Re: OOMA kill with vm.pfault_oom_attempts="-1" on RPi3 at r357147 (a vm_pfault_oom_attempts < 0 handling bug as of head -r357026)

On Tue, Jan 28, 2020 at 11:28:14AM -0800, Mark Millard wrote:
> 
> 
> On 2020-Jan-28, at 11:02, bob prohaska <fbsd at www.zefox.net> wrote:
> 
> > On Tue, Jan 28, 2020 at 09:42:17AM -0800, Mark Millard wrote:
> >> 
> >> 
> >> 
> > The (partly)modified kernel compiled and booted without
> > obvious trouble. It's trying to finish buildworld now.
> > 
Stopped already, with 
Jan 28 11:41:59 www kernel: pid 29909 (cc), jid 0, uid 0, was killed: fault's page allocation failed



> >> If you are testing with vm.pfault_oom_attempts="-1" then
> >> the vm_fault printf message should never happen anyway.
> >> 
> > Would it not be interesting if the message appeared in that
> > case? 
> 
> Thanks for the question: looking at the new code found a bug
> causing oom where it used to be avoided in head -r357025 and
> before.


Glad to be of service, even if inadvertently 8-)

 
> After vm_waitpfault(dset, vm_pfault_oom_wait * hz)
> the -r357026 code does a vm_pageout_oom(VM_OOM_MEM_PF) no
> matter what, even when vm_pfault_oom_attempts < 0 ||
> fs->oom < vm_pfault_oom_attempts :
> 
> New code in head -r357026
> ( nothing to avoid the vm_pageout_oom(VM_OOM_MEM_PF)
> for vm_pfault_oom_attempts < 0 ||
> fs->oom < vm_pfault_oom_attempts ):
> 
> 	if (fs->m == NULL) {
> 		unlock_and_deallocate(fs);
> 		if (vm_pfault_oom_attempts < 0 ||
> 		    fs->oom < vm_pfault_oom_attempts) {
> 			fs->oom++;
> 			vm_waitpfault(dset, vm_pfault_oom_wait * hz);
> 		}
> 		if (bootverbose)
> 			printf(
> "proc %d (%s) failed to alloc page on fault, starting OOM\n",
> 			    curproc->p_pid, curproc->p_comm);
> 		vm_pageout_oom(VM_OOM_MEM_PF);
> 		return (KERN_RESOURCE_SHORTAGE);
> 	}
> 
> Old code in head -r357025
> ( has the goto RetryFault_oom after vm_waitpfault(. . .),
> thereby avoiding the vm_pageout_oom(VM_OOM_MEM_PF) for
> vm_pfault_oom_attempts < 0 || fs->oom < vm_pfault_oom_attempts ) :
> 
> 			if (fs.m == NULL) {
> 				unlock_and_deallocate(&fs);
> 				if (vm_pfault_oom_attempts < 0 ||
> 				    oom < vm_pfault_oom_attempts) {
> 					oom++;
> 					vm_waitpfault(dset,
> 					    vm_pfault_oom_wait * hz);
> 					goto RetryFault_oom;
> 				}
> 				if (bootverbose)
> 					printf(
> 	"proc %d (%s) failed to alloc page on fault, starting OOM\n",
> 					    curproc->p_pid, curproc->p_comm);
> 				vm_pageout_oom(VM_OOM_MEM_PF);
> 				goto RetryFault;
> 			}
> 
> I expect this is the source of the behavioral
> difference folks have been seeing for OOM kills.
> 
> 
> As for "gather evidence" messages . . .
> 
> >> You may be able to just look and manually delete or
> >> comment out the bootverbose line in the more modern
> >> source that currently looks like:
> >> 
> >> 		if (bootverbose)
> >> 			printf(
> >> "proc %d (%s) failed to alloc page on fault, starting OOM\n",
> >> 			    curproc->p_pid, curproc->p_comm);
> >> 		vm_pageout_oom(VM_OOM_MEM_PF);
> >> 		return (KERN_RESOURCE_SHORTAGE);
> >> 
> > 
> > I can find those lines in /usr/src/sys/vm/vm_fault.c, but
> > unclear on the motivation to comment the lines out. Perhaps 
> > to eliminate the return(...) ?  Anyway, is it sufficient 
> > to insert /* before and */ after? 
> 
> The only line to delete or comment out in that
> code block is:
> 
> 		if (bootverbose)
> 
> Disabling that line makes the following printf
> always happen, even when a verbose boot was not
> done.
Oops, it's commented out now and the kernel is rebuilding.

> 
> Based on the above reported code change, having
> a message before vm_pageout_oom(VM_OOM_MEM_PF) is
> important to getting a report of the kill being
> via that code.
>

Thank you!

bob prohaska