[prev in list] [next in list] [prev in thread] [next in thread]
List: dragonfly-users
Subject: Re: Porting sbcl to DragonFly
From: vasily postnicov <shamaz.mazum () gmail ! com>
Date: 2013-04-24 17:26:13
Message-ID: CADnZ6BnfANww+i_gX8jaiO42hE5un4bOByCAvnC1akeNQQjdbg () mail ! gmail ! com
[Download RAW message or body]
2013/4/21 Markus Pfeiffer <markus.pfeiffer@morphism.de>
> Hi,
>
> On Sat, Apr 13, 2013 at 12:14:15AM +0400, vasily postnicov wrote:
> > I've started sourceforge project, where I put sbcl-1-1.6 patched for
> DragonFly.
> > I almost gave up fixing threads support, so it is switched off by now.
> >
> > Here is pkgsrc package (based on original lang/sbcl):
> > http://shamazmazum.users.sourceforge.net/sbcl.tar
> >
> > Can anyone upload it to DragonFly's pkgsrc tree? I can maintain this
> package
> > and send new versions on regular basis.
>
> Out of interest: What stops threads from working? Is it a DragonFly issue
> or
> is it just that SBCLs thread model is not easily ported?
>
> Markus
>
Playing with commenting lines out. The latest thing I found is this:
There is perform_thread_post_mortem function inside
path_to_sbcl/src/runtime/thread.c
>static void*
>perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
>{
>#ifdef CREATE_POST_MORTEM_THREAD
> pthread_detach(pthread_self());
>#endif
> if (post_mortem) {
> gc_assert(!pthread_join(post_mortem->os_thread, NULL));
> gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
> free(post_mortem->os_attr);
> // NOTE NEXT LINE
> os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
> free(post_mortem);
> }
> return NULL;
>}
If you remove call of os_invalidate, SBCL works without crashes.
os_invalidate is opposing/freeing call after we os_validate
thread->os_address. os_validate/os_invalidate are wrappers for mmap/munmap
and can be found in path_to_sbcl/src/runtime/bsd-os.c
os_validate is called in create_thread_struct. There is a note on this:
> /* May as well allocate all the spaces at once: it saves us from
> * having to decide what to do if only some of the allocations
> * succeed. SPACES must be appropriately aligned, since the GC
> * expects the control stack to start at a page boundary -- and
> * the OS may have even more rigorous requirements. We can't rely
> * on the alignment passed from os_validate, since that might
> * assume the current (e.g. 4k) pagesize, while we calculate with
> * the biggest (e.g. 64k) pagesize allowed by the ABI. */
> spaces=os_validate(0, THREAD_STRUCT_SIZE);
> if(!spaces)
> return NULL;
> /* Aligning up is safe as THREAD_STRUCT_SIZE has
> * THREAD_ALIGNMENT_BYTES padding. */
> aligned_spaces = (void *)((((uword_t)(char *)spaces)
> + THREAD_ALIGNMENT_BYTES-1)
> &~(uword_t)(THREAD_ALIGNMENT_BYTES-1));
Still have no idea why sometimes it works and sometimes crashes.
[Attachment #3 (text/html)]
<div dir="ltr"><br><div class="gmail_extra"><br><br><div \
class="gmail_quote">2013/4/21 Markus Pfeiffer <span dir="ltr"><<a \
href="mailto:markus.pfeiffer@morphism.de" \
target="_blank">markus.pfeiffer@morphism.de</a>></span><br> <blockquote \
class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid \
rgb(204,204,204);padding-left:1ex">Hi,<br> <div><div class="h5"><br>
On Sat, Apr 13, 2013 at 12:14:15AM +0400, vasily postnicov wrote:<br>
> I've started sourceforge project, where I put sbcl-1-1.6 patched for \
DragonFly.<br> > I almost gave up fixing threads support, so it is switched off by \
now.<br> ><br>
> Here is pkgsrc package (based on original lang/sbcl):<br>
> <a href="http://shamazmazum.users.sourceforge.net/sbcl.tar" \
target="_blank">http://shamazmazum.users.sourceforge.net/sbcl.tar</a><br> ><br>
> Can anyone upload it to DragonFly's pkgsrc tree? I can maintain this \
package<br> > and send new versions on regular basis.<br>
<br>
</div></div>Out of interest: What stops threads from working? Is it a DragonFly issue \
or<br> is it just that SBCLs thread model is not easily ported?<br>
<span class=""><font color="#888888"><br>
Markus<br>
</font></span></blockquote></div><br></div><div class="gmail_extra">Playing with \
commenting lines out. The latest thing I found is this:<br><br></div><div \
class="gmail_extra">There is perform_thread_post_mortem function inside \
path_to_sbcl/src/runtime/thread.c<br> <br>>static \
void*<br>>perform_thread_post_mortem(struct thread_post_mortem \
*post_mortem)<br>>{<br>>#ifdef CREATE_POST_MORTEM_THREAD<br>> \
pthread_detach(pthread_self());<br>>#endif<br>> if (post_mortem) {<br> > \
gc_assert(!pthread_join(post_mortem->os_thread, NULL));<br>> \
gc_assert(!pthread_attr_destroy(post_mortem->os_attr));<br>> \
free(post_mortem->os_attr);<br></div><div class="gmail_extra"> > // NOTE \
NEXT LINE<br></div><div class="gmail_extra">> \
os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);<br>> \
free(post_mortem);<br>> }<br>> return NULL;<br>>}<br> <br></div><div \
class="gmail_extra">If you remove call of os_invalidate, SBCL works without crashes. \
os_invalidate is opposing/freeing call after we os_validate thread->os_address. \
os_validate/os_invalidate are wrappers for mmap/munmap and can be found in \
path_to_sbcl/src/runtime/bsd-os.c<br> <br></div><div class="gmail_extra">os_validate \
is called in create_thread_struct. There is a note on this:<br><br>> /* May as \
well allocate all the spaces at once: it saves us from<br>> * having to decide \
what to do if only some of the allocations<br> > * succeed. SPACES must be \
appropriately aligned, since the GC<br>> * expects the control stack to start \
at a page boundary -- and<br>> * the OS may have even more rigorous \
requirements. We can't rely<br> > * on the alignment passed from \
os_validate, since that might<br>> * assume the current (e.g. 4k) pagesize, \
while we calculate with<br>> * the biggest (e.g. 64k) pagesize allowed by the \
ABI. */<br>> spaces=os_validate(0, THREAD_STRUCT_SIZE);<br> > \
if(!spaces)<br>> return NULL;<br>> /* Aligning up is safe as \
THREAD_STRUCT_SIZE has<br>> * THREAD_ALIGNMENT_BYTES padding. */<br>> \
aligned_spaces = (void *)((((uword_t)(char *)spaces)<br> > \
+ THREAD_ALIGNMENT_BYTES-1)<br>> \
&~(uword_t)(THREAD_ALIGNMENT_BYTES-1));<br><br></div><div \
class="gmail_extra">Still have no idea why sometimes it works and sometimes \
crashes.<br> </div></div>
[prev in list] [next in list] [prev in thread] [next in thread]
Configure |
About |
News |
Add a list |
Sponsored by KoreLogic