Hi Manivannan,
On 2025/01/28 23:12, Manivannan Sadhasivam wrote:
On Wed, Jan 22, 2025 at 11:24:44AM +0900, Kunihiko Hayashi wrote:
After devm_request_irq() fails with error,
pci_endpoint_test_free_irq_vectors() is called to free allocated vectors
with pci_free_irq_vectors().
You should mention the function name which you are referring to. Here it is,
pci_endpoint_test_request_irq().
I see. I'll make the commit message more clear.
However some requested IRQs are still allocated, so there are still
This is confusing. Are you saying that the previously requested IRQs are not
freed when an error happens during the for loop in
pci_endpoint_test_request_irq()?
Yes, after jumping to "fail:" label, it just prints an error message and
returns the function.
The pci_endpoint_test_request_irq() is called from the following functions:
- pci_endpoint_test_probe()
- pci_endpoint_test_set_irq()
Both call pci_endpoint_test_free_irq_vectors() after the error, though,
requested IRQs are not freed anywhere.
/proc/irq/* entries remaining and we encounters WARN() with the following
message:
remove_proc_entry: removing non-empty directory 'irq/30', leaking at
least 'pci-endpoint-test.0'
WARNING: CPU: 0 PID: 80 at fs/proc/generic.c:717 remove_proc_entry
+0x190/0x19c
When did you encounter this WARN?
Usually request_irq() can successfully get an interrupt.
If request_irq() returned an error, pci_endpoint_test_free_irq_vectors() was
called and the following call-trace was obtained:
[ 18.772522] Call trace:
[ 18.773743] remove_proc_entry+0x190/0x19c
[ 18.775789] unregister_irq_proc+0xd0/0x104
[ 18.777881] free_desc+0x4c/0xcc
[ 18.779495] irq_free_descs+0x68/0x8c
[ 18.781325] irq_domain_free_irqs+0x15c/0x1bc
[ 18.783502] msi_domain_free_locked.part.0+0x184/0x1d4
[ 18.786069] msi_domain_free_irqs_all_locked+0x64/0x8c
[ 18.788637] pci_msi_teardown_msi_irqs+0x48/0x54
[ 18.790947] pci_free_msi_irqs+0x18/0x38
[ 18.792907] pci_free_irq_vectors+0x64/0x8c
[ 18.794997] pci_endpoint_test_ioctl+0x7e8/0xf40
[ 18.797304] __arm64_sys_ioctl+0xa4/0xe8
[ 18.799265] invoke_syscall+0x48/0x110
[ 18.801139] el0_svc_common.constprop.0+0x40/0xe8
[ 18.803489] do_el0_svc+0x20/0x2c
[ 18.805145] el0_svc+0x30/0xd0
[ 18.806673] el0t_64_sync_handler+0x13c/0x158
[ 18.808850] el0t_64_sync+0x190/0x194
[ 18.810680] ---[ end trace 0000000000000000 ]---
To solve this issue, set the number of remaining IRQs and release the IRQs
in advance by calling pci_endpoint_test_release_irq().
First of all, using devm_request_irq() is wrong here as
pci_endpoint_test_request_irq() might be called multiple times by the userspace.
So we cannot use managed version of request_irq(). It is infact pointless since
pci_endpoint_test_clear_irq() would free them anyway. Instead we should just use
request_irq() and call pci_endpoint_test_clear_irq() in the error path as like
this patch does.
The pci_endpoint_test_clear_irq() will free all the number of IRQs.
However, in this error case, since only some IRQs are allocated, need to change the
number of IRQs that need to be freed.
And pci_endpoint_test_free_irq_vectors() is called just before return from the function,
call only pci_endpoint_test_release_irq() in the error case of
pci_endpoint_test_request_irq().
For example:
pci_endpoint_test_set_irq()
pci_endpoint_test_request_irq()
request_irq()
fail:
>> test->num_irqs = i // add in this patch
>> pci_endpoint_test_release_irq() // add in this patch
err:
pci_endpoint_test_free_irq_vectors()
But this should be a separate patch.
Agreed. I'll add another patch to use non-managed version.
Thank you,
---
Best Regards
Kunihiko Hayashi
