linux-sgx.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH V2] x86/sgx: Reduce delay and interference of enclave release
@ 2022-10-31 17:29 Reinette Chatre
  2022-10-31 20:42 ` Dave Hansen
  2022-11-01  1:27 ` Jarkko Sakkinen
  0 siblings, 2 replies; 4+ messages in thread
From: Reinette Chatre @ 2022-10-31 17:29 UTC (permalink / raw)
  To: dave.hansen, jarkko, tglx, bp, mingo, hpa, md.iqbal.hossain,
	haitao.huang, linux-sgx, x86, reinette.chatre
  Cc: linux-kernel

commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
releasing large enclaves") introduced a cond_resched() during enclave
release where the EREMOVE instruction is applied to every 4k enclave
page. Giving other tasks an opportunity to run while tearing down a
large enclave placates the soft lockup detector but Iqbal found
that the fix causes a 25% performance degradation of a workload
run using Gramine.

Gramine maintains a 1:1 mapping between processes and SGX enclaves.
That means if a workload in an enclave creates a subprocess then
Gramine creates a duplicate enclave for that subprocess to run in.
The consequence is that the release of the enclave used to run
the subprocess can impact the performance of the workload that is
run in the original enclave, especially in large enclaves when
SGX2 is not in use.

The workload run by Iqbal behaves as follows:
Create enclave (enclave "A")
/* Initialize workload in enclave "A" */
Create enclave (enclave "B")
/* Run subprocess in enclave "B" and send result to enclave "A" */
Release enclave (enclave "B")
/* Run workload in enclave "A" */
Release enclave (enclave "A")

The performance impact of releasing enclave "B" in the above scenario
is amplified when there is a lot of SGX memory and the enclave size
matches the SGX memory. When there is 128GB SGX memory and an enclave
size of 128GB, from the time enclave "B" starts the 128GB SGX memory
is oversubscribed with a combined demand for 256GB from the two
enclaves.

Before commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
releasing large enclaves") enclave release was done in a tight loop
without giving other tasks a chance to run. Even though the system
experienced soft lockups the workload (run in enclave "A") obtained
good performance numbers because when the workload started running
there was no interference.

Commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
releasing large enclaves") gave other tasks opportunity to run while an
enclave is released. The impact of this in this scenario is that while
enclave "B" is released and needing to access each page that belongs
to it in order to run the SGX EREMOVE instruction on it, enclave "A"
is attempting to run the workload needing to access the enclave
pages that belong to it. This causes a lot of swapping due to the
demand for the oversubscribed SGX memory. Longer latencies are
experienced by the workload in enclave "A" while enclave "B" is
released.

Improve the performance of enclave release while still avoiding the
soft lockup detector with two enhancements:
- Only call cond_resched() after XA_CHECK_SCHED iterations.
- Use the xarray advanced API to keep the xarray locked for
  XA_CHECK_SCHED iterations instead of locking and unlocking
  at every iteration.

This batching solution is copied from sgx_encl_may_map() that
also iterates through all enclave pages using this technique.

With this enhancement the workload experiences a 5%
performance degradation when compared to a kernel without
commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
releasing large enclaves"), an improvement to the reported 25%
degradation, while still placating the soft lockup detector.

Scenarios with poor performance are still possible even with these
enhancements. For example, short workloads creating sub processes
while running in large enclaves. Further performance improvements
are pursued in user space through avoiding to create duplicate enclaves
for certain sub processes, and using SGX2 that will do lazy allocation
of pages as needed so enclaves created for sub processes start quickly
and release quickly.

Fixes: 8795359e35bc ("x86/sgx: Silence softlockup detection when releasing large enclaves")
Reported-by: Md Iqbal Hossain <md.iqbal.hossain@intel.com>
Tested-by: Md Iqbal Hossain <md.iqbal.hossain@intel.com>
Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
---
I do not know if this qualifies as stable material.

Changes since V1:
- V1: https://lore.kernel.org/lkml/06a5f478d3bfaa57954954c82dd5d4040450171d.1666130846.git.reinette.chatre@intel.com/
- Use local variable for max index instead of open code in loop. (Jarkko)
- Send to broader X86 audience.

 arch/x86/kernel/cpu/sgx/encl.c | 23 +++++++++++++++++++----
 1 file changed, 19 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kernel/cpu/sgx/encl.c b/arch/x86/kernel/cpu/sgx/encl.c
index 1ec20807de1e..2c258255a629 100644
--- a/arch/x86/kernel/cpu/sgx/encl.c
+++ b/arch/x86/kernel/cpu/sgx/encl.c
@@ -680,11 +680,15 @@ const struct vm_operations_struct sgx_vm_ops = {
 void sgx_encl_release(struct kref *ref)
 {
 	struct sgx_encl *encl = container_of(ref, struct sgx_encl, refcount);
+	unsigned long max_page_index = PFN_DOWN(encl->base + encl->size - 1);
 	struct sgx_va_page *va_page;
 	struct sgx_encl_page *entry;
-	unsigned long index;
+	unsigned long count = 0;
+
+	XA_STATE(xas, &encl->page_array, PFN_DOWN(encl->base));
 
-	xa_for_each(&encl->page_array, index, entry) {
+	xas_lock(&xas);
+	xas_for_each(&xas, entry, max_page_index) {
 		if (entry->epc_page) {
 			/*
 			 * The page and its radix tree entry cannot be freed
@@ -699,9 +703,20 @@ void sgx_encl_release(struct kref *ref)
 		}
 
 		kfree(entry);
-		/* Invoke scheduler to prevent soft lockups. */
-		cond_resched();
+		/*
+		 * Invoke scheduler on every XA_CHECK_SCHED iteration
+		 * to prevent soft lockups.
+		 */
+		if (!(++count % XA_CHECK_SCHED)) {
+			xas_pause(&xas);
+			xas_unlock(&xas);
+
+			cond_resched();
+
+			xas_lock(&xas);
+		}
 	}
+	xas_unlock(&xas);
 
 	xa_destroy(&encl->page_array);
 
-- 
2.34.1


^ permalink raw reply related	[flat|nested] 4+ messages in thread

* Re: [PATCH V2] x86/sgx: Reduce delay and interference of enclave release
  2022-10-31 17:29 [PATCH V2] x86/sgx: Reduce delay and interference of enclave release Reinette Chatre
@ 2022-10-31 20:42 ` Dave Hansen
  2022-11-01  1:28   ` Jarkko Sakkinen
  2022-11-01  1:27 ` Jarkko Sakkinen
  1 sibling, 1 reply; 4+ messages in thread
From: Dave Hansen @ 2022-10-31 20:42 UTC (permalink / raw)
  To: Reinette Chatre, dave.hansen, jarkko, tglx, bp, mingo, hpa,
	md.iqbal.hossain, haitao.huang, linux-sgx, x86
  Cc: linux-kernel

On 10/31/22 10:29, Reinette Chatre wrote:
> I do not know if this qualifies as stable material.

I kept this unmarked for stable.  If anyone feels strongly about it,
they're more than welcome to submit it themselves.

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH V2] x86/sgx: Reduce delay and interference of enclave release
  2022-10-31 17:29 [PATCH V2] x86/sgx: Reduce delay and interference of enclave release Reinette Chatre
  2022-10-31 20:42 ` Dave Hansen
@ 2022-11-01  1:27 ` Jarkko Sakkinen
  1 sibling, 0 replies; 4+ messages in thread
From: Jarkko Sakkinen @ 2022-11-01  1:27 UTC (permalink / raw)
  To: Reinette Chatre
  Cc: dave.hansen, tglx, bp, mingo, hpa, md.iqbal.hossain,
	haitao.huang, linux-sgx, x86, linux-kernel

On Mon, Oct 31, 2022 at 10:29:58AM -0700, Reinette Chatre wrote:
> commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
> releasing large enclaves") introduced a cond_resched() during enclave
> release where the EREMOVE instruction is applied to every 4k enclave
> page. Giving other tasks an opportunity to run while tearing down a
> large enclave placates the soft lockup detector but Iqbal found
> that the fix causes a 25% performance degradation of a workload
> run using Gramine.
> 
> Gramine maintains a 1:1 mapping between processes and SGX enclaves.
> That means if a workload in an enclave creates a subprocess then
> Gramine creates a duplicate enclave for that subprocess to run in.
> The consequence is that the release of the enclave used to run
> the subprocess can impact the performance of the workload that is
> run in the original enclave, especially in large enclaves when
> SGX2 is not in use.
> 
> The workload run by Iqbal behaves as follows:
> Create enclave (enclave "A")
> /* Initialize workload in enclave "A" */
> Create enclave (enclave "B")
> /* Run subprocess in enclave "B" and send result to enclave "A" */
> Release enclave (enclave "B")
> /* Run workload in enclave "A" */
> Release enclave (enclave "A")
> 
> The performance impact of releasing enclave "B" in the above scenario
> is amplified when there is a lot of SGX memory and the enclave size
> matches the SGX memory. When there is 128GB SGX memory and an enclave
> size of 128GB, from the time enclave "B" starts the 128GB SGX memory
> is oversubscribed with a combined demand for 256GB from the two
> enclaves.
> 
> Before commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
> releasing large enclaves") enclave release was done in a tight loop
> without giving other tasks a chance to run. Even though the system
> experienced soft lockups the workload (run in enclave "A") obtained
> good performance numbers because when the workload started running
> there was no interference.
> 
> Commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
> releasing large enclaves") gave other tasks opportunity to run while an
> enclave is released. The impact of this in this scenario is that while
> enclave "B" is released and needing to access each page that belongs
> to it in order to run the SGX EREMOVE instruction on it, enclave "A"
> is attempting to run the workload needing to access the enclave
> pages that belong to it. This causes a lot of swapping due to the
> demand for the oversubscribed SGX memory. Longer latencies are
> experienced by the workload in enclave "A" while enclave "B" is
> released.
> 
> Improve the performance of enclave release while still avoiding the
> soft lockup detector with two enhancements:
> - Only call cond_resched() after XA_CHECK_SCHED iterations.
> - Use the xarray advanced API to keep the xarray locked for
>   XA_CHECK_SCHED iterations instead of locking and unlocking
>   at every iteration.
> 
> This batching solution is copied from sgx_encl_may_map() that
> also iterates through all enclave pages using this technique.
> 
> With this enhancement the workload experiences a 5%
> performance degradation when compared to a kernel without
> commit 8795359e35bc ("x86/sgx: Silence softlockup detection when
> releasing large enclaves"), an improvement to the reported 25%
> degradation, while still placating the soft lockup detector.
> 
> Scenarios with poor performance are still possible even with these
> enhancements. For example, short workloads creating sub processes
> while running in large enclaves. Further performance improvements
> are pursued in user space through avoiding to create duplicate enclaves
> for certain sub processes, and using SGX2 that will do lazy allocation
> of pages as needed so enclaves created for sub processes start quickly
> and release quickly.
> 
> Fixes: 8795359e35bc ("x86/sgx: Silence softlockup detection when releasing large enclaves")
> Reported-by: Md Iqbal Hossain <md.iqbal.hossain@intel.com>
> Tested-by: Md Iqbal Hossain <md.iqbal.hossain@intel.com>
> Signed-off-by: Reinette Chatre <reinette.chatre@intel.com>
> ---
> I do not know if this qualifies as stable material.
> 
> Changes since V1:
> - V1: https://lore.kernel.org/lkml/06a5f478d3bfaa57954954c82dd5d4040450171d.1666130846.git.reinette.chatre@intel.com/
> - Use local variable for max index instead of open code in loop. (Jarkko)
> - Send to broader X86 audience.
> 
>  arch/x86/kernel/cpu/sgx/encl.c | 23 +++++++++++++++++++----
>  1 file changed, 19 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/sgx/encl.c b/arch/x86/kernel/cpu/sgx/encl.c
> index 1ec20807de1e..2c258255a629 100644
> --- a/arch/x86/kernel/cpu/sgx/encl.c
> +++ b/arch/x86/kernel/cpu/sgx/encl.c
> @@ -680,11 +680,15 @@ const struct vm_operations_struct sgx_vm_ops = {
>  void sgx_encl_release(struct kref *ref)
>  {
>  	struct sgx_encl *encl = container_of(ref, struct sgx_encl, refcount);
> +	unsigned long max_page_index = PFN_DOWN(encl->base + encl->size - 1);
>  	struct sgx_va_page *va_page;
>  	struct sgx_encl_page *entry;
> -	unsigned long index;
> +	unsigned long count = 0;
> +
> +	XA_STATE(xas, &encl->page_array, PFN_DOWN(encl->base));
>  
> -	xa_for_each(&encl->page_array, index, entry) {
> +	xas_lock(&xas);
> +	xas_for_each(&xas, entry, max_page_index) {
>  		if (entry->epc_page) {
>  			/*
>  			 * The page and its radix tree entry cannot be freed
> @@ -699,9 +703,20 @@ void sgx_encl_release(struct kref *ref)
>  		}
>  
>  		kfree(entry);
> -		/* Invoke scheduler to prevent soft lockups. */
> -		cond_resched();
> +		/*
> +		 * Invoke scheduler on every XA_CHECK_SCHED iteration
> +		 * to prevent soft lockups.
> +		 */
> +		if (!(++count % XA_CHECK_SCHED)) {
> +			xas_pause(&xas);
> +			xas_unlock(&xas);
> +
> +			cond_resched();
> +
> +			xas_lock(&xas);
> +		}
>  	}
> +	xas_unlock(&xas);
>  
>  	xa_destroy(&encl->page_array);
>  
> -- 
> 2.34.1
> 

Tested-by: Jarkko Sakkinen <jarkko@kernel.org>
Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>

BR, Jarkko

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH V2] x86/sgx: Reduce delay and interference of enclave release
  2022-10-31 20:42 ` Dave Hansen
@ 2022-11-01  1:28   ` Jarkko Sakkinen
  0 siblings, 0 replies; 4+ messages in thread
From: Jarkko Sakkinen @ 2022-11-01  1:28 UTC (permalink / raw)
  To: Dave Hansen
  Cc: Reinette Chatre, dave.hansen, tglx, bp, mingo, hpa,
	md.iqbal.hossain, haitao.huang, linux-sgx, x86, linux-kernel

On Mon, Oct 31, 2022 at 01:42:02PM -0700, Dave Hansen wrote:
> On 10/31/22 10:29, Reinette Chatre wrote:
> > I do not know if this qualifies as stable material.
> 
> I kept this unmarked for stable.  If anyone feels strongly about it,
> they're more than welcome to submit it themselves.

I think this is the right choice.

BR, Jarkko

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2022-11-01  1:28 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-10-31 17:29 [PATCH V2] x86/sgx: Reduce delay and interference of enclave release Reinette Chatre
2022-10-31 20:42 ` Dave Hansen
2022-11-01  1:28   ` Jarkko Sakkinen
2022-11-01  1:27 ` Jarkko Sakkinen

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).