The N32H7xx branch of drv_hard_i2c.c had never been compiled. The file takes
its peripheral declarations from board.h, and only n32h497zgl7-evb includes
n32h49x_i2c.h, so on H7xx every I2C symbol was undefined and the option could
not build at all. Include the peripheral header from the driver itself, the way
drv_sdram.c already does.
I2C_ABORT_ON_TIMEOUT also degraded to a bare `state = I2C_READY` on H7xx, at the
four rt_completion_wait() timeouts in n32_i2c_master_xfer(). The H49x
implementation is written against the older IP - I2C_ClrIntPendingBit,
I2C_STS2_BUSY, I2C_CTRL1_STOPGEN - none of which exist in the H7xx SDK, so give
H7xx its own teardown. It disables the interrupt and DMA request sources,
disables the DMA channel that would otherwise stay armed against the caller's
buffer, releases the bus and clears the status residue.
Only flags with a matching INTCLR bit may be passed to I2C_ClrFlag(), which is a
plain "INTCLR |= flag" write; TFC (0x40) and BUSY (0x8000) have none.
I2C_Configuration() was called from n32_i2c_init() but defined nowhere in the
tree. It is a board hook for pin mux, which is not the driver's job - the H49x
path configures no pins either - so drop the call.
Verified by building n32h760zil7-stb with BSP_USING_HARD_I2C1 and its TX/RX DMA
enabled, which now compiles and links with no warnings, and n32h497zgl7-evb with
I2C1/I2C2 DMA enabled, whose image size is unchanged. Neither board routes I2C
pins yet, so this is build-verified only, not hardware-verified.
drv_sdram.c includes <sdram_port.h> unconditionally under BSP_USING_SDRAM,
and n32h497zgl7-evb had no such header, so the option could not build at
all. Add one for the M12L64164A fitted on this board: 4 banks x 1M x 16
bits = 8 MB, 12 row address bits, 8 column address bits, SDRAM1 mapped at
0x98000000. The timing values are the ones validated on this board in
hardware.
Three of the Kconfig defaults described a different part:
BSP_SDRAM_DATA_WIDTH_SEL default 32 -> 16
BSP_SDRAM_COLUMN_BITS_SEL default 9 -> 8
BSP_SDRAM_SIZE default 32 -> 8
BANK1, ROW_BIT_12 and CAS_LATENCY_3 already matched and are unchanged.
Verified by building n32h497zgl7-evb with BSP_PERIPH_PIN_CFG_READY and
BSP_USING_SDRAM enabled: drv_sdram.c and the SDK's n32h49x_sdram.c compile
and link with no warnings.
- qspi: both XSPI variants poll the status flags with no bound, so an
unresponsive bus hangs the caller and silently defeats the timeout it
passed in. Route every wait through the timeout-carrying helpers,
return -RT_ETIMEOUT and disable the controller on failure, and replace
the "poll then send" loops with one bounded wait per element.
- hard_i2c: the interrupt receive path programs the BYTENUM window once
and never reloads it, so a longer message was truncated and still
reported as success; refuse it with -RT_EINVAL (enable RX DMA for
longer reads) instead of capping XferSize. Stop capping the transmit
byte count at the window size as well: BYTENUM counts received bytes
only, and the cap stopped feeding DAT after byte 255 while TXDATE
stayed asserted. Reject RT_I2C_NO_START/RT_I2C_NO_STOP on the series
without frame chaining, where they would silently become a fresh START
and a STOP, and tear the DMA channel down when the address phase times
out instead of leaving it armed against a buffer about to unwind.
- spi: the H7xx LLI chain advanced the memory side by SPI_DMA_BLOCK_MAX
bytes although BlkTfrSize counts elements, so a 16-bit transfer walked
half a word early per node; advance by elements and scale both the
chunk offset and the staging buffer by the data width. Propagate the
DMA arm failure that was ignored.
- nand: an odd data length has no halfword-only encoding on the 16-bit
bus and padding would program a byte the caller never asked for, so
refuse it before the command phase; issue the empty-page dummy store
as a halfword there too, where 8-bit AHB writes are unsupported.
- Kconfig: gate the Ethernet/SDRAM/LCD/QSPI/NAND options behind
BSP_PERIPH_PIN_CFG_READY on all three boards. The peripheral clock and
IO are not configured for these boards yet, so selecting one of them
builds a driver whose pins stay in the reset function.
The link points to launchpad.net, which shut down its gcc-arm-embedded
downloads years ago (the URL now returns 503), and the tarball it names
is a 2016 Q3 build (gcc 5.4).
- Point to Arm GNU Toolchain Downloads, matching the link already used
in bsp/nxp/imx/imx91/README.md. The old gnu-rm page covers the
GNU Arm Embedded Toolchain, which Arm has superseded.
- Mention installing the toolchain from the distribution package
manager, with Debian/Ubuntu as an example, since rtconfig.py already
defaults EXEC_PATH to /usr/bin. libstdc++-arm-none-eabi-newlib is
listed because LFLAGS links with -lsupc++.
The dependency list in the QEMU (Ubuntu) quick start no longer works on
current Debian/Ubuntu releases:
- gdb-arm-none-eabi no longer exists (dropped since Debian 11 /
Ubuntu 22.04); gdb-multiarch replaces it
- the qemu meta-package has been split up and can no longer be installed
- libncurses5-dev has been removed; libncurses-dev replaces it
- libstdc++-arm-none-eabi-newlib was missing, so linking fails with
"cannot find -lsupc++", which does not point at the missing package
- kconfiglib was not mentioned at all, although scons --menuconfig needs
it; python3-kconfiglib is packaged, and pip install kconfiglib is
rejected by PEP 668 on recent releases
Merge the eleven single-package commands into one, add a table mapping
the old package names to the current ones, and note the kconfiglib
requirement where menuconfig is first used.
Verified with apt-get -s install on Debian 13.
NUM_MEM_REGIONS / NUM_CONFIGURABLE_REGIONS / NUM_EXCLUSIVE_REGIONS had no
Kconfig defaults. Enabling RT_USING_HW_STACK_GUARD without explicitly
setting them generated empty macros in rtconfig.h (e.g.
'#define NUM_CONFIGURABLE_REGIONS'), which expanded into invalid C
expressions like '(2 + )' and 'for (i = 0U; i < ; i++)', breaking
compilation for cortex-m4/m7/m33 MPU code.
Add safe defaults: NUM_MEM_REGIONS=8, NUM_CONFIGURABLE_REGIONS=0,
NUM_EXCLUSIVE_REGIONS=0. Also document the MemManage_Handler strong-symbol
conflict with CubeMX-generated stm32f4xx_it.c and how to resolve it.
Reported-by: wdfk-prog
Without CXX = rtconfig.CXX the scons Environment falls back to the
default host 'g++', and CXXFLAGS stays an empty CLVar, so the C++
group flags appended by DefineGroup end up as a single unquoted
argument and the build fails as soon as RT_USING_CPLUSPLUS is on:
g++: error: -fno-exceptions -fno-rtti ... -Wl,--gc-sections:
linker input file not found: No such file or directory
Found by the CI C++ ctor (.init_array) check.
The terminal subsystem's own headers (terminal.h, tty_config.h, ...) are
resolved by the compiler's quoted-include directory lookup for sources
inside components/lwp/terminal; nothing else depends on the path.
Stop polluting the global CPPPATH with this internal directory,
following the convention discussed in #10980.
When sal_socket() rejects a request, DFS v2 descriptor cleanup destroys the dfs_file but leaves its separately allocated vnode behind.
Destroy the vnode before releasing the failed descriptor and add a regression test that verifies repeated failures do not consume heap.
Fixes#11710
Assisted-by: OpenAI Codex:GPT-5
Grow maps without a 256 VMA cap, skip RSS walks for maps, report
matching status text, real RSS/statm, MemAvailable after page-pool
free pages, load averages plus last PID, cumulative processes, RV32
cpuinfo, and per-character mount escaping.
NULL read/write buffers now return -EFAULT. tmpfs setattr uses the
vnode-backed object after unlink, keeps S_ISUID/S_ISGID/S_ISVTX, and
the tmpfs utest checks ENAMETOOLONG plus post-unlink fchmod.
luchenxu
Leeto
wdfk-prog
OnlyoutzZ
yans6666
CYFS
guanxi.cgh
moment-NEW
moment-NEW
Cheng Li
CYFS3
yandld
zhangyuyang
hanzhijian
kaidegit
TTLLDD
wyf-777
wyf-777
bernard