mmap(2) / madvise(2) / msync(2) / mincore(2) for node.js revisited.
I needed shared memory mapping and came across @bnoordhuis module node-mmap, only to find that it didn't work with later versions of io.js, node.js and compatibles. So out of need I threw this together along with the functionality I found was missing in the node-mmap: advice and sync.
Strong temptations to re-order arguments to something more sane was kept at bay, and I kept it as mmap(2) and node-mmap for compatibility. Notable difference is the additional optional argument to pass a usage advise in the mapping stage. I've given advise and sync more practical arguments, out of a node.js perspective, compared to their C/C++ counterparts.
The flag constants have crooked names from C/C++ retained in order to make it straight forward for the user to search the net, and relate to man-pages.
This is my first node.js addon and after hours wasted reading up on V8 API I luckily stumbled upon Native Abstractions for Node. Makes life so much easier. Hot tip!
mmap-io is written in C++11 and LiveScript — although I love LS, it's more prudent to use TypeScript for a library, so I've rewritten that code.
It should be noted that mem-mapping is by nature potentially blocking, and should not be used in concurrent serving/processing applications, but rather has it's niche where multiple processes are working on the same giant sets of data (thereby sparing physical memory, and load times if the kernel does it's job for read ahead), preferably multiple readers and single or none concurrent writer, to not spoil the gains by shitloads of spin-locks, mutexes or such. And your noble specific use case of course.
- rewritten the C++-code to catch up with V8/Nan breaking changes for node.js 12.*, which also removes all warnings in earlier versions.
- refactored in to wrapper functions for extracting values, so should new breaking changes come in later versions, it will be quicker to adjust.
- major "package.json"-fuckups. Unless running build manually, the js-files where never packaged, nor built. Now whitelisted in package.json.
- major fuckup 2: when they finally where packaged, they we're killed off when the C++ module was rebuilt on installment. So. Finally: js-files are now built to "dist", and the binary into "build". This way the TypeScript and LiveScript aren't required for end user
- when replacing GNU Make, for some reason I used
yarn
in "package.json" — which may have failed builds for those not having it installed (and then not building "es-release"), and completely missing the point of getting rid of Make - updated README to reflect new build command (
npm run build
) (should only ever be needed if you clone from git and contribute) - added back the "main" entry in package.json. Hell of a blunder! Tests changed to import from root so they fail without it.
- the never before tested example code here in the README, has been ran and corrected, thanks to @LiamKarlMitchell
- rewrote the es part of the lib code from LiveScript to TypeScript. The prudent thing to do in a lib. The test remains in LS.
offs_t
changed tosize_t
because of bitwidth goofyness. Thanks to @bmarkiv- removed dependency on GNU Make by adding build commands to "package.json". Might help those on windows platform who didn't have it installed. However they still rely on a horde of "common posix utils", so your setup might be lacking anyway then.
- _note that there are some compile warnings because of changes in C++ API's in NAN/V8, ignored for now in contemplation whether to switch to node-addon-api (napi for C++), since call overhead isn't an issue in this library, everything considered.
- bumped the version to 1.0.0 since, well, why not.
- changed deprecated calls from
ForceSet
toDefineOwnProperty
- general source noise cleanups
- fixed compilation errors for newer nodejs. Thanks to @djulien
- windows-specific problems (#5, #6), and more, fixed. Thanks to @bkmartinjr
- updated README clarifying Contribution Guidelines
- compilation fixes 10.8 OSX fix. Thanks to @arrayjam
incore
fix for Mac OS. Thanks to @rustyconover
incore
added. Thanks to @rustyconover
- Compilation on Mac should work now. Thanks to @aeickhoff
- Windows compatibility added. Thanks to @toxicwolf
- Rewrote the bindings to Nan 2.0.9 API version (V8/io/Node hell...)
- Had to remove the error codes to get it working in the time I had available (or rather - didn't have..) — error messages are still there — with code in message instead. Though, usually nothing goes wrong, so only the test cares ;-)
- Added some helpful targets in Makefile
human_errors
,ls
only,build
only, etc. (useful if you wanna hack the module) - Since all functionality that can possibly be is in place, I bumped all the way to 0.8. Not battle tested enough to warrant higher.
- Commented away experimental async read-ahead caching when readahead hint was on. It hasn't broken, but it's an unnecessary risk. Plays safe. You can toy with it yourself if you want to try to milk out some ms performance.
- This is the first public commit, and the code has one day of development put into it as of now. More tests are needed so don't count on it being production ready just yet (but soon).
Use npm or git.
npm install mmap-io
git clone https://github.com/ozra/mmap-io.git
cd mmap-io
npm build
Note: All code in examples are in LiveScript
# Following code is plastic fruit; not t[ae]sted...
mmap = require "mmap-io"
fs = require "fs"
some-file = "./foo.bar"
fd = fs.open-sync some-file, "r"
fd-w = fs.open-sync some-file, "r+"
# In the following comments:
# - `[blah]` denotes optional argument
# - `foo = x` denotes default value for argument
size = fs.fstat-sync(fd).size
rx-prot = mmap.PROT_READ .|. mmap.PROT_EXECUTE
priv = mmap.MAP_SHARED
# map( size, protection, privacy, fd [, offset = 0 [, advise = 0]] ) -> Buffer
buffer = mmap.map size, rx-prot, priv, fd
buffer2 = mmap.map size, mmap.PROT_READ, priv, fd, 0, mmap.MADV_SEQUENTIAL
w-buffer = mmap.map size, mmap.PROT_WRITE, priv, fd-w
# advise( buffer, advise ) -> void
# advise( buffer, offset, length, advise ) -> void
mmap.advise w-buffer, mmap.MADV_RANDOM
# sync( buffer ) -> void
# sync( buffer, offset, length ) -> void
# sync( buffer, is-blocking-sync[, do-page-invalidation = false] ) -> void
# sync( buffer, offset = 0, length = buffer.length [, is-blocking-sync = false [, do-page-invalidation = false]] ) -> void
mmap.sync w-buffer
mmap.sync w-buffer, true
mmap.sync w-buffer, 0, size
mmap.sync w-buffer, 0, size, true
mmap.sync w-buffer, 0, size, true, false
# incore( buffer ) -> [ unmapped-pages Int, mapped-pages Int ]
core-stats = mmap.incore buffer
- Checkout man pages mmap(2), madvise(2), msync(2), mincore(2) for more detailed intell.
- The mappings are automatically unmapped when the buffer is garbage collected.
- Write-mappings need the fd to be opened with "r+", or you'll get a permission error (13).
- If you make a read-only mapping and then ignorantly set a value in the buffer, all hell previously unknown to a JS'er breaks loose (segmentation fault). It is possible to write some devilous code to intercept the SIGSEGV and throw an exception, but let's not do that!
Offset
, and in some caseslength
needs to be a multiple of mmap-io.PAGESIZE (which commonly is 4096)- Huge pages are only supported for anonymous / private mappings (well, in Linux), so I didn't throw in flags for that since I found no use.
- As Ben Noordhuis previously has stated: Supplying hint for a fixed virtual memory adress is kinda moot point in JS, so not supported.
- If you miss a feature - contribute! Or request it in an issue.
- If documentation isn't clear, make an issue.
- Please create a new branch like
fix-{ISSUE_NUMBER}-human-readable-descr-here
orfeature-this-and-that
for your work - PR using that branch
- Have a look at surrounding code for style and follow that
- 4 spaces indent
- 12 spaces initial indent for function bodies in the C++ code
- Yeah it was a "feel it out" of the moment, I'll leave it there.
- Please, when writing an issue or making a PR: tag my handle
@ozra
, so that GitHub adds a notification to the "participating" list (less risk I miss it) `
npm test