SW ISR table gets mangled when enabling SHARED_INTERRUPTS

[davedesro]

Describe the bug

tl;dr zephyr_pre0.elf does not have z_shared_isr() in its image, but should. See bottom for my work-around. Does the linker need modification?

When enabling the new shared interrupts feature (Great feature!! long overdue!! thank you!!) in v3.5.0-rc3, The interrupt pointers set in build/zephyr/isr_tables.c (generated by gen_isr_tables.py) are wrong.

After a bit of objdump'ing, I found that the linker for build/zephyr/zephyr_pre0.elf removes z_shared_isr() function since it's not used (due to a combo of -ffunction-sections at compile time, and -Wl,--gc-sections at link time). It's not used because zephyr_pre0.elf does not have isr_tables.c in it.

... however, build/zephyr/zephyr.elf does have build/zephyr/isr_tables.c, so z_shared_isr() is part of the final image.

This causes a shift of all ISR addresses when comparing zephyr_pre0.elf to zephyr.elf. When loaded, hang is bound to happen on the first couple of interrupts.

To Reproduce

Using the environment below,

1. set CONFIG_SHARED_INTERRUPTS=y and CONFIG_ADC_STM32_SHARED_IRQ=n
2. On NUCLEO_G474RE, enable &adc1 and &adc2 to exercise the new feature.
3. flash it. You should see a hang pretty quickly.. likely on the first interrupt.

Impact

Without a fix, the feature is not usable at least with the setup below.

Environment (please complete the following information):**

• Hardware: NUCLEO_G474RE
• SDK: 0.16.3
• Relevent config: CONFIG_SHARED_INTERRUPTS=y
• Zephyr tag v3.5.0-rc3
• Fedora Linux OS

Additional context

I'm working around this issue by modifying the linker.

Modify the linker to keep z_shared_isr() no matter what

> cat app/shared_irq.ld                                                                                                                       
SECTION_PROLOGUE(.text.z_shared_isr,,)                                                                                                                                                                                                  
{
    KEEP(*(.text.z_shared_isr))
} GROUP_ROM_LINK_IN(0, FLASH)

Add new .ld file to the build

> git diff app/CMakeLists.txt
diff --git a/app/CMakeLists.txt b/app/CMakeLists.txt
index 38c15f6..0196245 100644
--- a/app/CMakeLists.txt
+++ b/app/CMakeLists.txt
@@ -9,3 +9,6 @@ add_dependencies(app app_version)

 add_subdirectory(src)
 add_subdirectory_ifdef(CONFIG_SHELL src/cli)
+
+zephyr_linker_sources(SECTIONS shared_irq.ld)
+

@LaurentiuM1234 @carlescufi

[github-actions]

Hi @davedesro! We appreciate you submitting your first issue for our open-source project. 🌟

Even though I'm a bot, I can assure you that the whole community is genuinely grateful for your time and effort. 🤖💙

[henrikbrixandersen]

CC: @LaurentiuM1234

[LaurentiuM1234]

Hi @davedesro

First of all tyvm for the feedback! Glad to see people messing around with this new feature.

I've noticed this also happens on the qemu_cortex_a53_xip board. For instance, arm_arch_timer_compare_isr changes its address from 0x3248 (in zephyr_pre0.elf) to 0x329c, which seems to reach to the hang you've mentioned (since now something else is located at 0x3248). At the moment I'm not sure why this shift happens. My intuition is that this happens because z_shared_isr is added to .text region later on which changes the addresses of the symbols in .text. @andyross @carlocaione could you please confirm this or let me know if I've got something wrong? Thanks!

Regarding your fix: I've also tested it with qemu_cortex_a53_xip and it works fine. One suggestion I have: you can probably get away with adding your fix to the general common-rom-kernel-devices.ld and common-ram.ld. I guess it would look something like:

#if defined(CONFIG_SHARED_INTERRUPTS)

SECTION_PROLOGUE(text.z_shared_isr,,)
{
    KEEP(*(text.z_shared_isr))
}

... some more code goes here ...

Would you be willing to submit a PR with your fix? If not, I can take care of it once I understand why the shift happens.

[davedesro]

Hi @LaurentiuM1234. I am happy for the PR.. and it couldn't have come at a better time for our project :)

Yes, I have a good understand why the address shift is happening.

1. zephyr_pre0.elf is linked. This does not have the final interrupt vector table
2. objcopy is run to extract the .intList section, which contains all interrupts
3. scripts/build/gen_isr_tables.py is run, output is isr_tables.c, which contains the final "hw" vector table and "sw" vector table
4. zephyr.elf is linked including isr_tables.c

The problem is that the linker optimizes the zephyr_pre0.elf image by removing z_shared_isr() since the function is not used.. however, zephyr.elf does use this function via the new isr_tables.c file and is included in the final image.

This difference causes a shift in all symbol addresses by the size of the z_shared_isr() function... isr_tables.c has pointers for zephyr_pre0.elf. The build system assumes the ISR addresses are the same between both .elf files, when in fact they are not.

Anyways.. @LaurentiuM1234 I will submit a PR.

thanks!

[carlescufi]

Fixed by #64152

[davedesro]

Hi @carlescufi I see 3.5.0 was released on Friday, but this got merged afterwards.
Is this something that can be cherry picked for the v3.5-branch?

[henrikbrixandersen]

Hi @carlescufi I see 3.5.0 was released on Friday, but this got merged afterwards. Is this something that can be cherry picked for the v3.5-branch?

I have added the backport label for v3.5.x.