1 <html><head><title>The design of toybox</title></head> 2 <!--#include file="header.html" --> 3 4 <b><h2>Design goals</h2></b> 5 6 <p>Toybox should be simple, small, fast, and full featured. Often, these 7 things need to be balanced off against each other. In general, keeping the 8 code simple the most important (and hardest) goal, and small is slightly more 9 important than fast. Features are the reason we write code in the first 10 place but this has all been implemented before so if we can't do a better 11 job why bother? It should be possible to get 80% of the way to each goal 12 before they really start to fight.</p> 13 14 <p>Here they are in reverse order of importance:</p> 15 16 <b><h3>Features</h3></b> 17 18 <p>The <a href=roadmap.html>roadmap</a> has the list of features we're 19 trying to implement, and the reasons for them. After the 1.0 release 20 some of that material may get moved here.</p> 21 22 <p>Some things are simply outside the scope of the project: even though 23 posix defines commands for compiling and linking, we're not going to include 24 a compiler or linker (and support for a potentially infinite number of hardware 25 targets). And until somebody comes up with a ~30k ssh implementation, we're 26 going to point you at dropbear or polarssl.</p> 27 28 <p>Environmental dependencies are a type of complexity, so needing other 29 packages to build or run is a big downside. For example, we don't use curses 30 when we can simply output ansi escape sequences and trust all terminal 31 programs written in the past 30 years to be able to support them. (A common 32 use case is to download a statically linked toybox binary to an arbitrary 33 Linux system, and use it in an otherwise unknown environment; being 34 self-contained helps support this.)</p> 35 36 <b><h3>Speed</h3></b> 37 38 <p>It's easy to say lots about optimizing for speed (which is why this section 39 is so long), but at the same time it's the optimization we care the least about. 40 The essence of speed is being as efficient as possible, which means doing as 41 little work as possible. A design that's small and simple gets you 90% of the 42 way there, and most of the rest is either fine-tuning or more trouble than 43 it's worth (and often actually counterproductive). Still, here's some 44 advice:</p> 45 46 <p>First, understand the darn problem you're trying to solve. You'd think 47 I wouldn't have to say this, but I do. Trying to find a faster sorting 48 algorithm is no substitute for figuring out a way to skip the sorting step 49 entirely. The fastest way to do anything is not to have to do it at all, 50 and _all_ optimization boils down to avoiding unnecessary work.</p> 51 52 <p>Speed is easy to measure; there are dozens of profiling tools for Linux 53 (although personally I find the "time" command a good starting place). 54 Don't waste too much time trying to optimize something you can't measure, 55 and there's no much point speeding up things you don't spend much time doing 56 anyway.</p> 57 58 <p>Understand the difference between throughput and latency. Faster 59 processors improve throughput, but don't always do much for latency. 60 After 30 years of Moore's Law, most of the remaining problems are latency, 61 not throughput. (There are of course a few exceptions, like data compression 62 code, encryption, rsync...) Worry about throughput inside long-running 63 loops, and worry about latency everywhere else. (And don't worry too much 64 about avoiding system calls or function calls or anything else in the name 65 of speed unless you are in the middle of a tight loop that's you've already 66 proven isn't running fast enough.)</p> 67 68 <p>"Locality of reference" is generally nice, in all sorts of contexts. 69 It's obvious that waiting for disk access is 1000x slower than doing stuff in 70 RAM (and making the disk seek is 10x slower than sequential reads/writes), 71 but it's just as true that a loop which stays in L1 cache is many times faster 72 than a loop that has to wait for a DRAM fetch on each iteration. Don't worry 73 about whether "&" is faster than "%" until your executable loop stays in L1 74 cache and the data access is fetching cache lines intelligently. (To 75 understand DRAM, L1, and L2 cache, read Hannibal's marvelous ram guide at Ars 76 Technica: 77 <a href=http://arstechnica.com/paedia/r/ram_guide/ram_guide.part1-2.html>part one</a>, 78 <a href=http://arstechnica.com/paedia/r/ram_guide/ram_guide.part2-1.html>part two</a>, 79 <a href=http://arstechnica.com/paedia/r/ram_guide/ram_guide.part3-1.html>part three</a>, 80 plus this 81 <a href=http://arstechnica.com/articles/paedia/cpu/caching.ars/1>article on 82 cacheing</a>, and this one on 83 <a href=http://arstechnica.com/articles/paedia/cpu/bandwidth-latency.ars>bandwidth 84 and latency</a>. 85 And there's <a href=http://arstechnica.com/paedia/index.html>more where that came from</a>.) 86 Running out of L1 cache can execute one instruction per clock cycle, going 87 to L2 cache costs a dozen or so clock cycles, and waiting for a worst case dram 88 fetch (round trip latency with a bank switch) can cost thousands of 89 clock cycles. (Historically, this disparity has gotten worse with time, 90 just like the speed hit for swapping to disk. These days, a _big_ L1 cache 91 is 128k and a big L2 cache is a couple of megabytes. A cheap low-power 92 embedded processor may have 8k of L1 cache and no L2.)</p> 93 94 <p>Learn how virtual memory and memory managment units work. Don't touch 95 memory you don't have to. Even just reading memory evicts stuff from L1 and L2 96 cache, which may have to be read back in later. Writing memory can force the 97 operating system to break copy-on-write, which allocates more memory. (The 98 memory returned by malloc() is only a virtual allocation, filled with lots of 99 copy-on-write mappings of the zero page. Actual physical pages get allocated 100 when the copy-on-write gets broken by writing to the virtual page. This 101 is why checking the return value of malloc() isn't very useful anymore, it 102 only detects running out of virtual memory, not physical memory. Unless 103 you're using a NOMMU system, where all bets are off.)</p> 104 105 <p>Don't think that just because you don't have a swap file the system can't 106 start swap thrashing: any file backed page (ala mmap) can be evicted, and 107 there's a reason all running programs require an executable file (they're 108 mmaped, and can be flushed back to disk when memory is short). And long 109 before that, disk cache gets reclaimed and has to be read back in. When the 110 operating system really can't free up any more pages it triggers the out of 111 memory killer to free up pages by killing processes (the alternative is the 112 entire OS freezing solid). Modern operating systems seldom run out of 113 memory gracefully.</p> 114 115 <p>Also, it's better to be simple than clever. Many people think that mmap() 116 is faster than read() because it avoids a copy, but twiddling with the memory 117 management is itself slow, and can cause unnecessary CPU cache flushes. And 118 if a read faults in dozens of pages sequentially, but your mmap iterates 119 backwards through a file (causing lots of seeks, each of which your program 120 blocks waiting for), the read can be many times faster. On the other hand, the 121 mmap can sometimes use less memory, since the memory provided by mmap 122 comes from the page cache (allocated anyway), and it can be faster if you're 123 doing a lot of different updates to the same area. The moral? Measure, then 124 try to speed things up, and measure again to confirm it actually _did_ speed 125 things up rather than made them worse. (And understanding what's really going 126 on underneath is a big help to making it happen faster.)</p> 127 128 <p>In general, being simple is better than being clever. Optimization 129 strategies change with time. For example, decades ago precalculating a table 130 of results (for things like isdigit() or cosine(int degrees)) was clearly 131 faster because processors were so slow. Then processors got faster and grew 132 math coprocessors, and calculating the value each time became faster than 133 the table lookup (because the calculation fit in L1 cache but the lookup 134 had to go out to DRAM). Then cache sizes got bigger (the Pentium M has 135 2 megabytes of L2 cache) and the table fit in cache, so the table became 136 fast again... Predicting how changes in hardware will affect your algorithm 137 is difficult, and using ten year old optimization advice and produce 138 laughably bad results. But being simple and efficient is always going to 139 give at least a reasonable result.</p> 140 141 <p>The famous quote from Ken Thompson, "When in doubt, use brute force", 142 applies to toybox. Do the simple thing first, do as little of it as possible, 143 and make sure it's right. You can always speed it up later.</p> 144 145 <b><h3>Size</h3></b> 146 <p>Again, simple gives you most of this. An algorithm that does less work 147 is generally smaller. Understand the problem, treat size as a cost, and 148 get a good bang for the byte.</p> 149 150 <p>Understand the difference between binary size, heap size, and stack size. 151 Your binary is the executable file on disk, your heap is where malloc() memory 152 lives, and your stack is where local variables (and function call return 153 addresses) live. Optimizing for binary size is generally good: executing 154 fewer instructions makes your program run faster (and fits more of it in 155 cache). On embedded systems, binary size is especially precious because 156 flash is expensive (and its successor, MRAM, even more so). Small stack size 157 is important for nommu systems because they have to preallocate their stack 158 and can't make it bigger via page fault. And everybody likes a small heap.</p> 159 160 <p>Measure the right things. Especially with modern optimizers, expecting 161 something to be smaller is no guarantee it will be after the compiler's done 162 with it. Binary size isn't the most accurate indicator of the impact of a 163 given change, because lots of things get combined and rounded during 164 compilation and linking. Matt Mackall's bloat-o-meter is a python script 165 which compares two versions of a program, and shows size changes in each 166 symbol (using the "nm" command behind the scenes). To use this, run 167 "make baseline" to build a baseline version to compare against, and 168 then "make bloatometer" to compare that baseline version against the current 169 code.</p> 170 171 <p>Avoid special cases. Whenever you see similar chunks of code in more than 172 one place, it might be possible to combine them and have the users call shared 173 code. (This is the most commonly cited trick, which doesn't make it easy. If 174 seeing two lines of code do the same thing makes you slightly uncomfortable, 175 you've got the right mindset.)</p> 176 177 <p>Some specific advice: Using a char in place of an int when doing math 178 produces significantly larger code on some platforms (notably arm), 179 because each time the compiler has to emit code to convert it to int, do the 180 math, and convert it back. Bitfields have this problem on most platforms. 181 Because of this, using char to index a for() loop is probably not a net win, 182 although using char (or a bitfield) to store a value in a structure that's 183 repeated hundreds of times can be a good tradeoff of binary size for heap 184 space.</p> 185 186 <b><h3>Simple</h3></b> 187 188 <p>Complexity is a cost, just like code size or runtime speed. Treat it as 189 a cost, and spend your complexity budget wisely. (Sometimes this means you 190 can't afford a feature because it complicates the code too much to be 191 worth it.)</p> 192 193 <p>Simplicity has lots of benefits. Simple code is easy to maintain, easy to 194 port to new processors, easy to audit for security holes, and easy to 195 understand.</p> 196 197 <p>Simplicity itself can have subtle non-obvious aspects requiring a tradeoff 198 between one kind of simplicity and another: simple for the computer to 199 execute and simple for a human reader to understand aren't always the 200 same thing. A compact and clever algorithm that does very little work may 201 not be as easy to explain or understand as a larger more explicit version 202 requiring more code, memory, and CPU time. When balancing these, err on the 203 side of doing less work, but add comments describing how you 204 could be more explicit.</p> 205 206 <p>In general, comments are not a substitute for good code (or well chosen 207 variable or function names). Commenting "x += y;" with "/* add y to x */" 208 can actually detract from the program's readability. If you need to describe 209 what the code is doing (rather than _why_ it's doing it), that means the 210 code itself isn't very clear.</p> 211 212 <p>Prioritizing simplicity tends to serve our other goals: simplifying code 213 generally reduces its size (both in terms of binary size and runtime memory 214 usage), and avoiding unnecessary work makes code run faster. Smaller code 215 also tends to run faster on modern hardware due to CPU cacheing: fitting your 216 code into L1 cache is great, and staying in L2 cache is still pretty good.</p> 217 218 <p><a href=http://www.joelonsoftware.com/articles/fog0000000069.html>Joel 219 Spolsky argues against throwing code out and starting over</a>, and he has 220 good points: an existing debugged codebase contains a huge amount of baked 221 in knowledge about strange real-world use cases that the designers didn't 222 know about until users hit the bugs, and most of this knowledge is never 223 explicitly stated anywhere except in the source code.</p> 224 225 <p>That said, the Mythical Man-Month's "build one to throw away" advice points 226 out that until you've solved the problem you don't properly understand it, and 227 about the time you finish your first version is when you've finally figured 228 out what you _should_ have done. (The corrolary is that if you build one 229 expecting to throw it away, you'll actually wind up throwing away two. You 230 don't understand the problem until you _have_ solved it.)</p> 231 232 <p>Joel is talking about what closed source software can afford to do: Code 233 that works and has been paid for is a corporate asset not lightly abandoned. 234 Open source software can afford to re-implement code that works, over and 235 over from scratch, for incremental gains. Before toybox, the unix command line 236 has already been reimplemented from scratch several times in a row (the 237 original AT&T Unix command line in assembly and then in C, the BSD 238 versions, the GNU tools, BusyBox...) but maybe toybox can do a better job. :)</p> 239 240 <p>P.S. How could I resist linking to an article about 241 <a href=http://blog.outer-court.com/archive/2005-08-24-n14.html>why 242 programmers should strive to be lazy and dumb</a>?</p> 243 244 <b><h2>Portability issues</h2></b> 245 246 <b><h3>Platforms</h3></b> 247 <p>Toybox should run on Android (all commands with musl-libc, as large a subset 248 as practical with bionic), and every other hardware platform Linux runs on. 249 Other posix/susv4 environments (perhaps MacOS X or newlib+libgloss) are vaguely 250 interesting but only if they're easy to support; I'm not going to spend much 251 effort on them.</p> 252 253 <p>I don't do windows.</p> 254 255 <b><h3>32/64 bit</h3></b> 256 <p>Toybox should work on both 32 bit and 64 bit systems. By the end of 2008 257 64 bit hardware will be the new desktop standard, but 32 bit hardware will 258 continue to be important in embedded devices for years to come.</p> 259 260 <p>Toybox relies on the fact that on any Unix-like platform, pointer and long 261 are always the same size (on both 32 and 64 bit). Pointer and int are _not_ 262 the same size on 64 bit systems, but pointer and long are.</p> 263 264 <p>This is guaranteed by the LP64 memory model, a Unix standard (which Linux 265 and MacOS X both implement, and which modern 64 bit processors such as 266 x86-64 were <a href=http://www.pagetable.com/?p=6>designed for</a>). See 267 <a href=http://www.unix.org/whitepapers/64bit.html>the LP64 standard</a> and 268 <a href=http://www.unix.org/version2/whatsnew/lp64_wp.html>the LP64 269 rationale</a> for details.</p> 270 271 <p>Note that Windows doesn't work like this, and I don't care. 272 <a href=http://blogs.msdn.com/oldnewthing/archive/2005/01/31/363790.aspx>The 273 insane legacy reasons why this is broken on Windows are explained here.</a></p> 274 275 <b><h3>Signedness of char</h3></b> 276 <p>On platforms like x86, variables of type char default to unsigned. On 277 platforms like arm, char defaults to signed. This difference can lead to 278 subtle portability bugs, and to avoid them we specify which one we want by 279 feeding the compiler -funsigned-char.</p> 280 281 <p>The reason to pick "unsigned" is that way we're 8-bit clean by default.</p> 282 283 <p><h3>Error messages and internationalization:</h3></p> 284 <p>Error messages are extremely terse not just to save bytes, but because we 285 don't use any sort of _("string") translation infrastructure.</p> 286 287 <p>Thus "bad -A '%c'" is 288 preferable to "Unrecognized address base '%c'", because a non-english speaker 289 can see that -A was the problem, and with a ~20 word english vocabulary is 290 more likely to know (or guess) "bad" than the longer message.</p> 291 292 <p>The help text might someday have translated versions, and strerror() 293 messages produced by perror_exit() and friends can be expected to be 294 localized by libc. Our error functions also prepend the command name, 295 which non-english speakers can presumably recognize already.</p> 296 297 <p>An enventual goal is <a href=http://yarchive.net/comp/linux/utf8.html>UTF-8</a> support, although it isn't a priority for the 298 first pass of each command. (All commands should at least be 8-bit clean.)</p> 299 300 <p>Locale support isn't currently a goal; that's a presentation layer issue, 301 X11 or Dalvik's problem.</p> 302 303 <a name="codestyle" /> 304 <h2>Coding style</h2> 305 306 <p>The real coding style holy wars are over things that don't matter 307 (whitespace, indentation, curly bracket placement...) and thus have no 308 obviously correct answer. As in academia, "the fighting is so vicious because 309 the stakes are so small". That said, being consistent makes the code readable, 310 so here's how to make toybox code look like other toybox code.</p> 311 312 <p>Toybox source uses two spaces per indentation level, and wraps at 80 313 columns. (Indentation of continuation lines is awkward no matter what 314 you do, sometimes two spaces looks better, sometimes indenting to the 315 contents of a parentheses looks better.)</p> 316 317 <p>There's a space after C flow control statements that look like functions, so 318 "if (blah)" instead of "if(blah)". (Note that sizeof is actually an 319 operator, so we don't give it a space for the same reason ++ doesn't get 320 one. Yeah, it doesn't need the parentheses either, but it gets them. 321 These rules are mostly to make the code look consistent, and thus easier 322 to read.) We also put a space around assignment operators (on both sides), 323 so "int x = 0;".</p> 324 325 <p>Blank lines (vertical whitespace) go between thoughts. "We were doing that, 326 now we're doing this. (Not a hard and fast rule about _where_ it goes, 327 but there should be some.)"</p> 328 329 <p>Variable declarations go at the start of blocks, with a blank line between 330 them and other code. Yes, c99 allows you to put them anywhere, but they're 331 harder to find if you do that. If there's a large enough distance between 332 the declaration and the code using it to make you uncomfortable, maybe the 333 function's too big, or is there an if statement or something you can 334 use as an excuse to start a new closer block?</p> 335 336 <p>If statments with a single line body go on the same line if the result 337 fits in 80 columns, on a second line if it doesn't. We usually only use 338 curly brackets if we need to, either because the body is multiple lines or 339 because we need to distinguish which if an else binds to. Curly brackets go 340 on the same line as the test/loop statement. The exception to both cases is 341 if the test part of an if statement is long enough to split into multiple 342 lines, then we put the curly bracket on its own line afterwards (so it doesn't 343 get lost in the multple line variably indented mess), and we put it there 344 even if it's only grouping one line (because the indentation level is not 345 providing clear information in that case).</p> 346 347 <p>I.E.</p> 348 349 <blockquote> 350 <pre> 351 if (thingy) thingy; 352 else thingy; 353 354 if (thingy) { 355 thingy; 356 thingy; 357 } else thingy; 358 359 if (blah blah blah... 360 && blah blah blah) 361 { 362 thingy; 363 } 364 </pre></blockquote> 365 366 <p>Gotos are allowed for error handling, and for breaking out of 367 nested loops. In general, a goto should only jump forward (not back), and 368 should either jump to the end of an outer loop, or to error handling code 369 at the end of the function. Goto labels are never indented: they override the 370 block structure of the file. Putting them at the left edge makes them easy 371 to spot as overrides to the normal flow of control, which they are.</p> 372 373 <p>When there's a shorter way to say something, we tend to do that for 374 consistency. For example, we tend to say "*blah" instead of "blah[0]" unless 375 we're referring to more than one element of blah. Similarly, NULL is 376 really just 0 (and C will automatically typecast 0 to anything, except in 377 varargs), "if (function() != NULL)" is the same as "if (function())", 378 "x = (blah == NULL);" is "x = !blah;", and so on.</p> 379 380 <p>The goal is to be 381 concise, not cryptic: if you're worried about the code being hard to 382 understand, splitting it to multiple steps on multiple lines is 383 better than a NOP operation like "!= NULL". A common sign of trying to 384 hard is nesting ? : three levels deep, sometimes if/else and a temporary 385 variable is just plain easier to read. If you think you need a comment, 386 you may be right.</p> 387 388 <p>Comments are nice, but don't overdo it. Comments should explain _why_, 389 not how. If the code doesn't make the how part obvious, that's a problem with 390 the code. Sometimes choosing a better variable name is more revealing than a 391 comment. Comments on their own line are better than comments on the end of 392 lines, and they usually have a blank line before them. Most of toybox's 393 comments are c99 style // single line comments, even when there's more than 394 one of them. The /* multiline */ style is used at the start for the metadata, 395 but not so much in the code itself. They don't nest cleanly, are easy to leave 396 accidentally unterminated, need extra nonfunctional * to look right, and if 397 you need _that_ much explanation maybe what you really need is a URL citation 398 linking to a standards document? Long comments can fall out of sync with what 399 the code is doing. Comments do not get regression tested. There's no such 400 thing as self-documenting code (if nothing else, code with _no_ comments 401 is a bit unfriendly to new readers), but "chocolate sauce isn't the answer 402 to bad cooking" either. Don't use comments as a crutch to explain unclear 403 code if the code can be fixed.</p> 404 405 <!--#include file="footer.html" --> 406