PageRenderTime 214ms CodeModel.GetById 25ms app.highlight 148ms RepoModel.GetById 1ms app.codeStats 1ms

/drivers/infiniband/hw/ipath/ipath_file_ops.c

https://bitbucket.org/cresqo/cm7-p500-kernel
C | 2617 lines | 1894 code | 278 blank | 445 comment | 306 complexity | f6751c75abf34c137f8b85179b5485de MD5 | raw file
Possible License(s): LGPL-2.0, AGPL-1.0, GPL-2.0

Large files files are truncated, but you can click here to view the full file

   1/*
   2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        provided with the distribution.
  23 *
  24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34#include <linux/pci.h>
  35#include <linux/poll.h>
  36#include <linux/cdev.h>
  37#include <linux/swap.h>
  38#include <linux/vmalloc.h>
  39#include <linux/slab.h>
  40#include <linux/highmem.h>
  41#include <linux/io.h>
  42#include <linux/jiffies.h>
  43#include <linux/smp_lock.h>
  44#include <asm/pgtable.h>
  45
  46#include "ipath_kernel.h"
  47#include "ipath_common.h"
  48#include "ipath_user_sdma.h"
  49
  50static int ipath_open(struct inode *, struct file *);
  51static int ipath_close(struct inode *, struct file *);
  52static ssize_t ipath_write(struct file *, const char __user *, size_t,
  53			   loff_t *);
  54static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
  55			    unsigned long , loff_t);
  56static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  57static int ipath_mmap(struct file *, struct vm_area_struct *);
  58
  59static const struct file_operations ipath_file_ops = {
  60	.owner = THIS_MODULE,
  61	.write = ipath_write,
  62	.aio_write = ipath_writev,
  63	.open = ipath_open,
  64	.release = ipath_close,
  65	.poll = ipath_poll,
  66	.mmap = ipath_mmap
  67};
  68
  69/*
  70 * Convert kernel virtual addresses to physical addresses so they don't
  71 * potentially conflict with the chip addresses used as mmap offsets.
  72 * It doesn't really matter what mmap offset we use as long as we can
  73 * interpret it correctly.
  74 */
  75static u64 cvt_kvaddr(void *p)
  76{
  77	struct page *page;
  78	u64 paddr = 0;
  79
  80	page = vmalloc_to_page(p);
  81	if (page)
  82		paddr = page_to_pfn(page) << PAGE_SHIFT;
  83
  84	return paddr;
  85}
  86
  87static int ipath_get_base_info(struct file *fp,
  88			       void __user *ubase, size_t ubase_size)
  89{
  90	struct ipath_portdata *pd = port_fp(fp);
  91	int ret = 0;
  92	struct ipath_base_info *kinfo = NULL;
  93	struct ipath_devdata *dd = pd->port_dd;
  94	unsigned subport_cnt;
  95	int shared, master;
  96	size_t sz;
  97
  98	subport_cnt = pd->port_subport_cnt;
  99	if (!subport_cnt) {
 100		shared = 0;
 101		master = 0;
 102		subport_cnt = 1;
 103	} else {
 104		shared = 1;
 105		master = !subport_fp(fp);
 106	}
 107
 108	sz = sizeof(*kinfo);
 109	/* If port sharing is not requested, allow the old size structure */
 110	if (!shared)
 111		sz -= 7 * sizeof(u64);
 112	if (ubase_size < sz) {
 113		ipath_cdbg(PROC,
 114			   "Base size %zu, need %zu (version mismatch?)\n",
 115			   ubase_size, sz);
 116		ret = -EINVAL;
 117		goto bail;
 118	}
 119
 120	kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 121	if (kinfo == NULL) {
 122		ret = -ENOMEM;
 123		goto bail;
 124	}
 125
 126	ret = dd->ipath_f_get_base_info(pd, kinfo);
 127	if (ret < 0)
 128		goto bail;
 129
 130	kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 131	kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 132	kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 133	kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 134	/*
 135	 * have to mmap whole thing
 136	 */
 137	kinfo->spi_rcv_egrbuftotlen =
 138		pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 139	kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 140	kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 141		pd->port_rcvegrbuf_chunks;
 142	kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 143	if (master)
 144		kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 145	/*
 146	 * for this use, may be ipath_cfgports summed over all chips that
 147	 * are are configured and present
 148	 */
 149	kinfo->spi_nports = dd->ipath_cfgports;
 150	/* unit (chip/board) our port is on */
 151	kinfo->spi_unit = dd->ipath_unit;
 152	/* for now, only a single page */
 153	kinfo->spi_tid_maxsize = PAGE_SIZE;
 154
 155	/*
 156	 * Doing this per port, and based on the skip value, etc.  This has
 157	 * to be the actual buffer size, since the protocol code treats it
 158	 * as an array.
 159	 *
 160	 * These have to be set to user addresses in the user code via mmap.
 161	 * These values are used on return to user code for the mmap target
 162	 * addresses only.  For 32 bit, same 44 bit address problem, so use
 163	 * the physical address, not virtual.  Before 2.6.11, using the
 164	 * page_address() macro worked, but in 2.6.11, even that returns the
 165	 * full 64 bit address (upper bits all 1's).  So far, using the
 166	 * physical addresses (or chip offsets, for chip mapping) works, but
 167	 * no doubt some future kernel release will change that, and we'll be
 168	 * on to yet another method of dealing with this.
 169	 */
 170	kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 171	kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 172	kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 173	kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 174	kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 175		(void *) dd->ipath_statusp -
 176		(void *) dd->ipath_pioavailregs_dma;
 177	if (!shared) {
 178		kinfo->spi_piocnt = pd->port_piocnt;
 179		kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 180		kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 181			dd->ipath_ureg_align * pd->port_port;
 182	} else if (master) {
 183		kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
 184				    (pd->port_piocnt % subport_cnt);
 185		/* Master's PIO buffers are after all the slave's */
 186		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 187			dd->ipath_palign *
 188			(pd->port_piocnt - kinfo->spi_piocnt);
 189	} else {
 190		unsigned slave = subport_fp(fp) - 1;
 191
 192		kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
 193		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 194			dd->ipath_palign * kinfo->spi_piocnt * slave;
 195	}
 196
 197	if (shared) {
 198		kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 199			dd->ipath_ureg_align * pd->port_port;
 200		kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 201		kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 202		kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 203
 204		kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 205			PAGE_SIZE * subport_fp(fp));
 206
 207		kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 208			pd->port_rcvhdrq_size * subport_fp(fp));
 209		kinfo->spi_rcvhdr_tailaddr = 0;
 210		kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 211			pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 212			subport_fp(fp));
 213
 214		kinfo->spi_subport_uregbase =
 215			cvt_kvaddr(pd->subport_uregbase);
 216		kinfo->spi_subport_rcvegrbuf =
 217			cvt_kvaddr(pd->subport_rcvegrbuf);
 218		kinfo->spi_subport_rcvhdr_base =
 219			cvt_kvaddr(pd->subport_rcvhdr_base);
 220		ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 221			kinfo->spi_port, kinfo->spi_runtime_flags,
 222			(unsigned long long) kinfo->spi_subport_uregbase,
 223			(unsigned long long) kinfo->spi_subport_rcvegrbuf,
 224			(unsigned long long) kinfo->spi_subport_rcvhdr_base);
 225	}
 226
 227	/*
 228	 * All user buffers are 2KB buffers.  If we ever support
 229	 * giving 4KB buffers to user processes, this will need some
 230	 * work.
 231	 */
 232	kinfo->spi_pioindex = (kinfo->spi_piobufbase -
 233		(dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
 234	kinfo->spi_pioalign = dd->ipath_palign;
 235
 236	kinfo->spi_qpair = IPATH_KD_QP;
 237	/*
 238	 * user mode PIO buffers are always 2KB, even when 4KB can
 239	 * be received, and sent via the kernel; this is ibmaxlen
 240	 * for 2K MTU.
 241	 */
 242	kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 243	kinfo->spi_mtu = dd->ipath_ibmaxlen;	/* maxlen, not ibmtu */
 244	kinfo->spi_port = pd->port_port;
 245	kinfo->spi_subport = subport_fp(fp);
 246	kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 247	kinfo->spi_hw_version = dd->ipath_revision;
 248
 249	if (master) {
 250		kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 251	}
 252
 253	sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 254	if (copy_to_user(ubase, kinfo, sz))
 255		ret = -EFAULT;
 256
 257bail:
 258	kfree(kinfo);
 259	return ret;
 260}
 261
 262/**
 263 * ipath_tid_update - update a port TID
 264 * @pd: the port
 265 * @fp: the ipath device file
 266 * @ti: the TID information
 267 *
 268 * The new implementation as of Oct 2004 is that the driver assigns
 269 * the tid and returns it to the caller.   To make it easier to
 270 * catch bugs, and to reduce search time, we keep a cursor for
 271 * each port, walking the shadow tid array to find one that's not
 272 * in use.
 273 *
 274 * For now, if we can't allocate the full list, we fail, although
 275 * in the long run, we'll allocate as many as we can, and the
 276 * caller will deal with that by trying the remaining pages later.
 277 * That means that when we fail, we have to mark the tids as not in
 278 * use again, in our shadow copy.
 279 *
 280 * It's up to the caller to free the tids when they are done.
 281 * We'll unlock the pages as they free them.
 282 *
 283 * Also, right now we are locking one page at a time, but since
 284 * the intended use of this routine is for a single group of
 285 * virtually contiguous pages, that should change to improve
 286 * performance.
 287 */
 288static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 289			    const struct ipath_tid_info *ti)
 290{
 291	int ret = 0, ntids;
 292	u32 tid, porttid, cnt, i, tidcnt, tidoff;
 293	u16 *tidlist;
 294	struct ipath_devdata *dd = pd->port_dd;
 295	u64 physaddr;
 296	unsigned long vaddr;
 297	u64 __iomem *tidbase;
 298	unsigned long tidmap[8];
 299	struct page **pagep = NULL;
 300	unsigned subport = subport_fp(fp);
 301
 302	if (!dd->ipath_pageshadow) {
 303		ret = -ENOMEM;
 304		goto done;
 305	}
 306
 307	cnt = ti->tidcnt;
 308	if (!cnt) {
 309		ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 310			  (unsigned long long) ti->tidlist);
 311		/*
 312		 * Should we treat as success?  likely a bug
 313		 */
 314		ret = -EFAULT;
 315		goto done;
 316	}
 317	porttid = pd->port_port * dd->ipath_rcvtidcnt;
 318	if (!pd->port_subport_cnt) {
 319		tidcnt = dd->ipath_rcvtidcnt;
 320		tid = pd->port_tidcursor;
 321		tidoff = 0;
 322	} else if (!subport) {
 323		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 324			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 325		tidoff = dd->ipath_rcvtidcnt - tidcnt;
 326		porttid += tidoff;
 327		tid = tidcursor_fp(fp);
 328	} else {
 329		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 330		tidoff = tidcnt * (subport - 1);
 331		porttid += tidoff;
 332		tid = tidcursor_fp(fp);
 333	}
 334	if (cnt > tidcnt) {
 335		/* make sure it all fits in port_tid_pg_list */
 336		dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 337			 "TIDs, only trying max (%u)\n", cnt, tidcnt);
 338		cnt = tidcnt;
 339	}
 340	pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 341	tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 342
 343	memset(tidmap, 0, sizeof(tidmap));
 344	/* before decrement; chip actual # */
 345	ntids = tidcnt;
 346	tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 347				   dd->ipath_rcvtidbase +
 348				   porttid * sizeof(*tidbase));
 349
 350	ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 351		   pd->port_port, cnt, tid, tidbase);
 352
 353	/* virtual address of first page in transfer */
 354	vaddr = ti->tidvaddr;
 355	if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 356		       cnt * PAGE_SIZE)) {
 357		ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 358			  (void *)vaddr, cnt);
 359		ret = -EFAULT;
 360		goto done;
 361	}
 362	ret = ipath_get_user_pages(vaddr, cnt, pagep);
 363	if (ret) {
 364		if (ret == -EBUSY) {
 365			ipath_dbg("Failed to lock addr %p, %u pages "
 366				  "(already locked)\n",
 367				  (void *) vaddr, cnt);
 368			/*
 369			 * for now, continue, and see what happens but with
 370			 * the new implementation, this should never happen,
 371			 * unless perhaps the user has mpin'ed the pages
 372			 * themselves (something we need to test)
 373			 */
 374			ret = 0;
 375		} else {
 376			dev_info(&dd->pcidev->dev,
 377				 "Failed to lock addr %p, %u pages: "
 378				 "errno %d\n", (void *) vaddr, cnt, -ret);
 379			goto done;
 380		}
 381	}
 382	for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 383		for (; ntids--; tid++) {
 384			if (tid == tidcnt)
 385				tid = 0;
 386			if (!dd->ipath_pageshadow[porttid + tid])
 387				break;
 388		}
 389		if (ntids < 0) {
 390			/*
 391			 * oops, wrapped all the way through their TIDs,
 392			 * and didn't have enough free; see comments at
 393			 * start of routine
 394			 */
 395			ipath_dbg("Not enough free TIDs for %u pages "
 396				  "(index %d), failing\n", cnt, i);
 397			i--;	/* last tidlist[i] not filled in */
 398			ret = -ENOMEM;
 399			break;
 400		}
 401		tidlist[i] = tid + tidoff;
 402		ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 403			   "vaddr %lx\n", i, tid + tidoff, vaddr);
 404		/* we "know" system pages and TID pages are same size */
 405		dd->ipath_pageshadow[porttid + tid] = pagep[i];
 406		dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 407			dd->pcidev, pagep[i], 0, PAGE_SIZE,
 408			PCI_DMA_FROMDEVICE);
 409		/*
 410		 * don't need atomic or it's overhead
 411		 */
 412		__set_bit(tid, tidmap);
 413		physaddr = dd->ipath_physshadow[porttid + tid];
 414		ipath_stats.sps_pagelocks++;
 415		ipath_cdbg(VERBOSE,
 416			   "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 417			   tid, vaddr, (unsigned long long) physaddr,
 418			   pagep[i]);
 419		dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 420				    physaddr);
 421		/*
 422		 * don't check this tid in ipath_portshadow, since we
 423		 * just filled it in; start with the next one.
 424		 */
 425		tid++;
 426	}
 427
 428	if (ret) {
 429		u32 limit;
 430	cleanup:
 431		/* jump here if copy out of updated info failed... */
 432		ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 433			  -ret, i, cnt);
 434		/* same code that's in ipath_free_tid() */
 435		limit = sizeof(tidmap) * BITS_PER_BYTE;
 436		if (limit > tidcnt)
 437			/* just in case size changes in future */
 438			limit = tidcnt;
 439		tid = find_first_bit((const unsigned long *)tidmap, limit);
 440		for (; tid < limit; tid++) {
 441			if (!test_bit(tid, tidmap))
 442				continue;
 443			if (dd->ipath_pageshadow[porttid + tid]) {
 444				ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 445					   tid);
 446				dd->ipath_f_put_tid(dd, &tidbase[tid],
 447						    RCVHQ_RCV_TYPE_EXPECTED,
 448						    dd->ipath_tidinvalid);
 449				pci_unmap_page(dd->pcidev,
 450					dd->ipath_physshadow[porttid + tid],
 451					PAGE_SIZE, PCI_DMA_FROMDEVICE);
 452				dd->ipath_pageshadow[porttid + tid] = NULL;
 453				ipath_stats.sps_pageunlocks++;
 454			}
 455		}
 456		ipath_release_user_pages(pagep, cnt);
 457	} else {
 458		/*
 459		 * Copy the updated array, with ipath_tid's filled in, back
 460		 * to user.  Since we did the copy in already, this "should
 461		 * never fail" If it does, we have to clean up...
 462		 */
 463		if (copy_to_user((void __user *)
 464				 (unsigned long) ti->tidlist,
 465				 tidlist, cnt * sizeof(*tidlist))) {
 466			ret = -EFAULT;
 467			goto cleanup;
 468		}
 469		if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 470				 tidmap, sizeof tidmap)) {
 471			ret = -EFAULT;
 472			goto cleanup;
 473		}
 474		if (tid == tidcnt)
 475			tid = 0;
 476		if (!pd->port_subport_cnt)
 477			pd->port_tidcursor = tid;
 478		else
 479			tidcursor_fp(fp) = tid;
 480	}
 481
 482done:
 483	if (ret)
 484		ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 485			  ti->tidcnt, -ret);
 486	return ret;
 487}
 488
 489/**
 490 * ipath_tid_free - free a port TID
 491 * @pd: the port
 492 * @subport: the subport
 493 * @ti: the TID info
 494 *
 495 * right now we are unlocking one page at a time, but since
 496 * the intended use of this routine is for a single group of
 497 * virtually contiguous pages, that should change to improve
 498 * performance.  We check that the TID is in range for this port
 499 * but otherwise don't check validity; if user has an error and
 500 * frees the wrong tid, it's only their own data that can thereby
 501 * be corrupted.  We do check that the TID was in use, for sanity
 502 * We always use our idea of the saved address, not the address that
 503 * they pass in to us.
 504 */
 505
 506static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 507			  const struct ipath_tid_info *ti)
 508{
 509	int ret = 0;
 510	u32 tid, porttid, cnt, limit, tidcnt;
 511	struct ipath_devdata *dd = pd->port_dd;
 512	u64 __iomem *tidbase;
 513	unsigned long tidmap[8];
 514
 515	if (!dd->ipath_pageshadow) {
 516		ret = -ENOMEM;
 517		goto done;
 518	}
 519
 520	if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 521			   sizeof tidmap)) {
 522		ret = -EFAULT;
 523		goto done;
 524	}
 525
 526	porttid = pd->port_port * dd->ipath_rcvtidcnt;
 527	if (!pd->port_subport_cnt)
 528		tidcnt = dd->ipath_rcvtidcnt;
 529	else if (!subport) {
 530		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 531			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 532		porttid += dd->ipath_rcvtidcnt - tidcnt;
 533	} else {
 534		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 535		porttid += tidcnt * (subport - 1);
 536	}
 537	tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 538				   dd->ipath_rcvtidbase +
 539				   porttid * sizeof(*tidbase));
 540
 541	limit = sizeof(tidmap) * BITS_PER_BYTE;
 542	if (limit > tidcnt)
 543		/* just in case size changes in future */
 544		limit = tidcnt;
 545	tid = find_first_bit(tidmap, limit);
 546	ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 547		   "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 548		   limit, tid, porttid);
 549	for (cnt = 0; tid < limit; tid++) {
 550		/*
 551		 * small optimization; if we detect a run of 3 or so without
 552		 * any set, use find_first_bit again.  That's mainly to
 553		 * accelerate the case where we wrapped, so we have some at
 554		 * the beginning, and some at the end, and a big gap
 555		 * in the middle.
 556		 */
 557		if (!test_bit(tid, tidmap))
 558			continue;
 559		cnt++;
 560		if (dd->ipath_pageshadow[porttid + tid]) {
 561			struct page *p;
 562			p = dd->ipath_pageshadow[porttid + tid];
 563			dd->ipath_pageshadow[porttid + tid] = NULL;
 564			ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 565				   pid_nr(pd->port_pid), tid);
 566			dd->ipath_f_put_tid(dd, &tidbase[tid],
 567					    RCVHQ_RCV_TYPE_EXPECTED,
 568					    dd->ipath_tidinvalid);
 569			pci_unmap_page(dd->pcidev,
 570				dd->ipath_physshadow[porttid + tid],
 571				PAGE_SIZE, PCI_DMA_FROMDEVICE);
 572			ipath_release_user_pages(&p, 1);
 573			ipath_stats.sps_pageunlocks++;
 574		} else
 575			ipath_dbg("Unused tid %u, ignoring\n", tid);
 576	}
 577	if (cnt != ti->tidcnt)
 578		ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 579			  ti->tidcnt, cnt);
 580done:
 581	if (ret)
 582		ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 583			  ti->tidcnt, -ret);
 584	return ret;
 585}
 586
 587/**
 588 * ipath_set_part_key - set a partition key
 589 * @pd: the port
 590 * @key: the key
 591 *
 592 * We can have up to 4 active at a time (other than the default, which is
 593 * always allowed).  This is somewhat tricky, since multiple ports may set
 594 * the same key, so we reference count them, and clean up at exit.  All 4
 595 * partition keys are packed into a single infinipath register.  It's an
 596 * error for a process to set the same pkey multiple times.  We provide no
 597 * mechanism to de-allocate a pkey at this time, we may eventually need to
 598 * do that.  I've used the atomic operations, and no locking, and only make
 599 * a single pass through what's available.  This should be more than
 600 * adequate for some time. I'll think about spinlocks or the like if and as
 601 * it's necessary.
 602 */
 603static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 604{
 605	struct ipath_devdata *dd = pd->port_dd;
 606	int i, any = 0, pidx = -1;
 607	u16 lkey = key & 0x7FFF;
 608	int ret;
 609
 610	if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 611		/* nothing to do; this key always valid */
 612		ret = 0;
 613		goto bail;
 614	}
 615
 616	ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 617		   "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 618		   pd->port_port, key, dd->ipath_pkeys[0],
 619		   atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 620		   atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 621		   atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 622		   atomic_read(&dd->ipath_pkeyrefs[3]));
 623
 624	if (!lkey) {
 625		ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 626			   pd->port_port);
 627		ret = -EINVAL;
 628		goto bail;
 629	}
 630
 631	/*
 632	 * Set the full membership bit, because it has to be
 633	 * set in the register or the packet, and it seems
 634	 * cleaner to set in the register than to force all
 635	 * callers to set it. (see bug 4331)
 636	 */
 637	key |= 0x8000;
 638
 639	for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 640		if (!pd->port_pkeys[i] && pidx == -1)
 641			pidx = i;
 642		if (pd->port_pkeys[i] == key) {
 643			ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 644				   "(%x) more than once\n",
 645				   pd->port_port, key);
 646			ret = -EEXIST;
 647			goto bail;
 648		}
 649	}
 650	if (pidx == -1) {
 651		ipath_dbg("All pkeys for port %u already in use, "
 652			  "can't set %x\n", pd->port_port, key);
 653		ret = -EBUSY;
 654		goto bail;
 655	}
 656	for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 657		if (!dd->ipath_pkeys[i]) {
 658			any++;
 659			continue;
 660		}
 661		if (dd->ipath_pkeys[i] == key) {
 662			atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 663
 664			if (atomic_inc_return(pkrefs) > 1) {
 665				pd->port_pkeys[pidx] = key;
 666				ipath_cdbg(VERBOSE, "p%u set key %x "
 667					   "matches #%d, count now %d\n",
 668					   pd->port_port, key, i,
 669					   atomic_read(pkrefs));
 670				ret = 0;
 671				goto bail;
 672			} else {
 673				/*
 674				 * lost race, decrement count, catch below
 675				 */
 676				atomic_dec(pkrefs);
 677				ipath_cdbg(VERBOSE, "Lost race, count was "
 678					   "0, after dec, it's %d\n",
 679					   atomic_read(pkrefs));
 680				any++;
 681			}
 682		}
 683		if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 684			/*
 685			 * It makes no sense to have both the limited and
 686			 * full membership PKEY set at the same time since
 687			 * the unlimited one will disable the limited one.
 688			 */
 689			ret = -EEXIST;
 690			goto bail;
 691		}
 692	}
 693	if (!any) {
 694		ipath_dbg("port %u, all pkeys already in use, "
 695			  "can't set %x\n", pd->port_port, key);
 696		ret = -EBUSY;
 697		goto bail;
 698	}
 699	for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 700		if (!dd->ipath_pkeys[i] &&
 701		    atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 702			u64 pkey;
 703
 704			/* for ipathstats, etc. */
 705			ipath_stats.sps_pkeys[i] = lkey;
 706			pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 707			pkey =
 708				(u64) dd->ipath_pkeys[0] |
 709				((u64) dd->ipath_pkeys[1] << 16) |
 710				((u64) dd->ipath_pkeys[2] << 32) |
 711				((u64) dd->ipath_pkeys[3] << 48);
 712			ipath_cdbg(PROC, "p%u set key %x in #%d, "
 713				   "portidx %d, new pkey reg %llx\n",
 714				   pd->port_port, key, i, pidx,
 715				   (unsigned long long) pkey);
 716			ipath_write_kreg(
 717				dd, dd->ipath_kregs->kr_partitionkey, pkey);
 718
 719			ret = 0;
 720			goto bail;
 721		}
 722	}
 723	ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 724		  "can't set %x\n", pd->port_port, key);
 725	ret = -EBUSY;
 726
 727bail:
 728	return ret;
 729}
 730
 731/**
 732 * ipath_manage_rcvq - manage a port's receive queue
 733 * @pd: the port
 734 * @subport: the subport
 735 * @start_stop: action to carry out
 736 *
 737 * start_stop == 0 disables receive on the port, for use in queue
 738 * overflow conditions.  start_stop==1 re-enables, to be used to
 739 * re-init the software copy of the head register
 740 */
 741static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 742			     int start_stop)
 743{
 744	struct ipath_devdata *dd = pd->port_dd;
 745
 746	ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 747		   start_stop ? "en" : "dis", dd->ipath_unit,
 748		   pd->port_port, subport);
 749	if (subport)
 750		goto bail;
 751	/* atomically clear receive enable port. */
 752	if (start_stop) {
 753		/*
 754		 * On enable, force in-memory copy of the tail register to
 755		 * 0, so that protocol code doesn't have to worry about
 756		 * whether or not the chip has yet updated the in-memory
 757		 * copy or not on return from the system call. The chip
 758		 * always resets it's tail register back to 0 on a
 759		 * transition from disabled to enabled.  This could cause a
 760		 * problem if software was broken, and did the enable w/o
 761		 * the disable, but eventually the in-memory copy will be
 762		 * updated and correct itself, even in the face of software
 763		 * bugs.
 764		 */
 765		if (pd->port_rcvhdrtail_kvaddr)
 766			ipath_clear_rcvhdrtail(pd);
 767		set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 768			&dd->ipath_rcvctrl);
 769	} else
 770		clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 771			  &dd->ipath_rcvctrl);
 772	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 773			 dd->ipath_rcvctrl);
 774	/* now be sure chip saw it before we return */
 775	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 776	if (start_stop) {
 777		/*
 778		 * And try to be sure that tail reg update has happened too.
 779		 * This should in theory interlock with the RXE changes to
 780		 * the tail register.  Don't assign it to the tail register
 781		 * in memory copy, since we could overwrite an update by the
 782		 * chip if we did.
 783		 */
 784		ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 785	}
 786	/* always; new head should be equal to new tail; see above */
 787bail:
 788	return 0;
 789}
 790
 791static void ipath_clean_part_key(struct ipath_portdata *pd,
 792				 struct ipath_devdata *dd)
 793{
 794	int i, j, pchanged = 0;
 795	u64 oldpkey;
 796
 797	/* for debugging only */
 798	oldpkey = (u64) dd->ipath_pkeys[0] |
 799		((u64) dd->ipath_pkeys[1] << 16) |
 800		((u64) dd->ipath_pkeys[2] << 32) |
 801		((u64) dd->ipath_pkeys[3] << 48);
 802
 803	for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 804		if (!pd->port_pkeys[i])
 805			continue;
 806		ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 807			   pd->port_pkeys[i]);
 808		for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 809			/* check for match independent of the global bit */
 810			if ((dd->ipath_pkeys[j] & 0x7fff) !=
 811			    (pd->port_pkeys[i] & 0x7fff))
 812				continue;
 813			if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 814				ipath_cdbg(VERBOSE, "p%u clear key "
 815					   "%x matches #%d\n",
 816					   pd->port_port,
 817					   pd->port_pkeys[i], j);
 818				ipath_stats.sps_pkeys[j] =
 819					dd->ipath_pkeys[j] = 0;
 820				pchanged++;
 821			}
 822			else ipath_cdbg(
 823				VERBOSE, "p%u key %x matches #%d, "
 824				"but ref still %d\n", pd->port_port,
 825				pd->port_pkeys[i], j,
 826				atomic_read(&dd->ipath_pkeyrefs[j]));
 827			break;
 828		}
 829		pd->port_pkeys[i] = 0;
 830	}
 831	if (pchanged) {
 832		u64 pkey = (u64) dd->ipath_pkeys[0] |
 833			((u64) dd->ipath_pkeys[1] << 16) |
 834			((u64) dd->ipath_pkeys[2] << 32) |
 835			((u64) dd->ipath_pkeys[3] << 48);
 836		ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 837			   "new pkey reg %llx\n", pd->port_port,
 838			   (unsigned long long) oldpkey,
 839			   (unsigned long long) pkey);
 840		ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 841				 pkey);
 842	}
 843}
 844
 845/*
 846 * Initialize the port data with the receive buffer sizes
 847 * so this can be done while the master port is locked.
 848 * Otherwise, there is a race with a slave opening the port
 849 * and seeing these fields uninitialized.
 850 */
 851static void init_user_egr_sizes(struct ipath_portdata *pd)
 852{
 853	struct ipath_devdata *dd = pd->port_dd;
 854	unsigned egrperchunk, egrcnt, size;
 855
 856	/*
 857	 * to avoid wasting a lot of memory, we allocate 32KB chunks of
 858	 * physically contiguous memory, advance through it until used up
 859	 * and then allocate more.  Of course, we need memory to store those
 860	 * extra pointers, now.  Started out with 256KB, but under heavy
 861	 * memory pressure (creating large files and then copying them over
 862	 * NFS while doing lots of MPI jobs), we hit some allocation
 863	 * failures, even though we can sleep...  (2.6.10) Still get
 864	 * failures at 64K.  32K is the lowest we can go without wasting
 865	 * additional memory.
 866	 */
 867	size = 0x8000;
 868	egrperchunk = size / dd->ipath_rcvegrbufsize;
 869	egrcnt = dd->ipath_rcvegrcnt;
 870	pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 871	pd->port_rcvegrbufs_perchunk = egrperchunk;
 872	pd->port_rcvegrbuf_size = size;
 873}
 874
 875/**
 876 * ipath_create_user_egr - allocate eager TID buffers
 877 * @pd: the port to allocate TID buffers for
 878 *
 879 * This routine is now quite different for user and kernel, because
 880 * the kernel uses skb's, for the accelerated network performance
 881 * This is the user port version
 882 *
 883 * Allocate the eager TID buffers and program them into infinipath
 884 * They are no longer completely contiguous, we do multiple allocation
 885 * calls.
 886 */
 887static int ipath_create_user_egr(struct ipath_portdata *pd)
 888{
 889	struct ipath_devdata *dd = pd->port_dd;
 890	unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 891	size_t size;
 892	int ret;
 893	gfp_t gfp_flags;
 894
 895	/*
 896	 * GFP_USER, but without GFP_FS, so buffer cache can be
 897	 * coalesced (we hope); otherwise, even at order 4,
 898	 * heavy filesystem activity makes these fail, and we can
 899	 * use compound pages.
 900	 */
 901	gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 902
 903	egrcnt = dd->ipath_rcvegrcnt;
 904	/* TID number offset for this port */
 905	egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 906	egrsize = dd->ipath_rcvegrbufsize;
 907	ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 908		   "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 909
 910	chunk = pd->port_rcvegrbuf_chunks;
 911	egrperchunk = pd->port_rcvegrbufs_perchunk;
 912	size = pd->port_rcvegrbuf_size;
 913	pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 914				     GFP_KERNEL);
 915	if (!pd->port_rcvegrbuf) {
 916		ret = -ENOMEM;
 917		goto bail;
 918	}
 919	pd->port_rcvegrbuf_phys =
 920		kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 921			GFP_KERNEL);
 922	if (!pd->port_rcvegrbuf_phys) {
 923		ret = -ENOMEM;
 924		goto bail_rcvegrbuf;
 925	}
 926	for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 927
 928		pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 929			&dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 930			gfp_flags);
 931
 932		if (!pd->port_rcvegrbuf[e]) {
 933			ret = -ENOMEM;
 934			goto bail_rcvegrbuf_phys;
 935		}
 936	}
 937
 938	pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 939
 940	for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 941		dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 942		unsigned i;
 943
 944		for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 945			dd->ipath_f_put_tid(dd, e + egroff +
 946					    (u64 __iomem *)
 947					    ((char __iomem *)
 948					     dd->ipath_kregbase +
 949					     dd->ipath_rcvegrbase),
 950					    RCVHQ_RCV_TYPE_EAGER, pa);
 951			pa += egrsize;
 952		}
 953		cond_resched();	/* don't hog the cpu */
 954	}
 955
 956	ret = 0;
 957	goto bail;
 958
 959bail_rcvegrbuf_phys:
 960	for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 961		pd->port_rcvegrbuf[e]; e++) {
 962		dma_free_coherent(&dd->pcidev->dev, size,
 963				  pd->port_rcvegrbuf[e],
 964				  pd->port_rcvegrbuf_phys[e]);
 965
 966	}
 967	kfree(pd->port_rcvegrbuf_phys);
 968	pd->port_rcvegrbuf_phys = NULL;
 969bail_rcvegrbuf:
 970	kfree(pd->port_rcvegrbuf);
 971	pd->port_rcvegrbuf = NULL;
 972bail:
 973	return ret;
 974}
 975
 976
 977/* common code for the mappings on dma_alloc_coherent mem */
 978static int ipath_mmap_mem(struct vm_area_struct *vma,
 979	struct ipath_portdata *pd, unsigned len, int write_ok,
 980	void *kvaddr, char *what)
 981{
 982	struct ipath_devdata *dd = pd->port_dd;
 983	unsigned long pfn;
 984	int ret;
 985
 986	if ((vma->vm_end - vma->vm_start) > len) {
 987		dev_info(&dd->pcidev->dev,
 988		         "FAIL on %s: len %lx > %x\n", what,
 989			 vma->vm_end - vma->vm_start, len);
 990		ret = -EFAULT;
 991		goto bail;
 992	}
 993
 994	if (!write_ok) {
 995		if (vma->vm_flags & VM_WRITE) {
 996			dev_info(&dd->pcidev->dev,
 997				 "%s must be mapped readonly\n", what);
 998			ret = -EPERM;
 999			goto bail;
1000		}
1001
1002		/* don't allow them to later change with mprotect */
1003		vma->vm_flags &= ~VM_MAYWRITE;
1004	}
1005
1006	pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1007	ret = remap_pfn_range(vma, vma->vm_start, pfn,
1008			      len, vma->vm_page_prot);
1009	if (ret)
1010		dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1011			 "bytes r%c failed: %d\n", what, pd->port_port,
1012			 pfn, len, write_ok?'w':'o', ret);
1013	else
1014		ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1015			   "r%c\n", what, pd->port_port, pfn, len,
1016			   write_ok?'w':'o');
1017bail:
1018	return ret;
1019}
1020
1021static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1022		     u64 ureg)
1023{
1024	unsigned long phys;
1025	int ret;
1026
1027	/*
1028	 * This is real hardware, so use io_remap.  This is the mechanism
1029	 * for the user process to update the head registers for their port
1030	 * in the chip.
1031	 */
1032	if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1033		dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1034			 "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1035		ret = -EFAULT;
1036	} else {
1037		phys = dd->ipath_physaddr + ureg;
1038		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1039
1040		vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1041		ret = io_remap_pfn_range(vma, vma->vm_start,
1042					 phys >> PAGE_SHIFT,
1043					 vma->vm_end - vma->vm_start,
1044					 vma->vm_page_prot);
1045	}
1046	return ret;
1047}
1048
1049static int mmap_piobufs(struct vm_area_struct *vma,
1050			struct ipath_devdata *dd,
1051			struct ipath_portdata *pd,
1052			unsigned piobufs, unsigned piocnt)
1053{
1054	unsigned long phys;
1055	int ret;
1056
1057	/*
1058	 * When we map the PIO buffers in the chip, we want to map them as
1059	 * writeonly, no read possible.   This prevents access to previous
1060	 * process data, and catches users who might try to read the i/o
1061	 * space due to a bug.
1062	 */
1063	if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1064		dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1065			 "reqlen %lx > PAGE\n",
1066			 vma->vm_end - vma->vm_start);
1067		ret = -EINVAL;
1068		goto bail;
1069	}
1070
1071	phys = dd->ipath_physaddr + piobufs;
1072
1073#if defined(__powerpc__)
1074	/* There isn't a generic way to specify writethrough mappings */
1075	pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1076	pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1077	pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1078#endif
1079
1080	/*
1081	 * don't allow them to later change to readable with mprotect (for when
1082	 * not initially mapped readable, as is normally the case)
1083	 */
1084	vma->vm_flags &= ~VM_MAYREAD;
1085	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1086
1087	ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1088				 vma->vm_end - vma->vm_start,
1089				 vma->vm_page_prot);
1090bail:
1091	return ret;
1092}
1093
1094static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1095			   struct ipath_portdata *pd)
1096{
1097	struct ipath_devdata *dd = pd->port_dd;
1098	unsigned long start, size;
1099	size_t total_size, i;
1100	unsigned long pfn;
1101	int ret;
1102
1103	size = pd->port_rcvegrbuf_size;
1104	total_size = pd->port_rcvegrbuf_chunks * size;
1105	if ((vma->vm_end - vma->vm_start) > total_size) {
1106		dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1107			 "reqlen %lx > actual %lx\n",
1108			 vma->vm_end - vma->vm_start,
1109			 (unsigned long) total_size);
1110		ret = -EINVAL;
1111		goto bail;
1112	}
1113
1114	if (vma->vm_flags & VM_WRITE) {
1115		dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1116			 "writable (flags=%lx)\n", vma->vm_flags);
1117		ret = -EPERM;
1118		goto bail;
1119	}
1120	/* don't allow them to later change to writeable with mprotect */
1121	vma->vm_flags &= ~VM_MAYWRITE;
1122
1123	start = vma->vm_start;
1124
1125	for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1126		pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1127		ret = remap_pfn_range(vma, start, pfn, size,
1128				      vma->vm_page_prot);
1129		if (ret < 0)
1130			goto bail;
1131	}
1132	ret = 0;
1133
1134bail:
1135	return ret;
1136}
1137
1138/*
1139 * ipath_file_vma_fault - handle a VMA page fault.
1140 */
1141static int ipath_file_vma_fault(struct vm_area_struct *vma,
1142					struct vm_fault *vmf)
1143{
1144	struct page *page;
1145
1146	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1147	if (!page)
1148		return VM_FAULT_SIGBUS;
1149	get_page(page);
1150	vmf->page = page;
1151
1152	return 0;
1153}
1154
1155static const struct vm_operations_struct ipath_file_vm_ops = {
1156	.fault = ipath_file_vma_fault,
1157};
1158
1159static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1160		       struct ipath_portdata *pd, unsigned subport)
1161{
1162	unsigned long len;
1163	struct ipath_devdata *dd;
1164	void *addr;
1165	size_t size;
1166	int ret = 0;
1167
1168	/* If the port is not shared, all addresses should be physical */
1169	if (!pd->port_subport_cnt)
1170		goto bail;
1171
1172	dd = pd->port_dd;
1173	size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1174
1175	/*
1176	 * Each process has all the subport uregbase, rcvhdrq, and
1177	 * rcvegrbufs mmapped - as an array for all the processes,
1178	 * and also separately for this process.
1179	 */
1180	if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1181		addr = pd->subport_uregbase;
1182		size = PAGE_SIZE * pd->port_subport_cnt;
1183	} else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1184		addr = pd->subport_rcvhdr_base;
1185		size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1186	} else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1187		addr = pd->subport_rcvegrbuf;
1188		size *= pd->port_subport_cnt;
1189        } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1190                                        PAGE_SIZE * subport)) {
1191                addr = pd->subport_uregbase + PAGE_SIZE * subport;
1192                size = PAGE_SIZE;
1193        } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1194                                pd->port_rcvhdrq_size * subport)) {
1195                addr = pd->subport_rcvhdr_base +
1196                        pd->port_rcvhdrq_size * subport;
1197                size = pd->port_rcvhdrq_size;
1198        } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1199                               size * subport)) {
1200                addr = pd->subport_rcvegrbuf + size * subport;
1201                /* rcvegrbufs are read-only on the slave */
1202                if (vma->vm_flags & VM_WRITE) {
1203                        dev_info(&dd->pcidev->dev,
1204                                 "Can't map eager buffers as "
1205                                 "writable (flags=%lx)\n", vma->vm_flags);
1206                        ret = -EPERM;
1207                        goto bail;
1208                }
1209                /*
1210                 * Don't allow permission to later change to writeable
1211                 * with mprotect.
1212                 */
1213                vma->vm_flags &= ~VM_MAYWRITE;
1214	} else {
1215		goto bail;
1216	}
1217	len = vma->vm_end - vma->vm_start;
1218	if (len > size) {
1219		ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1220		ret = -EINVAL;
1221		goto bail;
1222	}
1223
1224	vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1225	vma->vm_ops = &ipath_file_vm_ops;
1226	vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1227	ret = 1;
1228
1229bail:
1230	return ret;
1231}
1232
1233/**
1234 * ipath_mmap - mmap various structures into user space
1235 * @fp: the file pointer
1236 * @vma: the VM area
1237 *
1238 * We use this to have a shared buffer between the kernel and the user code
1239 * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1240 * buffers in the chip.  We have the open and close entries so we can bump
1241 * the ref count and keep the driver from being unloaded while still mapped.
1242 */
1243static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1244{
1245	struct ipath_portdata *pd;
1246	struct ipath_devdata *dd;
1247	u64 pgaddr, ureg;
1248	unsigned piobufs, piocnt;
1249	int ret;
1250
1251	pd = port_fp(fp);
1252	if (!pd) {
1253		ret = -EINVAL;
1254		goto bail;
1255	}
1256	dd = pd->port_dd;
1257
1258	/*
1259	 * This is the ipath_do_user_init() code, mapping the shared buffers
1260	 * into the user process. The address referred to by vm_pgoff is the
1261	 * file offset passed via mmap().  For shared ports, this is the
1262	 * kernel vmalloc() address of the pages to share with the master.
1263	 * For non-shared or master ports, this is a physical address.
1264	 * We only do one mmap for each space mapped.
1265	 */
1266	pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1267
1268	/*
1269	 * Check for 0 in case one of the allocations failed, but user
1270	 * called mmap anyway.
1271	 */
1272	if (!pgaddr)  {
1273		ret = -EINVAL;
1274		goto bail;
1275	}
1276
1277	ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1278		   (unsigned long long) pgaddr, vma->vm_start,
1279		   vma->vm_end - vma->vm_start, dd->ipath_unit,
1280		   pd->port_port, subport_fp(fp));
1281
1282	/*
1283	 * Physical addresses must fit in 40 bits for our hardware.
1284	 * Check for kernel virtual addresses first, anything else must
1285	 * match a HW or memory address.
1286	 */
1287	ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1288	if (ret) {
1289		if (ret > 0)
1290			ret = 0;
1291		goto bail;
1292	}
1293
1294	ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1295	if (!pd->port_subport_cnt) {
1296		/* port is not shared */
1297		piocnt = pd->port_piocnt;
1298		piobufs = pd->port_piobufs;
1299	} else if (!subport_fp(fp)) {
1300		/* caller is the master */
1301		piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1302			 (pd->port_piocnt % pd->port_subport_cnt);
1303		piobufs = pd->port_piobufs +
1304			dd->ipath_palign * (pd->port_piocnt - piocnt);
1305	} else {
1306		unsigned slave = subport_fp(fp) - 1;
1307
1308		/* caller is a slave */
1309		piocnt = pd->port_piocnt / pd->port_subport_cnt;
1310		piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1311	}
1312
1313	if (pgaddr == ureg)
1314		ret = mmap_ureg(vma, dd, ureg);
1315	else if (pgaddr == piobufs)
1316		ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1317	else if (pgaddr == dd->ipath_pioavailregs_phys)
1318		/* in-memory copy of pioavail registers */
1319		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1320			      	     (void *) dd->ipath_pioavailregs_dma,
1321				     "pioavail registers");
1322	else if (pgaddr == pd->port_rcvegr_phys)
1323		ret = mmap_rcvegrbufs(vma, pd);
1324	else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1325		/*
1326		 * The rcvhdrq itself; readonly except on HT (so have
1327		 * to allow writable mapping), multiple pages, contiguous
1328		 * from an i/o perspective.
1329		 */
1330		ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1331				     pd->port_rcvhdrq,
1332				     "rcvhdrq");
1333	else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1334		/* in-memory copy of rcvhdrq tail register */
1335		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1336				     pd->port_rcvhdrtail_kvaddr,
1337				     "rcvhdrq tail");
1338	else
1339		ret = -EINVAL;
1340
1341	vma->vm_private_data = NULL;
1342
1343	if (ret < 0)
1344		dev_info(&dd->pcidev->dev,
1345			 "Failure %d on off %llx len %lx\n",
1346			 -ret, (unsigned long long)pgaddr,
1347			 vma->vm_end - vma->vm_start);
1348bail:
1349	return ret;
1350}
1351
1352static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1353{
1354	unsigned pollflag = 0;
1355
1356	if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1357	    pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1358		pollflag |= POLLIN | POLLRDNORM;
1359		pd->port_hdrqfull_poll = pd->port_hdrqfull;
1360	}
1361
1362	return pollflag;
1363}
1364
1365static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1366				      struct file *fp,
1367				      struct poll_table_struct *pt)
1368{
1369	unsigned pollflag = 0;
1370	struct ipath_devdata *dd;
1371
1372	dd = pd->port_dd;
1373
1374	/* variable access in ipath_poll_hdrqfull() needs this */
1375	rmb();
1376	pollflag = ipath_poll_hdrqfull(pd);
1377
1378	if (pd->port_urgent != pd->port_urgent_poll) {
1379		pollflag |= POLLIN | POLLRDNORM;
1380		pd->port_urgent_poll = pd->port_urgent;
1381	}
1382
1383	if (!pollflag) {
1384		/* this saves a spin_lock/unlock in interrupt handler... */
1385		set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1386		/* flush waiting flag so don't miss an event... */
1387		wmb();
1388		poll_wait(fp, &pd->port_wait, pt);
1389	}
1390
1391	return pollflag;
1392}
1393
1394static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1395				    struct file *fp,
1396				    struct poll_table_struct *pt)
1397{
1398	u32 head;
1399	u32 tail;
1400	unsigned pollflag = 0;
1401	struct ipath_devdata *dd;
1402
1403	dd = pd->port_dd;
1404
1405	/* variable access in ipath_poll_hdrqfull() needs this */
1406	rmb();
1407	pollflag = ipath_poll_hdrqfull(pd);
1408
1409	head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1410	if (pd->port_rcvhdrtail_kvaddr)
1411		tail = ipath_get_rcvhdrtail(pd);
1412	else
1413		tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1414
1415	if (head != tail)
1416		pollflag |= POLLIN | POLLRDNORM;
1417	else {
1418		/* this saves a spin_lock/unlock in interrupt handler */
1419		set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1420		/* flush waiting flag so we don't miss an event */
1421		wmb();
1422
1423		set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1424			&dd->ipath_rcvctrl);
1425
1426		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1427				 dd->ipath_rcvctrl);
1428
1429		if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1430			ipath_write_ureg(dd, ur_rcvhdrhead,
1431					 dd->ipath_rhdrhead_intr_off | head,
1432					 pd->port_port);
1433
1434		poll_wait(fp, &pd->port_wait, pt);
1435	}
1436
1437	return pollflag;
1438}
1439
1440static unsigned int ipath_poll(struct file *fp,
1441			       struct poll_table_struct *pt)
1442{
1443	struct ipath_portdata *pd;
1444	unsigned pollflag;
1445
1446	pd = port_fp(fp);
1447	if (!pd)
1448		pollflag = 0;
1449	else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1450		pollflag = ipath_poll_urgent(pd, fp, pt);
1451	else
1452		pollflag = ipath_poll_next(pd, fp, pt);
1453
1454	return pollflag;
1455}
1456
1457static int ipath_supports_subports(int user_swmajor, int user_swminor)
1458{
1459	/* no subport implementation prior to software version 1.3 */
1460	return (user_swmajor > 1) || (user_swminor >= 3);
1461}
1462
1463static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1464{
1465	/* this code is written long-hand for clarity */
1466	if (IPATH_USER_SWMAJOR != user_swmajor) {
1467		/* no promise of compatibility if major mismatch */
1468		return 0;
1469	}
1470	if (IPATH_USER_SWMAJOR == 1) {
1471		switch (IPATH_USER_SWMINOR) {
1472		case 0:
1473		case 1:
1474		case 2:
1475			/* no subport implementation so cannot be compatible */
1476			return 0;
1477		case 3:
1478			/* 3 is only compatible with itself */
1479			return user_swminor == 3;
1480		default:
1481			/* >= 4 are compatible (or are expected to be) */
1482			return user_swminor >= 4;
1483		}
1484	}
1485	/* make no promises yet for future major versions */
1486	return 0;
1487}
1488
1489static int init_subports(struct ipath_devdata *dd,
1490			 struct ipath_portdata *pd,
1491			 const struct ipath_user_info *uinfo)
1492{
1493	int ret = 0;
1494	unsigned num_subports;
1495	size_t size;
1496
1497	/*
1498	 * If the user is requesting zero subports,
1499	 * skip the subport allocation.
1500	 */
1501	if (uinfo->spu_subport_cnt <= 0)
1502		goto bail;
1503
1504	/* Self-consistency check for ipath_compatible_subports() */
1505	if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1506	    !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1507				       IPATH_USER_SWMINOR)) {
1508		dev_info(&dd->pcidev->dev,
1509			 "Inconsistent ipath_compatible_subports()\n");
1510		goto bail;
1511	}
1512
1513	/* Check for subport compatibility */
1514	if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1515				       uinfo->spu_userversion & 0xffff)) {
1516		dev_info(&dd->pcidev->dev,
1517			 "Mismatched user version (%d.%d) and driver "
1518			 "version (%d.%d) while port sharing. Ensure "
1519                         "that driver and library are from the same "
1520                         "release.\n",
1521			 (int) (uinfo->spu_userversion >> 16),
1522                         (int) (uinfo->spu_userversion & 0xffff),
1523			 IPATH_USER_SWMAJOR,
1524	                 IPATH_USER_SWMINOR);
1525		goto bail;
1526	}
1527	if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1528		ret = -EINVAL;
1529		goto bail;
1530	}
1531
1532	num_subports = uinfo->spu_subport_cnt;
1533	pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1534	if (!pd->subport_uregbase) {
1535		ret = -ENOMEM;
1536		goto bail;
1537	}
1538	/* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1539	size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1540		     sizeof(u32), PAGE_SIZE) * num_subports;
1541	pd->subport_rcvhdr_base = vmalloc(size);
1542	if (!pd->subport_rcvhdr_base) {
1543		ret = -ENOMEM;
1544		goto bail_ureg;
1545	}
1546
1547	pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1548					pd->port_rcvegrbuf_size *
1549					num_subports);
1550	if (!pd->subport_rcvegrbuf) {
1551		ret = -ENOMEM;
1552		goto bail_rhdr;
1553	}
1554
1555	pd->port_subport_cnt = uinfo->spu_subport_cnt;
1556	pd->port_subport_id = uinfo->spu_subport_id;
1557	pd->active_slaves = 1;
1558	set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1559	memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1560	memset(pd->subport_rcvhdr_base, 0, size);
1561	memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1562				         pd->port_rcvegrbuf_size *
1563				         num_subports);
1564	goto bail;
1565
1566bail_rhdr:
1567	vfree(pd->subport_rcvhdr_base);
1568bail_ureg:
1569	vfree(pd->subport_uregbase);
1570	pd->subport_uregbase = NULL;
1571bail:
1572	return ret;
1573}
1574
1575static int try_alloc_port(struct ipath_devdata *dd, int port,
1576			  struct file *fp,
1577			  const struct ipath_user_info *uinfo)
1578{
1579	struct ipath_portdata *pd;
1580	int ret;
1581
1582	if (!(pd = dd->ipath_pd[port])) {
1583		void *ptmp;
1584
1585		pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1586
1587		/*
1588		 * Allocate memory for use in ipath_tid_update() just once
1589		 * at open, not per call.  Reduces cost of expected send
1590		 * setup.
1591		 */
1592		ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1593			       dd->ipath_rcvtidcnt * sizeof(struct page **),
1594			       GFP_KERNEL);
1595		if (!pd || !ptmp) {
1596			ipath_dev_err(dd, "Unable to allocate portdata "
1597				      "memory, failing open\n");
1598			ret = -ENOMEM;
1599			kfree(pd);
1600			kfree(ptmp);
1601			goto bail;
1602		}
1603		dd->ipath_pd[port] = pd;
1604		dd->ipath_pd[port]->port_port = port;
1605		dd->ipath_pd[port]->port_dd = dd;
1606		dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1607		init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1608	}
1609	if (!pd->port_cnt) {
1610		pd->userversion = uinfo->spu_userversion;
1611		init_user_egr_sizes(pd);
1612		if ((ret = init_subports(dd, pd, uinfo)) != 0)
1613			goto bail;
1614		ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1615			   current->comm, current->pid, dd->ipath_unit,
1616			   port);
1617		pd->port_cnt = 1;
1618		port_fp(fp) = pd;
1619		pd->port_pid = get_pid(task_pid(current));
1620		strlcpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1621		ipath_stats.sps_ports++;
1622		ret = 0;
1623	} else
1624		ret = -EBUSY;
1625
1626bail:
1627	return ret;
1628}
1629
1630static inline int usable(struct ipath_devdata *dd)
1631{
1632	return dd &&
1633		(dd->ipath_flags & IPATH_PRESENT) &&
1634		dd->ipath_kregbase &&
1635		dd->ipath_lid &&
1636		!(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1637				     | IPATH_LINKUNK));
1638}
1639
1640static int find_free_port(int unit, struct file *fp,
1641			  const struct ipath_user_info *uinfo)
1642{
1643	struct ipath_devdata *dd = ipath_lookup(unit);
1644	int ret, i;
1645
1646	if (!dd) {
1647		ret = -ENODEV;
1648		goto bail;
1649	}
1650
1651	if (!usable(dd)) {
1652		ret = -ENETDOWN;
1653		goto bail;
1654	}
1655
1656	for (i = 1; i < dd->ipath_cfgports; i++) {
1657		ret = try_alloc_port(dd, i, fp, uinfo);
1658		if (ret != -EBUSY)
1659			goto bail;
1660	}
1661	ret = -EBUSY;
1662
1663bail:
1664	return ret;
1665}
1666
1667static int find_best_unit(struct file *fp,
1668			  const struct ipath_user_info *uinfo)
1669{
1670	int ret = 0, i, prefunit = -1, devmax;
1671	int maxofallports, npresent, nup;
1672	int ndev;
1673
1674	devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1675
1676	/*
1677	 * This code is present to allow a knowledgeable person to
1678	 * specify the layout of processes to processors before opening
1679	 * this driver, and then we'll assign the process to the "closest"
1680	 * InfiniPath chip to that processor (we assume reasonable connectivity,
1681	 * for now).  This code assumes that if affinity has been set
1682	 * before this point, that at most one cpu is set; for now this
1683	 * is reasonable.  I check for both cpumask_empty() and cpumask_full(),
1684	 * in case some kernel variant sets none of the bits when no
1685	 * affinity is set.  2.6.11 and 12 kernels have all present
1686	 * cpus set.  Some day we'll have to fix it up further to handle
1687	 * a cpu subset.  This algorithm fails for two HT chips connected
1688	 * in …

Large files files are truncated, but you can click here to view the full file