Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Andy Lutomirski <luto@amacapital.net>
Mime-Version: 1.0 (1.0)
Subject: Re: [RFC PATCH 13/22] x86/fpu/xstate: Expand dynamic user state area on first use
Date:   Tue, 13 Oct 2020 18:11:55 -0700
Message-Id: <78F221DB-CB61-40DB-9C6F-6C86D0F1BCDF@amacapital.net>
References: <c4669d5f-11b8-3879-562c-78a791b86229@intel.com>
Cc:     "Brown, Len" <len.brown@intel.com>,
        Andy Lutomirski <luto@kernel.org>,
        "Bae, Chang Seok" <chang.seok.bae@intel.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@kernel.org>, Borislav Petkov <bp@suse.de>,
        X86 ML <x86@kernel.org>, "Liu, Jing2" <jing2.liu@intel.com>,
        "Shankar, Ravi V" <ravi.v.shankar@intel.com>,
        LKML <linux-kernel@vger.kernel.org>
In-Reply-To: <c4669d5f-11b8-3879-562c-78a791b86229@intel.com>
To:     Dave Hansen <dave.hansen@intel.com>
Precedence: bulk


> On Oct 13, 2020, at 3:44 PM, Dave Hansen <dave.hansen@intel.com> wrote:
>=20
> =EF=BB=BFOn 10/13/20 3:31 PM, Brown, Len wrote:
>> vmalloc() does not fail, and does not return an error, and so there is no=
 concept
>> of returning a signal.
>=20
> Well, the order-0 allocations are no-fail, as are the vmalloc kernel
> structures and the page tables that might have to be allocated.  But,
> that's not guaranteed to be in place *forever*.  I think we still need
> to check for and handle allocation failures, even if they're not known
> to be possible today.
>=20
>> If we got to the point where vmalloc() sleeps, then the system
>> has bigger OOM issues, and the OOM killer would be on the prowl.
>=20
> vmalloc() can *certainly* sleep.  Allocation failures mean returning
> NULL from the allocator, and the very way we avoid doing that is by
> sleeping to go reclaim some memory from some other allocation.
>=20
> Sleeping is a normal and healthy part of handling allocation requests,
> including vmalloc().
>=20
>> If we were concerned about using vmalloc for a couple of pages in the tas=
k structure,
>> Then we could implement a routine to harvest unused buffers and free them=
 --
>> but that didn't seem worth the complexity.  Note that this feature is 64-=
bit only.
>=20
> IMNHO, vmalloc() is overkill for ~10k, which is roughly the size of the
> XSAVE buffer for the first AMX implementation.  But, it's not overkill
> for the ~66k of space that will be needed if some CPU implementation
> comes along and uses all of the architectural space AMX provides.

I have no problem with vmalloc(), but I do have a problem with vfree() due t=
o the IPIs that result. We need a cache or something.

I have to say: this mechanism is awful. Can we get away with skipping the dy=
namic XSAVES mess entirely?  What if we instead allocate however much space w=
e need as an array of pages and have one percpu contiguous region. To save, w=
e XSAVE(S or C) just the AMX state to the percpu area and then copy it.  To r=
estore, we do the inverse.  Or would this kill the modified optimization and=
 thus be horrible?=