IB/ipath: Driver support for userspace sharing of HW contexts

This allows multiple userspace processes to share a single hardware
context in a master/slave arrangement.  It is backwards binary compatible
with existing userspace.

Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

5 files changed, 754 insertions, 241 deletions

diff --git a/drivers/infiniband/hw/ipath/ipath_common.h b/drivers/infiniband/hw/ipath/ipath_common.h
index f577905e3ac..46b7b20d8a9 100644
--- a/drivers/infiniband/hw/ipath/ipath_common.h
+++ b/drivers/infiniband/hw/ipath/ipath_common.h
@@ -185,6 +185,7 @@ typedef enum _ipath_ureg {
 #define IPATH_RUNTIME_PCIE	0x2
 #define IPATH_RUNTIME_FORCE_WC_ORDER	0x4
 #define IPATH_RUNTIME_RCVHDR_COPY	0x8
+#define IPATH_RUNTIME_MASTER	0x10
 
 /*
  * This structure is returned by ipath_userinit() immediately after
@@ -202,7 +203,8 @@ struct ipath_base_info {
 	/* version of software, for feature checking. */
 	__u32 spi_sw_version;
 	/* InfiniPath port assigned, goes into sent packets */
-	__u32 spi_port;
+	__u16 spi_port;
+	__u16 spi_subport;
 	/*
 	 * IB MTU, packets IB data must be less than this.
 	 * The MTU is in bytes, and will be a multiple of 4 bytes.
@@ -218,7 +220,7 @@ struct ipath_base_info {
 	__u32 spi_tidcnt;
 	/* size of the TID Eager list in infinipath, in entries */
 	__u32 spi_tidegrcnt;
-	/* size of a single receive header queue entry. */
+	/* size of a single receive header queue entry in words. */
 	__u32 spi_rcvhdrent_size;
 	/*
 	 * Count of receive header queue entries allocated.
@@ -310,6 +312,12 @@ struct ipath_base_info {
 	__u32 spi_filler_for_align;
 	/* address of readonly memory copy of the rcvhdrq tail register. */
 	__u64 spi_rcvhdr_tailaddr;
+
+	/* shared memory pages for subports if IPATH_RUNTIME_MASTER is set */
+	__u64 spi_subport_uregbase;
+	__u64 spi_subport_rcvegrbuf;
+	__u64 spi_subport_rcvhdr_base;
+
 } __attribute__ ((aligned(8)));
 
 
@@ -328,12 +336,12 @@ struct ipath_base_info {
 
 /*
  * Minor version differences are always compatible
- * a within a major version, however if if user software is larger
+ * a within a major version, however if user software is larger
  * than driver software, some new features and/or structure fields
  * may not be implemented; the user code must deal with this if it
- * cares, or it must abort after initialization reports the difference
+ * cares, or it must abort after initialization reports the difference.
  */
-#define IPATH_USER_SWMINOR 2
+#define IPATH_USER_SWMINOR 3
 
 #define IPATH_USER_SWVERSION ((IPATH_USER_SWMAJOR<<16) | IPATH_USER_SWMINOR)
 
@@ -379,7 +387,16 @@ struct ipath_user_info {
 	 */
 	__u32 spu_rcvhdrsize;
 
-	__u64 spu_unused; /* kept for compatible layout */
+	/*
+	 * If two or more processes wish to share a port, each process
+	 * must set the spu_subport_cnt and spu_subport_id to the same
+	 * values.  The only restriction on the spu_subport_id is that
+	 * it be unique for a given node.
+	 */
+	__u16 spu_subport_cnt;
+	__u16 spu_subport_id;
+
+	__u32 spu_unused; /* kept for compatible layout */
 
 	/*
 	 * address of struct base_info to write to
@@ -398,13 +415,17 @@ struct ipath_user_info {
 #define IPATH_CMD_TID_UPDATE	19	/* update expected TID entries */
 #define IPATH_CMD_TID_FREE	20	/* free expected TID entries */
 #define IPATH_CMD_SET_PART_KEY	21	/* add partition key */
+#define IPATH_CMD_SLAVE_INFO	22	/* return info on slave processes */
 
-#define IPATH_CMD_MAX		21
+#define IPATH_CMD_MAX		22
 
 struct ipath_port_info {
 	__u32 num_active;	/* number of active units */
 	__u32 unit;		/* unit (chip) assigned to caller */
-	__u32 port;		/* port on unit assigned to caller */
+	__u16 port;		/* port on unit assigned to caller */
+	__u16 subport;		/* subport on unit assigned to caller */
+	__u16 num_ports;	/* number of ports available on unit */
+	__u16 num_subports;	/* number of subport slaves opened on port */
 };
 
 struct ipath_tid_info {
@@ -435,6 +456,8 @@ struct ipath_cmd {
 		__u32 recv_ctrl;
 		/* partition key to set */
 		__u16 part_key;
+		/* user address of __u32 bitmask of active slaves */
+		__u64 slave_mask_addr;
 	} cmd;
 };
 
@@ -596,6 +619,10 @@ struct infinipath_counters {
 
 /* K_PktFlags bits */
 #define INFINIPATH_KPF_INTR 0x1
+#define INFINIPATH_KPF_SUBPORT_MASK 0x3
+#define INFINIPATH_KPF_SUBPORT_SHIFT 1
+
+#define INFINIPATH_MAX_SUBPORT	4
 
 /* SendPIO per-buffer control */
 #define INFINIPATH_SP_TEST    0x40
@@ -610,7 +637,7 @@ struct ipath_header {
 	/*
 	 * Version - 4 bits, Port - 4 bits, TID - 10 bits and Offset -
 	 * 14 bits before ECO change ~28 Dec 03.  After that, Vers 4,
-	 * Port 3, TID 11, offset 14.
+	 * Port 4, TID 11, offset 13.
 	 */
 	__le32 ver_port_tid_offset;
 	__le16 chksum;
diff --git a/drivers/infiniband/hw/ipath/ipath_driver.c b/drivers/infiniband/hw/ipath/ipath_driver.c
index a01301d0753..0fe37c5467a 100644
--- a/drivers/infiniband/hw/ipath/ipath_driver.c
+++ b/drivers/infiniband/hw/ipath/ipath_driver.c
@@ -1827,9 +1827,9 @@ void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
 			dma_free_coherent(&dd->pcidev->dev, size,
 				base, pd->port_rcvegrbuf_phys[e]);
 		}
-		vfree(pd->port_rcvegrbuf);
+		kfree(pd->port_rcvegrbuf);
 		pd->port_rcvegrbuf = NULL;
-		vfree(pd->port_rcvegrbuf_phys);
+		kfree(pd->port_rcvegrbuf_phys);
 		pd->port_rcvegrbuf_phys = NULL;
 		pd->port_rcvegrbuf_chunks = 0;
 	} else if (pd->port_port == 0 && dd->ipath_port0_skbs) {
@@ -1845,6 +1845,9 @@ void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
 		vfree(skbs);
 	}
 	kfree(pd->port_tid_pg_list);
+	vfree(pd->subport_uregbase);
+	vfree(pd->subport_rcvegrbuf);
+	vfree(pd->subport_rcvhdr_base);
 	kfree(pd);
 }
 
diff --git a/drivers/infiniband/hw/ipath/ipath_file_ops.c b/drivers/infiniband/hw/ipath/ipath_file_ops.c
index 29930e22318..caf8cb891da 100644
--- a/drivers/infiniband/hw/ipath/ipath_file_ops.c
+++ b/drivers/infiniband/hw/ipath/ipath_file_ops.c
@@ -41,6 +41,12 @@
 #include "ipath_kernel.h"
 #include "ipath_common.h"
 
+/*
+ * mmap64 doesn't allow all 64 bits for 32-bit applications
+ * so only use the low 43 bits.
+ */
+#define MMAP64_MASK	0x7FFFFFFFFFFUL
+
 static int ipath_open(struct inode *, struct file *);
 static int ipath_close(struct inode *, struct file *);
 static ssize_t ipath_write(struct file *, const char __user *, size_t,
@@ -57,18 +63,35 @@ static struct file_operations ipath_file_ops = {
 	.mmap = ipath_mmap
 };
 
-static int ipath_get_base_info(struct ipath_portdata *pd,
+static int ipath_get_base_info(struct file *fp,
 			       void __user *ubase, size_t ubase_size)
 {
+	struct ipath_portdata *pd = port_fp(fp);
 	int ret = 0;
 	struct ipath_base_info *kinfo = NULL;
 	struct ipath_devdata *dd = pd->port_dd;
+	unsigned subport_cnt;
+	int shared, master;
+	size_t sz;
+
+	subport_cnt = pd->port_subport_cnt;
+	if (!subport_cnt) {
+		shared = 0;
+		master = 0;
+		subport_cnt = 1;
+	} else {
+		shared = 1;
+		master = !subport_fp(fp);
+	}
 
-	if (ubase_size < sizeof(*kinfo)) {
+	sz = sizeof(*kinfo);
+	/* If port sharing is not requested, allow the old size structure */
+	if (!shared)
+		sz -= 3 * sizeof(u64);
+	if (ubase_size < sz) {
 		ipath_cdbg(PROC,
-			   "Base size %lu, need %lu (version mismatch?)\n",
-			   (unsigned long) ubase_size,
-			   (unsigned long) sizeof(*kinfo));
+			   "Base size %zu, need %zu (version mismatch?)\n",
+			   ubase_size, sz);
 		ret = -EINVAL;
 		goto bail;
 	}
@@ -95,7 +118,9 @@ static int ipath_get_base_info(struct ipath_portdata *pd,
 	kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 	kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 		pd->port_rcvegrbuf_chunks;
-	kinfo->spi_tidcnt = dd->ipath_rcvtidcnt;
+	kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
+	if (master)
+		kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 	/*
 	 * for this use, may be ipath_cfgports summed over all chips that
 	 * are are configured and present
@@ -118,31 +143,76 @@ static int ipath_get_base_info(struct ipath_portdata *pd,
 	 * page_address() macro worked, but in 2.6.11, even that returns the
 	 * full 64 bit address (upper bits all 1's).  So far, using the
 	 * physical addresses (or chip offsets, for chip mapping) works, but
-	 * no doubt some future kernel release will chang that, and we'll be
-	 * on to yet another method of dealing with this
+	 * no doubt some future kernel release will change that, and we'll be
+	 * on to yet another method of dealing with this.
 	 */
 	kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
-	kinfo->spi_rcvhdr_tailaddr = (u64)pd->port_rcvhdrqtailaddr_phys;
+	kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 	kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 	kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 	kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 		(void *) dd->ipath_statusp -
 		(void *) dd->ipath_pioavailregs_dma;
-	kinfo->spi_piobufbase = (u64) pd->port_piobufs;
-	kinfo->__spi_uregbase =
-		dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
+	if (!shared) {
+		kinfo->spi_piocnt = dd->ipath_pbufsport;
+		kinfo->spi_piobufbase = (u64) pd->port_piobufs;
+		kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
+			dd->ipath_palign * pd->port_port;
+	} else if (master) {
+		kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
+				    (dd->ipath_pbufsport % subport_cnt);
+		/* Master's PIO buffers are after all the slave's */
+		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
+			dd->ipath_palign *
+			(dd->ipath_pbufsport - kinfo->spi_piocnt);
+		kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
+			dd->ipath_palign * pd->port_port;
+	} else {
+		unsigned slave = subport_fp(fp) - 1;
 
-	kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1);
-	kinfo->spi_piocnt = dd->ipath_pbufsport;
+		kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
+		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
+			dd->ipath_palign * kinfo->spi_piocnt * slave;
+		kinfo->__spi_uregbase = ((u64) pd->subport_uregbase +
+			PAGE_SIZE * slave) & MMAP64_MASK;
+
+		kinfo->spi_rcvhdr_base = ((u64) pd->subport_rcvhdr_base +
+			pd->port_rcvhdrq_size * slave) & MMAP64_MASK;
+		kinfo->spi_rcvhdr_tailaddr =
+			(u64) pd->port_rcvhdrqtailaddr_phys & MMAP64_MASK;
+		kinfo->spi_rcv_egrbufs = ((u64) pd->subport_rcvegrbuf +
+			dd->ipath_rcvegrcnt * dd->ipath_rcvegrbufsize * slave) &
+			MMAP64_MASK;
+	}
+
+	kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
+		dd->ipath_palign;
 	kinfo->spi_pioalign = dd->ipath_palign;
 
 	kinfo->spi_qpair = IPATH_KD_QP;
 	kinfo->spi_piosize = dd->ipath_ibmaxlen;
 	kinfo->spi_mtu = dd->ipath_ibmaxlen;	/* maxlen, not ibmtu */
 	kinfo->spi_port = pd->port_port;
+	kinfo->spi_subport = subport_fp(fp);
 	kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 	kinfo->spi_hw_version = dd->ipath_revision;
 
+	if (master) {
+		kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
+		kinfo->spi_subport_uregbase =
+			(u64) pd->subport_uregbase & MMAP64_MASK;
+		kinfo->spi_subport_rcvegrbuf =
+			(u64) pd->subport_rcvegrbuf & MMAP64_MASK;
+		kinfo->spi_subport_rcvhdr_base =
+			(u64) pd->subport_rcvhdr_base & MMAP64_MASK;
+		ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
+			kinfo->spi_port,
+			kinfo->spi_runtime_flags,
+			kinfo->spi_subport_uregbase,
+			kinfo->spi_subport_rcvegrbuf,
+			kinfo->spi_subport_rcvhdr_base);
+	}
+
 	if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
 		ret = -EFAULT;
 
@@ -154,6 +224,7 @@ bail:
 /**
  * ipath_tid_update - update a port TID
  * @pd: the port
+ * @fp: the ipath device file
  * @ti: the TID information
  *
  * The new implementation as of Oct 2004 is that the driver assigns
@@ -176,11 +247,11 @@ bail:
  * virtually contiguous pages, that should change to improve
  * performance.
  */
-static int ipath_tid_update(struct ipath_portdata *pd,
+static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 			    const struct ipath_tid_info *ti)
 {
 	int ret = 0, ntids;
-	u32 tid, porttid, cnt, i, tidcnt;
+	u32 tid, porttid, cnt, i, tidcnt, tidoff;
 	u16 *tidlist;
 	struct ipath_devdata *dd = pd->port_dd;
 	u64 physaddr;
@@ -188,6 +259,7 @@ static int ipath_tid_update(struct ipath_portdata *pd,
 	u64 __iomem *tidbase;
 	unsigned long tidmap[8];
 	struct page **pagep = NULL;
+	unsigned subport = subport_fp(fp);
 
 	if (!dd->ipath_pageshadow) {
 		ret = -ENOMEM;
@@ -204,20 +276,34 @@ static int ipath_tid_update(struct ipath_portdata *pd,
 		ret = -EFAULT;
 		goto done;
 	}
-	tidcnt = dd->ipath_rcvtidcnt;
-	if (cnt >= tidcnt) {
+	porttid = pd->port_port * dd->ipath_rcvtidcnt;
+	if (!pd->port_subport_cnt) {
+		tidcnt = dd->ipath_rcvtidcnt;
+		tid = pd->port_tidcursor;
+		tidoff = 0;
+	} else if (!subport) {
+		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
+			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
+		tidoff = dd->ipath_rcvtidcnt - tidcnt;
+		porttid += tidoff;
+		tid = tidcursor_fp(fp);
+	} else {
+		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
+		tidoff = tidcnt * (subport - 1);
+		porttid += tidoff;
+		tid = tidcursor_fp(fp);
+	}
+	if (cnt > tidcnt) {
 		/* make sure it all fits in port_tid_pg_list */
 		dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 			 "TIDs, only trying max (%u)\n", cnt, tidcnt);
 		cnt = tidcnt;
 	}
-	pagep = (struct page **)pd->port_tid_pg_list;
-	tidlist = (u16 *) (&pagep[cnt]);
+	pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
+	tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 
 	memset(tidmap, 0, sizeof(tidmap));
-	tid = pd->port_tidcursor;
 	/* before decrement; chip actual # */
-	porttid = pd->port_port * tidcnt;
 	ntids = tidcnt;
 	tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 				   dd->ipath_rcvtidbase +
@@ -274,9 +360,9 @@ static int ipath_tid_update(struct ipath_portdata *pd,
 			ret = -ENOMEM;
 			break;
 		}
-		tidlist[i] = tid;
+		tidlist[i] = tid + tidoff;
 		ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
-			   "vaddr %lx\n", i, tid, vaddr);
+			   "vaddr %lx\n", i, tid + tidoff, vaddr);
 		/* we "know" system pages and TID pages are same size */
 		dd->ipath_pageshadow[porttid + tid] = pagep[i];
 		/*
@@ -341,7 +427,10 @@ static int ipath_tid_update(struct ipath_portdata *pd,
 		}
 		if (tid == tidcnt)
 			tid = 0;
-		pd->port_tidcursor = tid;
+		if (!pd->port_subport_cnt)
+			pd->port_tidcursor = tid;
+		else
+			tidcursor_fp(fp) = tid;
 	}
 
 done:
@@ -354,6 +443,7 @@ done:
 /**
  * ipath_tid_free - free a port TID
  * @pd: the port
+ * @subport: the subport
  * @ti: the TID info
  *
  * right now we are unlocking one page at a time, but since
@@ -367,7 +457,7 @@ done:
  * they pass in to us.
  */
 
-static int ipath_tid_free(struct ipath_portdata *pd,
+static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 			  const struct ipath_tid_info *ti)
 {
 	int ret = 0;
@@ -388,11 +478,20 @@ static int ipath_tid_free(struct ipath_portdata *pd,
 	}
 
 	porttid = pd->port_port * dd->ipath_rcvtidcnt;
+	if (!pd->port_subport_cnt)
+		tidcnt = dd->ipath_rcvtidcnt;
+	else if (!subport) {
+		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
+			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
+		porttid += dd->ipath_rcvtidcnt - tidcnt;
+	} else {
+		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
+		porttid += tidcnt * (subport - 1);
+	}
 	tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 				   dd->ipath_rcvtidbase +
 				   porttid * sizeof(*tidbase));
 
-	tidcnt = dd->ipath_rcvtidcnt;
 	limit = sizeof(tidmap) * BITS_PER_BYTE;
 	if (limit > tidcnt)
 		/* just in case size changes in future */
@@ -581,20 +680,24 @@ bail:
 /**
  * ipath_manage_rcvq - manage a port's receive queue
  * @pd: the port
+ * @subport: the subport
  * @start_stop: action to carry out
  *
  * start_stop == 0 disables receive on the port, for use in queue
  * overflow conditions.  start_stop==1 re-enables, to be used to
  * re-init the software copy of the head register
  */
-static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop)
+static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
+			     int start_stop)
 {
 	struct ipath_devdata *dd = pd->port_dd;
 	u64 tval;
 
-	ipath_cdbg(PROC, "%sabling rcv for unit %u port %u\n",
+	ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 		   start_stop ? "en" : "dis", dd->ipath_unit,
-		   pd->port_port);
+		   pd->port_port, subport);
+	if (subport)
+		goto bail;
 	/* atomically clear receive enable port. */
 	if (start_stop) {
 		/*
@@ -630,6 +733,7 @@ static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop)
 		tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 	}
 	/* always; new head should be equal to new tail; see above */
+bail:
 	return 0;
 }
 
@@ -687,6 +791,36 @@ static void ipath_clean_part_key(struct ipath_portdata *pd,
 	}
 }
 
+/*
+ * Initialize the port data with the receive buffer sizes
+ * so this can be done while the master port is locked.
+ * Otherwise, there is a race with a slave opening the port
+ * and seeing these fields uninitialized.
+ */
+static void init_user_egr_sizes(struct ipath_portdata *pd)
+{
+	struct ipath_devdata *dd = pd->port_dd;
+	unsigned egrperchunk, egrcnt, size;
+
+	/*
+	 * to avoid wasting a lot of memory, we allocate 32KB chunks of
+	 * physically contiguous memory, advance through it until used up
+	 * and then allocate more.  Of course, we need memory to store those
+	 * extra pointers, now.  Started out with 256KB, but under heavy
+	 * memory pressure (creating large files and then copying them over
+	 * NFS while doing lots of MPI jobs), we hit some allocation
+	 * failures, even though we can sleep...  (2.6.10) Still get
+	 * failures at 64K.  32K is the lowest we can go without wasting
+	 * additional memory.
+	 */
+	size = 0x8000;
+	egrperchunk = size / dd->ipath_rcvegrbufsize;
+	egrcnt = dd->ipath_rcvegrcnt;
+	pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
+	pd->port_rcvegrbufs_perchunk = egrperchunk;
+	pd->port_rcvegrbuf_size = size;
+}
+
 /**
  * ipath_create_user_egr - allocate eager TID buffers
  * @pd: the port to allocate TID buffers for
@@ -702,7 +836,7 @@ static void ipath_clean_part_key(struct ipath_portdata *pd,
 static int ipath_create_user_egr(struct ipath_portdata *pd)
 {
 	struct ipath_devdata *dd = pd->port_dd;
-	unsigned e, egrcnt, alloced, egrperchunk, chunk, egrsize, egroff;
+	unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 	size_t size;
 	int ret;
 	gfp_t gfp_flags;
@@ -722,31 +856,18 @@ static int ipath_create_user_egr(struct ipath_portdata *pd)
 	ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 		   "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 
-	/*
-	 * to avoid wasting a lot of memory, we allocate 32KB chunks of
-	 * physically contiguous memory, advance through it until used up
-	 * and then allocate more.  Of course, we need memory to store those
-	 * extra pointers, now.  Started out with 256KB, but under heavy
-	 * memory pressure (creating large files and then copying them over
-	 * NFS while doing lots of MPI jobs), we hit some allocation
-	 * failures, even though we can sleep...  (2.6.10) Still get
-	 * failures at 64K.  32K is the lowest we can go without wasting
-	 * additional memory.
-	 */
-	size = 0x8000;
-	alloced = ALIGN(egrsize * egrcnt, size);
-	egrperchunk = size / egrsize;
-	chunk = (egrcnt + egrperchunk - 1) / egrperchunk;
-	pd->port_rcvegrbuf_chunks = chunk;
-	pd->port_rcvegrbufs_perchunk = egrperchunk;
-	pd->port_rcvegrbuf_size = size;
-	pd->port_rcvegrbuf = vmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]));
+	chunk = pd->port_rcvegrbuf_chunks;
+	egrperchunk = pd->port_rcvegrbufs_perchunk;
+	size = pd->port_rcvegrbuf_size;
+	pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
+				     GFP_KERNEL);
 	if (!pd->port_rcvegrbuf) {
 		ret = -ENOMEM;
 		goto bail;
 	}
 	pd->port_rcvegrbuf_phys =
-		vmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]));
+		kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
+			GFP_KERNEL);
 	if (!pd->port_rcvegrbuf_phys) {
 		ret = -ENOMEM;
 		goto bail_rcvegrbuf;
@@ -791,97 +912,15 @@ bail_rcvegrbuf_phys:
 				  pd->port_rcvegrbuf_phys[e]);
 
 	}
-	vfree(pd->port_rcvegrbuf_phys);
+	kfree(pd->port_rcvegrbuf_phys);
 	pd->port_rcvegrbuf_phys = NULL;
 bail_rcvegrbuf:
-	vfree(pd->port_rcvegrbuf);
+	kfree(pd->port_rcvegrbuf);
 	pd->port_rcvegrbuf = NULL;
 bail:
 	return ret;
 }
 
-static int ipath_do_user_init(struct ipath_portdata *pd,
-			      const struct ipath_user_info *uinfo)
-{
-	int ret = 0;
-	struct ipath_devdata *dd = pd->port_dd;
-	u32 head32;
-
-	/* for now, if major version is different, bail */
-	if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
-		dev_info(&dd->pcidev->dev,
-			 "User major version %d not same as driver "
-			 "major %d\n", uinfo->spu_userversion >> 16,
-			 IPATH_USER_SWMAJOR);
-		ret = -ENODEV;
-		goto done;
-	}
-
-	if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
-		ipath_dbg("User minor version %d not same as driver "
-			  "minor %d\n", uinfo->spu_userversion & 0xffff,
-			  IPATH_USER_SWMINOR);
-
-	if (uinfo->spu_rcvhdrsize) {
-		ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
-		if (ret)
-			goto done;
-	}
-
-	/* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
-
-	/* for right now, kernel piobufs are at end, so port 1 is at 0 */
-	pd->port_piobufs = dd->ipath_piobufbase +
-		dd->ipath_pbufsport * (pd->port_port -
-				       1) * dd->ipath_palign;
-	ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
-		   pd->port_port, pd->port_piobufs);
-
-	/*
-	 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
-	 * array for time being.  If pd->port_port > chip-supported,
-	 * we need to do extra stuff here to handle by handling overflow
-	 * through port 0, someday
-	 */
-	ret = ipath_create_rcvhdrq(dd, pd);
-	if (!ret)
-		ret = ipath_create_user_egr(pd);
-	if (ret)
-		goto done;
-
-	/*
-	 * set the eager head register for this port to the current values
-	 * of the tail pointers, since we don't know if they were
-	 * updated on last use of the port.
-	 */
-	head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
-	ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
-	dd->ipath_lastegrheads[pd->port_port] = -1;
-	dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
-	ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
-		pd->port_port, head32);
-	pd->port_tidcursor = 0;	/* start at beginning after open */
-	/*
-	 * now enable the port; the tail registers will be written to memory
-	 * by the chip as soon as it sees the write to
-	 * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
-	 * transition from 0 to 1, so clear it first, then set it as part of
-	 * enabling the port.  This will (very briefly) affect any other
-	 * open ports, but it shouldn't be long enough to be an issue.
-	 * We explictly set the in-memory copy to 0 beforehand, so we don't
-	 * have to wait to be sure the DMA update has happened.
-	 */
-	*pd->port_rcvhdrtail_kvaddr = 0ULL;
-	set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
-		&dd->ipath_rcvctrl);
-	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
-			 dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
-	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
-			 dd->ipath_rcvctrl);
-done:
-	return ret;
-}
-
 
 /* common code for the mappings on dma_alloc_coherent mem */
 static int ipath_mmap_mem(struct vm_area_struct *vma,
@@ -957,7 +996,8 @@ static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
 
 static int mmap_piobufs(struct vm_area_struct *vma,
 			struct ipath_devdata *dd,
-			struct ipath_portdata *pd)
+			struct ipath_portdata *pd,
+			unsigned piobufs, unsigned piocnt)
 {
 	unsigned long phys;
 	int ret;
@@ -968,16 +1008,15 @@ static int mmap_piobufs(struct vm_area_struct *vma,
 	 * process data, and catches users who might try to read the i/o
 	 * space due to a bug.
 	 */
-	if ((vma->vm_end - vma->vm_start) >
-	    (dd->ipath_pbufsport * dd->ipath_palign)) {
+	if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
 		dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
 			 "reqlen %lx > PAGE\n",
 			 vma->vm_end - vma->vm_start);
-		ret = -EFAULT;
+		ret = -EINVAL;
 		goto bail;
 	}
 
-	phys = dd->ipath_physaddr + pd->port_piobufs;
+	phys = dd->ipath_physaddr + piobufs;
 
 	/*
 	 * Don't mark this as non-cached, or we don't get the
@@ -1021,7 +1060,7 @@ static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 			 "reqlen %lx > actual %lx\n",
 			 vma->vm_end - vma->vm_start,
 			 (unsigned long) total_size);
-		ret = -EFAULT;
+		ret = -EINVAL;
 		goto bail;
 	}
 
@@ -1049,6 +1088,122 @@ bail:
 	return ret;
 }
 
+/*
+ * ipath_file_vma_nopage - handle a VMA page fault.
+ */
+static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
+					  unsigned long address, int *type)
+{
+	unsigned long offset = address - vma->vm_start;
+	struct page *page = NOPAGE_SIGBUS;
+	void *pageptr;
+
+	/*
+	 * Convert the vmalloc address into a struct page.
+	 */
+	pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
+	page = vmalloc_to_page(pageptr);
+	if (!page)
+		goto out;
+
+	/* Increment the reference count. */
+	get_page(page);
+	if (type)
+		*type = VM_FAULT_MINOR;
+out:
+	return page;
+}
+
+static struct vm_operations_struct ipath_file_vm_ops = {
+	.nopage = ipath_file_vma_nopage,
+};
+
+static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
+		       struct ipath_portdata *pd, unsigned subport)
+{
+	unsigned long len;
+	struct ipath_devdata *dd;
+	void *addr;
+	size_t size;
+	int ret;
+
+	/* If the port is not shared, all addresses should be physical */
+	if (!pd->port_subport_cnt) {
+		ret = -EINVAL;
+		goto bail;
+	}
+
+	dd = pd->port_dd;
+	size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
+
+	/*
+	 * Master has all the slave uregbase, rcvhdrq, and
+	 * rcvegrbufs mmapped.
+	 */
+	if (subport == 0) {
+		unsigned num_slaves = pd->port_subport_cnt - 1;
+
+		if (pgaddr == ((u64) pd->subport_uregbase & MMAP64_MASK)) {
+			addr = pd->subport_uregbase;
+			size = PAGE_SIZE * num_slaves;
+		} else if (pgaddr == ((u64) pd->subport_rcvhdr_base &
+				      MMAP64_MASK)) {
+			addr = pd->subport_rcvhdr_base;
+			size = pd->port_rcvhdrq_size * num_slaves;
+		} else if (pgaddr == ((u64) pd->subport_rcvegrbuf &
+				      MMAP64_MASK)) {
+			addr = pd->subport_rcvegrbuf;
+			size *= num_slaves;
+		} else {
+			ret = -EINVAL;
+			goto bail;
+		}
+	} else if (pgaddr == (((u64) pd->subport_uregbase +
+			       PAGE_SIZE * (subport - 1)) & MMAP64_MASK)) {
+		addr = pd->subport_uregbase + PAGE_SIZE * (subport - 1);
+		size = PAGE_SIZE;
+	} else if (pgaddr == (((u64) pd->subport_rcvhdr_base +
+			       pd->port_rcvhdrq_size * (subport - 1)) &
+			      MMAP64_MASK)) {
+		addr = pd->subport_rcvhdr_base +
+			pd->port_rcvhdrq_size * (subport - 1);
+		size = pd->port_rcvhdrq_size;
+	} else if (pgaddr == (((u64) pd->subport_rcvegrbuf +
+			       size * (subport - 1)) & MMAP64_MASK)) {
+		addr = pd->subport_rcvegrbuf + size * (subport - 1);
+		/* rcvegrbufs are read-only on the slave */
+		if (vma->vm_flags & VM_WRITE) {
+			dev_info(&dd->pcidev->dev,
+				 "Can't map eager buffers as "
+				 "writable (flags=%lx)\n", vma->vm_flags);
+			ret = -EPERM;
+			goto bail;
+		}
+		/*
+		 * Don't allow permission to later change to writeable
+		 * with mprotect.
+		 */
+		vma->vm_flags &= ~VM_MAYWRITE;
+	} else {
+		ret = -EINVAL;
+		goto bail;
+	}
+	len = vma->vm_end - vma->vm_start;
+	if (len > size) {
+		ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
+		ret = -EINVAL;
+		goto bail;
+	}
+
+	vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
+	vma->vm_ops = &ipath_file_vm_ops;
+	vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
+	ret = 0;
+
+bail:
+	return ret;
+}
+
 /**
  * ipath_mmap - mmap various structures into user space
  * @fp: the file pointer
@@ -1064,73 +1219,99 @@ static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
 	struct ipath_portdata *pd;
 	struct ipath_devdata *dd;
 	u64 pgaddr, ureg;
+	unsigned piobufs, piocnt;
 	int ret;
 
 	pd = port_fp(fp);
+	if (!pd) {
+		ret = -EINVAL;
+		goto bail;
+	}
 	dd = pd->port_dd;
 
 	/*
 	 * This is the ipath_do_user_init() code, mapping the shared buffers
 	 * into the user process. The address referred to by vm_pgoff is the
-	 * virtual, not physical, address; we only do one mmap for each
-	 * space mapped.
+	 * file offset passed via mmap().  For shared ports, this is the
+	 * kernel vmalloc() address of the pages to share with the master.
+	 * For non-shared or master ports, this is a physical address.
+	 * We only do one mmap for each space mapped.
 	 */
 	pgaddr = vma->vm_pgoff << PAGE_SHIFT;
 
 	/*
-	 * Must fit in 40 bits for our hardware; some checked elsewhere,
-	 * but we'll be paranoid.  Check for 0 is mostly in case one of the
-	 * allocations failed, but user called mmap anyway.   We want to catch
-	 * that before it can match.
+	 * Check for 0 in case one of the allocations failed, but user
+	 * called mmap anyway.
 	 */
-	if (!pgaddr || pgaddr >= (1ULL<<40))  {
-		ipath_dev_err(dd, "Bad phys addr %llx, start %lx, end %lx\n",
-			(unsigned long long)pgaddr, vma->vm_start, vma->vm_end);
-		return -EINVAL;
+	if (!pgaddr)  {
+		ret = -EINVAL;
+		goto bail;
 	}
 
-	/* just the offset of the port user registers, not physical addr */
-	ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
-
-	ipath_cdbg(MM, "ushare: pgaddr %llx vm_start=%lx, vmlen %lx\n",
+	ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
 		   (unsigned long long) pgaddr, vma->vm_start,
-		   vma->vm_end - vma->vm_start);
+		   vma->vm_end - vma->vm_start, dd->ipath_unit,
+		   pd->port_port, subport_fp(fp));
 
-	if (vma->vm_start & (PAGE_SIZE-1)) {
-		ipath_dev_err(dd,
-			"vm_start not aligned: %lx, end=%lx phys %lx\n",
-			vma->vm_start, vma->vm_end, (unsigned long)pgaddr);
-		ret = -EINVAL;
+	/*
+	 * Physical addresses must fit in 40 bits for our hardware.
+	 * Check for kernel virtual addresses first, anything else must
+	 * match a HW or memory address.
+	 */
+	if (pgaddr >= (1ULL<<40)) {
+		ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
+		goto bail;
+	}
+
+	if (!pd->port_subport_cnt) {
+		/* port is not shared */
+		ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
+		piocnt = dd->ipath_pbufsport;
+		piobufs = pd->port_piobufs;
+	} else if (!subport_fp(fp)) {
+		/* caller is the master */
+		ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
+		piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
+			 (dd->ipath_pbufsport % pd->port_subport_cnt);
+		piobufs = pd->port_piobufs +
+			dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
+	} else {
+		unsigned slave = subport_fp(fp) - 1;
+
+		/* caller is a slave */
+		ureg = 0;
+		piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
+		piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
 	}
-	else if (pgaddr == ureg)
+
+	if (pgaddr == ureg)
 		ret = mmap_ureg(vma, dd, ureg);
-	else if (pgaddr == pd->port_piobufs)
-		ret = mmap_piobufs(vma, dd, pd);
-	else if (pgaddr == (u64) pd->port_rcvegr_phys)
+	else if (pgaddr == piobufs)
+		ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
+	else if (pgaddr == dd->ipath_pioavailregs_phys)
+		/* in-memory copy of pioavail registers */
+		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
+			      	     dd->ipath_pioavailregs_phys,
+				     "pioavail registers");
+	else if (subport_fp(fp))
+		/* Subports don't mmap the physical receive buffers */
+		ret = -EINVAL;
+	else if (pgaddr == pd->port_rcvegr_phys)
 		ret = mmap_rcvegrbufs(vma, pd);
-	else if (pgaddr == (u64) pd->port_rcvhdrq_phys) {
+	else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
 		/*
 		 * The rcvhdrq itself; readonly except on HT (so have
 		 * to allow writable mapping), multiple pages, contiguous
 		 * from an i/o perspective.
 		 */
-		unsigned total_size =
-			ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize
-			   * sizeof(u32), PAGE_SIZE);
-		ret = ipath_mmap_mem(vma, pd, total_size, 1,
+		ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
 				     pd->port_rcvhdrq_phys,
 				     "rcvhdrq");
-	}
-	else if (pgaddr == (u64)pd->port_rcvhdrqtailaddr_phys)
+	else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
 		/* in-memory copy of rcvhdrq tail register */
 		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
 				     pd->port_rcvhdrqtailaddr_phys,
 				     "rcvhdrq tail");
-	else if (pgaddr == dd->ipath_pioavailregs_phys)
-		/* in-memory copy of pioavail registers */
-		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
-				     dd->ipath_pioavailregs_phys,
-				     "pioavail registers");
 	else
 		ret = -EINVAL;
 
@@ -1138,9 +1319,10 @@ static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
 
 	if (ret < 0)
 		dev_info(&dd->pcidev->dev,
-			 "Failure %d on addr %lx, off %lx\n",
-			 -ret, vma->vm_start, vma->vm_pgoff);
-
+			 "Failure %d on off %llx len %lx\n",
+			 -ret, (unsigned long long)pgaddr,
+			 vma->vm_end - vma->vm_start);
+bail:
 	return ret;
 }
 
@@ -1154,6 +1336,8 @@ static unsigned int ipath_poll(struct file *fp,
 	struct ipath_devdata *dd;
 
 	pd = port_fp(fp);
+	if (!pd)
+		goto bail;
 	dd = pd->port_dd;
 
 	bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
@@ -1176,7 +1360,7 @@ static unsigned int ipath_poll(struct file *fp,
 
 	if (tail == head) {
 		set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
-		if(dd-&