Date: Sat, 19 Nov 2005 23:40:26 GMT From: Peter Jeremy <PeterJeremy@optushome.com.au> To: freebsd-bugs@FreeBSD.org Subject: Re: kern/78179: bus_dmamem_alloc() with BUS_DMA_NOWAIT can block Message-ID: <200511192340.jAJNeQLU095805@freefall.freebsd.org>
next in thread | raw e-mail | index | archive | help
The following reply was made to PR kern/78179; it has been noted by GNATS.
From: Peter Jeremy <PeterJeremy@optushome.com.au>
To: Mark Tinguely <tinguely@casselton.net>
Cc: bug-followup@freebsd.org
Subject: Re: kern/78179: bus_dmamem_alloc() with BUS_DMA_NOWAIT can block
Date: Sun, 20 Nov 2005 10:39:09 +1100
On Thu, 2005-Nov-17 16:04:14 -0600, Mark Tinguely wrote:
>I took a stab at the problem that NOWAIT is not being honored in
>contigmalloc() by making the vm_contig_launder_page() honor
>the flag.
Thank you. Of course, this doesn't help drivers like bktr(4) that
use contigmalloc(M_NOWAIT) and panic() if the memory isn't available.
I've gone through your changes and suspect you may have missed a
few possibilities that can sleep.
>*** vm/vm_contig.c.orig Thu Nov 17 15:10:10 2005
>--- vm/vm_contig.c Thu Nov 17 13:59:47 2005
>***************
>*** 95,100 ****
>--- 95,105 ----
> object = m->object;
> if (!VM_OBJECT_TRYLOCK(object))
> return (EAGAIN);
>+ if (flags & M_NOWAIT && (m->flags & PG_BUSY || m->busy)) {
>+ VM_OBJECT_UNLOCK(object);
>+ printf("vm_contig_launder_page: would sleep (busy)\n");
>+ return (EWOULDBLOCK);
I suspect all your EWOULDBLOCK should be EAGAIN for consistency.
>+ }
> if (vm_page_sleep_if_busy(m, TRUE, "vpctw0")) {
> VM_OBJECT_UNLOCK(object);
> vm_page_lock_queues();
>***************
>*** 104,126 ****
...
>--- 109,149 ----
> if (m->dirty == 0 && m->hold_count == 0)
> pmap_remove_all(m);
This can sleep because it seizes a mutex for each entry.
> if (m->dirty) {
>+ /* Both paths use vm_pageout_flush() which forces
>+ * a syncronous putpage for the kernel_object.
>+ */
>+ if (flags & M_NOWAIT && object == kernel_object) {
>+ VM_OBJECT_UNLOCK(object);
>+ printf("vm_contig_launder_page: would sleep (kobject)\n");
>+ return (EWOULDBLOCK);
>+ }
> if (object->type == OBJT_VNODE) {
> vm_page_unlock_queues();
> vp = object->handle;
> VM_OBJECT_UNLOCK(object);
> vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
vn_lock(9) implies this can sleep unless the 2nd argument includes LK_NOWAIT.
> VM_OBJECT_LOCK(object);
This will sleep if something else has grabbed the object mutex.
>! vm_object_page_clean(object, 0, 0,
>! (flags & M_NOWAIT) ? 0 : OBJPC_SYNC);
> VM_OBJECT_UNLOCK(object);
> VOP_UNLOCK(vp, 0, curthread);
> vm_page_lock_queues();
This will sleep if something else is holding vm_page_queue_mtx.
>! if ((flags & M_NOWAIT) && m->dirty) {
>! printf("vm_contig_launder_page: would sleep (dirty)\n");
>! return (EWOULDBLOCK);
>! } else
>! return (0);
> } else if (object->type == OBJT_SWAP ||
> object->type == OBJT_DEFAULT) {
> m_tmp = m;
>! vm_pageout_flush(&m_tmp, 1,
>! (flags & M_NOWAIT) ? 0 : VM_PAGER_PUT_SYNC);
> VM_OBJECT_UNLOCK(object);
>! if ((flags & M_NOWAIT) && m->dirty) {
>! printf("vm_contig_launder_page: would sleep (dirty)\n");
>! return (EWOULDBLOCK);
>! } else
>! return (0);
> }
> } else if (m->hold_count == 0)
> vm_page_cache(m);
>***************
--
Peter Jeremy
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200511192340.jAJNeQLU095805>