Re: [PATCH v2 1/5] selftests/sgx: Retry the ioctl()'s returned with EAGAIN

[Jarkko Sakkinen]

On Thu, Sep 08, 2022 at 05:06:58PM -0700, Reinette Chatre wrote:
> Hi Jarkko,
> 
> On 9/8/2022 4:19 PM, Jarkko Sakkinen wrote:
> > On Thu, Sep 08, 2022 at 03:43:06PM -0700, Reinette Chatre wrote:
> >> Hi Jarkko and Haitao,
> >>
> >> On 9/4/2022 7:04 PM, Jarkko Sakkinen wrote:
> >>> From: Haitao Huang <haitao.huang@xxxxxxxxxxxxxxx>
> >>>
> >>> For EMODT and EREMOVE ioctl()'s with a large range, kernel
> >>> may not finish in one shot and return EAGAIN error code
> >>> and count of bytes of EPC pages on that operations are
> >>> finished successfully.
> >>>
> >>> Change the unclobbered_vdso_oversubscribed_remove test
> >>> to rerun the ioctl()'s in a loop, updating offset and length
> >>> using the byte count returned in each iteration.
> >>>
> >>> Fixes: 6507cce561b4 ("selftests/sgx: Page removal stress test")
> >>
> >> Should this patch be moved to the "critical fixes for v6.0" series?
> > 
> > I think not because it does not risk stability of the
> > kernel itself. It's "nice to have" but not mandatory.
> 
> ok, thank you for considering it.
> 
> ...
> 
> >>> @@ -453,16 +454,30 @@ TEST_F_TIMEOUT(enclave, unclobbered_vdso_oversubscribed_remove, 900)
> >>>  	modt_ioc.offset = heap->offset;
> >>>  	modt_ioc.length = heap->size;
> >>>  	modt_ioc.page_type = SGX_PAGE_TYPE_TRIM;
> >>> -
> >>> +	count = 0;
> >>>  	TH_LOG("Changing type of %zd bytes to trimmed may take a while ...",
> >>>  	       heap->size);
> >>> -	ret = ioctl(self->encl.fd, SGX_IOC_ENCLAVE_MODIFY_TYPES, &modt_ioc);
> >>> -	errno_save = ret == -1 ? errno : 0;
> >>> +	do {
> >>> +		ret = ioctl(self->encl.fd, SGX_IOC_ENCLAVE_MODIFY_TYPES, &modt_ioc);
> >>> +
> >>> +		errno_save = ret == -1 ? errno : 0;
> >>> +		if (errno_save != EAGAIN)
> >>> +			break;
> >>> +
> >>> +		EXPECT_EQ(modt_ioc.result, 0);
> >>
> >> If this check triggers then there is something seriously wrong and in that case
> >> it may also be that this loop may be unable to terminate or the error condition would
> >> keep appearing until the loop terminates (which may be many iterations). Considering
> >> the severity and risk I do think that ASSERT_EQ() would be more appropriate,
> >> similar to how ASSERT_EQ() is used in patch 5/5.
> >>
> >> Apart from that I think that this looks good.
> >>
> >> Thank you very much for adding this.
> >>
> >> Reinette
> > 
> > Hmm... I could along the lines:
> > 
> > /*
> >  * Get time since Epoch is milliseconds.
> >  */
> > unsigned long get_time(void)
> > {
> >     struct timeval start;
> > 
> >     gettimeofday(&start, NULL);
> > 
> >     return (unsigneg long)start.tv_sec * 1000L + (unsigned long)start.tv_usec / 1000L;
> > }
> > 
> > and
> > 
> > #define IOCTL_RETRY_TIMEOUT 100
> > 
> > In the test function:
> > 
> >         unsigned long start_time;
> > 
> >         /* ... */
> > 
> >         start_time = get_time();
> >         do {
> >                 EXPECT_LT(get_time() - start_time(), IOCTL_RETRY_TIMEOUT);
> > 
> >                 /* ... */
> >         }
> > 
> >         /* ... */
> > 
> > What do you think?
> 
> I do think that your proposal can be considered for an additional check in this
> test but the way I understand it it does not address my feedback.
> 
> In this patch the flow is:
> 
> 	do {
> 		ret = ioctl(self->encl.fd, SGX_IOC_ENCLAVE_MODIFY_TYPES, &modt_ioc);
> 
> 		errno_save = ret == -1 ? errno : 0;
> 		if (errno_save != EAGAIN)
> 			break;
> 
> 		EXPECT_EQ(modt_ioc.result, 0);
> 		...
> 	} while ...
> 
> 
> If this EXPECT_EQ() check fails then it means that errno_save is EAGAIN
> and modt_ioc.result != 0. This should never happen because in the kernel
> (sgx_enclave_modify_types()) the only time modt_ioc.result can be set is
> when the ioctl() returns EFAULT.
> 
> In my opinion this check should be changed to:
> 		ASSERT_EQ(modt_ioc.result, 0);

Right, I missed this. It should be definitely ASSERT_EQ(().

BR, Jarkko