]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
IB/ipath: Driver support for userspace sharing of HW contexts
authorBryan O'Sullivan <bos@pathscale.com>
Thu, 28 Sep 2006 15:59:59 +0000 (08:59 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 28 Sep 2006 18:16:25 +0000 (11:16 -0700)
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>
drivers/infiniband/hw/ipath/ipath_common.h
drivers/infiniband/hw/ipath/ipath_driver.c
drivers/infiniband/hw/ipath/ipath_file_ops.c
drivers/infiniband/hw/ipath/ipath_kernel.h
drivers/infiniband/hw/ipath/ipath_sysfs.c

index f577905e3acaafc3d624c04a00dca92b6f08be35..46b7b20d8a91728270eeea9dbd55e29e354ba53c 100644 (file)
@@ -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;
index a01301d0753cec154dd3e86f350903a455e4733d..0fe37c5467ac0ba4427ef861d0db7b905b29083c 100644 (file)
@@ -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);
 }
 
index 29930e22318e5a09f78349fe4c8d9e052db9aa62..caf8cb891da4116ea062e111966826fd38f81922 100644 (file)
 #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->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
+               if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
                        (void)ipath_write_ureg(dd, ur_rcvhdrhead,
                                               dd->ipath_rhdrhead_intr_off
                                               | head, pd->port_port);
@@ -1200,18 +1384,80 @@ static unsigned int ipath_poll(struct file *fp,
        ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
                         dd->ipath_rcvctrl);
 
+bail:
        return pollflag;
 }
 
+static int init_subports(struct ipath_devdata *dd,
+                        struct ipath_portdata *pd,
+                        const struct ipath_user_info *uinfo)
+{
+       int ret = 0;
+       unsigned num_slaves;
+       size_t size;
+
+       /* Old user binaries don't know about subports */
+       if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
+               goto bail;
+       /*
+        * If the user is requesting zero or one port,
+        * skip the subport allocation.
+        */
+       if (uinfo->spu_subport_cnt <= 1)
+               goto bail;
+       if (uinfo->spu_subport_cnt > 4) {
+               ret = -EINVAL;
+               goto bail;
+       }
+
+       num_slaves = uinfo->spu_subport_cnt - 1;
+       pd->subport_uregbase = vmalloc(PAGE_SIZE * num_slaves);
+       if (!pd->subport_uregbase) {
+               ret = -ENOMEM;
+               goto bail;
+       }
+       /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
+       size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
+                    sizeof(u32), PAGE_SIZE) * num_slaves;
+       pd->subport_rcvhdr_base = vmalloc(size);
+       if (!pd->subport_rcvhdr_base) {
+               ret = -ENOMEM;
+               goto bail_ureg;
+       }
+
+       pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
+                                       pd->port_rcvegrbuf_size *
+                                       num_slaves);
+       if (!pd->subport_rcvegrbuf) {
+               ret = -ENOMEM;
+               goto bail_rhdr;
+       }
+
+       pd->port_subport_cnt = uinfo->spu_subport_cnt;
+       pd->port_subport_id = uinfo->spu_subport_id;
+       pd->active_slaves = 1;
+       goto bail;
+
+bail_rhdr:
+       vfree(pd->subport_rcvhdr_base);
+bail_ureg:
+       vfree(pd->subport_uregbase);
+       pd->subport_uregbase = NULL;
+bail:
+       return ret;
+}
+
 static int try_alloc_port(struct ipath_devdata *dd, int port,
-                         struct file *fp)
+                         struct file *fp,
+                         const struct ipath_user_info *uinfo)
 {
+       struct ipath_portdata *pd;
        int ret;
 
-       if (!dd->ipath_pd[port]) {
-               void *p, *ptmp;
+       if (!(pd = dd->ipath_pd[port])) {
+               void *ptmp;
 
-               p = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
+               pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
 
                /*
                 * Allocate memory for use in ipath_tid_update() just once
@@ -1221,34 +1467,36 @@ static int try_alloc_port(struct ipath_devdata *dd, int port,
                ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
                               dd->ipath_rcvtidcnt * sizeof(struct page **),
                               GFP_KERNEL);
-               if (!p || !ptmp) {
+               if (!pd || !ptmp) {
                        ipath_dev_err(dd, "Unable to allocate portdata "
                                      "memory, failing open\n");
                        ret = -ENOMEM;
-                       kfree(p);
+                       kfree(pd);
                        kfree(ptmp);
                        goto bail;
                }
-               dd->ipath_pd[port] = p;
+               dd->ipath_pd[port] = pd;
                dd->ipath_pd[port]->port_port = port;
                dd->ipath_pd[port]->port_dd = dd;
                dd->ipath_pd[port]->port_tid_pg_list = ptmp;
                init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
        }
-       if (!dd->ipath_pd[port]->port_cnt) {
-               dd->ipath_pd[port]->port_cnt = 1;
-               fp->private_data = (void *) dd->ipath_pd[port];
+       if (!pd->port_cnt) {
+               pd->userversion = uinfo->spu_userversion;
+               init_user_egr_sizes(pd);
+               if ((ret = init_subports(dd, pd, uinfo)) != 0)
+                       goto bail;
                ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
                           current->comm, current->pid, dd->ipath_unit,
                           port);
-               dd->ipath_pd[port]->port_pid = current->pid;
-               strncpy(dd->ipath_pd[port]->port_comm, current->comm,
-                       sizeof(dd->ipath_pd[port]->port_comm));
+               pd->port_cnt = 1;
+               port_fp(fp) = pd;
+               pd->port_pid = current->pid;
+               strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
                ipath_stats.sps_ports++;
                ret = 0;
-               goto bail;
-       }
-       ret = -EBUSY;
+       } else
+               ret = -EBUSY;
 
 bail:
        return ret;
@@ -1264,7 +1512,8 @@ static inline int usable(struct ipath_devdata *dd)
                                     | IPATH_LINKUNK));
 }
 
-static int find_free_port(int unit, struct file *fp)
+static int find_free_port(int unit, struct file *fp,
+                         const struct ipath_user_info *uinfo)
 {
        struct ipath_devdata *dd = ipath_lookup(unit);
        int ret, i;
@@ -1279,8 +1528,8 @@ static int find_free_port(int unit, struct file *fp)
                goto bail;
        }
 
-       for (i = 0; i < dd->ipath_cfgports; i++) {
-               ret = try_alloc_port(dd, i, fp);
+       for (i = 1; i < dd->ipath_cfgports; i++) {
+               ret = try_alloc_port(dd, i, fp, uinfo);
                if (ret != -EBUSY)
                        goto bail;
        }
@@ -1290,13 +1539,14 @@ bail:
        return ret;
 }
 
-static int find_best_unit(struct file *fp)
+static int find_best_unit(struct file *fp,
+                         const struct ipath_user_info *uinfo)
 {
        int ret = 0, i, prefunit = -1, devmax;
        int maxofallports, npresent, nup;
        int ndev;
 
-       (void) ipath_count_units(&npresent, &nup, &maxofallports);
+       devmax = ipath_count_units(&npresent, &nup, &maxofallports);
 
        /*
         * This code is present to allow a knowledgeable person to
@@ -1343,8 +1593,6 @@ static int find_best_unit(struct file *fp)
 
        if (prefunit != -1)
                devmax = prefunit + 1;
-       else
-               devmax = ipath_count_units(NULL, NULL, NULL);
 recheck:
        for (i = 1; i < maxofallports; i++) {
                for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
@@ -1359,7 +1607,7 @@ recheck:
                                 * next.
                                 */
                                continue;
-                       ret = try_alloc_port(dd, i, fp);
+                       ret = try_alloc_port(dd, i, fp, uinfo);
                        if (!ret)
                                goto done;
                }
@@ -1395,22 +1643,174 @@ done:
        return ret;
 }
 
+static int find_shared_port(struct file *fp,
+                           const struct ipath_user_info *uinfo)
+{
+       int devmax, ndev, i;
+       int ret = 0;
+
+       devmax = ipath_count_units(NULL, NULL, NULL);
+
+       for (ndev = 0; ndev < devmax; ndev++) {
+               struct ipath_devdata *dd = ipath_lookup(ndev);
+
+               if (!dd)
+                       continue;
+               for (i = 1; i < dd->ipath_cfgports; i++) {
+                       struct ipath_portdata *pd = dd->ipath_pd[i];
+
+                       /* Skip ports which are not yet open */
+                       if (!pd || !pd->port_cnt)
+                               continue;
+                       /* Skip port if it doesn't match the requested one */
+                       if (pd->port_subport_id != uinfo->spu_subport_id)
+                               continue;
+                       /* Verify the sharing process matches the master */
+                       if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
+                           pd->userversion != uinfo->spu_userversion ||
+                           pd->port_cnt >= pd->port_subport_cnt) {
+                               ret = -EINVAL;
+                               goto done;
+                       }
+                       port_fp(fp) = pd;
+                       subport_fp(fp) = pd->port_cnt++;
+                       tidcursor_fp(fp) = 0;
+                       pd->active_slaves |= 1 << subport_fp(fp);
+                       ipath_cdbg(PROC,
+                                  "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
+                                  current->comm, current->pid,
+                                  subport_fp(fp),
+                                  pd->port_comm, pd->port_pid,
+                                  dd->ipath_unit, pd->port_port);
+                       ret = 1;
+                       goto done;
+               }
+       }
+
+done:
+       return ret;
+}
+
 static int ipath_open(struct inode *in, struct file *fp)
 {
-       int ret, user_minor;
+       /* The real work is performed later in ipath_do_user_init() */
+       fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
+       return fp->private_data ? 0 : -ENOMEM;
+}
+
+static int ipath_do_user_init(struct file *fp,
+                             const struct ipath_user_info *uinfo)
+{
+       int ret;
+       struct ipath_portdata *pd;
+       struct ipath_devdata *dd;
+       u32 head32;
+       int i_minor;
+       unsigned swminor;
+
+       /* Check to be sure we haven't already initialized this file */
+       if (port_fp(fp)) {
+               ret = -EINVAL;
+               goto done;
+       }
+
+       /* for now, if major version is different, bail */
+       if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
+               ipath_dbg("User major version %d not same as driver "
+                         "major %d\n", uinfo->spu_userversion >> 16,
+                         IPATH_USER_SWMAJOR);
+               ret = -ENODEV;
+               goto done;
+       }
+
+       swminor = uinfo->spu_userversion & 0xffff;
+       if (swminor != IPATH_USER_SWMINOR)
+               ipath_dbg("User minor version %d not same as driver "
+                         "minor %d\n", swminor, IPATH_USER_SWMINOR);
 
        mutex_lock(&ipath_mutex);
 
-       user_minor = iminor(in) - IPATH_USER_MINOR_BASE;
+       if (swminor == IPATH_USER_SWMINOR && uinfo->spu_subport_cnt &&
+           (ret = find_shared_port(fp, uinfo))) {
+               mutex_unlock(&ipath_mutex);
+               if (ret > 0)
+                       ret = 0;
+               goto done;
+       }
+
+       i_minor = iminor(fp->f_dentry->d_inode) - IPATH_USER_MINOR_BASE;
        ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
-                  (long)in->i_rdev, user_minor);
+                  (long)fp->f_dentry->d_inode->i_rdev, i_minor);
 
-       if (user_minor)
-               ret = find_free_port(user_minor - 1, fp);
+       if (i_minor)
+               ret = find_free_port(i_minor - 1, fp, uinfo);
        else
-               ret = find_best_unit(fp);
+               ret = find_best_unit(fp, uinfo);
 
        mutex_unlock(&ipath_mutex);
+
+       if (ret)
+               goto done;
+
+       pd = port_fp(fp);
+       dd = pd->port_dd;
+
+       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;
 }
 
@@ -1453,6 +1853,7 @@ static void unlock_expected_tids(struct ipath_portdata *pd)
 static int ipath_close(struct inode *in, struct file *fp)
 {
        int ret = 0;
+       struct ipath_filedata *fd;
        struct ipath_portdata *pd;
        struct ipath_devdata *dd;
        unsigned port;
@@ -1462,9 +1863,24 @@ static int ipath_close(struct inode *in, struct file *fp)
 
        mutex_lock(&ipath_mutex);
 
-       pd = port_fp(fp);
-       port = pd->port_port;
+       fd = (struct ipath_filedata *) fp->private_data;
        fp->private_data = NULL;
+       pd = fd->pd;
+       if (!pd) {
+               mutex_unlock(&ipath_mutex);
+               goto bail;
+       }
+       if (--pd->port_cnt) {
+               /*
+                * XXX If the master closes the port before the slave(s),
+                * revoke the mmap for the eager receive queue so
+                * the slave(s) don't wait for receive data forever.
+                */
+               pd->active_slaves &= ~(1 << fd->subport);
+               mutex_unlock(&ipath_mutex);
+               goto bail;
+       }
+       port = pd->port_port;
        dd = pd->port_dd;
 
        if (pd->port_hdrqfull) {
@@ -1503,8 +1919,6 @@ static int ipath_close(struct inode *in, struct file *fp)
 
                /* clean up the pkeys for this port user */
                ipath_clean_part_key(pd, dd);
-
-
                /*
                 * be paranoid, and never write 0's to these, just use an
                 * unused part of the port 0 tail page.  Of course,
@@ -1533,29 +1947,39 @@ static int ipath_close(struct inode *in, struct file *fp)
                dd->ipath_f_clear_tids(dd, pd->port_port);
        }
 
-       pd->port_cnt = 0;
        pd->port_pid = 0;
-
        dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
        mutex_unlock(&ipath_mutex);
        ipath_free_pddata(dd, pd); /* after releasing the mutex */
 
+bail:
+       kfree(fd);
        return ret;
 }
 
-static int ipath_port_info(struct ipath_portdata *pd,
+static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
                           struct ipath_port_info __user *uinfo)
 {
        struct ipath_port_info info;
        int nup;
        int ret;
+       size_t sz;
 
        (void) ipath_count_units(NULL, &nup, NULL);
        info.num_active = nup;
        info.unit = pd->port_dd->ipath_unit;
        info.port = pd->port_port;
+       info.subport = subport;
+       /* Don't return new fields if old library opened the port. */
+       if ((pd->userversion & 0xffff) == IPATH_USER_SWMINOR) {
+               /* Number of user ports available for this device. */
+               info.num_ports = pd->port_dd->ipath_cfgports - 1;
+               info.num_subports = pd->port_subport_cnt;
+               sz = sizeof(info);
+       } else
+               sz = sizeof(info) - 2 * sizeof(u16);
 
-       if (copy_to_user(uinfo, &info, sizeof(info))) {
+       if (copy_to_user(uinfo, &info, sz)) {
                ret = -EFAULT;
                goto bail;
        }
@@ -1565,6 +1989,16 @@ bail:
        return ret;
 }
 
+static int ipath_get_slave_info(struct ipath_portdata *pd,
+                               void __user *slave_mask_addr)
+{
+       int ret = 0;
+
+       if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
+               ret = -EFAULT;
+       return ret;
+}
+
 static ssize_t ipath_write(struct file *fp, const char __user *data,
                           size_t count, loff_t *off)
 {
@@ -1617,6 +2051,11 @@ static ssize_t ipath_write(struct file *fp, const char __user *data,
                dest = &cmd.cmd.part_key;
                src = &ucmd->cmd.part_key;
                break;
+       case IPATH_CMD_SLAVE_INFO:
+               copy = sizeof(cmd.cmd.slave_mask_addr);
+               dest = &cmd.cmd.slave_mask_addr;
+               src = &ucmd->cmd.slave_mask_addr;
+               break;
        default:
                ret = -EINVAL;
                goto bail;
@@ -1634,34 +2073,43 @@ static ssize_t ipath_write(struct file *fp, const char __user *data,
 
        consumed += copy;
        pd = port_fp(fp);
+       if (!pd && cmd.type != IPATH_CMD_USER_INIT) {
+               ret = -EINVAL;
+               goto bail;
+       }
 
        switch (cmd.type) {
        case IPATH_CMD_USER_INIT:
-               ret = ipath_do_user_init(pd, &cmd.cmd.user_info);
-               if (ret < 0)
+               ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
+               if (ret)
                        goto bail;
                ret = ipath_get_base_info(
-                       pd, (void __user *) (unsigned long)
+                       fp, (void __user *) (unsigned long)
                        cmd.cmd.user_info.spu_base_info,
                        cmd.cmd.user_info.spu_base_info_size);
                break;
        case IPATH_CMD_RECV_CTRL:
-               ret = ipath_manage_rcvq(pd, cmd.cmd.recv_ctrl);
+               ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
                break;
        case IPATH_CMD_PORT_INFO:
-               ret = ipath_port_info(pd,
+               ret = ipath_port_info(pd, subport_fp(fp),
                                      (struct ipath_port_info __user *)
                                      (unsigned long) cmd.cmd.port_info);
                break;
        case IPATH_CMD_TID_UPDATE:
-               ret = ipath_tid_update(pd, &cmd.cmd.tid_info);
+               ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
                break;
        case IPATH_CMD_TID_FREE:
-               ret = ipath_tid_free(pd, &cmd.cmd.tid_info);
+               ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
                break;
        case IPATH_CMD_SET_PART_KEY:
                ret = ipath_set_part_key(pd, cmd.cmd.part_key);
                break;
+       case IPATH_CMD_SLAVE_INFO:
+               ret = ipath_get_slave_info(pd,
+                                          (void __user *) (unsigned long)
+                                          cmd.cmd.slave_mask_addr);
+               break;
        }
 
        if (ret >= 0)
@@ -1858,4 +2306,3 @@ void ipath_user_remove(struct ipath_devdata *dd)
 bail:
        return;
 }
-
index a8a56276ff1dbe6dd591a8756f8f0171a8d24063..0cabd4f16234a6fa093f71f7caf5d06f5eb15baf 100644 (file)
@@ -79,8 +79,8 @@ struct ipath_portdata {
        dma_addr_t port_rcvhdrq_phys;
        dma_addr_t port_rcvhdrqtailaddr_phys;
        /*
-        * number of opens on this instance (0 or 1; ignoring forks, dup,
-        * etc. for now)
+        * number of opens (including slave subports) on this instance
+        * (ignoring forks, dup, etc. for now)
         */
        int port_cnt;
        /*
@@ -89,6 +89,10 @@ struct ipath_portdata {
         */
        /* instead of calculating it */
        unsigned port_port;
+       /* non-zero if port is being shared. */
+       u16 port_subport_cnt;
+       /* non-zero if port is being shared. */
+       u16 port_subport_id;
        /* chip offset of PIO buffers for this port */
        u32 port_piobufs;
        /* how many alloc_pages() chunks in port_rcvegrbuf_pages */
@@ -121,6 +125,16 @@ struct ipath_portdata {
        u16 port_pkeys[4];
        /* so file ops can get at unit */
        struct ipath_devdata *port_dd;
+       /* A page of memory for rcvhdrhead, rcvegrhead, rcvegrtail * N */
+       void *subport_uregbase;
+       /* An array of pages for the eager receive buffers * N */
+       void *subport_rcvegrbuf;
+       /* An array of pages for the eager header queue entries * N */
+       void *subport_rcvhdr_base;
+       /* The version of the library which opened this port */
+       u32 userversion;
+       /* Bitmask of active slaves */
+       u32 active_slaves;
 };
 
 struct sk_buff;
@@ -512,6 +526,12 @@ struct ipath_devdata {
        u32 ipath_lli_errors;
 };
 
+/* Private data for file operations */
+struct ipath_filedata {
+       struct ipath_portdata *pd;
+       unsigned subport;
+       unsigned tidcursor;
+};
 extern struct list_head ipath_dev_list;
 extern spinlock_t ipath_devs_lock;
 extern struct ipath_devdata *ipath_lookup(int unit);
@@ -572,7 +592,11 @@ int ipath_set_lid(struct ipath_devdata *, u32, u8);
 int ipath_set_rx_pol_inv(struct ipath_devdata *dd, u8 new_pol_inv);
 
 /* for use in system calls, where we want to know device type, etc. */
-#define port_fp(fp) ((struct ipath_portdata *) (fp)->private_data)
+#define port_fp(fp) ((struct ipath_filedata *)(fp)->private_data)->pd
+#define subport_fp(fp) \
+       ((struct ipath_filedata *)(fp)->private_data)->subport
+#define tidcursor_fp(fp) \
+       ((struct ipath_filedata *)(fp)->private_data)->tidcursor
 
 /*
  * values for ipath_flags
index e299148c4b68ec4ae793f0805a339d3c75bbb081..b48ebf62e5dd57e2a5a6405539deeda2a4c96647 100644 (file)
@@ -297,6 +297,16 @@ static ssize_t show_nguid(struct device *dev,
        return scnprintf(buf, PAGE_SIZE, "%u\n", dd->ipath_nguid);
 }
 
+static ssize_t show_nports(struct device *dev,
+                          struct device_attribute *attr,
+                          char *buf)
+{
+       struct ipath_devdata *dd = dev_get_drvdata(dev);
+
+       /* Return the number of user ports available. */
+       return scnprintf(buf, PAGE_SIZE, "%u\n", dd->ipath_cfgports - 1);
+}
+
 static ssize_t show_serial(struct device *dev,
                           struct device_attribute *attr,
                           char *buf)
@@ -608,6 +618,7 @@ static DEVICE_ATTR(mlid, S_IWUSR | S_IRUGO, show_mlid, store_mlid);
 static DEVICE_ATTR(mtu, S_IWUSR | S_IRUGO, show_mtu, store_mtu);
 static DEVICE_ATTR(enabled, S_IWUSR | S_IRUGO, show_enabled, store_enabled);
 static DEVICE_ATTR(nguid, S_IRUGO, show_nguid, NULL);
+static DEVICE_ATTR(nports, S_IRUGO, show_nports, NULL);
 static DEVICE_ATTR(reset, S_IWUSR, NULL, store_reset);
 static DEVICE_ATTR(serial, S_IRUGO, show_serial, NULL);
 static DEVICE_ATTR(status, S_IRUGO, show_status, NULL);
@@ -623,6 +634,7 @@ static struct attribute *dev_attributes[] = {
        &dev_attr_mlid.attr,
        &dev_attr_mtu.attr,
        &dev_attr_nguid.attr,
+       &dev_attr_nports.attr,
        &dev_attr_serial.attr,
        &dev_attr_status.attr,
        &dev_attr_status_str.attr,