Date: Thu, 27 May 2010 19:56:26 +0530 From: "C. Jayachandran" <c.jayachandran@gmail.com> To: Randall Stewart <rrs@lakerest.net>, Neel Natu <neelnatu@gmail.com>, freebsd-mips@freebsd.org Subject: Few more fixes to the new pagetable page allocation code. Message-ID: <AANLkTimVM41XCZDN5pFmkr68D21i1JRwKF3m0IFJvYH8@mail.gmail.com>
index | next in thread | raw e-mail
[-- Attachment #1 --] The attached changes (http://people.freebsd.org/~jchandra/for-review/pmap-free-pte-page.diff) finally seems to get rid of the occasional hangs I saw during testing. The first part fixes a lock order reversal (in pmap_remove_pages...pmap_release_pte_page...uma_zfree_arg...kmem_malloc). The pmap and pagequeue locks have to be released before calling uma_zfree(). The second part takes the the pagequeue lock before changing the page fields of the pte page, I'm not sure this is necessary. Please let me know you comments. With these changes, I can get 'make -j128 buildworld' consistently work on a 32 cpu setup. Thanks, JC. [-- Attachment #2 --] Index: sys/mips/mips/pmap.c =================================================================== --- sys/mips/mips/pmap.c (revision 208589) +++ sys/mips/mips/pmap.c (working copy) @@ -886,8 +886,12 @@ /* * If the page is finally unwired, simply free it. */ + atomic_subtract_int(&cnt.v_wire_count, 1); + PMAP_UNLOCK(pmap); + vm_page_unlock_queues(); pmap_release_pte_page(m); - atomic_subtract_int(&cnt.v_wire_count, 1); + vm_page_lock_queues(); + PMAP_LOCK(pmap); return (1); } @@ -1008,9 +1012,14 @@ paddr = MIPS_KSEG0_TO_PHYS(va); m = PHYS_TO_VM_PAGE(paddr); + if (!locked) + vm_page_lock_queues(); m->pindex = index; m->valid = VM_PAGE_BITS_ALL; m->wire_count = 1; + if (!locked) + vm_page_unlock_queues(); + atomic_add_int(&cnt.v_wire_count, 1); *vap = (vm_offset_t)va; return (m);help
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?AANLkTimVM41XCZDN5pFmkr68D21i1JRwKF3m0IFJvYH8>