Date: Mon, 3 Aug 2015 14:32:16 +0300 From: Slawa Olhovchenkov <slw@zxy.spb.ru> To: Konstantin Belousov <kostikbel@gmail.com> Cc: Peter Wemm <peter@wemm.org>, svn-src-all@freebsd.org, Steven Hartland <steven@multiplay.co.uk>, svn-src-head@freebsd.org, src-committers@freebsd.org Subject: Re: svn commit: r286223 - head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs Message-ID: <20150803113215.GA8792@zxy.spb.ru> In-Reply-To: <20150803111942.GB2072@kib.kiev.ua> References: <201508030934.t739YAkT026668@repo.freebsd.org> <20150803094730.GA24698@zxy.spb.ru> <55BF431E.3020601@freebsd.org> <2757800.HIDNx1G49O@overcee.wemm.org> <20150803111942.GB2072@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Aug 03, 2015 at 02:19:42PM +0300, Konstantin Belousov wrote:
> > The whole thing has missing the point.
> >
> > Changing the default for the entire kernel just because the zfs compat
> > wrappers can't be bothered requesting a suitable value is.. unfortunate..
> > particularly when it is in freebsd-provided code, not upstream zfs code.
> >
> > Fix the kproc_kthread_add() calls in do_thread_create() and zvol_geom_run()
> > instead. Enforce a lower bound there for zfs threads instead of making the
> > entire rest of the kernel use more memory.
> >
> > eg: I'm thinking along these lines:
> > Index: cddl/compat/opensolaris/sys/proc.h
> > ====================================================
> > --- cddl/compat/opensolaris/sys/proc.h (revision 286224)
> > +++ cddl/compat/opensolaris/sys/proc.h (working copy)
> > @@ -77,6 +77,8 @@
> > ASSERT(state == TS_RUN);
> > ASSERT(pp == &p0);
> >
> > + if (stksize < 16384)
> > + stksize = 16384; /* Enforce lower bound on ZFS threads */
> > error = kproc_kthread_add(proc, arg, &zfsproc, &td, RFSTOPPED,
> > stksize / PAGE_SIZE, "zfskern", "solthread %p", proc);
> > if (error == 0) {
> >
> >
> > Beware, some platforms have large pages (eg: ia64 in -stable has 8k, 16k or
> > 32k pages, from memory). Specifying an arbitrary number of pages in code
> > that's supposed to be portable isn't a good idea.
>
> This would not help. Issue is the size of the thread0 stack, which
> overflows in this case.
>
> I looked at the possibility of making default kernel stack size
> configurable by a loader tunable, and the issue is that thread0 gets its
> stack set up too early (locore for i386, hammer_time() for amd64). I.e.,
> it is possible to make the setting effective for all threads after thread0,
> but not for the one which causes the issue.
>
> I do not want to modify ABI between loader and kernel to pass the parameter.
Parsing kenv by asm too complex?
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20150803113215.GA8792>