usbehci: initial support for usb on zynq, remove uncached.h

the following hooks have been added to the ehci Ctlr
structore to handle cache coherency (on arm):

	void*	(*tdalloc)(ulong,int,ulong);
	void*	(*dmaalloc)(ulong);
	void	(*dmafree)(void*);
	void	(*dmaflush)(int,void*,ulong);

tdalloc() is used to allocate descriptors and the periodic
frame schedule array. on arm, this needs to return uncached
memory. tdalloc()ed memory is never freed.

dmaalloc()/dmafree() is used for io buffers. this can return
cached memory when when hardware maintains cache coherency (pc)
or dmaflush() is provided to flush/invalidate the cache (zynq),
otherwise needs to return uncached memory.

dmaflush() is used to flush/invalidate the cache. the first
argument tells us if we need to flush (non zero) or
invalidate (zero).

uncached.h is gone now. this change makes the handling explicit.

20 files changed, 178 insertions, 172 deletions

diff --git a/sys/src/9/kw/fns.h b/sys/src/9/kw/fns.h
index b2f4d0f99..b862dd835 100644
--- a/sys/src/9/kw/fns.h
+++ b/sys/src/9/kw/fns.h
@@ -110,7 +110,7 @@ uintptr mmukmap(uintptr, uintptr, usize);
 uintptr mmukunmap(uintptr, uintptr, usize);
 extern void* mmuuncache(void*, usize);
 extern void* ucalloc(usize);
-extern void* ucallocalign(usize size, int align, int span);
+extern void* ucallocalign(usize size, int align, usize span);
 extern Block* ucallocb(int);
 extern void ucfree(void*);
 extern void ucfreeb(Block*);
diff --git a/sys/src/9/kw/mkfile b/sys/src/9/kw/mkfile
index 455dd2a37..219d06ffa 100644
--- a/sys/src/9/kw/mkfile
+++ b/sys/src/9/kw/mkfile
@@ -132,7 +132,7 @@ l.$O lexception.$O lproc.$O mmu.$O: arm.s arm.h mem.h
 main.$O:	errstr.h init.h reboot.h
 mouse.$O:	screen.h
 devusb.$O:	../port/usb.h
-usbehci.$O usbohci.$O usbuhci.$O: ../port/usb.h usbehci.h uncached.h
+usbehci.$O usbohci.$O usbuhci.$O: ../port/usb.h usbehci.h
 
 init.h:D:	../port/initcode.c init9.s
 	$CC ../port/initcode.c
diff --git a/sys/src/9/kw/mmu.c b/sys/src/9/kw/mmu.c
index 5128e2cee..377b1a596 100644
--- a/sys/src/9/kw/mmu.c
+++ b/sys/src/9/kw/mmu.c
@@ -261,7 +261,7 @@ mmurelease(Proc* proc)
 	for(page = proc->mmul2cache; page != nil; page = next){
 		next = page->next;
 		if(--page->ref)
-			panic("mmurelease: page->ref %d", page->ref);
+			panic("mmurelease: page->ref %lud", page->ref);
 		pagechainhead(page);
 	}
 	if(proc->mmul2cache && palloc.r.p)
diff --git a/sys/src/9/kw/ucalloc.c b/sys/src/9/kw/ucalloc.c
index 37e8564c0..f7d142bad 100644
--- a/sys/src/9/kw/ucalloc.c
+++ b/sys/src/9/kw/ucalloc.c
@@ -127,7 +127,7 @@ ucalloc(usize size)
 }
 
 void*
-ucallocalign(usize size, int align, int span)
+ucallocalign(usize size, int align, usize span)
 {
 	assert(size < ucpool.minarena-128);
 
diff --git a/sys/src/9/kw/uncached.h b/sys/src/9/kw/uncached.h
deleted file mode 100644
index 9fffd31fc..000000000
--- a/sys/src/9/kw/uncached.h
+++ /dev/null
@@ -1,36 +0,0 @@
-/*
- * running the l2 cache as write-back and using cached memory for
- * usb data structures yields spurious errors such as
- *
- *	qhintr: td 0x60ee3d80 csw 0x8824a error 0x48 transaction error
- *
- * from usbehci.  so, at least for now, we will use uncached memory until
- * we sort out the write-back problems.
- */
-#define free			ucfree
-#define malloc			myucalloc
-#define mallocz			ucallocz
-#define smalloc			myucalloc
-#define xspanalloc		ucallocalign
-
-#define allocb			ucallocb
-#define iallocb			uciallocb
-#define freeb			ucfreeb
-
-static void *
-ucallocz(uint n, int)
-{
-	char *p = ucalloc(n);
-
-	if (p)
-		memset(p, 0, n);
-	else
-		panic("ucalloc: out of memory");
-	return p;
-}
-
-static void *
-myucalloc(uint n)
-{
-	return ucallocz(n, 1);
-}
diff --git a/sys/src/9/kw/usbehci.h b/sys/src/9/kw/usbehci.h
index cd1015a8c..57d172279 100644
--- a/sys/src/9/kw/usbehci.h
+++ b/sys/src/9/kw/usbehci.h
@@ -172,6 +172,11 @@ struct Ctlr
 	Ecapio*	capio;		/* Capability i/o regs */
 	Eopio*	opio;		/* Operational i/o regs */
 
+	void*	(*tdalloc)(ulong,int,ulong);
+	void*	(*dmaalloc)(ulong);
+	void	(*dmafree)(void*);
+	void	(*dmaflush)(int,void*,ulong len);
+
 	int	nframes;	/* 1024, 512, or 256 frames in the list */
 	ulong*	frames;		/* periodic frame list (hw) */
 	Qh*	qhs;		/* async Qh circular list for bulk/ctl */
diff --git a/sys/src/9/kw/usbehcikw.c b/sys/src/9/kw/usbehcikw.c
index 83e6f1e56..41b8051fa 100644
--- a/sys/src/9/kw/usbehcikw.c
+++ b/sys/src/9/kw/usbehcikw.c
@@ -13,7 +13,6 @@
 #include	"../port/error.h"
 #include	"../port/usb.h"
 #include	"usbehci.h"
-//#include	"uncached.h"
 
 #define WINTARG(ctl)	(((ctl) >> 4) & 017)
 #define WINATTR(ctl)	(((ctl) >> 8) & 0377)
@@ -331,6 +330,10 @@ reset(Hci *hp)
 		capio->parms & 0x40 ? "explicit" : "automatic",
 		capio->parms & 0x10 ? "" : "no ", hp->nports);
 
+	ctlr->tdalloc = ucallocalign;
+	ctlr->dmaalloc = ucalloc;
+	ctlr->dmafree = ucfree;
+
 	ehcireset(ctlr);
 	ehcimeminit(ctlr);
 
diff --git a/sys/src/9/omap/mkfile b/sys/src/9/omap/mkfile
index 321d2971f..1b6fdef81 100644
--- a/sys/src/9/omap/mkfile
+++ b/sys/src/9/omap/mkfile
@@ -111,7 +111,7 @@ l.$O rebootcode.$O: cache.v7.s
 main.$O: errstr.h init.h reboot.h
 devdss.$O devmouse.$O mouse.$O screen.$O: screen.h
 devusb.$O: ../port/usb.h
-usbehci.$O usbohci.$O usbuhci.$O: ../port/usb.h usbehci.h uncached.h
+usbehci.$O usbohci.$O usbuhci.$O: ../port/usb.h usbehci.h
 
 init.h:D:	../port/initcode.c init9.s
 	$CC ../port/initcode.c
diff --git a/sys/src/9/omap/uncached.h b/sys/src/9/omap/uncached.h
deleted file mode 100644
index 9fffd31fc..000000000
--- a/sys/src/9/omap/uncached.h
+++ /dev/null
@@ -1,36 +0,0 @@
-/*
- * running the l2 cache as write-back and using cached memory for
- * usb data structures yields spurious errors such as
- *
- *	qhintr: td 0x60ee3d80 csw 0x8824a error 0x48 transaction error
- *
- * from usbehci.  so, at least for now, we will use uncached memory until
- * we sort out the write-back problems.
- */
-#define free			ucfree
-#define malloc			myucalloc
-#define mallocz			ucallocz
-#define smalloc			myucalloc
-#define xspanalloc		ucallocalign
-
-#define allocb			ucallocb
-#define iallocb			uciallocb
-#define freeb			ucfreeb
-
-static void *
-ucallocz(uint n, int)
-{
-	char *p = ucalloc(n);
-
-	if (p)
-		memset(p, 0, n);
-	else
-		panic("ucalloc: out of memory");
-	return p;
-}
-
-static void *
-myucalloc(uint n)
-{
-	return ucallocz(n, 1);
-}
diff --git a/sys/src/9/omap/usbehci.h b/sys/src/9/omap/usbehci.h
index b43d4774b..45e88169b 100644
--- a/sys/src/9/omap/usbehci.h
+++ b/sys/src/9/omap/usbehci.h
@@ -39,6 +39,11 @@ struct Ctlr
 	Ecapio*	capio;		/* Capability i/o regs */
 	Eopio*	opio;		/* Operational i/o regs */
 
+	void*	(*tdalloc)(ulong,int,ulong);
+	void*	(*dmaalloc)(ulong);
+	void	(*dmafree)(void*);
+	void	(*dmaflush)(int,void*,ulong len);
+
 	int	nframes;	/* 1024, 512, or 256 frames in the list */
 	ulong*	frames;		/* periodic frame list (hw) */
 	Qh*	qhs;		/* async Qh circular list for bulk/ctl */
diff --git a/sys/src/9/omap/usbehciomap.c b/sys/src/9/omap/usbehciomap.c
index 089dd525b..0634db69c 100644
--- a/sys/src/9/omap/usbehciomap.c
+++ b/sys/src/9/omap/usbehciomap.c
@@ -178,6 +178,10 @@ reset(Hci *hp)
 		capio->parms & 0x40 ? "explicit" : "automatic",
 		capio->parms & 0x10 ? "" : "no ", hp->nports);
 
+	ctlr->tdalloc = ucallocalign;
+	ctlr->dmaalloc = ucalloc;
+	cltr->dmafree = ucfree;
+
 	ehcireset(ctlr);
 	ehcimeminit(ctlr);
 
diff --git a/sys/src/9/pc/uncached.h b/sys/src/9/pc/uncached.h
deleted file mode 100644
index c8aff3ac8..000000000
--- a/sys/src/9/pc/uncached.h
+++ /dev/null
@@ -1,4 +0,0 @@
-/*
- * On the PC, processor accesses, memory caches and DMA are all
- * coherent, so we don't need to use uncached memory.
- */
diff --git a/sys/src/9/pc/usbehci.h b/sys/src/9/pc/usbehci.h
index 1c833b06b..885bdc606 100644
--- a/sys/src/9/pc/usbehci.h
+++ b/sys/src/9/pc/usbehci.h
@@ -174,6 +174,11 @@ struct Ctlr
 	Ecapio*	capio;		/* Capability i/o regs */
 	Eopio*	opio;		/* Operational i/o regs */
 
+	void*	(*tdalloc)(ulong,int,ulong);
+	void*	(*dmaalloc)(ulong);
+	void	(*dmafree)(void*);
+	void	(*dmaflush)(int,void*,ulong len);
+
 	int	nframes;	/* 1024, 512, or 256 frames in the list */
 	ulong*	frames;		/* periodic frame list (hw) */
 	Qh*	qhs;		/* async Qh circular list for bulk/ctl */
diff --git a/sys/src/9/pc64/mkfile b/sys/src/9/pc64/mkfile
index 9dd5f4c78..fc052d291 100644
--- a/sys/src/9/pc64/mkfile
+++ b/sys/src/9/pc64/mkfile
@@ -83,7 +83,7 @@ install:V:	$p$CONF
 
 
 # copies generated by the rule below
-PCHEADERS=wifi.h uncached.h usbehci.h screen.h etherif.h ethermii.h mp.h io.h ahci.h
+PCHEADERS=wifi.h usbehci.h screen.h etherif.h ethermii.h mp.h io.h ahci.h
 
 REPCH=`{echo $PCHEADERS | sed 's/\.h//g; s/ /|/g'}
 ^($REPCH)\.h:R:	'../pc/\1.h'
@@ -117,7 +117,7 @@ devaoe.$O sdaoe.$O:		../port/aoe.h
 main.$O:			init.h reboot.h
 
 devusb.$O usbuhci.$O usbohci.$O usbehci.$O: ../port/usb.h
-usbehci.$O:			usbehci.h uncached.h
+usbehci.$O:			usbehci.h
 trap.$O:			/sys/include/tos.h
 ethermii.$O:			ethermii.h
 etheriwl.$O:			wifi.h
diff --git a/sys/src/9/port/usbehci.c b/sys/src/9/port/usbehci.c
index 2db838780..d9fd74d15 100644
--- a/sys/src/9/port/usbehci.c
+++ b/sys/src/9/port/usbehci.c
@@ -20,7 +20,6 @@
 #include	"../port/error.h"
 #include	"../port/usb.h"
 #include	"usbehci.h"
-#include	"uncached.h"
 
 #define diprint		if(ehcidebug || iso->debug)print
 #define ddiprint	if(ehcidebug>1 || iso->debug>1)print
@@ -399,14 +398,14 @@ ehcirun(Ctlr *ctlr, int on)
 }
 
 static void*
-edalloc(void)
+edalloc(Ctlr *ctlr)
 {
 	Ed *ed, *pool;
 	int i;
 
 	lock(&edpool);
 	if(edpool.free == nil){
-		pool = xspanalloc(Incr*sizeof(Ed), Align, 0);
+		pool = (*ctlr->tdalloc)(Incr*sizeof(Ed), Align, 0);
 		if(pool == nil)
 			panic("edalloc");
 		for(i=Incr; --i>=0;){
@@ -448,53 +447,54 @@ edfree(void *a)
  */
 
 static Itd*
-itdalloc(void)
+itdalloc(Ctlr *ctlr)
 {
 	Itd *td;
 
-	td = edalloc();
+	td = edalloc(ctlr);
 	td->link = Lterm;
 	return td;
 }
 
 static void
-itdfree(Itd *td)
+itdfree(Ctlr*, Itd *td)
 {
 	edfree(td);
 }
 
 static Sitd*
-sitdalloc(void)
+sitdalloc(Ctlr *ctlr)
 {
 	Sitd *td;
 
-	td = edalloc();
+	td = edalloc(ctlr);
 	td->link = td->blink = Lterm;
 	return td;
 }
 
 static void
-sitdfree(Sitd *td)
+sitdfree(Ctlr*, Sitd *td)
 {
 	edfree(td);
 }
 
 static Td*
-tdalloc(void)
+tdalloc(Ctlr *ctlr)
 {
 	Td *td;
 
-	td = edalloc();
+	td = edalloc(ctlr);
 	td->nlink = td->alink = Lterm;
 	return td;
 }
 
 static void
-tdfree(Td *td)
+tdfree(Ctlr *ctlr, Td *td)
 {
 	if(td == nil)
 		return;
-	free(td->buff);
+	if(td->buff != nil)
+		(*ctlr->dmafree)(td->buff);
 	edfree(td);
 }
 
@@ -661,7 +661,7 @@ qhalloc(Ctlr *ctlr, Ep *ep, Qio *io, char* tag)
 	Qh *qh;
 	int ttype;
 
-	qh = edalloc();
+	qh = edalloc(ctlr);
 	qh->nlink = Lterm;
 	qh->alink = Lterm;
 	qh->csw = Tdhalt;
@@ -777,7 +777,7 @@ qhfree(Ctlr *ctlr, Qh *qh)
 
 	while((td = qh->tds) != nil){
 		qh->tds = td->next;
-		tdfree(td);
+		tdfree(ctlr, td);
 	}
 
 	edfree(qh);
@@ -1236,7 +1236,7 @@ isocanwrite(void *a)
 }
 
 static void
-itdinit(Isoio *iso, Itd *td)
+itdinit(Ctlr *ctlr, Isoio *iso, Itd *td)
 {
 	int p, t;
 	ulong pa, tsize, size;
@@ -1248,6 +1248,8 @@ itdinit(Isoio *iso, Itd *td)
 	 */
 	p = 0;
 	size = td->ndata = td->mdata;
+	if(ctlr->dmaflush != nil)
+		(*ctlr->dmaflush)(1, td->data, size);
 	pa = PADDR(td->data);
 	for(t = 0; size > 0 && t < 8; t++){
 		tsize = size;
@@ -1265,8 +1267,10 @@ itdinit(Isoio *iso, Itd *td)
 }
 
 static void
-sitdinit(Isoio *iso, Sitd *td)
+sitdinit(Ctlr *ctlr, Isoio *iso, Sitd *td)
 {
+	if(ctlr->dmaflush != nil)
+		(*ctlr->dmaflush)(1, td->data, td->mdata);
 	td->ndata = td->mdata & Stdlenmask;
 	td->buffer[0] = PADDR(td->data);
 	td->buffer[1] = (td->buffer[0] & ~0xFFF) + 0x1000;
@@ -1379,7 +1383,7 @@ isohsinterrupt(Ctlr *ctlr, Isoio *iso)
 			tdi->ndata = 0;
 		if(tdi->next == iso->tdu || tdi->next->next == iso->tdu){
 			memset(iso->tdu->data, 0, iso->tdu->mdata);
-			itdinit(iso, iso->tdu);
+			itdinit(ctlr, iso, iso->tdu);
 			iso->tdu = iso->tdu->next;
 			iso->nleft = 0;
 		}
@@ -1444,8 +1448,7 @@ isofsinterrupt(Ctlr *ctlr, Isoio *iso)
 
 		if(stdi->next == iso->stdu || stdi->next->next == iso->stdu){
 			memset(iso->stdu->data, 0, iso->stdu->mdata);
-			coherence();
-			sitdinit(iso, iso->stdu);
+			sitdinit(ctlr, iso, iso->stdu);
 			iso->stdu = iso->stdu->next;
 			iso->nleft = 0;
 		}
@@ -1883,6 +1886,8 @@ episohscpy(Ctlr *ctlr, Ep *ep, Isoio* iso, uchar *b, long count)
 				ep->dev->nb, ep->nb);
 		else{
 			iunlock(ctlr);		/* We could page fault here */
+			if(ctlr->dmaflush != nil)
+				(*ctlr->dmaflush)(0, tdu->data, tdu->mdata);
 			memmove(b+tot, tdu->data, nr);
 			ilock(ctlr);
 			if(iso->tdu != tdu)
@@ -1893,7 +1898,7 @@ episohscpy(Ctlr *ctlr, Ep *ep, Isoio* iso, uchar *b, long count)
 			coherence();
 		}
 		if(tdu->ndata == 0){
-			itdinit(iso, tdu);
+			itdinit(ctlr, iso, tdu);
 			iso->tdu = tdu->next;
 		}
 	}
@@ -1921,6 +1926,8 @@ episofscpy(Ctlr *ctlr, Ep *ep, Isoio* iso, uchar *b, long count)
 				ep->dev->nb, ep->nb);
 		else{
 			iunlock(ctlr);		/* We could page fault here */
+			if(ctlr->dmaflush != nil)
+				(*ctlr->dmaflush)(0, stdu->data, stdu->mdata);
 			memmove(b+tot, stdu->data, nr);
 			ilock(ctlr);
 			if(iso->stdu != stdu)
@@ -1932,7 +1939,7 @@ episofscpy(Ctlr *ctlr, Ep *ep, Isoio* iso, uchar *b, long count)
 			coherence();
 		}
 		if(stdu->ndata == 0){
-			sitdinit(iso, stdu);
+			sitdinit(ctlr, iso, stdu);
 			iso->stdu = stdu->next;
 		}
 	}
@@ -2024,7 +2031,7 @@ putsamples(Ctlr *ctlr, Isoio *iso, uchar *b, long count)
 				continue;
 			iso->nleft += n;
 			if(iso->nleft == tdu->mdata){
-				itdinit(iso, tdu);
+				itdinit(ctlr, iso, tdu);
 				iso->tdu = tdu->next;
 				iso->nleft = 0;
 			}
@@ -2039,7 +2046,7 @@ putsamples(Ctlr *ctlr, Isoio *iso, uchar *b, long count)
 				continue;
 			iso->nleft += n;
 			if(iso->nleft == stdu->mdata){
-				sitdinit(iso, stdu);
+				sitdinit(ctlr, iso, stdu);
 				iso->stdu = stdu->next;
 				iso->nleft = 0;
 			}
@@ -2137,7 +2144,7 @@ nexttoggle(int toggle, int count, int maxpkt)
 }
 
 static Td*
-epgettd(Qio *io, int flags, void *a, int count, int maxpkt)
+epgettd(Ctlr *ctlr, Qio *io, int flags, void *a, int count, int maxpkt)
 {
 	Td *td;
 	ulong pa;
@@ -2145,7 +2152,7 @@ epgettd(Qio *io, int flags, void *a, int count, int maxpkt)
 
 	if(count > Tdmaxpkt)
 		panic("ehci: epgettd: too many bytes");
-	td = tdalloc();
+	td = tdalloc(ctlr);
 	td->csw = flags | io->toggle | io->tok | count << Tdlenshift |
 		Tderr2 | Tderr1;
 	coherence();
@@ -2159,12 +2166,16 @@ epgettd(Qio *io, int flags, void *a, int count, int maxpkt)
 		td->data = td->sbuff;
 		td->buff = nil;
 	} else if(count <= 0x4000){
-		td->buff = td->data = smalloc(count);
+		td->buff = td->data = (*ctlr->dmaalloc)(count);
 	} else {
-		td->buff = smalloc(count + 0x1000);
+		td->buff = (*ctlr->dmaalloc)(count+0x1000);
 		td->data = (uchar*)ROUND((uintptr)td->buff, 0x1000);
 	}
-
+	if(a != nil && count > 0){
+		memmove(td->data, a, count);
+		if(ctlr->dmaflush != nil && td->buff != nil)
+			(*ctlr->dmaflush)(1, td->data, count);
+	}
 	pa = PADDR(td->data);
 	for(i = 0; i < nelem(td->buffer); i++){
 		td->buffer[i] = pa;
@@ -2172,8 +2183,6 @@ epgettd(Qio *io, int flags, void *a, int count, int maxpkt)
 		pa += 0x1000;
 	}
 	td->ndata = count;
-	if(a != nil && count > 0)
-		memmove(td->data, a, count);
 	coherence();
 	io->toggle = nexttoggle(io->toggle, count, maxpkt);
 	coherence();
@@ -2386,9 +2395,9 @@ epio(Ep *ep, Qio *io, void *a, long count, int mustlock)
 		if(count-tot < n)
 			n = count-tot;
 		if(c != nil && io->tok != Tdtokin)
-			td = epgettd(io, Tdactive, c+tot, n, ep->maxpkt);
+			td = epgettd(ctlr, io, Tdactive, c+tot, n, ep->maxpkt);
 		else
-			td = epgettd(io, Tdactive, nil, n, ep->maxpkt);
+			td = epgettd(ctlr, io, Tdactive, nil, n, ep->maxpkt);
 		if(td0 == nil)
 			td0 = td;
 		else
@@ -2452,6 +2461,8 @@ epio(Ep *ep, Qio *io, void *a, long count, int mustlock)
 				if((tot + n) > count)
 					n = count - tot;
 				if(c != nil && (td->csw & Tdtok) == Tdtokin){
+					if(ctlr->dmaflush != nil && td->buff != nil)
+						(*ctlr->dmaflush)(0, td->data, n);
 					memmove(c, td->data, n);
 					c += n;
 				}
@@ -2459,7 +2470,7 @@ epio(Ep *ep, Qio *io, void *a, long count, int mustlock)
 			}
 		}
 		ntd = td->next;
-		tdfree(td);
+		tdfree(ctlr, td);
 	}
 	if(mustlock){
 		qunlock(io);
@@ -2679,12 +2690,14 @@ isofsinit(Ep *ep, Isoio *iso)
 	Sitd *td, *ltd;
 	int i;
 	ulong frno;
+	Ctlr *ctlr;
 
+	ctlr = ep->hp->aux;
 	left = 0;
 	ltd = nil;
 	frno = iso->td0frno;
 	for(i = 0; i < iso->nframes; i++){
-		td = sitdalloc();
+		td = sitdalloc(ctlr);
 		td->data = iso->data + i * ep->maxpkt;
 		td->epc = ep->dev->port << Stdportshift;
 		td->epc |= ep->dev->hub << Stdhubshift;
@@ -2710,7 +2723,7 @@ isofsinit(Ep *ep, Isoio *iso)
 
 		iso->sitdps[frno] = td;
 		coherence();
-		sitdinit(iso, td);
+		sitdinit(ctlr, iso, td);
 		if(ltd != nil)
 			ltd->next = td;
 		ltd = td;
@@ -2727,16 +2740,18 @@ isohsinit(Ep *ep, Isoio *iso)
 	long left;
 	ulong frno, i, pa;
 	Itd *ltd, *td;
+	Ctlr *ctlr;
 
 	iso->hs = 1;
 	ival = 1;
 	if(ep->pollival > 8)
 		ival = ep->pollival/8;
+	ctlr = ep->hp->aux;
 	left = 0;
 	ltd = nil;
 	frno = iso->td0frno;
 	for(i = 0; i < iso->nframes; i++){
-		td = itdalloc();
+		td = itdalloc(ctlr);
 		td->data = iso->data + i * 8 * iso->maxsize;
 		pa = PADDR(td->data) & ~0xFFF;
 		for(p = 0; p < nelem(td->buffer); p++){
@@ -2761,7 +2776,7 @@ isohsinit(Ep *ep, Isoio *iso)
 		coherence();
 		iso->itdps[frno] = td;
 		coherence();
-		itdinit(iso, td);
+		itdinit(ctlr, iso, td);
 		if(ltd != nil)
 			ltd->next = td;
 		ltd = td;
@@ -2826,8 +2841,8 @@ isoopen(Ctlr *ctlr, Ep *ep)
 	 */
 	assert(ep->maxpkt > 0 && ep->ntds > 0 && ep->ntds < 4);
 	assert(ep->maxpkt <= 1024);
-	iso->tdps = smalloc(sizeof(uintptr) * Nisoframes);
-	iso->data = smalloc(iso->nframes * tpf * ep->ntds * ep->maxpkt);
+	iso->tdps = smalloc(sizeof(void*) * Nisoframes);
+	iso->data = (*ctlr->dmaalloc)(iso->nframes * tpf * ep->ntds * ep->maxpkt);
 	iso->td0frno = TRUNC(ctlr->opio->frno + 10, Nisoframes);
 	/* read: now; write: 1s ahead */
 
@@ -3067,9 +3082,9 @@ cancelisoio(Ctlr *ctlr, Isoio *iso, int pollival, ulong load)
 	frno = iso->td0frno;
 	for(i = 0; i < iso->nframes; i++){
 		if(iso->hs != 0)
-			itdfree(iso->itdps[frno]);
+			itdfree(ctlr, iso->itdps[frno]);
 		else
-			sitdfree(iso->sitdps[frno]);
+			sitdfree(ctlr, iso->sitdps[frno]);
 		iso->tdps[frno] = nil;
 		frno = TRUNC(frno+pollival, Nisoframes);
 	}
@@ -3165,7 +3180,7 @@ mkqhtree(Ctlr *ctlr)
 	if(qt->bw == nil || tree == nil)
 		panic("ehci: mkqhtree: no memory");
 	for(i = 0; i < n; i++){
-		tree[i] = qh = edalloc();
+		tree[i] = qh = edalloc(ctlr);
 		if(qh == nil)
 			panic("ehci: mkqhtree: no memory");
 		qh->nlink = qh->alink = qh->link = Lterm;
@@ -3208,10 +3223,17 @@ ehcimeminit(Ctlr *ctlr)
 	int i, frsize;
 	Eopio *opio;
 
+	if(ctlr->tdalloc == nil)
+		ctlr->tdalloc = xspanalloc;
+	if(ctlr->dmaalloc == nil)
+		ctlr->dmaalloc = smalloc;
+	if(ctlr->dmafree == nil)
+		ctlr->dmafree = free;
+
 	opio = ctlr->opio;
 	frsize = ctlr->nframes * sizeof(ulong);
 	assert((frsize & 0xFFF) == 0);		/* must be 4k aligned */
-	ctlr->frames = xspanalloc(frsize, frsize, 0);
+	ctlr->frames = (*ctlr->tdalloc)(frsize, frsize, 0);
 	if(ctlr->frames == nil)
 		panic("ehci reset: no memory");
 
@@ -3223,7 +3245,7 @@ ehcimeminit(Ctlr *ctlr)
 
 	qhalloc(ctlr, nil, nil, nil);	/* init async list */
 	mkqhtree(ctlr);			/* init sync list */
-	edfree(edalloc());		/* try to get some ones pre-allocated */
+	edfree(edalloc(ctlr));		/* try to get some ones pre-allocated */
 
 	dprint("ehci %#p flb %#lux frno %#lux\n",
 		ctlr->capio, opio->frbase, opio->frno);
diff --git a/sys/src/9/zynq/mkfile b/sys/src/9/zynq/mkfile
index 079207241..5ba3628e9 100644
--- a/sys/src/9/zynq/mkfile
+++ b/sys/src/9/zynq/mkfile
@@ -86,5 +86,5 @@ install:V:	$p$CONF
 	for(i in $EXTRACOPIES)
 		import $i / /n/$i && cp $p$CONF $p$CONF.gz /n/$i/$objtype/
 
-devusb.$O usbehci.$O usbehcizynq.$O: ../port/usb.h uncached.h
+devusb.$O usbehci.$O usbehcizynq.$O: ../port/usb.h
 usbehci.$O usbehcizynq.$O: usbehci.h
diff --git a/sys/src/9/zynq/uncached.h b/sys/src/9/zynq/uncached.h
deleted file mode 100644
index 0a7b50883..000000000
--- a/sys/src/9/zynq/uncached.h
+++ /dev/null
@@ -1,27 +0,0 @@
-#define free			ucfree
-#define malloc			myucalloc
-#define mallocz			ucallocz
-#define smalloc			myucalloc
-#define xspanalloc		ucallocalign
-
-#define allocb			ucallocb
-#define iallocb			uciallocb
-#define freeb			ucfreeb
-
-static void *
-ucallocz(uint n, int)
-{
-	char *p = ucalloc(n);
-
-	if (p)
-		memset(p, 0, n);
-	else
-		panic("ucalloc: out of memory");
-	return p;
-}
-
-static void *
-myucalloc(uint n)
-{
-	return ucallocz(n, 1);
-}
diff --git a/sys/src/9/zynq/usbehci.h b/sys/src/9/zynq/usbehci.h
index 2c71d12bc..6b8b32fd0 100644
--- a/sys/src/9/zynq/usbehci.h
+++ b/sys/src/9/zynq/usbehci.h
@@ -84,16 +84,30 @@ typedef struct Qtree Qtree;
 
 struct Eopio
 {
-	ulong cmd;
-	ulong sts;
-	ulong intr;
-	ulong frno;
-	ulong dummy1[2];
-	ulong frbase;
-	ulong link;
-	ulong dummy2[9];
-	ulong config;
-	ulong portsc[1];
+/*140*/	ulong cmd;
+/*144*/	ulong sts;
+/*148*/	ulong intr;
+/*14c*/	ulong frno;
+
+/*150*/	ulong reserved1;
+
+/*154*/	ulong frbase;
+/*158*/	ulong link;
+
+/*15c*/	ulong _ttctrl;
+/*160*/	ulong _burstsize;
+/*164*/	ulong _txfilltuning;
+/*168*/ ulong _txttfilltuning;
+/*16c*/	ulong _ic_usb;
+/*170*/	ulong _ulpi_viewport;
+
+/*174*/	ulong reserved2;
+
+/*178*/	ulong _endptnak;
+/*17c*/	ulong _endptnaken;
+
+/*180*/	ulong config;
+/*184*/	ulong portsc[1];
 };
 
 struct Poll
@@ -113,6 +127,11 @@ struct Ctlr
 	void*	capio;		/* base address for debug info */
 	Eopio*	opio;		/* Operational i/o regs */
 
+	void*	(*tdalloc)(ulong,int,ulong);
+	void*	(*dmaalloc)(ulong);
+	void	(*dmafree)(void*);
+	void	(*dmaflush)(int,void*,ulong len);
+
 	int	nframes;	/* 1024, 512, or 256 frames in the list */
 	ulong*	frames;		/* periodic frame list (hw) */
 	Qh*	qhs;		/* async Qh circular list for bulk/ctl */
diff --git a/sys/src/9/zynq/usbehcizynq.c b/sys/src/9/zynq/usbehcizynq.c
index e233fb429..d5ec5b005 100644
--- a/sys/src/9/zynq/usbehcizynq.c
+++ b/sys/src/9/zynq/usbehcizynq.c
@@ -61,6 +61,47 @@ ehcireset(Ctlr *ctlr)
 	iunlock(ctlr);
 }
 
+enum {
+	Cls = 64,
+};
+
+/* descriptors need to be allocated in uncached memory */
+static void*
+tdalloc(ulong size, int, ulong)
+{
+	return ucalloc(size);
+}
+
+static void*
+dmaalloc(ulong len)
+{
+	return mallocalign(ROUND(len, Cls), Cls, 0, 0);
+}
+static void
+dmafree(void *data)
+{
+	free(data);
+}
+
+static void
+dmaflush(int clean, void *data, ulong len)
+{
+	uintptr va, pa;
+
+	va = (uintptr)data & ~(Cls-1);
+	pa = PADDR(va);
+	len = ROUND(len, Cls);
+	if(clean){
+		/* flush cache before write */
+		cleandse((uchar*)va, (uchar*)va+len);
+		clean2pa(pa, pa+len);
+	} else {
+		/* invalidate cache before read */
+		invaldse((uchar*)va, (uchar*)va+len);
+		inval2pa(pa, pa+len);
+	}
+}
+
 static int
 reset(Hci *hp)
 {
@@ -84,12 +125,17 @@ reset(Hci *hp)
 	ctlr->opio = (Eopio *) ((uchar *) ctlr->r + 0x140);
 	ctlr->capio = (void *) ctlr->base;
 	hp->nports = 1;
-	ctlr->r[USBMODE] |= USBHOST;
-	
+
+	ctlr->tdalloc = tdalloc;
+	ctlr->dmaalloc = dmaalloc;
+	ctlr->dmafree = dmafree;
+	ctlr->dmaflush = dmaflush;
+
 	ehcireset(ctlr);
+	ctlr->r[USBMODE] |= USBHOST;
+	ctlr->r[ULPI] = 1<<30 | 1<<29 | 0x0B << 16 | 3<<5;
 	ehcimeminit(ctlr);
 	ehcilinkage(hp);
-	ctlr->r[ULPI] = 1<<30 | 1<<29 | 0x0B << 16 | 3<<5;
 	if(hp->interrupt != nil)
 		intrenable(hp->irq, hp->interrupt, hp, LEVEL, hp->type);
 	return 0;
@@ -98,6 +144,6 @@ reset(Hci *hp)
 void
 usbehcilink(void)
 {
-	ehcidebug = 2;
+//	ehcidebug = 2;
 	addhcitype("ehci", reset);
 }
diff --git a/sys/src/9/zynq/zynq b/sys/src/9/zynq/zynq
index 75d244ab7..7b67c902b 100644
--- a/sys/src/9/zynq/zynq
+++ b/sys/src/9/zynq/zynq
@@ -26,7 +26,7 @@ link
 	etherzynq
 	ethermedium
 	loopbackmedium
-#	usbehci usbehcizynq
+	usbehci usbehcizynq
 
 misc
 	uartzynq