Provide a detailed description of design considerations and future plans

README.md

@@ -21,8 +21,14 @@ Args:
    -V, --version       Print version
 ```
 
+We also provide a manpage, which targets OpenBSD-level quality. Incorrect or
+outdated information is a bug.
+
 ## Testing
 
+For testing, there is a shell script "test.sh" included. It sets up a block dev
+and writes (and verifies) a bunch of nonsense to it. It boils down to:
+
 ```
 dd if=/dev/zero of=./disk.img bs=1M count=1024
 losetup -fP ./disk.img
@@ -30,6 +36,75 @@ losetup -a
 losedup -d /dev/loop0
 ```
 
+Writes to a regular files is not currently in scope, although it would simplify
+testing.
+
+## Design Considerations
+
+Most of the sanity checking is currently highly Linux specific. We should
+prefer general/posix solutions that reach **at least** FreeBSD, preferably
+OpenBSD and Solaris as well. FreeBSD did implement procfs, but its a Linux-ism
+and it has since been deprecated. I would prefer not to turn this rather simple
+code into macro-mozaic, as i've seen other similar projects do. After all, this
+is just a juiced-up dd-rewrite at its core.  
+
+**Apple products are unsupported.** Im simply not interested in ensuring
+compatibility with a walled-garden ecosystem. If *you* are, we can change that.  
+
+At the time of writing, my FreeBSD server is down for maintenance, which means
+all of my development and testing is focused on AMD64, AArch64 and AArch32
+Linux.  
+
+In the setup phase of the program, we can absolutely afford to do lots of
+sanity checking via syscalls. Between each block write, we flush the disk
+buffers. These flushes are larger (1 MiB when buffer is full) than the sector
+size (often 4k or 64k) of the flash, so as far as i know, this is a gentle way
+to write flash, and should not incur any significant performance overhead. 1
+MiB is also a multiple of the most common sector sizes. We could write it all
+with no flushing, but that would mean the progress indicator will measure
+buffered writes, which is useless.  
+
+It should be possible to induce optimal block size, but this ide has not been
+explored yet.  
+
+Currently when verifying the written data, we read from both the input file and
+the output block device and do a byte-by-byte comparison. A CRC32 is also
+calculated in the first pass of the input file read. In the second pass (the
+verification stage), we calculate the CRC32 of the block device data and
+compare that to our previous result. **This means that we currently use two
+separate methods of verification**. The program allocates **two** separate
+buffers in the startup phase for comparisons. This will change in the coming
+releases, and we will rely only on the CRC.  
+
+De-allocation is handled in the interrupt vector as well as in the 'sad-paths',
+but ultimately this program can be regarded as samurai-principled. We try to
+handle deallocation, but exit on failure and let the kernel handle the rest.  
+
+The use of a crypto-grade checksumming algorithm was considered, but was
+ultimately rejected in favour of a lookup based CRC32. Its simpler, faster and
+easier to understand (See: [crc32.h](./crc32.h)). We may include a compile-time
+option to disable the lookup table to reduce size for really small targets, but
+we speculate that those targets are already satisfied with busybox-dd.  
+
+We also considered using libudev or any of its analogues, to determine the type
+of block device (spinning or flash), but my somewhat inconclusive research
+indicates that it does not include functionality to determine the medium type
+(usb/sd/sata), which is ultimately what i would like to warn the user about.
+The libudev library is also Linux specific.  
+
+We also need completion scripts for the most common shells. This includes csh,
+bash and zsh. Should be easy enough when we set our minds to it.  
+
+See:
+ - `grep -nE 'BLK[A-Za-z0-9]+' /usr/include/linux/fs.h`  
+   In particular, we're interested in BLKGETSIZE64 and BLKFLSBUF for now.
+
+We can read the device info from:
+ - /sys/class/block/[name]/*
+Like:
+ - /sys/class/block/[name]/device/model
+ - /sys/class/block/[name]/size
+
 ## Inspiration
 
 See: