Re: git: 5a7e48dddfb5 - main - tests: Add MAP_32BIT flag test

Konstantin Belousov <kostikbel_at_gmail.com> wrote on
Date: Thu, 03 Aug 2023 20:58:47 UTC :

> On Thu, Aug 03, 2023 at 02:50:57PM +0300, Dmitry Chagin wrote:
> > On Wed, Aug 02, 2023 at 12:57:36AM +0300, Konstantin Belousov wrote:
> > > On Tue, Aug 01, 2023 at 09:29:32PM +0100, Jessica Clarke wrote:
> > > > On 1 Aug 2023, at 21:24, Dmitry Chagin <dchagin@FreeBSD.org> wrote:
> > > > > 
> > > > > The branch main has been updated by dchagin:
> > > > > 
> > > > > URL: https://cgit.FreeBSD.org/src/commit/?id=5a7e48dddfb5a668ded4742b79e6f6f88b647e6a
> > > > > 
> > > > > commit 5a7e48dddfb5a668ded4742b79e6f6f88b647e6a
> > > > > Author: Dmitry Chagin <dchagin@FreeBSD.org>
> > > > > AuthorDate: 2023-08-01 20:23:15 +0000
> > > > > Commit: Dmitry Chagin <dchagin@FreeBSD.org>
> > > > > CommitDate: 2023-08-01 20:23:15 +0000
> > > > > 
> > > > > tests: Add MAP_32BIT flag test
> > > > > 
> > > > > Reviewed by: kib, markj
> > > > > Differential Revision: https://reviews.freebsd.org/D41236
> > > > > MFC after: 1 month
> > > > > ---
> > > > > tests/sys/vm/Makefile | 8 ++++++
> > > > > tests/sys/vm/mmap_map_32bit_helper.c | 51 ++++++++++++++++++++++++++++++++++++
> > > > > tests/sys/vm/mmap_map_32bit_test.sh | 37 ++++++++++++++++++++++++++
> > > > > 3 files changed, 96 insertions(+)
> > > > > 
> > > > > diff --git a/tests/sys/vm/Makefile b/tests/sys/vm/Makefile
> > > > > index 8ef8a45e5f39..9aac49bc5320 100644
> > > > > --- a/tests/sys/vm/Makefile
> > > > > +++ b/tests/sys/vm/Makefile
> > > > > @@ -9,4 +9,12 @@ ATF_TESTS_C+= mlock_test \
> > > > > page_fault_signal \
> > > > > shared_shadow_inval_test
> > > > > 
> > > > > +.if ${MACHINE_ARCH} != "i386" && ${MACHINE} != "arm" && \
> > > > > + (${MACHINE} != "powerpc" || ${MACHINE_ARCH} != "powerpc")
> > > > > + # MAP_32BIT is only available on 64-bit platforms
> > > > 
> > > > So use ${MACHINE_ABI:Mlong64} after including src.opts.mk?
> > > 
> > > May be we should enable the MAP_32BIT flag on 32bit arches, instead?
> > > It requests allocation below 2G, which is half of VA on i386, or 2/3
> > > on arm. Interestingly, it should work already from compat32.
> > 
> > It's disturbing that no none has requested this, I can do that
> Please do. The largest part of the work is to recheck the man page and tests.


If I guess correctly, the above implies restricting the process size
on native armv7 to a strict 2 GiBytes or so.

The OrangePi+2Ed (cortex-a7) that I use has a somewhat larger effective
process address space (RAM+SWAP usable by one process) than the armv7
chroot/jail use on any of the cortex-a72/cortex-a53/cortex-x1c/cortex-a78c
aarch64 systems that I've access to. I have actually built something on
the OrangePi+2Ed that I could not build on the aarch64 systems in an armv7
poudriere-devel jail --because of the process size limitation on aarch64
systems was smaller. (Not that I remember the details. I do not do such
builds regularly. But I might have notes about it someplace.)

So a ~2 GiByte process size limitation would make armv7 chroot/jail/lib32 
on aarch64 vs. native armv7 more uniform --but some native armv7's could
be somewhat less capable.

Note: It is not obvious to me that keeping armv7 going until FreeBSD 15's
release avoids being too restrictive, even as tier 2. Even the somewhat
larger address space for native could easily end up being too much of a
constraint, for example, much like the process size build problems that
happen now for armv7 on aarch64 for port builds in poudriere(-devel)
jails.


For reference:

CPU: ARM Cortex-A7 r0p5 (ECO: 0x00000000)
CPU Features: 
  Multiprocessing, Thumb2, Security, Virtualization, Generic Timer, VMSAv7,
  PXN, LPAE, Coherent Walk
Optional instructions: 
  SDIV/UDIV, UMULL, SMULL, SIMD(ext)
LoUU:2 LoC:3 LoUIS:2 
Cache level 1:
 32KB/64B 4-way data cache WB Read-Alloc Write-Alloc
 32KB/32B 2-way instruction cache Read-Alloc
Cache level 2:
 512KB/64B 8-way unified cache WB Read-Alloc Write-Alloc
real memory  = 2113257472 (2015 MB)
avail memory = 2054692864 (1959 MB)


===
Mark Millard
marklmi at yahoo.com