Hi Jon, Sorry for the (very) late reply. I had some long holidays in between and since this is a difficult topic for me, I kept pushing this to the end of my backlog.
On mer., déc. 18, 2024 at 17:09, Jon Humphreys <[email protected]> wrote: > Mattijs Korpershoek <[email protected]> writes: > >> Hi Jonathan, >> >> Thank you for the patch. >> >> On mar., déc. 17, 2024 at 14:48, Jonathan Humphreys <[email protected]> >> wrote: >> >>> If CONFIG_SET_DFU_ALT_INFO is enabled, the dfu_alt_info environment >>> variable is dynamically set when initializing the DFU entities, which is >>> done as part of normal flow of a DFU operation. >>> >>> The USB DFU boot support will set it's own specific value for dfu_alt_info >>> before performing the DFU operation. This means that if >>> CONFIG_SET_DFU_ALT_INFO is enabled, the dfu_alt_info environment variable >>> that the USB DFU boot path had set is overwritten, causing USB DFU boot to >>> fail. >>> >>> Likewise, if the user sets their own value for dfu_alt_info, say at the >>> U-Boot prompt, it get's overwritten if CONFIG_SET_DFU_ALT_INFO is enabled. >>> >>> This patch will first check that dfu_alt_info isn't already set before >>> calling set_dfu_alt_info(), when CONFIG_SET_DFU_ALT_INFO is enabled. >> >> To me, this is a policy change: before we could override the environment >> via set_dfu_alt_info(). Now we cannot anymore (if "dfu_alt_info" is already >> set in the environment). >> >> Also, it seems that this change goes against the uefi doc which states: >> >> """ >> A string is defined which is to be used for populating the >> dfu_alt_info variable. This string is used by the function >> set_dfu_alt_info. Instead of taking the variable from the environment, >> the capsule update feature requires that the variable be set through >> the function, since that is more robust. Allowing the user to change >> the location of the firmware updates is not a very secure >> practice. Getting this information from the firmware itself is more >> secure, assuming the firmware has been verified by a previous stage >> boot loader. >> """ >> >> See: >> https://docs.u-boot.org/en/latest/develop/uefi/uefi.html#performing-the-update >> >> Moreover, looking at various boards that implement >> set_dfu_alt_info(), we can see different behaviours: >> >> Boards examples that won't override "dfu_alt_info" via >> set_dfu_alt_info() if "dfu_alt_info" is already set via environment >> >> * board/xilinx/zynq/board.c >> * board/emulation/common/qemu_dfu.c >> >> Boards examplesthat will always override the "dfu_alt_info" via >> set_dfu_alt_info(): >> >> * board/libre-computer/aml-a311d-cc/aml-a311d-cc.c >> * board/ti/am62px/evm.c >> >> Since set_dfu_alt_info() is a board specific callback, why can't this >> logic be implemented for boards that want this behaviour change? > > Because I would then need to duplicate the same logic for every board that > wanted both USB DFU boot and EFI capsules to work. And the paramters > passed in do not allow the function to know the use case (am I DFU booting > or updating EFI capsules?). See more below. I understand that duplicating logic for every board you maintain is not optimal, however, it gives each vendor the freedom of implementing their policy. I've added a couple of folks who I think could help giving their opinion on EFI capsules/policy. Heinrich, Ilias, Sugosh, do you have any opinion on this patch? > >> >> Regards, >> >> Mattijs >> >>> >>> Signed-off-by: Jonathan Humphreys <[email protected]> >>> --- >>> drivers/dfu/dfu.c | 7 +++++-- >>> 1 file changed, 5 insertions(+), 2 deletions(-) >>> >>> diff --git a/drivers/dfu/dfu.c b/drivers/dfu/dfu.c >>> index 756569217bb..ab8abae1d89 100644 >>> --- a/drivers/dfu/dfu.c >>> +++ b/drivers/dfu/dfu.c >>> @@ -169,10 +169,13 @@ int dfu_init_env_entities(char *interface, char >>> *devstr) >>> dfu_reinit_needed = false; >>> dfu_alt_info_changed = false; >>> >>> + str_env = env_get("dfu_alt_info"); >>> #ifdef CONFIG_SET_DFU_ALT_INFO >>> - set_dfu_alt_info(interface, devstr); >>> + if (!str_env) { >>> + set_dfu_alt_info(interface, devstr); >>> + str_env = env_get("dfu_alt_info"); >>> + } >>> #endif >>> - str_env = env_get("dfu_alt_info"); >>> if (!str_env) { >>> pr_err("\"dfu_alt_info\" env variable not defined!\n"); >>> return -EINVAL; >>> -- >>> 2.34.1 > > Mattijs, thanks for the thorough reply. I did wrestle a lot with how wide > of a fix to propose for this problem, and in the end, decided on the narrow > fix of simply preventing the overwriting of the variable. > > Yes it is a policy change, but the policy is already unclear, inconsistent, > and confusing, IMO. > > For example: > 1) EFI capsule update wants to very strictly control the dfu alt values > by setting it in set_dfu_alt_info(), but then any other DFU use > case breaks. USB DFU boot is now broken. > 2) The behavior the user sees wrt the dfu_alt_info env variable is very > confusing and non-intuitive. Take this example: > > => env set dfu_alt_info "sf 0:0=exe1.bin raw 0 88000;exe2.bin raw 88000 > 100000" > => env print dfu_alt_info > dfu_alt_info=sf 0:0=exe1.bin raw 0 88000;exe2.bin raw 88000 100000 > => dfu 0 list > DFU alt settings list: > dev: SF alt: 0 name: tiboot3.bin layout: RAW_ADDR > dev: SF alt: 1 name: tispl.bin layout: RAW_ADDR > dev: SF alt: 2 name: u-boot.img layout: RAW_ADDR > => env print dfu_alt_info > dfu_alt_info=sf 0:0=tiboot3.bin raw 0 80000;tispl.bin raw 80000 > 200000;u-boot.img raw 280000 400000 > => > > As you can see, the user set's the dfu_alt_info variable according to their > specific use case, then simply tries to list the DFU alt settings, and > because this code goes through the dfu_init_env_entities() path, it gets > changed to the EFI capsule settings. > > I was hoping to get a simpler fix in now so we can get USB DFU boot working > again, and we can visit the overall policy design next. As you suggest, I > could also push the testing of overwriting into the board specific > set_dfu_alt_info() function, but then I need to duplicate the code in 8 > different places for the TI boards, and other vendors may still have the > problem. I agree that the above behaviour is confusing and I'm reconsidering to take up this patch. I'd like some buy-in from either Heinrich, Ilias or Sughosh on this since I'm not 100% confortable with the "policy change" > > Looking to the longer term solution, here are my thoughts. > 1) We need to decouple CONFIG_SET_DFU_ALT_INFO from EFI capsules. The only > reason TI boards are now setting CONFIG_SET_DFU_ALT_INFO is because EFI > capsule update is enabled. Outside of a few legacy uses (I think - it > appears they were introduced prior to supporting multi-interface dfu alt > strings), I think this is true for other vendor's boards as well. > 2) Have EFI capsule support do as USB DFU boot does today, and set the > dfu alt string it wants used *before* initiating the DFU operation. With > CONFIG_SET_DFU_ALT_INFO no longer enabled, the value it set will not get > overridden. > 3) Have the actual value of the dfu alt string used in the DFU operation be > passed in, rather than read from the dfu_alt_info environment variable. > The USB DFU and EFI capsule use case will pass in the dfu alt string > they want. The standard 'dfu' command can pass in the value of the > dfu_alt_info env variable. Note that this effectively decouples the dfu > command from the alt settings that USB DFU boot and EFI capsules use, > but I think this is what we want. > > This then allows both USB DFU boot and EFI capsule use cases to work as > intended and allows the dfu command to operate on the user defined > dfu_alt_info value. > > I welcome comments from those that have the history and intended behavior > background of the DFU support :). I do as well. I have taken over maintaince on this subsystem a year ago and have not had much patches/work done on the subsystem. Therefore I'm not as knowledgeable as I would have liked to be. I'm sorry about that. > > I also welcome comments on how to proceed for 2025.01. Should we live with > USB DFU boot broken until we get the long term fix in, or ok with the patch > posted here. The patch posted here does allow for a user to change EFI > capsule's dfu alt settings, as Mattijs says, but especially given capsules > can be authenticated, I'm not sure how this would be exploited, and if that > risk is worse that broken DFU boot. > > thank > Jon
