sgx_ioc_enclave_add_page is interrupted and returns -EINVAL about linux-sgx-driver

Not related to bug itself but sgx_page_alloc() should return -EINTR (note to self mainly) and call site should decide how to deal with it.

In the case of sgx_ioc_enclave_create() the return value should be checked and if it is -EINTR it should return -ERESTARTSYS.

It's a driver bug. Returning -EINVAL makes zero sense.

[haitaohuang]

I think there are more cases where sgx_init_page/sgx_alloc_page used and possibly return -ERESTARTSYS, which should have returned -EINTR. And callers of those functions should handle properly. That may be the reason why you see 1% failures in create_enclave

Now you are mixing two things. The regression here is that the call site returns -EINVAL. The comment about -EINTR was completely unrelated to the regression.

[haitaohuang]

here is an attempt to track down those pending signal situations, where -EINTR returned from either mmu_notifier_register or sgx_alloc_page(0):
haitaohuang@f2ba8fb
Please review or give it a try.

[haitaohuang]

should einit IOC also return -ERESTARTSYS at https://github.com/01org/linux-sgx-driver/blob/master/sgx_ioctl.c#L849?
It is idempotent, isn't it?

I don't know how this relates to the bug in question. Here the problem is converting the return value to -EINVAL and that's the only problem here really.

[haitaohuang]

Yes, we won't be sure to find fix of the original failures until we can reproduce. The patch only attempt to fix those EINTR cases in driver, which may help application side to isolate the issue.

[deeglaze]

Your patch worked for me, @haitaohuang, except now I hit yet another problem that appears unrelated since it's in the uae service. I've opened an issue in linux-sgx.

[haitaohuang]

fixed