there may or may not be continuation chunk(s) present. | |
| 312 | |
| 313 Continuation chunks | |
| 314 | |
| 315 The descendant pointer of each file head object (the object of type F1, the one | |
| 316 reached by traversing the directory tree) indicates whether or not there are | |
| 317 any continuation chunks present. If this descendant pointer is nil, there are | |
| 318 no continuation chunks; otherwise it points to the first continuation chunk | |
| 319 object. File continuation objects have type F4, don't have any siblings (the | |
| 320 sibling pointer is nil - but see below regarding relocated chunks), and the | |
| 321 descendant pointer of each continuation object points to the next continuation | |
| 322 object, if there is one - nil otherwise. | |
| 323 | |
| 324 Payload data delineation | |
| 325 | |
| 326 Each chunk, whether head or continuation, always has a length that is a nonzero | |
| 327 multiple of 16 bytes. The length of the chunk here means the amount of flash | |
| 328 space it occupies in its data sector - which is NOT equal to the payload data | |
| 329 length. | |
| 330 | |
| 331 The head chunk of each file begins with the filename, terminated by a NUL byte. | |
| 332 If there are any payload data bytes present in this head chunk (I'll explain | |
| 333 momentarily how you would tell), the byte immediately after the NUL that | |
| 334 terminates the filename is the first byte of the payload. In the case of a | |
| 335 continuation chunk, there is no filename and the first byte of the chunk is the | |
| 336 first byte of that chunk's portion of the user data payload. | |
| 337 | |
| 338 Each data-containing chunk (head or continuation) has the following termination | |
| 339 after the last byte of that chunk's payload data: one byte of 00, followed by | |
| 340 however many bytes are needed ([0,15] range) of FFs to pad to a 16-byte | |
| 341 boundary. A file head chunk that has no payload data has the same format as a | |
| 342 directory name chunk: filename followed by its terminating NUL followed by | |
| 343 [0,15] bytes of FFs to pad to the next 16-byte boundary. | |
| 344 | |
| 345 When working with a head chunk, find the beginning of possible payload data (1 | |
| 346 byte after the filename terminating NUL) and find the end per the standard | |
| 347 termination logic: scanning from the end of the chunk, skip FFs until 00 is | |
| 348 found (encountering anything else is an error). If the head chunk has no data, | |
| 349 the effective data length (end_pointer - start_pointer) will be 0 or -1. (The | |
| 350 latter possibility is the most likely, as there will normally be a "shared" 00 | |
| 351 byte, serving as both the filename terminator and the 00 before the padding | |
| 352 FF bytes.) | |
| 353 | |
| 354 Relocated chunks | |
| 355 | |
| 356 Let's go back to the scenario in which a particular data sector is full (no more | |
| 357 usable free space left) and contains a mixture of active and dirty (deleted or | |
| 358 invalidated) data. How does the dirty flash space get reclaimed, so that the | |
| 359 amount of available space (blank flash ready to hold new data) becomes equal to | |
| 360 the total FFS size minus the total size of active files and overhead? It can | |
| 361 only be done by relocating the still-active objects from the full sector to a | |
| 362 new one, invalidating the old copies, and once the old sector consists of | |
| 363 nothing but invalidated data, subjecting it to flash erasure. | |
| 364 | |
| 365 So how do the active FFS objects get relocated from a "condemned" sector to a | |
| 366 new one? If the object is a directory, a new index entry is created, pointing | |
| 367 to the newly relocated name chunk, but it is then made to fit into the old tree | |
| 368 structure without disrupting the latter: the new index entry is added at the | |
| 369 tail of the sibling-chain of the parent directory's descendants, the old index | |
| 370 entry for the same directory is invalidated (as if the directory were rmdir'ed), | |
| 371 and the descendant pointer of the newly written index entry is set to a copy of | |
| 372 the descendant pointer from the old index entry for the same directory. The | |
| 373 same approach is used when the head chunk of a file needs to be relocated; in | |
| 374 both cases a read-only FFS implementation doesn't need to do anything special to | |
| 375 support reading file and directory objects that have been relocated in this | |
| 376 manner. | |
| 377 | |
| 378 However, if the relocated object is a file continuation chunk, then the manner | |
| 379 in which such objects get relocated does affect file reading code. What if a | |
| 380 chunk in the middle of a chain linked by "descend" pointers needs to be moved? | |
| 381 What happens in this case is that the old copy of the chunk gets invalidated | |
| 382 (the object type byte turned to 00) like in the other object relocating cases, | |
| 383 and the sibling pointer of that old index entry (which was originally FFFF as | |
| 384 continuation objects have no siblings) is set to point to the new index entry | |
| 385 for the same chunk. The "descend" pointer in the new index entry is a copy of | |
| 386 that pointer from the old index entry. | |
| 387 | |
| 388 The manner of chunk relocation just described has been observed in the FFS | |
| 389 images read out of my most recent batch of Pirelli phones - the same ones in | |
| 390 which the root directory object is not at index #1. Thinking about it as I | |
| 391 write this, I've realized that the way in which continuation objects get | |
| 392 relocated is exactly the same as for other object types - thus the compaction | |
| 393 code in the firmware doesn't need to examine what object type it is moving. | |
| 394 However, the case of continuation chunk relocation deserves special attention | |
| 395 because it affects a read-only implementation like ours - the utilities whose | |
| 396 source accompanies this document used to fail on these FFS images until I | |
| 397 implemented the following additional handling: | |
| 398 | |
| 399 When following the chunk chain of a file, normally the only object type that's | |
| 400 expected is F4 - any other object type is an error. However, as a result of | |
| 401 chunk relocation, one can also encounter deleted objects, i.e., type == 00. | |
| 402 If such a deleted object is encountered, follow its sibling pointer, which must | |
| 403 be non-nil. | |
| 404 | |
| 405 Journal file | |
| 406 | |
| 407 Every Mokopir-FFS image I've seen so far contains a special file named | |
| 408 /.journal; this file is special in the following ways: | |
| 409 | |
| 410 * The object type byte is E1 instead of F1; | |
| 411 * Unlike regular files, this special file is internally-writable. | |
| 412 | |
| 413 What I mean by the above is that regular files are mostly immutable: once a | |
| 414 file has been created with some data content in the head chunk, it can only be | |
| 415 either appended to (one or more continuation chunks added), or overwritten by | |
| 416 creating a new file with the same name at the same level in the tree hierarchy | |
| 417 and invalidating the old one. But the special /.journal file is different: I | |
| 418 have never observed it to consist of more than the head chunk, and this head | |
| 419 chunk is pre-allocated with some largish and apparently fixed length (4 KiB on | |
| 420 my GTA02, 16 KiB on the Pirelli). This pre-allocated chunk contains what look | |
| 421 like 16-byte records at the beginning (on the first 4-byte boundary after the | |
| 422 NUL terminating the ".journal" name), followed by blank flash for the remainder | |
| 423 of the pre-allocated chunk - so it surely looks like new flash writes happen | |
| 424 within this chunk. | |
| 425 | |
| 426 I do not currently know the purpose of this /.journal file or the meaning of the | |
| 427 records it seems to contain. This understanding would surely be needed if one | |
| 428 wanted to create FFS images from scratch or to implement FFS write operations, | |
| 429 but I reason that a read-only implementation can get away with simply ignoring | |
| 430 this file. I reason that this file can't be necessary in order to parse an FFS | |
| 431 image for reading because one needs to parse the tree structure first in order | |
| 432 to locate this journal file itself. | |
| 433 | |
| 434 ------------------------------------------------------------------------------- | |
| 435 | |
| 436 That's all I can think of right now. If anything is unclear, see the | |
| 437 accompanying source code for the listing/extraction utilities: with the general | |
| 438 explanation given by this document, it should be clear what my code does and | |
| 439 why. And if a given piece of knowledge is found neither in this document nor | |
| 440 in my source code, then I don't know it myself either, and my read-only | |
| 441 Mokopir-FFS implementation makes do without it. | |
| 442 | |
| 443 All knowledge contained herein has been recovered by reverse engineering. | |
| 444 Believe it or not, I have figured it out by staring at the hex dump of FFS | |
| 445 sectors, reasoning about how one could possibly implement an FFS given the | |
| 446 requirement of dynamic writability and the physical constraints of flash memory, | |
| 447 and writing listing/extraction test code iteratively until I got something that | |
| 448 appears to correctly parse all FFS images available to me - the result is the | |
| 449 code in this package. | |
| 450 | |
| 451 I never got as far as attempting to locate the FFS implementation routines | |
| 452 within the proprietary firmware binary code images, and I haven't found an | |
| 453 implementation of this particular FFS in any of the leaked sources yet either. | |
| 454 The TSM30 code doesn't seem to be of any use as its FFS appears to be totally | |
| 455 different. As to the more recently found LoCosto code leak, I found that one a | |
| 456 few days *after* I got the Moko/Pirelli "MysteryFFS" reverse-engineered on my | |
| 457 own, and when I did look at the FFS in the LoCosto code later, I saw what seems | |
| 458 to be a different FFS as well. | |
| 459 | |
| 460 Michael Spacefalcon | |
| 461 SE 52 Mes 16 |
