Index | Thread | Search

From:
Alexander Bluhm <bluhm@openbsd.org>
Subject:
Re: SEV: busdma bounce buffer
To:
Mark Kettenis <mark.kettenis@xs4all.nl>
Cc:
Hans-J?rg H?xer <Hans-Joerg_Hoexer@genua.de>, tech@openbsd.org
Date:
Mon, 12 Aug 2024 15:30:23 +0200

Download raw body.

Thread
On Mon, Aug 12, 2024 at 12:26:35PM +0200, Mark Kettenis wrote:
> > Date: Wed, 7 Aug 2024 17:50:57 +0200
> > From: Hans-J?rg H?xer <Hans-Joerg_Hoexer@genua.de>
> > 
> > Hi,
> 
> Hi Hans-Joerg,
> 
> > the diff below introduces bounce buffers to busdma(9) when the
> > kernel is running as a SEV-guest.  This is determined by checking
> > cpu_sev_guestmode (set in locore0).
> > 
> > The goal is, to make virtio(4) work for SEV-guests.  With the recenct
> > fixes done by sf@ -- thanks! -- the next step is dma bounce buffering
> > for SEV-enabled guests.
> 
> Is this just for virtio(4)?  Did you consider PCI passthrough?

Next goal is to get virtio(4) on a vmm(4) host running.  There
viornd, vio, vioblk, vmmci attach at virtio.  Everything else can
be fixed later.  PCI passthrough is not supported by vmm(4).

I can confirm that the diff below works on vmm, KVM, and native
hardware without SEV.  This is expected as the if (cpu_sev_guestmode)
prevents any change in behavior in normal operation.

So I think a diff like this can be commited when kettenis@ agrees
that the general direction is correct.

If bounce buffers are actually activated, there is fallout.  Currently
I am testing with this diff where I enforce bounce buffer regardless
of cpu_sev_guestmode.  Then every missing bus_dmamap_sync() causes
problems.

On vmm guest with forced bounce buffers, disk through vioblk works
well.  Network with vio interfaces stalls under heavy load.  After
that ifconfig vio down/up causes a panic, but sf@ has a fix for
that.

KVM guest with forced bounce buffers hangs during attach.  More
virtual devices show up that may have bus_dmamap_sync() bugs.

bluhm

> > _bus_dmamap_create():  On map creation we allocate an array of
> > PAGE_SIZEd bounce buffers.  These pages are mapped with PMAP_NOCRYPT.
> > This means, they are unencrypted and shared with the hypervisor.
> > Note, that we allocate at least one page for each dma segment.  This
> > allows us to not have worry about a single shared page holding clear
> > text data from different dma segments.  On the other hand, for
> > segments of size less than a page (e.g. only a few bytes), we always
> > use a full page.
> 
> That is an interesting design choice.  I have a slight worry that this
> may lead to excessive allocation of bounce buffers if a driver creates
> a map with many more segments than what is actually needed.  The
> bus_dma API does provide a BUS_DMA_ALLOCNOW flag and maybe you should
> respect that flag and delay allocation of the bounce buffers until
> bus_dmamap_load() gets called.  That said it seems all the virtio
> drivers, except viomb(4), set that flag.
> 
> Speaking about viomb(4), how does the ballooning work if you have
> encrypted memory?
> 
> > _bus_dmamap_sync():  On sync data gets now copied from the original
> > encrypted pages to the unencrypted bounce buffer; and vice versa.
> > 
> > _bus_dmamap_load_buffer():  When loading actual pages (these are
> > always encrypted) into the map, we replace the ds_addr in the
> > descriptor with the (guest physical) address of the boune buffer.
> > The original address is stored in the new member ds_addr2.
> 
> So this means that you also encrypt the rings.  Is there any point in
> doing this?  I mean you're sharing the contents of the ring with the
> host, so you're not realling hiding anything.  So it might be worth
> changing bus_dmamem_alloc() to allocate non-encrypted memory instead.
> Or rather have bus_dmamem_map() map it with the PMAP_NOCRYPT flag and
> set a flag in the segments such that bus_dmamap_load() can't skip the
> bouncing.
> 
> Anyway, it's entirely reasonable to postpone that as a future
> optimization.
> 
> > When the kernel is not in SEV guest mode, all this code is bypassed;
> > except the check on cpu_sev_guestmode in _bus_dmamap_sync().  Thus
> > I guess, that performance impact in non-SEV mode is neglectable.
> 
> Should be.  And we might actually want to use bounce buffers for PCI.
> So having the bounce buffers integrated in the generic implementation
> isn't a bad idea.
> 
> > Note:  Another approach could be to define a dedicated bus_dma_tag_t for
> > SEV guests.  Maybe similar to isa_bus_dma_tag.  I have not explored this.
> 
> A separate bus_dma tag might make sense if there is a need for further
> differentiation between virtio and other busses.
> 
> > I think, this diff is quite straight forward and in the end rather simple.
> > And I guess, there is plenty of room for optimization. And I want to
> > mention, that this diff is based on some (ancient) code from the late
> > mickey@.
> > 
> > What do you think?
> 
> I see some differences with the approach taken by the isa bounce
> buffer implementation that I'd like to take a closer look at.
> 
> A few more comments below.
> 
> > -------------------------------------------------------------------
> > commit 06d4c480dc94ae8a4ca2e6a6810aecb2c9e8afc7
> > Author: Hans-Joerg Hoexer <hshoexer@genua.de>
> > Date:   Fri Jul 12 10:56:19 2024 +0200
> > 
> >     busdma: Bounce buffering for SEV
> >     
> >     When running as SEV guest (indicated by cpu_sev_guestmode != 0),
> >     we allocate additional pages for each segment on dma map creation.
> >     These pages are mapped with the PMAP_NOCRYPT attribute, i.e. the
> >     crypt bit is not set in the PTE.  Thus, these pages are shared with
> >     the hypervisor.
> >     
> >     When the map is loaded with actual pages, the address in the
> >     descriptor is replaced with the corresponding bounce buffer.  Using
> >     bus_dmamap_sync(), data is copied from the encrypted pages used by
> >     guest-drivers to the unencrypted bounce buffers shared with the
> >     hypervisor.  And vice versa.
> >     
> >     When the kernel is running in non-SEV guest mode (i.e. normal host,
> >     or non-SEV guest), no bounce buffers are used.
> > 
> > diff --git a/sys/arch/amd64/amd64/bus_dma.c b/sys/arch/amd64/amd64/bus_dma.c
> > index 19a366115f9..32f21f5166e 100644
> > --- a/sys/arch/amd64/amd64/bus_dma.c
> > +++ b/sys/arch/amd64/amd64/bus_dma.c
> > @@ -108,8 +108,13 @@ _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
> >      bus_size_t maxsegsz, bus_size_t boundary, int flags, bus_dmamap_t *dmamp)
> >  {
> >  	struct bus_dmamap *map;
> > +	struct pglist mlist;
> > +	struct vm_page **pg, *pgnext;
> > +	size_t mapsize, sz, ssize;
> > +	vaddr_t va, sva;
> >  	void *mapstore;
> > -	size_t mapsize;
> > +	int npages, error;
> > +	const struct kmem_dyn_mode *kd;
> >  
> >  	/*
> >  	 * Allocate and initialize the DMA map.  The end of the map
> > @@ -125,6 +130,16 @@ _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
> >  	 */
> >  	mapsize = sizeof(struct bus_dmamap) +
> >  	    (sizeof(bus_dma_segment_t) * (nsegments - 1));
> > +
> > +	/* allocate and use bounce buffers when running as SEV guest */
> > +	if (cpu_sev_guestmode) {
> > +		/* this many pages plus one in case we get split */
> > +		npages = round_page(size) / PAGE_SIZE + 1;
> > +		if (npages < nsegments)
> > +			npages = nsegments;
> > +		mapsize += sizeof(struct vm_page *) * npages;
> > +	}
> > +
> >  	if ((mapstore = malloc(mapsize, M_DEVBUF,
> >  	    (flags & BUS_DMA_NOWAIT) ?
> >  	        (M_NOWAIT|M_ZERO) : (M_WAITOK|M_ZERO))) == NULL)
> > @@ -135,8 +150,59 @@ _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
> >  	map->_dm_segcnt = nsegments;
> >  	map->_dm_maxsegsz = maxsegsz;
> >  	map->_dm_boundary = boundary;
> > +	if (cpu_sev_guestmode) {
> > +		map->_dm_pages = (void *)&map->dm_segs[nsegments];
> > +		map->_dm_npages = npages;
> > +	}
> >  	map->_dm_flags = flags & ~(BUS_DMA_WAITOK|BUS_DMA_NOWAIT);
> >  
> > +	if (!cpu_sev_guestmode) {
> > +		*dmamp = map;
> > +		return (0);
> > +	}
> > +
> > +	sz = npages << PGSHIFT;
> > +	kd = flags & BUS_DMA_NOWAIT ? &kd_trylock : &kd_waitok;
> > +	va = (vaddr_t)km_alloc(sz, &kv_any, &kp_none, kd);
> > +	if (va == 0) {
> > +		map->_dm_npages = 0;
> > +		free(map, M_DEVBUF, mapsize);
> > +		return (ENOMEM);
> > +	}
> > +
> > +	TAILQ_INIT(&mlist);
> > +	error = uvm_pglistalloc(sz, 0, -1, PAGE_SIZE, 0, &mlist, nsegments,
> > +	    (flags & BUS_DMA_NOWAIT) ? UVM_PLA_NOWAIT : UVM_PLA_WAITOK);
> 
> This allocates memory that isn't necessarily DMA-safe.  I guess that
> is fine for virtio(4) since it doesn't care.  But a potential issue if
> we want to use this for PCI passthrough?
> 
> > +	if (error) {
> > +		map->_dm_npages = 0;
> > +		km_free((void *)va, sz, &kv_any, &kp_none);
> > +		free(map, M_DEVBUF, mapsize);
> > +		return (ENOMEM);
> > +	}
> > +
> > +	sva = va;
> > +	ssize = sz;
> > +	pgnext = TAILQ_FIRST(&mlist);
> > +	for (pg = map->_dm_pages; npages--; va += PAGE_SIZE, pg++) {
> > +		*pg = pgnext;
> > +		error = pmap_enter(pmap_kernel(), va, VM_PAGE_TO_PHYS(*pg),
> > +		    PROT_READ | PROT_WRITE,
> > +		    PROT_READ | PROT_WRITE | PMAP_WIRED |
> > +		    PMAP_CANFAIL | PMAP_NOCRYPT);
> > +		if (error) {
> > +			pmap_update(pmap_kernel());
> > +			map->_dm_npages = 0;
> > +			km_free((void *)sva, ssize, &kv_any, &kp_none);
> > +			free(map, M_DEVBUF, mapsize);
> > +			uvm_pglistfree(&mlist);
> > +			return (ENOMEM);
> > +		}
> > +		pgnext = TAILQ_NEXT(*pg, pageq);
> > +		bzero((void *)va, PAGE_SIZE);
> > +	}
> > +	pmap_update(pmap_kernel());
> > +	map->_dm_pgva = sva;
> > +
> >  	*dmamp = map;
> >  	return (0);
> >  }
> > @@ -149,7 +215,25 @@ void
> >  _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map)
> >  {
> >  	size_t mapsize;
> > -	
> > +	struct vm_page **pg;
> > +	struct pglist mlist;
> > +	vaddr_t va;
> > +
> > +	if (map->_dm_pgva) {
> > +		km_free((void *)map->_dm_pgva, map->_dm_npages << PGSHIFT,
> > +		    &kv_any, &kp_none);
> > +	}
> > +
> > +	if (map->_dm_pages) {
> > +		TAILQ_INIT(&mlist);
> > +		for (pg = map->_dm_pages, va = map->_dm_pgva;
> > +		    map->_dm_npages--; va += PAGE_SIZE, pg++) {
> > +			pmap_remove(pmap_kernel(), va, va + PAGE_SIZE);
> > +			TAILQ_INSERT_TAIL(&mlist, *pg, pageq);
> > +		}
> > +		uvm_pglistfree(&mlist);
> > +	}
> > +
> >  	mapsize = sizeof(struct bus_dmamap) +
> >  		(sizeof(bus_dma_segment_t) * (map->_dm_segcnt - 1));
> >  	free(map, M_DEVBUF, mapsize);
> > @@ -383,6 +467,7 @@ _bus_dmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
> >  	 */
> >  	map->dm_mapsize = 0;
> >  	map->dm_nsegs = 0;
> > +	map->_dm_nused = 0;
> >  }
> >  
> >  /*
> > @@ -393,7 +478,40 @@ void
> >  _bus_dmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t addr,
> >      bus_size_t size, int op)
> >  {
> > -	/* Nothing to do here. */
> > +	bus_dma_segment_t *sg;
> > +	int i, off = addr;
> > +	bus_size_t l;
> > +
> > +	if (!cpu_sev_guestmode)
> > +		return;
> > +
> > +	for (i = map->_dm_segcnt, sg = map->dm_segs; size && i--; sg++) {
> > +		if (off >= sg->ds_len) {
> > +			off -= sg->ds_len;
> > +			continue;
> > +		}
> > +
> > +		l = sg->ds_len - off;
> > +		if (l > size)
> > +			l = size;
> > +		size -= l;
> > +
> > +		/* PREREAD and POSTWRITE are no-ops. */
> > +
> > +		/* READ: device -> memory */
> > +		if (op & BUS_DMASYNC_POSTREAD) {
> > +			bcopy((void *)(sg->ds_bounce + off),
> > +			    (void *)(sg->ds_va + off), l);
> > +		}
> > +
> > +		/* WRITE: memory -> device */
> > +		if (op & BUS_DMASYNC_PREWRITE) {
> > +			bcopy((void *)(sg->ds_va + off),
> > +			    (void *)(sg->ds_bounce + off), l);
> > +		}
> > +
> > +		off = 0;
> > +	}
> >  }
> >  
> >  /*
> > @@ -565,10 +683,11 @@ _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
> >      int first)
> >  {
> >  	bus_size_t sgsize;
> > -	bus_addr_t curaddr, lastaddr, baddr, bmask;
> > -	vaddr_t vaddr = (vaddr_t)buf;
> > -	int seg;
> > +	bus_addr_t curaddr, lastaddr, baddr, bmask, oaddr = -1;
> > +	vaddr_t pgva = -1, vaddr = (vaddr_t)buf;
> > +	int seg, page, off;
> >  	pmap_t pmap;
> > +	struct vm_page *pg;
> >  
> >  	if (p != NULL)
> >  		pmap = p->p_vmspace->vm_map.pmap;
> > @@ -589,6 +708,19 @@ _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
> >  			panic("Non dma-reachable buffer at curaddr %#lx(raw)",
> >  			    curaddr);
> >  
> > +		if (cpu_sev_guestmode) {
> > +			/* use bounce buffer */
> > +			if (map->_dm_nused + 1 >= map->_dm_npages)
> > +				return (ENOMEM);
> > +
> > +			off = vaddr & PAGE_MASK;
> > +			pg = map->_dm_pages[page = map->_dm_nused++];
> > +			oaddr = curaddr;
> > +			curaddr = VM_PAGE_TO_PHYS(pg) + off;
> > +
> > +			pgva = map->_dm_pgva + (page << PGSHIFT) + off;
> > +		}
> > +
> >  		/*
> >  		 * Compute the segment size, and adjust counts.
> >  		 */
> > @@ -611,7 +743,10 @@ _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
> >  		 */
> >  		if (first) {
> >  			map->dm_segs[seg].ds_addr = curaddr;
> > +			map->dm_segs[seg].ds_addr2 = oaddr;
> >  			map->dm_segs[seg].ds_len = sgsize;
> > +			map->dm_segs[seg].ds_va = vaddr;
> > +			map->dm_segs[seg].ds_bounce = pgva;
> >  			first = 0;
> >  		} else {
> >  			if (curaddr == lastaddr &&
> > @@ -625,7 +760,10 @@ _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
> >  				if (++seg >= map->_dm_segcnt)
> >  					break;
> >  				map->dm_segs[seg].ds_addr = curaddr;
> > +				map->dm_segs[seg].ds_addr2 = oaddr;
> >  				map->dm_segs[seg].ds_len = sgsize;
> > +				map->dm_segs[seg].ds_va = vaddr;
> > +				map->dm_segs[seg].ds_bounce = pgva;
> >  			}
> >  		}
> >  
> > diff --git a/sys/arch/amd64/include/bus.h b/sys/arch/amd64/include/bus.h
> > index 33d6cd6eaeb..e3ceaf71ee7 100644
> > --- a/sys/arch/amd64/include/bus.h
> > +++ b/sys/arch/amd64/include/bus.h
> > @@ -551,7 +551,11 @@ typedef struct bus_dmamap		*bus_dmamap_t;
> >   */
> >  struct bus_dma_segment {
> >  	bus_addr_t	ds_addr;	/* DMA address */
> > +	bus_addr_t	ds_addr2;	/* replacement store */
> 
> A more descriptive name would be better.  Also, because this isn't
> part of the bus_dma API, it should probably start with an underscore.
> So perhaps _ds_bounce_addr?
> 
> >  	bus_size_t	ds_len;		/* length of transfer */
> > +	vaddr_t		ds_va;		/* mapped loaded data */
> > +	vaddr_t		ds_bounce;	/* mapped bounced data */
> > +
> 
> Same here.  For consitencey maybe _ds_va and _ds_bounce_va?
> 
> >  	/*
> >  	 * Ugh. need this so can pass alignment down from bus_dmamem_alloc
> >  	 * to scatter gather maps. only the first one is used so the rest is
> > @@ -655,6 +659,11 @@ struct bus_dmamap {
> >  
> >  	void		*_dm_cookie;	/* cookie for bus-specific functions */
> >  
> > +	struct vm_page **_dm_pages;	/* replacement pages */
> > +	vaddr_t		_dm_pgva;	/* those above -- mapped */
> > +	int		_dm_npages;	/* number of pages allocated */
> > +	int		_dm_nused;	/* number of pages replaced */
> > +
> >  	/*
> >  	 * PUBLIC MEMBERS: these are used by machine-independent code.
> >  	 */