Feature: Flash blank check #1971
Conversation
src/target/target_flash.c
Outdated
| platform_timeout_s timeout; | ||
| platform_timeout_set(&timeout, 500); | ||
|
|
||
| while (len) { |
There was a problem hiding this comment.
We think this would be better written as a for loop - for (size_t offset = 0U; offset < len; offset += flash->writebufsize) - so there aren't any weird end of loop conditions and src and len get to be const.
There was a problem hiding this comment.
That was for const size_t local_len = MIN(flash->writebufsize - offset, len); in case a flash bank capacity is not an even/integer multiple of writebufsize. If there is such a guarantee, then I think I can simplify into for-loop.
There was a problem hiding this comment.
That is a guarantee you can depend on. Even if it wasn't, MIN(len - offset, flash->writebufsize) for local_len would still give you the right answer with the suggested restructure.
There was a problem hiding this comment.
How about this form then? 8c8db53
There was a problem hiding this comment.
That looks considerably better!
src/target/target_flash.c
Outdated
| platform_timeout_set(&timeout, 500); | ||
|
|
||
| while (len) { | ||
| const size_t offset = src % flash->writebufsize; |
There was a problem hiding this comment.
Making this a for loop makes this calculation disappear and eliminates an expensive % call
ALTracer
force-pushed
the
feature/blank-check
branch
2 times, most recently
from
8c8db53 to
aeb352c
Compare
* Reuse flash write buffer for block reads * Compare against erased value and skip to next sector on mismatch * Indicate progress via tc_printf() to gdb_if.c or BMDA stdout
ALTracer
force-pushed
the
feature/blank-check
branch
from
aeb352c to
c1b2c0a
Compare
Detailed description
(gdb) x/16zw 0x08004000etc.)Many alternative gdbservers already implement a similar mechanism -- openocd flash, pyocd, Segger JLinkGDBServer, etc.
I opted to reuse the whatever flash write buffer is allocated in BMP heap for fast block reads (like bmd_crc32). The memcmp could be optimized to run in 4 bytes per comparison, but for a quick PoC 1 byte is enough, I get I/O bounded on swdptap.c (100-150 KiB/s for FS probes).
Tested on BMDA for dual blackpills, as usual, over JTAG transport, example output is
Notice how it walks the singly linked flash bank list backwards, but walks eraseblocks forwards. I erased the BMF from DUT, keeping only the BMD bootloader on it (in first 16 KiB). Output in GDB is similar, but there are progress dots printed before some lines.
Needs testing on smaller probes. Future usage could be skipping vFlashErase-triggered erase delays if reading memory-mapped flash is faster than waiting on it to erase (break-even is different for genuine STM32 with true simultaneous mass-erase vs. The Compatibles which back it with some sequential slow QuadSPI-like flash). It could also have interesting interactions with mass-erase -> load (skips all slow sector erases and starts writing to empty flash immediately), eliminating some unresponsiveness.
Your checklist for this pull request
Closing issues