On Tue, Oct 30, 2018 at 02:44:29PM +0100, Patrick Delaunay wrote:
- Update the function syscon_node_to_regmap() to force bound on syscon uclass and directly use the list of device from DM.
- Remove the static list syscon_list.
This patch avoid issue (crash) when syscon_node_to_regmap() is called before and after reallocation (list content is invalid).
Signed-off-by: Patrick Delaunay patrick.delaunay@st.com
Hi
This patch is a modified (after Simon Glass review) and rebased version on v2018.11-rc3 for [U-Boot] syscon: reset node list syscon_list after relocation http://patchwork.ozlabs.org/patch/983139/
This patch correct a crash see on v2018.11-rc1 with my board stm32mp1 for the command "reset", reproduced on v2018.11-rc3.
The crash is a side effect of 2 patches 1f6ca3f42f6edf143473159297f4c515b1cf36f6 sysreset: syscon: update regmap access to syscon
23471aed5c33e104d6fa64575932577618543bec board_f: Add reset status printing
With the first patch the syscon_node_to_regmap() is called each time that the class sysreset_syscon is probed.
=> in v2018.09 probe is done only when reset is requested
NB: for stm32mp1, the node rcc_reboot in tagged pre-relocated to support reset in pre-reloc phases (allow panic).
With the second patch, U-Boot probes all the sysreset uclass in preloc phase to allow to print the reset status, and the list is initialized in board_f / pre-reloc:
-> print_resetinfo --> uclass_first_device_err(UCLASS_SYSRESET, &dev); ---> syscon_reboot_probe() ----> syscon_node_to_regmap(node) -----> of_syscon_register() -------> list_add_tail(&syscon->list, &syscon_list);
During relocation, the content of syscon_list (static) is updated but the list use pointers allocated by malloc in pre-reloc phasis
And when I execute the reset command -> do_reset() --> reset_cpu() ---> sysreset_walk_halt(SYSRESET_COLD); ----> loop on device UCLASS_SYSRESET -----> syscon_reboot_probe() ------> syscon_node_to_regmap(node) -------> list_for_each_entry(entry, &syscon_list, list)
I have a crash here because the pointer syscon_list.next is invalid.
This patch solves the issue by using DM list for syscon uclass.
Tested on my board, the initial issue is solved and I check the behavior with dm_dump_all(): the RCC driver is not binded/probed as syscon uclass by default (checked with dm tree) and correctly bounded and probed when it is necessary (before relocation: print_resetinfo() or after relocation for command reset).
So this is a regression we should fix, to be clear, yes? Thanks!