On Wed, Nov 20, 2024 at 10:17:48AM +0100, Martin Pieuchot wrote:
> On 09/11/24(Sat) 13:23, Martin Pieuchot wrote:
> > The current swapping algorithm is broken for MP architectures.  However
> > the bouncing logic, and it sleeping point, in uvm_swap_io() hides it.
> > 
> > In order to see the degenerative behavior you need:
> > 
> > - A MP machine with more than 2 CPUs
> > - No high/low memory split and a lot of memory
> > - Disable swap encryption (sysctl vm.swapencrypt.enable=0)
> > 
> > When all these criteria are met the kernel ends up doing ping-pong for
> > every buffer between the page daemon, the aiodone daemon and the disk
> > interrupt handler, all currently needing the KERNEL_LOCK().
> > 
> > To fix this degenerative behavior, where the page daemon is starving the 
> > aiodone daemon and interrupt handler, we need to account for in-flight
> > pages.  Diff below does that by accounting for pages being currently
> > written to disk in the "shortage" calculation.
> > 
> > This diff makes swapping works as expected, no degenerative behavior, no
> > over-swapping on MP machine with the above mentioned specification.  I
> > believe it finally fixes swapping for 64 bits archs with a ton of memory.
> > 
> > Note that I deliberately did not use atomic_load_int() for reading
> > `uvmexp.paging'.  I do not want to introduce too many differences with
> > NetBSD for the moment and only a single thread (the page daemon) is
> > reading the counter.
> > 
> > This diff is a requirement for in-place encryption which, obviously,
> > does not use bouncing.
> > 
> > While here, release the KERNEL_LOCK() in the aiodone daemon when not
> > needed.
> 
> ok?
 
Some bikeshed comments below:

> > Index: uvm/uvm_pdaemon.c
> > ===================================================================
> > RCS file: /cvs/src/sys/uvm/uvm_pdaemon.c,v
> > diff -u -p -r1.129 uvm_pdaemon.c
> > --- uvm/uvm_pdaemon.c	7 Nov 2024 10:46:52 -0000	1.129
> > +++ uvm/uvm_pdaemon.c	9 Nov 2024 11:17:29 -0000
> > @@ -232,7 +232,7 @@ uvm_pageout(void *arg)
> >  		long size;
> >  
> >  		uvm_lock_fpageq();
> > -		if (TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
> > +		if (TAILQ_EMPTY(&uvm.pmr_control.allocs) || uvmexp.paging > 0) {
> >  			msleep_nsec(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
> >  			    "pgdaemon", INFSLP);
> >  			uvmexp.pdwoke++;
> > @@ -245,7 +245,8 @@ uvm_pageout(void *arg)
> >  			constraint = no_constraint;
> >  		}
> >  		/* How many pages do we need to free during this round? */
> > -		shortage = uvmexp.freetarg - uvmexp.free + BUFPAGES_DEFICIT;
> > +		shortage = uvmexp.freetarg -
> > +		    (uvmexp.free + uvmexp.paging) + BUFPAGES_DEFICIT;
> >  		uvm_unlock_fpageq();
> >  
> >  		/*
> > @@ -302,8 +303,7 @@ uvm_pageout(void *arg)
> >  		 * wake up any waiters.
> >  		 */
> >  		uvm_lock_fpageq();
> > -		if (uvmexp.free > uvmexp.reserve_kernel ||
> > -		    uvmexp.paging == 0) {
> > +		if (uvmexp.free > uvmexp.reserve_kernel || uvmexp.paging == 0) {
> >  			wakeup(&uvmexp.free);
> >  		}
> >  
> > @@ -339,10 +339,11 @@ uvm_pageout(void *arg)
> >  void
> >  uvm_aiodone_daemon(void *arg)
> >  {
> > -	int s, free;
> > +	int s, npages;
> >  	struct buf *bp, *nbp;
> >  
> >  	uvm.aiodoned_proc = curproc;
> > +	KERNEL_UNLOCK();
> >  
> >  	for (;;) {
> >  		/*
> > @@ -357,11 +358,13 @@ uvm_aiodone_daemon(void *arg)
> >  		TAILQ_INIT(&uvm.aio_done);
> >  		mtx_leave(&uvm.aiodoned_lock);
> >  
> > +

Extra new line.

> >  		/* process each i/o that's done. */
> > -		free = uvmexp.free;
> > +		npages = 0;
> > +		KERNEL_LOCK();
> >  		while (bp != NULL) {
> >  			if (bp->b_flags & B_PDAEMON) {
> > -				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
> > +				npages += bp->b_bufsize >> PAGE_SHIFT;
> >  			}
> >  			nbp = TAILQ_NEXT(bp, b_freelist);
> >  			s = splbio();	/* b_iodone must by called at splbio */
> > @@ -371,8 +374,11 @@ uvm_aiodone_daemon(void *arg)
> >  
> >  			sched_pause(yield);
> >  		}
> > +		KERNEL_UNLOCK();
> > +
> >  		uvm_lock_fpageq();
> > -		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
> > +		atomic_add_int(&uvmexp.paging, -npages);

Why not use atomic_sub_int here?

> > +		wakeup(uvmexp.free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
> >  		    &uvmexp.free);

Ugh. Not your issue but wakeup with a ternary operator looks crazy.

> >  		uvm_unlock_fpageq();
> >  	}
> > @@ -776,7 +782,7 @@ uvmpd_scan_inactive(struct uvm_pmalloc *
> >  		 */
> >  
> >  		if (result == VM_PAGER_PEND) {
> > -			uvmexp.paging += npages;
> > +			atomic_add_int(&uvmexp.paging, npages);
> >  			uvm_lock_pageq();
> >  			uvmexp.pdpending++;
> >  			if (p) {
> > Index: uvm/uvmexp.h
> > ===================================================================
> > RCS file: /cvs/src/sys/uvm/uvmexp.h,v
> > diff -u -p -r1.15 uvmexp.h
> > --- uvm/uvmexp.h	1 May 2024 12:54:27 -0000	1.15
> > +++ uvm/uvmexp.h	9 Nov 2024 11:11:44 -0000
> > @@ -60,7 +60,7 @@ struct uvmexp {
> >  	int free;       /* [F] number of free pages */
> >  	int active;     /* [L] # of active pages */
> >  	int inactive;   /* [L] # of pages that we free'd but may want back */
> > -	int paging;	/* number of pages in the process of being paged out */
> > +	int paging;	/* [a] # of pages in the process of being paged out */
> >  	int wired;      /* number of wired pages */
> >  
> >  	int zeropages;		/* [F] number of zero'd pages */
> > 
> > 
> 
> 

Diff looks good to me but I have no idea of UVM internals.

-- 
:wq Claudio