ipath_file_ops.c

来自「LINUX 2.6.17.4的源码」· C语言 代码 · 共 1,915 行 · 第 1/4 页

/*
 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include <linux/pci.h>
#include <linux/poll.h>
#include <linux/cdev.h>
#include <linux/swap.h>
#include <linux/vmalloc.h>
#include <asm/pgtable.h>

#include "ipath_kernel.h"
#include "ips_common.h"
#include "ipath_layer.h"

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,
			   loff_t *);
static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
static int ipath_mmap(struct file *, struct vm_area_struct *);

static struct file_operations ipath_file_ops = {
	.owner = THIS_MODULE,
	.write = ipath_write,
	.open = ipath_open,
	.release = ipath_close,
	.poll = ipath_poll,
	.mmap = ipath_mmap
};

static int ipath_get_base_info(struct ipath_portdata *pd,
			       void __user *ubase, size_t ubase_size)
{
	int ret = 0;
	struct ipath_base_info *kinfo = NULL;
	struct ipath_devdata *dd = pd->port_dd;

	if (ubase_size < sizeof(*kinfo)) {
		ipath_cdbg(PROC,
			   "Base size %lu, need %lu (version mismatch?)\n",
			   (unsigned long) ubase_size,
			   (unsigned long) sizeof(*kinfo));
		ret = -EINVAL;
		goto bail;
	}

	kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
	if (kinfo == NULL) {
		ret = -ENOMEM;
		goto bail;
	}

	ret = dd->ipath_f_get_base_info(pd, kinfo);
	if (ret < 0)
		goto bail;

	kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
	kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
	kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
	kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
	/*
	 * have to mmap whole thing
	 */
	kinfo->spi_rcv_egrbuftotlen =
		pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
	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;
	/*
	 * for this use, may be ipath_cfgports summed over all chips that
	 * are are configured and present
	 */
	kinfo->spi_nports = dd->ipath_cfgports;
	/* unit (chip/board) our port is on */
	kinfo->spi_unit = dd->ipath_unit;
	/* for now, only a single page */
	kinfo->spi_tid_maxsize = PAGE_SIZE;

	/*
	 * Doing this per port, and based on the skip value, etc.  This has
	 * to be the actual buffer size, since the protocol code treats it
	 * as an array.
	 *
	 * These have to be set to user addresses in the user code via mmap.
	 * These values are used on return to user code for the mmap target
	 * addresses only.  For 32 bit, same 44 bit address problem, so use
	 * the physical address, not virtual.  Before 2.6.11, using the
	 * 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
	 */
	kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_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;

	kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1);
	kinfo->spi_piocnt = dd->ipath_pbufsport;
	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_sw_version = IPATH_KERN_SWVERSION;
	kinfo->spi_hw_version = dd->ipath_revision;

	if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
		ret = -EFAULT;

bail:
	kfree(kinfo);
	return ret;
}

/**
 * ipath_tid_update - update a port TID
 * @pd: the port
 * @ti: the TID information
 *
 * The new implementation as of Oct 2004 is that the driver assigns
 * the tid and returns it to the caller.   To make it easier to
 * catch bugs, and to reduce search time, we keep a cursor for
 * each port, walking the shadow tid array to find one that's not
 * in use.
 *
 * For now, if we can't allocate the full list, we fail, although
 * in the long run, we'll allocate as many as we can, and the
 * caller will deal with that by trying the remaining pages later.
 * That means that when we fail, we have to mark the tids as not in
 * use again, in our shadow copy.
 *
 * It's up to the caller to free the tids when they are done.
 * We'll unlock the pages as they free them.
 *
 * Also, right now we are locking one page at a time, but since
 * the intended use of this routine is for a single group of
 * virtually contiguous pages, that should change to improve
 * performance.
 */
static int ipath_tid_update(struct ipath_portdata *pd,
			    const struct ipath_tid_info *ti)
{
	int ret = 0, ntids;
	u32 tid, porttid, cnt, i, tidcnt;
	u16 *tidlist;
	struct ipath_devdata *dd = pd->port_dd;
	u64 physaddr;
	unsigned long vaddr;
	u64 __iomem *tidbase;
	unsigned long tidmap[8];
	struct page **pagep = NULL;

	if (!dd->ipath_pageshadow) {
		ret = -ENOMEM;
		goto done;
	}

	cnt = ti->tidcnt;
	if (!cnt) {
		ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
			  (unsigned long long) ti->tidlist);
		/*
		 * Should we treat as success?  likely a bug
		 */
		ret = -EFAULT;
		goto done;
	}
	tidcnt = dd->ipath_rcvtidcnt;
	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]);

	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 +
				   porttid * sizeof(*tidbase));

	ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
		   pd->port_port, cnt, tid, tidbase);

	/* virtual address of first page in transfer */
	vaddr = ti->tidvaddr;
	if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
		       cnt * PAGE_SIZE)) {
		ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
			  (void *)vaddr, cnt);
		ret = -EFAULT;
		goto done;
	}
	ret = ipath_get_user_pages(vaddr, cnt, pagep);
	if (ret) {
		if (ret == -EBUSY) {
			ipath_dbg("Failed to lock addr %p, %u pages "
				  "(already locked)\n",
				  (void *) vaddr, cnt);
			/*
			 * for now, continue, and see what happens but with
			 * the new implementation, this should never happen,
			 * unless perhaps the user has mpin'ed the pages
			 * themselves (something we need to test)
			 */
			ret = 0;
		} else {
			dev_info(&dd->pcidev->dev,
				 "Failed to lock addr %p, %u pages: "
				 "errno %d\n", (void *) vaddr, cnt, -ret);
			goto done;
		}
	}
	for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
		for (; ntids--; tid++) {
			if (tid == tidcnt)
				tid = 0;
			if (!dd->ipath_pageshadow[porttid + tid])
				break;
		}
		if (ntids < 0) {
			/*
			 * oops, wrapped all the way through their TIDs,
			 * and didn't have enough free; see comments at
			 * start of routine
			 */
			ipath_dbg("Not enough free TIDs for %u pages "
				  "(index %d), failing\n", cnt, i);
			i--;	/* last tidlist[i] not filled in */
			ret = -ENOMEM;
			break;
		}
		tidlist[i] = tid;
		ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
			   "vaddr %lx\n", i, tid, vaddr);
		/* we "know" system pages and TID pages are same size */
		dd->ipath_pageshadow[porttid + tid] = pagep[i];
		/*
		 * don't need atomic or it's overhead
		 */
		__set_bit(tid, tidmap);
		physaddr = page_to_phys(pagep[i]);
		ipath_stats.sps_pagelocks++;
		ipath_cdbg(VERBOSE,
			   "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
			   tid, vaddr, (unsigned long long) physaddr,
			   pagep[i]);
		dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
		/*
		 * don't check this tid in ipath_portshadow, since we
		 * just filled it in; start with the next one.
		 */
		tid++;
	}

	if (ret) {
		u32 limit;
	cleanup:
		/* jump here if copy out of updated info failed... */
		ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
			  -ret, i, cnt);
		/* same code that's in ipath_free_tid() */
		limit = sizeof(tidmap) * BITS_PER_BYTE;
		if (limit > tidcnt)
			/* just in case size changes in future */
			limit = tidcnt;
		tid = find_first_bit((const unsigned long *)tidmap, limit);
		for (; tid < limit; tid++) {
			if (!test_bit(tid, tidmap))
				continue;
			if (dd->ipath_pageshadow[porttid + tid]) {
				ipath_cdbg(VERBOSE, "Freeing TID %u\n",
					   tid);
				dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
						    dd->ipath_tidinvalid);
				dd->ipath_pageshadow[porttid + tid] = NULL;
				ipath_stats.sps_pageunlocks++;
			}
		}
		ipath_release_user_pages(pagep, cnt);
	} else {
		/*
		 * Copy the updated array, with ipath_tid's filled in, back
		 * to user.  Since we did the copy in already, this "should
		 * never fail" If it does, we have to clean up...
		 */
		if (copy_to_user((void __user *)
				 (unsigned long) ti->tidlist,
				 tidlist, cnt * sizeof(*tidlist))) {
			ret = -EFAULT;
			goto cleanup;
		}
		if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
				 tidmap, sizeof tidmap)) {
			ret = -EFAULT;
			goto cleanup;
		}
		if (tid == tidcnt)
			tid = 0;
		pd->port_tidcursor = tid;
	}

done:
	if (ret)
		ipath_dbg("Failed to map %u TID pages, failing with %d\n",
			  ti->tidcnt, -ret);
	return ret;
}

/**
 * ipath_tid_free - free a port TID
 * @pd: the port
 * @ti: the TID info
 *
 * right now we are unlocking one page at a time, but since
 * the intended use of this routine is for a single group of
 * virtually contiguous pages, that should change to improve
 * performance.  We check that the TID is in range for this port
 * but otherwise don't check validity; if user has an error and
 * frees the wrong tid, it's only their own data that can thereby
 * be corrupted.  We do check that the TID was in use, for sanity
 * We always use our idea of the saved address, not the address that
 * they pass in to us.
 */

static int ipath_tid_free(struct ipath_portdata *pd,
			  const struct ipath_tid_info *ti)
{
	int ret = 0;
	u32 tid, porttid, cnt, limit, tidcnt;
	struct ipath_devdata *dd = pd->port_dd;
	u64 __iomem *tidbase;
	unsigned long tidmap[8];

	if (!dd->ipath_pageshadow) {
		ret = -ENOMEM;
		goto done;
	}

	if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
			   sizeof tidmap)) {
		ret = -EFAULT;
		goto done;
	}

	porttid = pd->port_port * dd->ipath_rcvtidcnt;
	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 */
		limit = tidcnt;
	tid = find_first_bit(tidmap, limit);
	ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
		   "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
		   limit, tid, porttid);
	for (cnt = 0; tid < limit; tid++) {
		/*
		 * small optimization; if we detect a run of 3 or so without
		 * any set, use find_first_bit again.  That's mainly to
		 * accelerate the case where we wrapped, so we have some at
		 * the beginning, and some at the end, and a big gap
		 * in the middle.
		 */
		if (!test_bit(tid, tidmap))
			continue;
		cnt++;
		if (dd->ipath_pageshadow[porttid + tid]) {
			ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
				   pd->port_pid, tid);
			dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
					    dd->ipath_tidinvalid);
			ipath_release_user_pages(
				&dd->ipath_pageshadow[porttid + tid], 1);
			dd->ipath_pageshadow[porttid + tid] = NULL;
			ipath_stats.sps_pageunlocks++;
		} else
			ipath_dbg("Unused tid %u, ignoring\n", tid);
	}
	if (cnt != ti->tidcnt)
		ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
			  ti->tidcnt, cnt);
done:
	if (ret)
		ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
			  ti->tidcnt, -ret);
	return ret;
}

/**
 * ipath_set_part_key - set a partition key
 * @pd: the port
 * @key: the key
 *
 * We can have up to 4 active at a time (other than the default, which is
 * always allowed).  This is somewhat tricky, since multiple ports may set
 * the same key, so we reference count them, and clean up at exit.  All 4
 * partition keys are packed into a single infinipath register.  It's an
 * error for a process to set the same pkey multiple times.  We provide no
 * mechanism to de-allocate a pkey at this time, we may eventually need to
 * do that.  I've used the atomic operations, and no locking, and only make
 * a single pass through what's available.  This should be more than
 * adequate for some time. I'll think about spinlocks or the like if and as
 * it's necessary.
 */
static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
{
	struct ipath_devdata *dd = pd->port_dd;
	int i, any = 0, pidx = -1;
	u16 lkey = key & 0x7FFF;
	int ret;

	if (lkey == (IPS_DEFAULT_P_KEY & 0x7FFF)) {
		/* nothing to do; this key always valid */
		ret = 0;
		goto bail;
	}

	ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
		   "%hx:%x %hx:%x %hx:%x %hx:%x\n",
		   pd->port_port, key, dd->ipath_pkeys[0],
		   atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
		   atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
		   atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
		   atomic_read(&dd->ipath_pkeyrefs[3]));

	if (!lkey) {
		ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
			   pd->port_port);
		ret = -EINVAL;
		goto bail;
	}