lkml.org 
[lkml]   [2017]   [Dec]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v6 06/11] intel_sgx: driver for Intel Software Guard Extensions
On Thu, Nov 30, 2017 at 09:32:01AM -0800, Sean Christopherson wrote:
> On Sat, Nov 25, 2017 at 09:29:24PM +0200, Jarkko Sakkinen wrote:
> > +static void *sgx_try_alloc_page(void)
> > +{
> > + struct sgx_epc_bank *bank;
> > + void *page = NULL;
> > + int i;
> > +
> > + for (i = 0; i < sgx_nr_epc_banks; i++) {
> > + bank = &sgx_epc_banks[i];
> > +
> > + down_write(&bank->lock);
>
> Is a R/W semaphore actually preferable to a spinlock? Concurrent
> free calls don't seem that interesting/beneficial because freeing
> an enclave's pages isn't multiplexed across multiple CPUs, unlike
> the allocation of EPC pages.

I get about ~10-15% performance increase on high stress. It is benefical
to spinlock.

> As a whole, I'm not a fan of packing the EPC page pointers into an
> array rather than encapsulating them in a struct+list. The primary
> benefit I see for the array approach is that it saves ~8 bytes per
> free EPC page, but at a cost of increased memory usage for in-use
> pages and severely restricting the ability to enhance/modify how
> EPC pages are tracked, reclaimed, etc...

This is not true. You can put EPC page descriptor to any structure you
want.

list_head can be dropped from struct sgx_enc_page too.

> The main issue is that the array approach relies on the caller to
> handle reclaim. This effectively makes it impossible to reclaim
> pages from multiple processes, requires other consumers e.g. KVM
> to implement their own reclaim logic and kthread, and prevents
> cgroup accounting because the cgroup can't initiate reclaim.

Not really following here.

/Jarkko

\
 
 \ /
  Last update: 2017-12-04 10:00    [W:0.104 / U:7.292 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site