Re: panic: pmap_page_protect: illegal for unmanaged page

["Simon 'corecode' Schubert" <corecode@xxxxxxxxxxxx>]

Matthew Dillon wrote:
> :device memory has pages of type PG_FICTITIOUS, but pmap_remove_all() complains:
> :        /*      
> :         * XXX this makes pmap_page_protect(NONE) illegal for non-managed
> :         * pages!
> :         */
>     The comment itself is correct in that fictitious pages are not managed
>     by PV lists, so its impossible for pmap_remove_all() to function on
>     a fictitiuos page.  pmap_remove_all() is typically called to remove
>     user mappings of pages undergoing write I/O to prevent the pages from
>     being modified by userland while the I/O is in progress.
>
>     If there is a code path calling that function on fictitious pages we
>     have to examine and fix it (and the 'fix' might be as you suggest, just
>     ignoring it for such pages).  I don't think its possible to put fictitious
>     pages under PV management due to the physical-to-virtual translations
>     that might be needed in that codepath... device pages tend to be at
>     physical addresses that just don't work with the way the translation code
>     was written.

It happens if you call msync() with an argument of MS_INVALIDATE.

So you suggest to simply return?

What I wonder is, will *any* page of a OBJT_DEVICE not be PG_FICTITIOUS?  In other words:  shouldn't we simply avoid the whole invalidate code path for OBJT_DEVICE objects?

I'm quite confused about the intentions of this, anyways:

from vm_map_clean:
                       int clean_only = 
                               (object->type == OBJT_DEVICE) ? FALSE : TRUE;
                       vm_object_reference(object);
                       switch(current->maptype) {
                       case VM_MAPTYPE_NORMAL:
                               vm_object_page_remove(object,
                                   OFF_TO_IDX(offset),
                                   OFF_TO_IDX(offset + size + PAGE_MASK),
                                   clean_only);

Obviously OBJT_DEVICE is special cased, and it seems that somebody explicitly wanted that OBJT_DEVICE mappings will be completely destroyed on this call.

But then, down this code path, pmap_remove_all() will panic on PG_FICTITOUS pages.  Interesting enough, pmap_clearbit() will simply ignore PG_FICTITIOUS pages.  I think I'll try the same for pmap_remove_all() now.

cheers
 simon

--
Serve - BSD     +++  RENT this banner advert  +++    ASCII Ribbon   /"\
Work - Mac      +++  space for low €€€ NOW!1  +++      Campaign     \ /
Party Enjoy Relax   |   http://dragonflybsd.org      Against  HTML   \
Dude 2c 2 the max   !   http://golden-apple.biz       Mail + News   / \