On Wed, 2024-06-12 at 17:32 +0200, Michal Simek wrote:
On 6/12/24 17:11, Vasileios Amoiridis wrote:
On Wed, 2024-06-12 at 16:47 +0200, Michal Simek wrote:
On 6/12/24 15:57, Vasileios Amoiridis wrote:
On Wed, 2024-06-12 at 15:42 +0200, Michal Simek wrote:
Hi Vasileios,
út 4. 6. 2024 v 15:21 odesílatel Vasileios Amoiridis vassilisamir@gmail.com napsal:
From: Vasileios Amoiridis vasileios.amoiridis@cern.ch
Changes in v2: - Remove duplication of custom hardcoded env_locations[] code. - Add implementation with general arch_env_get_location(op, prio)
v1: https://lore.kernel.org/u-boot/20240522174738.73522-1-vassilisamir@gmail.com...
Vasileios Amoiridis (1): xilinx: Add option to load environment from outside of boot media
board/xilinx/versal-net/board.c | 47 ++++++++++++++-----
board/xilinx/versal/board.c | 47 ++++++++++++++-----
board/xilinx/zynq/board.c | 49 +++++++++++++++----
board/xilinx/zynqmp/zynqmp.c | 55 +++++++++++++++++--
4 files changed, 101 insertions(+), 97 deletions(-)
-- 2.34.1
I have to remove this patch from my queue because it is actually breaking current behavior. CI is reporting an issue with it and I did a closer look and what is happening is that if one location has no valid data it goes and looks at another location from the list.
But that is the desired behavior, if the environment is not in one location to check to the others.
That's how u-boot is written but if device doesn't exist or not accessible by u-boot it is clear that it shouldn't be used and then behavior is correct.
Hmmm, how can a device be not-accessible by u-boot but u-boot still tries to use it? How could that happen?
That Xilinx virtual defconfig are setup in a way that all drivers are enabled via Kconfig but they are not present on all boards.
Ok, now I understand how that happens.
If you look below in defconfig we are have that variables can be in NAND but there is no NAND device on zc702 but still u-boot tries to reach nand and read variables from it.
So, the ideal scenario would be to U-Boot never try to access the device? Because from what I see, is that since the device is not there, U-Boot throws an error, and moves on.
Also, I think that this is a problem of the user, no? If for example you have configure ENV_IS_IN_NAND but there is no NAND device, then with the patch you get a visible error that U-Boot tries to access NAND which is not there.
For our platforms where you can burn boot.bin to any boot medium we had no choice that's why code was written like that that variables are saved all the time to boot medium. Also that I don't want to maintain other defconfigs for slightly different configurations. On products customers should look at our all in one defconfig and disable things which are not needed. And where variables are saved is definitely one of them.
Well, that makes sense.
If device exists and simple varibles are not yet saved there going to different device is IMHO problematic.
Maybe I am still a bit confused, but I think that if you have defined ENV_IS_IN_QSPI and ENV_IS_IN_NAND but the environment is not there then the behavior that we see is actually correct. U-Boot tries:
a) Get env from QSPI but fails because env is not there, b) Get env from NAND but fails because NAND is not there, c) Get env from nowhere.
To me, this looks like the correct behavior. Isn't it?
This sequence is correct but what it is not correct is when you reach nowhere and you run saveenv you can't really save variables even they can be saved to QSPI. They can't be saved to NAND because it is not present on the board.
This can be solved by using the env_select command (cmd/nvedit.c) no?
The issue is that none write initial variable to QSPI that's why there will be all the time bad CRC but location is correct.
On Zynq it behaves like this.
MMC: mmc@e0100000: 0 prio 0 Loading Environment from SPIFlash... SF: Detected n25q128a13 with page size 256 Bytes, erase size 64 KiB, total 16 MiB *** Warning - bad CRC, using default environment
prio 1 Loading Environment from NAND... *** Error - No Valid Environment Area found *** Warning - bad env area, using default environment
prio 2 Loading Environment from SPIFlash... *** Warning - bad CRC, using default environment
prio 3 Loading Environment from nowhere... OK
This is the message that we get as well when this patch is not added.
It means you are booting out of QSPI but qspi is not accessible by u- boot. Correct?
Well, we are booting from QSPI but environment is in eMMC. In our config, we have defined ENV_IS_IN_FAT and ENV_IS_NOWHERE but we have not defined ENV_IS_IN_QSPI.
ok.
prio is my print to show where code is. Qemu boots out in QSPI boot mode and SPI is tried first and because this is xilinx_zynq_virt defconfig drivers/env locations for other devices are present too. That's why it goes over the list and it always ends in nowhere which never fails.
If this runs on real HW then the same behavior is visible and I don't think it is right. I think this solution should be rethought. In product current behavior from our code is not the best and current U-Boot implementation is not allowing flexibility too. We are enabling redundant variables which can be only on the same device but not that you have one copy in QSPI and second in EMMC. That's why I think location should be pretty much read from DT. There is options/u-boot node where location of variables should be described and this should replace env_get_location().
You mean that there should be some type of new U-Boot node that describes where the environment should be located and go and search in this device?
Not sure if node or just dt property which points where that variables are stored.
This doesn't sound too difficult to do. Still, even if this was done, what would the priority be for looking for the environment? For now, the code is written in a way to check for the environment only in the bootmedia. My patch added a functionality that if the environment is not in the bootmedia to check somewhere else. This should be the order also if we had this u-boot node? Check first in the bootmedia and then check if there is a dt property?
I don't think this is going to be that simple. Especially to support redundant variables saved to two different devices.
To my understanding, I think it is enough to have a dt property that says "emmc" or "nand" etc.. that can be read when the env_get_location() is executed and the correct env_location is selected. What am I missing that makes you think that it is not so trivial?
I don't have that much experience with the U-Boot sources themselves so my question might be stupid, I know.
It is a lot of work that's why I think second solution is more reasonable for you which is simply create new Kconfig symbol to disable board env_get_location() implementation.
Thanks, Michal
You mean to basically disable the board env_get_location() that exists in board/xilinx/zynqmp/zynqmp.c for example?
yes. Something like this.
config BOARD_ENV_LOCATION bool "Use board defined ENV location" default y ...
if defined (BOARD_ENV_SETUP/LOCATION) enum env_location env_get_location(enum env_operation op, int prio) { ... } endif
M
This will be accepted by you in upstream?
I think it is reasonable compromise to pretty much disable bootmode media selection for storing variables.
Thanks, Michal
Ok, I can submit a v3 then.
Cheers, Vasilis