Whamcloud - gitweb
LU-14428 libcfs: discard cfs_trace_copyin_string()
[fs/lustre-release.git] / libcfs / libcfs / module.c
index 767bc07..d7c1bcf 100644 (file)
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
  * Lustre is a trademark of Sun Microsystems, Inc.
  */
+#include <linux/miscdevice.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/string.h>
+#include <linux/stat.h>
+#include <linux/errno.h>
+#include <linux/unistd.h>
+#include <net/sock.h>
+#include <linux/uio.h>
+#include <linux/uaccess.h>
+
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/list.h>
+
+#include <linux/sysctl.h>
+#include <linux/debugfs.h>
+#include <asm/div64.h>
 
 #define DEBUG_SUBSYSTEM S_LNET
 
 #include <libcfs/libcfs.h>
 #include <libcfs/libcfs_crypto.h>
 #include <lnet/lib-lnet.h>
-#include <lnet/lib-dlc.h>
-#include <lnet/lnet.h>
-#include <lnet/nidstr.h>
+#include <libcfs/crypto/llcrypt.h>
+#include "tracefile.h"
 
-static void
-kportal_memhog_free (struct libcfs_device_userstate *ldu)
-{
-       struct page **level0p = &ldu->ldu_memhog_root_page;
-       struct page **level1p;
-       struct page **level2p;
-       int           count1;
-       int           count2;
-
-       if (*level0p != NULL) {
-               level1p = (struct page **)page_address(*level0p);
-               count1 = 0;
-
-               while (count1 < PAGE_CACHE_SIZE/sizeof(struct page *) &&
-                      *level1p != NULL) {
-
-                       level2p = (struct page **)page_address(*level1p);
-                       count2 = 0;
-
-                       while (count2 < PAGE_CACHE_SIZE/sizeof(struct page *) &&
-                              *level2p != NULL) {
-
-                               __free_page(*level2p);
-                               ldu->ldu_memhog_pages--;
-                               level2p++;
-                               count2++;
-                       }
-
-                       __free_page(*level1p);
-                       ldu->ldu_memhog_pages--;
-                       level1p++;
-                       count1++;
-               }
+static struct dentry *lnet_debugfs_root;
 
-               __free_page(*level0p);
-               ldu->ldu_memhog_pages--;
+BLOCKING_NOTIFIER_HEAD(libcfs_ioctl_list);
+EXPORT_SYMBOL(libcfs_ioctl_list);
 
-               *level0p = NULL;
-       }
+static inline size_t libcfs_ioctl_packlen(struct libcfs_ioctl_data *data)
+{
+       size_t len = sizeof(*data);
 
-       LASSERT(ldu->ldu_memhog_pages == 0);
+       len += (data->ioc_inllen1 + 7) & ~7;
+       len += (data->ioc_inllen2 + 7) & ~7;
+       return len;
 }
 
-static int
-kportal_memhog_alloc(struct libcfs_device_userstate *ldu, int npages,
-                    gfp_t flags)
+static bool libcfs_ioctl_is_invalid(struct libcfs_ioctl_data *data)
 {
-       struct page **level0p;
-       struct page **level1p;
-       struct page **level2p;
-       int           count1;
-       int           count2;
-
-       LASSERT(ldu->ldu_memhog_pages == 0);
-       LASSERT(ldu->ldu_memhog_root_page == NULL);
-
-       if (npages < 0)
-               return -EINVAL;
-
-       if (npages == 0)
-               return 0;
+       const int maxlen = 1 << 30;
+       if (data->ioc_hdr.ioc_len > maxlen)
+               return true;
 
-       level0p = &ldu->ldu_memhog_root_page;
-       *level0p = alloc_page(flags);
-       if (*level0p == NULL)
-               return -ENOMEM;
-       ldu->ldu_memhog_pages++;
+       if (data->ioc_inllen1 > maxlen)
+               return true;
 
-       level1p = (struct page **)page_address(*level0p);
-       count1 = 0;
-       memset(level1p, 0, PAGE_CACHE_SIZE);
+       if (data->ioc_inllen2 > maxlen)
+               return true;
 
-       while (ldu->ldu_memhog_pages < npages &&
-              count1 < PAGE_CACHE_SIZE/sizeof(struct page *)) {
+       if (data->ioc_inlbuf1 && !data->ioc_inllen1)
+               return true;
 
-               if (cfs_signal_pending())
-                       return -EINTR;
+       if (data->ioc_inlbuf2 && !data->ioc_inllen2)
+               return true;
 
-               *level1p = alloc_page(flags);
-               if (*level1p == NULL)
-                       return -ENOMEM;
-               ldu->ldu_memhog_pages++;
+       if (data->ioc_pbuf1 && !data->ioc_plen1)
+               return true;
 
-               level2p = (struct page **)page_address(*level1p);
-               count2 = 0;
-               memset(level2p, 0, PAGE_CACHE_SIZE);
+       if (data->ioc_pbuf2 && !data->ioc_plen2)
+               return true;
 
-               while (ldu->ldu_memhog_pages < npages &&
-                      count2 < PAGE_CACHE_SIZE/sizeof(struct page *)) {
+       if (data->ioc_plen1 && !data->ioc_pbuf1)
+               return true;
 
-                       if (cfs_signal_pending())
-                               return -EINTR;
+       if (data->ioc_plen2 && !data->ioc_pbuf2)
+               return true;
 
-                       *level2p = alloc_page(flags);
-                       if (*level2p == NULL)
-                               return -ENOMEM;
-                       ldu->ldu_memhog_pages++;
+       if (libcfs_ioctl_packlen(data) != data->ioc_hdr.ioc_len)
+               return true;
 
-                       level2p++;
-                       count2++;
-               }
-
-               level1p++;
-               count1++;
-       }
+       if (data->ioc_inllen1 &&
+               data->ioc_bulk[((data->ioc_inllen1 + 7) & ~7) +
+                              data->ioc_inllen2 - 1] != '\0')
+               return true;
 
-       return 0;
+       return false;
 }
 
-/* called when opening /dev/device */
-static int libcfs_psdev_open(unsigned long flags, void *args)
+int libcfs_ioctl_data_adjust(struct libcfs_ioctl_data *data)
 {
-       struct libcfs_device_userstate *ldu;
        ENTRY;
 
-       try_module_get(THIS_MODULE);
-
-       LIBCFS_ALLOC(ldu, sizeof(*ldu));
-       if (ldu != NULL) {
-               ldu->ldu_memhog_pages = 0;
-               ldu->ldu_memhog_root_page = NULL;
+       if (libcfs_ioctl_is_invalid(data)) {
+               CERROR("libcfs ioctl: parameter not correctly formatted\n");
+               RETURN(-EINVAL);
        }
-       *(struct libcfs_device_userstate **)args = ldu;
+
+       if (data->ioc_inllen1 != 0)
+               data->ioc_inlbuf1 = &data->ioc_bulk[0];
+
+       if (data->ioc_inllen2 != 0)
+               data->ioc_inlbuf2 = &data->ioc_bulk[0] +
+                                   cfs_size_round(data->ioc_inllen1);
 
        RETURN(0);
 }
 
-/* called when closing /dev/device */
-static int libcfs_psdev_release(unsigned long flags, void *args)
+int libcfs_ioctl_getdata(struct libcfs_ioctl_hdr **hdr_pp,
+                        struct libcfs_ioctl_hdr __user *uhdr)
 {
-       struct libcfs_device_userstate *ldu;
-       ENTRY;
+       struct libcfs_ioctl_hdr hdr;
+       int err;
 
-       ldu = (struct libcfs_device_userstate *)args;
-       if (ldu != NULL) {
-               kportal_memhog_free(ldu);
-               LIBCFS_FREE(ldu, sizeof(*ldu));
+       ENTRY;
+       if (copy_from_user(&hdr, uhdr, sizeof(hdr)))
+               RETURN(-EFAULT);
+
+       if (hdr.ioc_version != LIBCFS_IOCTL_VERSION &&
+           hdr.ioc_version != LIBCFS_IOCTL_VERSION2) {
+               CERROR("libcfs ioctl: version mismatch expected %#x, got %#x\n",
+                      LIBCFS_IOCTL_VERSION, hdr.ioc_version);
+               RETURN(-EINVAL);
        }
 
-       module_put(THIS_MODULE);
-       RETURN(0);
-}
-
-static DECLARE_RWSEM(ioctl_list_sem);
-static LIST_HEAD(ioctl_list);
+       if (hdr.ioc_len < sizeof(struct libcfs_ioctl_hdr)) {
+               CERROR("libcfs ioctl: user buffer too small for ioctl\n");
+               RETURN(-EINVAL);
+       }
 
-int libcfs_register_ioctl(struct libcfs_ioctl_handler *hand)
-{
-       int rc = 0;
+       if (hdr.ioc_len > LIBCFS_IOC_DATA_MAX) {
+               CERROR("libcfs ioctl: user buffer is too large %d/%d\n",
+                      hdr.ioc_len, LIBCFS_IOC_DATA_MAX);
+               RETURN(-EINVAL);
+       }
 
-       down_write(&ioctl_list_sem);
-       if (!list_empty(&hand->item))
-               rc = -EBUSY;
-       else
-               list_add_tail(&hand->item, &ioctl_list);
-       up_write(&ioctl_list_sem);
+       LIBCFS_ALLOC(*hdr_pp, hdr.ioc_len);
+       if (*hdr_pp == NULL)
+               RETURN(-ENOMEM);
 
-       return rc;
-}
-EXPORT_SYMBOL(libcfs_register_ioctl);
+       if (copy_from_user(*hdr_pp, uhdr, hdr.ioc_len))
+               GOTO(free, err = -EFAULT);
 
-int libcfs_deregister_ioctl(struct libcfs_ioctl_handler *hand)
-{
-       int rc = 0;
+       if ((*hdr_pp)->ioc_version != hdr.ioc_version ||
+               (*hdr_pp)->ioc_len != hdr.ioc_len) {
+               GOTO(free, err = -EINVAL);
+       }
 
-       down_write(&ioctl_list_sem);
-       if (list_empty(&hand->item))
-               rc = -ENOENT;
-       else
-               list_del_init(&hand->item);
-       up_write(&ioctl_list_sem);
+       RETURN(0);
 
-       return rc;
+free:
+       LIBCFS_FREE(*hdr_pp, hdr.ioc_len);
+       RETURN(err);
 }
-EXPORT_SYMBOL(libcfs_deregister_ioctl);
 
-static int libcfs_ioctl(struct cfs_psdev_file *pfile,
-                       unsigned long cmd, void __user *uparam)
+static int libcfs_ioctl(unsigned long cmd, void __user *uparam)
 {
        struct libcfs_ioctl_data *data = NULL;
        struct libcfs_ioctl_hdr  *hdr;
@@ -251,10 +209,6 @@ static int libcfs_ioctl(struct cfs_psdev_file *pfile,
        case IOC_LIBCFS_CLEAR_DEBUG:
                libcfs_debug_clear_buffer();
                break;
-       /*
-        * case IOC_LIBCFS_PANIC:
-        * Handled in arch/cfs_module.c
-        */
        case IOC_LIBCFS_MARK_DEBUG:
                if (data == NULL ||
                    data->ioc_inlbuf1 == NULL ||
@@ -264,66 +218,653 @@ static int libcfs_ioctl(struct cfs_psdev_file *pfile,
                libcfs_debug_mark_buffer(data->ioc_inlbuf1);
                break;
 
-       case IOC_LIBCFS_MEMHOG:
-               if (data == NULL)
-                       GOTO(out, err = -EINVAL);
+       default:
+               err = blocking_notifier_call_chain(&libcfs_ioctl_list,
+                                                  cmd, hdr);
+               if (!(err & NOTIFY_STOP_MASK))
+                       /* No-one claimed the ioctl */
+                       err = -EINVAL;
+               else
+                       err = notifier_to_errno(err);
+               if (copy_to_user(uparam, hdr, hdr->ioc_len) && !err)
+                       err = -EFAULT;
+               break;
+       }
+out:
+       LIBCFS_FREE(hdr, hdr->ioc_len);
+       RETURN(err);
+}
 
-               if (pfile->private_data == NULL)
-                       GOTO(out, err = -EINVAL);
+static long
+libcfs_psdev_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+       if (!capable(CAP_SYS_ADMIN))
+               return -EACCES;
+
+       if (_IOC_TYPE(cmd) != IOC_LIBCFS_TYPE ||
+           _IOC_NR(cmd) < IOC_LIBCFS_MIN_NR  ||
+           _IOC_NR(cmd) > IOC_LIBCFS_MAX_NR) {
+               CDEBUG(D_IOCTL, "invalid ioctl ( type %d, nr %d, size %d )\n",
+                      _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
+               return -EINVAL;
+       }
 
-               kportal_memhog_free(pfile->private_data);
-               err = kportal_memhog_alloc(pfile->private_data,
-                                          data->ioc_count, data->ioc_flags);
-               if (err != 0)
-                       kportal_memhog_free(pfile->private_data);
-               break;
+       return libcfs_ioctl(cmd, (void __user *)arg);
+}
+
+static const struct file_operations libcfs_fops = {
+       .owner                  = THIS_MODULE,
+       .unlocked_ioctl         = libcfs_psdev_ioctl,
+};
 
-       default: {
-               struct libcfs_ioctl_handler *hand;
+static struct miscdevice libcfs_dev = {
+       .minor                  = MISC_DYNAMIC_MINOR,
+       .name                   = "lnet",
+       .fops                   = &libcfs_fops,
+};
+
+int lprocfs_call_handler(void *data, int write, loff_t *ppos,
+                        void __user *buffer, size_t *lenp,
+                        int (*handler)(void *data, int write, loff_t pos,
+                                       void __user *buffer, int len))
+{
+       int rc = handler(data, write, *ppos, buffer, *lenp);
+
+       if (rc < 0)
+               return rc;
 
-               err = -EINVAL;
-               down_read(&ioctl_list_sem);
-               list_for_each_entry(hand, &ioctl_list, item) {
-                       err = hand->handle_ioctl(cmd, hdr);
-                       if (err == -EINVAL)
-                               continue;
+       if (write) {
+               *ppos += *lenp;
+       } else {
+               *lenp = rc;
+               *ppos += rc;
+       }
+       return 0;
+}
+EXPORT_SYMBOL(lprocfs_call_handler);
 
-                       if (err == 0)
-                               err = libcfs_ioctl_popdata(hdr, uparam);
+static int __proc_dobitmasks(void *data, int write,
+                            loff_t pos, void __user *buffer, int nob)
+{
+       const int     tmpstrlen = 512;
+       char         *tmpstr = NULL;
+       int           rc;
+       unsigned int *mask = data;
+       int           is_subsys = (mask == &libcfs_subsystem_debug) ? 1 : 0;
+       int           is_printk = (mask == &libcfs_printk) ? 1 : 0;
+
+       if (!write) {
+               rc = cfs_trace_allocate_string_buffer(&tmpstr, tmpstrlen);
+               if (rc < 0)
+                       return rc;
+               libcfs_debug_mask2str(tmpstr, tmpstrlen, *mask, is_subsys);
+               rc = strlen(tmpstr);
+
+               if (pos >= rc) {
+                       rc = 0;
+               } else {
+                       rc = cfs_trace_copyout_string(buffer, nob,
+                                                     tmpstr + pos, "\n");
+               }
+       } else {
+               tmpstr = memdup_user_nul(buffer, nob);
+               if (!tmpstr)
+                       return -ENOMEM;
+
+               rc = libcfs_debug_str2mask(mask, strim(tmpstr), is_subsys);
+               /* Always print LBUG/LASSERT to console, so keep this mask */
+               if (is_printk)
+                       *mask |= D_EMERG;
+       }
+
+       kfree(tmpstr);
+       return rc;
+}
+
+static int proc_dobitmasks(struct ctl_table *table, int write,
+                          void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_dobitmasks);
+}
+
+static int min_watchdog_ratelimit;             /* disable ratelimiting */
+static int max_watchdog_ratelimit = (24*60*60); /* limit to once per day */
+
+static int __proc_dump_kernel(void *data, int write,
+                             loff_t pos, void __user *buffer, int nob)
+{
+       if (!write)
+               return 0;
+
+       return cfs_trace_dump_debug_buffer_usrstr(buffer, nob);
+}
+
+static int proc_dump_kernel(struct ctl_table *table, int write,
+                           void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_dump_kernel);
+}
+
+static int __proc_daemon_file(void *data, int write,
+                             loff_t pos, void __user *buffer, int nob)
+{
+       if (!write) {
+               int len = strlen(cfs_tracefile);
+
+               if (pos >= len)
+                       return 0;
+
+               return cfs_trace_copyout_string(buffer, nob,
+                                               cfs_tracefile + pos, "\n");
+       }
+
+       return cfs_trace_daemon_command_usrstr(buffer, nob);
+}
+
+static int proc_daemon_file(struct ctl_table *table, int write,
+                           void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_daemon_file);
+}
+
+static int libcfs_force_lbug(struct ctl_table *table, int write,
+                            void __user *buffer,
+                            size_t *lenp, loff_t *ppos)
+{
+       if (write)
+               LBUG();
+       return 0;
+}
+
+static int proc_fail_loc(struct ctl_table *table, int write,
+                        void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       int rc;
+       long old_fail_loc = cfs_fail_loc;
+
+       if (!*lenp || *ppos) {
+               *lenp = 0;
+               return 0;
+       }
+
+       if (write) {
+               char *kbuf = memdup_user_nul(buffer, *lenp);
+
+               if (IS_ERR(kbuf))
+                       return PTR_ERR(kbuf);
+               rc = kstrtoul(kbuf, 0, &cfs_fail_loc);
+               kfree(kbuf);
+               *ppos += *lenp;
+       } else {
+               char kbuf[64/3+3];
+
+               rc = scnprintf(kbuf, sizeof(kbuf), "%lu\n", cfs_fail_loc);
+               if (copy_to_user(buffer, kbuf, rc))
+                       rc = -EFAULT;
+               else {
+                       *lenp = rc;
+                       *ppos += rc;
+               }
+       }
+
+       if (old_fail_loc != cfs_fail_loc) {
+               cfs_race_state = 1;
+               wake_up(&cfs_race_waitq);
+       }
+       return rc;
+}
+
+int debugfs_doint(struct ctl_table *table, int write,
+                 void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       int rc;
+
+       if (!*lenp || *ppos) {
+               *lenp = 0;
+               return 0;
+       }
+
+       if (write) {
+               char *kbuf = memdup_user_nul(buffer, *lenp);
+               int val;
+
+               if (IS_ERR(kbuf))
+                       return PTR_ERR(kbuf);
+
+               rc = kstrtoint(kbuf, 0, &val);
+               kfree(kbuf);
+               if (!rc) {
+                       if (table->extra1 && val < *(int *)table->extra1)
+                               val = *(int *)table->extra1;
+                       if (table->extra2 && val > *(int *)table->extra2)
+                               val = *(int *)table->extra2;
+                       *(int *)table->data = val;
+               }
+               *ppos += *lenp;
+       } else {
+               char kbuf[64/3+3];
+
+               rc = scnprintf(kbuf, sizeof(kbuf), "%u\n", *(int *)table->data);
+               if (copy_to_user(buffer, kbuf, rc))
+                       rc = -EFAULT;
+               else {
+                       *lenp = rc;
+                       *ppos += rc;
+               }
+       }
+
+       return rc;
+}
+EXPORT_SYMBOL(debugfs_doint);
+
+static int debugfs_dou64(struct ctl_table *table, int write,
+                        void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       int rc;
+
+       if (!*lenp || *ppos) {
+               *lenp = 0;
+               return 0;
+       }
+
+       if (write) {
+               char *kbuf = memdup_user_nul(buffer, *lenp);
+               unsigned long long val;
+
+               if (IS_ERR(kbuf))
+                       return PTR_ERR(kbuf);
+
+               rc = kstrtoull(kbuf, 0, &val);
+               kfree(kbuf);
+               if (!rc)
+                       *(u64 *)table->data = val;
+               *ppos += *lenp;
+       } else {
+               char kbuf[64/3+3];
+
+               rc = scnprintf(kbuf, sizeof(kbuf), "%llu\n",
+                              (unsigned long long)*(u64 *)table->data);
+               if (copy_to_user(buffer, kbuf, rc))
+                       rc = -EFAULT;
+               else {
+                       *lenp = rc;
+                       *ppos += rc;
+               }
+       }
+
+       return rc;
+}
+
+static int debugfs_dostring(struct ctl_table *table, int write,
+                           void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       int len = *lenp;
+       char *kbuf = table->data;
+
+       if (!len || *ppos) {
+               *lenp = 0;
+               return 0;
+       }
+       if (len > table->maxlen)
+               len = table->maxlen;
+       if (write) {
+               if (copy_from_user(kbuf, buffer, len))
+                       return -EFAULT;
+               memset(kbuf+len, 0, table->maxlen - len);
+               *ppos = *lenp;
+       } else {
+               len = strnlen(kbuf, len);
+               if (copy_to_user(buffer, kbuf, len))
+                       return -EFAULT;
+               if (len < *lenp) {
+                       if (copy_to_user(buffer+len, "\n", 1))
+                               return -EFAULT;
+                       len += 1;
+               }
+               *ppos += len;
+               *lenp -= len;
+       }
+       return len;
+}
+
+static int __proc_cpt_table(void *data, int write,
+                           loff_t pos, void __user *buffer, int nob)
+{
+       char *buf = NULL;
+       int   len = 4096;
+       int   rc  = 0;
+
+       if (write)
+               return -EPERM;
+
+       while (1) {
+               LIBCFS_ALLOC(buf, len);
+               if (buf == NULL)
+                       return -ENOMEM;
+
+               rc = cfs_cpt_table_print(cfs_cpt_tab, buf, len);
+               if (rc >= 0)
                        break;
+
+               if (rc == -EFBIG) {
+                       LIBCFS_FREE(buf, len);
+                       len <<= 1;
+                       continue;
                }
-               up_read(&ioctl_list_sem);
-               break; }
+               goto out;
        }
+
+       if (pos >= rc) {
+               rc = 0;
+               goto out;
+       }
+
+       rc = cfs_trace_copyout_string(buffer, nob, buf + pos, NULL);
 out:
-       libcfs_ioctl_freedata(hdr);
-       RETURN(err);
+       if (buf != NULL)
+               LIBCFS_FREE(buf, len);
+       return rc;
+}
+
+static int proc_cpt_table(struct ctl_table *table, int write,
+                         void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_cpt_table);
+}
+
+static int __proc_cpt_distance(void *data, int write,
+                              loff_t pos, void __user *buffer, int nob)
+{
+       char *buf = NULL;
+       int   len = 4096;
+       int   rc  = 0;
+
+       if (write)
+               return -EPERM;
+
+       while (1) {
+               LIBCFS_ALLOC(buf, len);
+               if (buf == NULL)
+                       return -ENOMEM;
+
+               rc = cfs_cpt_distance_print(cfs_cpt_tab, buf, len);
+               if (rc >= 0)
+                       break;
+
+               if (rc == -EFBIG) {
+                       LIBCFS_FREE(buf, len);
+                       len <<= 1;
+                       continue;
+               }
+               goto out;
+       }
+
+       if (pos >= rc) {
+               rc = 0;
+               goto out;
+       }
+
+       rc = cfs_trace_copyout_string(buffer, nob, buf + pos, NULL);
+ out:
+       if (buf != NULL)
+               LIBCFS_FREE(buf, len);
+       return rc;
 }
 
-struct cfs_psdev_ops libcfs_psdev_ops = {
-        libcfs_psdev_open,
-        libcfs_psdev_release,
-        NULL,
-        NULL,
-        libcfs_ioctl
+static int proc_cpt_distance(struct ctl_table *table, int write,
+                            void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                    __proc_cpt_distance);
+}
+
+static struct ctl_table lnet_table[] = {
+       {
+               .procname       = "debug",
+               .data           = &libcfs_debug,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = &proc_dobitmasks,
+       },
+       {
+               .procname       = "subsystem_debug",
+               .data           = &libcfs_subsystem_debug,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = &proc_dobitmasks,
+       },
+       {
+               .procname       = "printk",
+               .data           = &libcfs_printk,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = &proc_dobitmasks,
+       },
+       {
+               .procname       = "cpu_partition_table",
+               .maxlen         = 128,
+               .mode           = 0444,
+               .proc_handler   = &proc_cpt_table,
+       },
+       {
+               .procname       = "cpu_partition_distance",
+               .maxlen         = 128,
+               .mode           = 0444,
+               .proc_handler   = &proc_cpt_distance,
+       },
+       {
+               .procname       = "debug_log_upcall",
+               .data           = lnet_debug_log_upcall,
+               .maxlen         = sizeof(lnet_debug_log_upcall),
+               .mode           = 0644,
+               .proc_handler   = &debugfs_dostring,
+       },
+       {
+               .procname       = "lnet_memused",
+               .data           = (u64 *)&libcfs_kmem.counter,
+               .maxlen         = sizeof(u64),
+               .mode           = 0444,
+               .proc_handler   = &debugfs_dou64,
+       },
+       {
+               .procname       = "catastrophe",
+               .data           = &libcfs_catastrophe,
+               .maxlen         = sizeof(int),
+               .mode           = 0444,
+               .proc_handler   = &debugfs_doint,
+       },
+       {
+               .procname       = "dump_kernel",
+               .maxlen         = 256,
+               .mode           = 0200,
+               .proc_handler   = &proc_dump_kernel,
+       },
+       {
+               .procname       = "daemon_file",
+               .mode           = 0644,
+               .maxlen         = 256,
+               .proc_handler   = &proc_daemon_file,
+       },
+       {
+               .procname       = "watchdog_ratelimit",
+               .data           = &libcfs_watchdog_ratelimit,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = &debugfs_doint,
+               .extra1         = &min_watchdog_ratelimit,
+               .extra2         = &max_watchdog_ratelimit,
+       },
+       {
+               .procname       = "force_lbug",
+               .data           = NULL,
+               .maxlen         = 0,
+               .mode           = 0200,
+               .proc_handler   = &libcfs_force_lbug
+       },
+       {
+               .procname       = "fail_loc",
+               .data           = &cfs_fail_loc,
+               .maxlen         = sizeof(cfs_fail_loc),
+               .mode           = 0644,
+               .proc_handler   = &proc_fail_loc
+       },
+       {
+               .procname       = "fail_val",
+               .data           = &cfs_fail_val,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = &debugfs_doint
+       },
+       {
+               .procname       = "fail_err",
+               .data           = &cfs_fail_err,
+               .maxlen         = sizeof(cfs_fail_err),
+               .mode           = 0644,
+               .proc_handler   = &debugfs_doint,
+       },
+       {
+       }
 };
 
-MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
-MODULE_DESCRIPTION("Portals v3.1");
-MODULE_LICENSE("GPL");
+static const struct lnet_debugfs_symlink_def lnet_debugfs_symlinks[] = {
+       { .name         = "console_ratelimit",
+         .target       = "../../../module/libcfs/parameters/libcfs_console_ratelimit" },
+       { .name         = "debug_path",
+         .target       = "../../../module/libcfs/parameters/libcfs_debug_file_path" },
+       { .name         = "panic_on_lbug",
+         .target       = "../../../module/libcfs/parameters/libcfs_panic_on_lbug" },
+       { .name         = "console_backoff",
+         .target       = "../../../module/libcfs/parameters/libcfs_console_backoff" },
+       { .name         = "debug_mb",
+         .target       = "../../../module/libcfs/parameters/libcfs_debug_mb" },
+       { .name         = "console_min_delay_centisecs",
+         .target       = "../../../module/libcfs/parameters/libcfs_console_min_delay" },
+       { .name         = "console_max_delay_centisecs",
+         .target       = "../../../module/libcfs/parameters/libcfs_console_max_delay" },
+       { .name         = NULL },
+};
+
+static ssize_t lnet_debugfs_read(struct file *filp, char __user *buf,
+                                size_t count, loff_t *ppos)
+{
+       struct ctl_table *table = filp->private_data;
+       ssize_t rc = -EINVAL;
+
+       if (table) {
+               rc = table->proc_handler(table, 0, buf, &count, ppos);
+               if (!rc)
+                       rc = count;
+       }
+
+       return rc;
+}
+
+static ssize_t lnet_debugfs_write(struct file *filp, const char __user *buf,
+                                 size_t count, loff_t *ppos)
+{
+       struct ctl_table *table = filp->private_data;
+       ssize_t rc = -EINVAL;
+
+       if (table) {
+               rc = table->proc_handler(table, 1, (void __user *)buf, &count,
+                                        ppos);
+               if (!rc)
+                       rc = count;
+       }
+
+       return rc;
+}
+
+static const struct file_operations lnet_debugfs_file_operations_rw = {
+       .open           = simple_open,
+       .read           = lnet_debugfs_read,
+       .write          = lnet_debugfs_write,
+       .llseek         = default_llseek,
+};
 
-static int init_libcfs_module(void)
+static const struct file_operations lnet_debugfs_file_operations_ro = {
+       .open           = simple_open,
+       .read           = lnet_debugfs_read,
+       .llseek         = default_llseek,
+};
+
+static const struct file_operations lnet_debugfs_file_operations_wo = {
+       .open           = simple_open,
+       .write          = lnet_debugfs_write,
+       .llseek         = default_llseek,
+};
+
+static const struct file_operations *lnet_debugfs_fops_select(umode_t mode)
+{
+       if (!(mode & S_IWUGO))
+               return &lnet_debugfs_file_operations_ro;
+
+       if (!(mode & S_IRUGO))
+               return &lnet_debugfs_file_operations_wo;
+
+       return &lnet_debugfs_file_operations_rw;
+}
+
+void lnet_insert_debugfs(struct ctl_table *table)
+{
+       if (!lnet_debugfs_root)
+               lnet_debugfs_root = debugfs_create_dir("lnet", NULL);
+
+       /* Even if we cannot create, just ignore it altogether) */
+       if (IS_ERR_OR_NULL(lnet_debugfs_root))
+               return;
+
+       /* We don't save the dentry returned in next two calls, because
+        * we don't call debugfs_remove() but rather remove_recursive()
+        */
+       for (; table && table->procname; table++)
+               debugfs_create_file(table->procname, table->mode,
+                                   lnet_debugfs_root, table,
+                                   lnet_debugfs_fops_select(table->mode));
+}
+EXPORT_SYMBOL_GPL(lnet_insert_debugfs);
+
+static void lnet_insert_debugfs_links(
+               const struct lnet_debugfs_symlink_def *symlinks)
+{
+       for (; symlinks && symlinks->name; symlinks++)
+               debugfs_create_symlink(symlinks->name, lnet_debugfs_root,
+                                      symlinks->target);
+}
+
+void lnet_remove_debugfs(struct ctl_table *table)
+{
+       for (; table && table->procname; table++) {
+               struct qstr dname = QSTR_INIT(table->procname,
+                                             strlen(table->procname));
+               struct dentry *dentry;
+
+               dentry = d_hash_and_lookup(lnet_debugfs_root, &dname);
+               debugfs_remove(dentry);
+       }
+}
+EXPORT_SYMBOL_GPL(lnet_remove_debugfs);
+
+static int __init libcfs_init(void)
 {
        int rc;
 
-       libcfs_arch_init();
+       cfs_arch_init();
+
+       init_libcfs_vfree_atomic();
 
        rc = libcfs_debug_init(5 * 1024 * 1024);
        if (rc < 0) {
-               printk(KERN_ERR "LustreError: libcfs_debug_init: %d\n", rc);
+               pr_err("LustreError: libcfs_debug_init: rc = %d\n", rc);
                return (rc);
        }
 
+       cfs_debug_init();
+
        rc = cfs_cpu_init();
        if (rc != 0)
                goto cleanup_debug;
@@ -340,12 +881,11 @@ static int init_libcfs_module(void)
                goto cleanup_deregister;
        }
 
-       /* max to 4 threads, should be enough for rehash */
-       rc = min(cfs_cpt_weight(cfs_cpt_table, CFS_CPT_ANY), 4);
-       rc = cfs_wi_sched_create("cfs_rh", cfs_cpt_table, CFS_CPT_ANY,
-                                rc, &cfs_sched_rehash);
-       if (rc != 0) {
-               CERROR("Startup workitem scheduler: error: %d\n", rc);
+       cfs_rehash_wq = alloc_workqueue("cfs_rh", WQ_SYSFS, 4);
+       if (!cfs_rehash_wq) {
+               rc = -ENOMEM;
+               CERROR("libcfs: failed to start rehash workqueue: rc = %d\n",
+                      rc);
                goto cleanup_deregister;
        }
 
@@ -355,17 +895,18 @@ static int init_libcfs_module(void)
                goto cleanup_wi;
        }
 
+       lnet_insert_debugfs(lnet_table);
+       if (!IS_ERR_OR_NULL(lnet_debugfs_root))
+               lnet_insert_debugfs_links(lnet_debugfs_symlinks);
 
-       rc = insert_proc();
+       rc = llcrypt_init();
        if (rc) {
-               CERROR("insert_proc: error %d\n", rc);
-               goto cleanup_crypto;
+               CERROR("llcrypt_init: error %d\n", rc);
+               goto cleanup_wi;
        }
 
        CDEBUG (D_OTHER, "portals setup OK\n");
        return 0;
-cleanup_crypto:
-       cfs_crypto_unregister();
 cleanup_wi:
        cfs_wi_shutdown();
 cleanup_deregister:
@@ -377,39 +918,47 @@ cleanup_debug:
        return rc;
 }
 
-static void exit_libcfs_module(void)
+static void __exit libcfs_exit(void)
 {
        int rc;
 
-       remove_proc();
+       /* Remove everthing */
+       debugfs_remove_recursive(lnet_debugfs_root);
+       lnet_debugfs_root = NULL;
 
-       CDEBUG(D_MALLOC, "before Portals cleanup: kmem %d\n",
-              atomic_read(&libcfs_kmemory));
+       CDEBUG(D_MALLOC, "before Portals cleanup: kmem %lld\n",
+              libcfs_kmem_read());
 
-       if (cfs_sched_rehash != NULL) {
-               cfs_wi_sched_destroy(cfs_sched_rehash);
-               cfs_sched_rehash = NULL;
+       llcrypt_exit();
+
+       if (cfs_rehash_wq) {
+               destroy_workqueue(cfs_rehash_wq);
+               cfs_rehash_wq = NULL;
        }
 
        cfs_crypto_unregister();
        cfs_wi_shutdown();
 
-       rc = misc_deregister(&libcfs_dev);
-       if (rc)
-               CERROR("misc_deregister error %d\n", rc);
+       misc_deregister(&libcfs_dev);
 
        cfs_cpu_fini();
 
-       if (atomic_read(&libcfs_kmemory) != 0)
-               CERROR("Portals memory leaked: %d bytes\n",
-                      atomic_read(&libcfs_kmemory));
+       /* the below message is checked in test-framework.sh check_mem_leak() */
+       if (libcfs_kmem_read() != 0)
+               CERROR("Portals memory leaked: %lld bytes\n",
+                      libcfs_kmem_read());
 
        rc = libcfs_debug_cleanup();
        if (rc)
-               printk(KERN_ERR "LustreError: libcfs_debug_cleanup: %d\n",
-                      rc);
+               pr_err("LustreError: libcfs_debug_cleanup: rc = %d\n", rc);
 
-       libcfs_arch_cleanup();
+       exit_libcfs_vfree_atomic();
 }
 
-cfs_module(libcfs, "1.0.0", init_libcfs_module, exit_libcfs_module);
+MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
+MODULE_DESCRIPTION("Lustre helper library");
+MODULE_VERSION(LIBCFS_VERSION);
+MODULE_LICENSE("GPL");
+
+module_init(libcfs_init);
+module_exit(libcfs_exit);