fs/ioprio.c

   1 /*
   2  * fs/ioprio.c
   3  *
   4  * Copyright (C) 2004 Jens Axboe <axboe@kernel.dk>
   5  *
   6  * Helper functions for setting/querying io priorities of processes. The
   7  * system calls closely mimmick getpriority/setpriority, see the man page for
   8  * those. The prio argument is a composite of prio class and prio data, where
   9  * the data argument has meaning within that class. The standard scheduling
  10  * classes have 8 distinct prio levels, with 0 being the highest prio and 7
  11  * being the lowest.
  12  *
  13  * IOW, setting BE scheduling class with prio 2 is done ala:
  14  *
  15  * unsigned int prio = (IOPRIO_CLASS_BE << IOPRIO_CLASS_SHIFT) | 2;
  16  *
  17  * ioprio_set(PRIO_PROCESS, pid, prio);
  18  *
  19  * See also Documentation/block/ioprio.txt
  20  *
  21  */
  22 #include <linux/gfp.h>
  23 #include <linux/kernel.h>
  24 #include <linux/export.h>
  25 #include <linux/ioprio.h>
  26 #include <linux/blkdev.h>
  27 #include <linux/capability.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/security.h>
  30 #include <linux/pid_namespace.h>
  31
  32 int set_task_ioprio(struct task_struct *task, int ioprio)
  33 {
  34         int err;
  35         struct io_context *ioc;
  36         const struct cred *cred = current_cred(), *tcred;
  37
  38         rcu_read_lock();
  39         tcred = __task_cred(task);
  40         if (tcred->uid != cred->euid &&
  41             tcred->uid != cred->uid && !capable(CAP_SYS_NICE)) {
  42                 rcu_read_unlock();
  43                 return -EPERM;
  44         }
  45         rcu_read_unlock();
  46
  47         err = security_task_setioprio(task, ioprio);
  48         if (err)
  49                 return err;
  50
  51         task_lock(task);
  52         do {
  53                 ioc = task->io_context;
  54                 /* see wmb() in current_io_context() */
  55                 smp_read_barrier_depends();
  56                 if (ioc)
  57                         break;
  58
  59                 ioc = alloc_io_context(GFP_ATOMIC, -1);
  60                 if (!ioc) {
  61                         err = -ENOMEM;
  62                         break;
  63                 }
  64                 task->io_context = ioc;
  65         } while (1);
  66
  67         if (!err) {
  68                 ioc->ioprio = ioprio;
  69                 ioc->ioprio_changed = 1;
  70         }
  71
  72         task_unlock(task);
  73         return err;
  74 }
  75 EXPORT_SYMBOL_GPL(set_task_ioprio);
  76
  77 SYSCALL_DEFINE3(ioprio_set, int, which, int, who, int, ioprio)
  78 {
  79         int class = IOPRIO_PRIO_CLASS(ioprio);
  80         int data = IOPRIO_PRIO_DATA(ioprio);
  81         struct task_struct *p, *g;
  82         struct user_struct *user;
  83         struct pid *pgrp;
  84         int ret;
  85
  86         switch (class) {
  87                 case IOPRIO_CLASS_RT:
  88                         if (!capable(CAP_SYS_ADMIN))
  89                                 return -EPERM;
  90                         /* fall through, rt has prio field too */
  91                 case IOPRIO_CLASS_BE:
  92                         if (data >= IOPRIO_BE_NR || data < 0)
  93                                 return -EINVAL;
  94
  95                         break;
  96                 case IOPRIO_CLASS_IDLE:
  97                         break;
  98                 case IOPRIO_CLASS_NONE:
  99                         if (data)
 100                                 return -EINVAL;
 101                         break;
 102                 default:
 103                         return -EINVAL;
 104         }
 105
 106         ret = -ESRCH;
 107         rcu_read_lock();
 108         switch (which) {
 109                 case IOPRIO_WHO_PROCESS:
 110                         if (!who)
 111                                 p = current;
 112                         else
 113                                 p = find_task_by_vpid(who);
 114                         if (p)
 115                                 ret = set_task_ioprio(p, ioprio);
 116                         break;
 117                 case IOPRIO_WHO_PGRP:
 118                         if (!who)
 119                                 pgrp = task_pgrp(current);
 120                         else
 121                                 pgrp = find_vpid(who);
 122                         do_each_pid_thread(pgrp, PIDTYPE_PGID, p) {
 123                                 ret = set_task_ioprio(p, ioprio);
 124                                 if (ret)
 125                                         break;
 126                         } while_each_pid_thread(pgrp, PIDTYPE_PGID, p);
 127                         break;
 128                 case IOPRIO_WHO_USER:
 129                         if (!who)
 130                                 user = current_user();
 131                         else
 132                                 user = find_user(who);
 133
 134                         if (!user)
 135                                 break;
 136
 137                         do_each_thread(g, p) {
 138                                 if (__task_cred(p)->uid != who)
 139                                         continue;
 140                                 ret = set_task_ioprio(p, ioprio);
 141                                 if (ret)
 142                                         goto free_uid;
 143                         } while_each_thread(g, p);
 144 free_uid:
 145                         if (who)
 146                                 free_uid(user);
 147                         break;
 148                 default:
 149                         ret = -EINVAL;
 150         }
 151
 152         rcu_read_unlock();
 153         return ret;
 154 }
 155
 156 static int get_task_ioprio(struct task_struct *p)
 157 {
 158         int ret;
 159
 160         ret = security_task_getioprio(p);
 161         if (ret)
 162                 goto out;
 163         ret = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, IOPRIO_NORM);
 164         if (p->io_context)
 165                 ret = p->io_context->ioprio;
 166 out:
 167         return ret;
 168 }
 169
 170 int ioprio_best(unsigned short aprio, unsigned short bprio)
 171 {
 172         unsigned short aclass;
 173         unsigned short bclass;
 174
 175         if (!ioprio_valid(aprio))
 176                 aprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, IOPRIO_NORM);
 177         if (!ioprio_valid(bprio))
 178                 bprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, IOPRIO_NORM);
 179
 180         aclass = IOPRIO_PRIO_CLASS(aprio);
 181         bclass = IOPRIO_PRIO_CLASS(bprio);
 182         if (aclass == bclass)
 183                 return min(aprio, bprio);
 184         if (aclass > bclass)
 185                 return bprio;
 186         else
 187                 return aprio;
 188 }
 189
 190 SYSCALL_DEFINE2(ioprio_get, int, which, int, who)
 191 {
 192         struct task_struct *g, *p;
 193         struct user_struct *user;
 194         struct pid *pgrp;
 195         int ret = -ESRCH;
 196         int tmpio;
 197
 198         rcu_read_lock();
 199         switch (which) {
 200                 case IOPRIO_WHO_PROCESS:
 201                         if (!who)
 202                                 p = current;
 203                         else
 204                                 p = find_task_by_vpid(who);
 205                         if (p)
 206                                 ret = get_task_ioprio(p);
 207                         break;
 208                 case IOPRIO_WHO_PGRP:
 209                         if (!who)
 210                                 pgrp = task_pgrp(current);
 211                         else
 212                                 pgrp = find_vpid(who);
 213                         do_each_pid_thread(pgrp, PIDTYPE_PGID, p) {
 214                                 tmpio = get_task_ioprio(p);
 215                                 if (tmpio < 0)
 216                                         continue;
 217                                 if (ret == -ESRCH)
 218                                         ret = tmpio;
 219                                 else
 220                                         ret = ioprio_best(ret, tmpio);
 221                         } while_each_pid_thread(pgrp, PIDTYPE_PGID, p);
 222                         break;
 223                 case IOPRIO_WHO_USER:
 224                         if (!who)
 225                                 user = current_user();
 226                         else
 227                                 user = find_user(who);
 228
 229                         if (!user)
 230                                 break;
 231
 232                         do_each_thread(g, p) {
 233                                 if (__task_cred(p)->uid != user->uid)
 234                                         continue;
 235                                 tmpio = get_task_ioprio(p);
 236                                 if (tmpio < 0)
 237                                         continue;
 238                                 if (ret == -ESRCH)
 239                                         ret = tmpio;
 240                                 else
 241                                         ret = ioprio_best(ret, tmpio);
 242                         } while_each_thread(g, p);
 243
 244                         if (who)
 245                                 free_uid(user);
 246                         break;
 247                 default:
 248                         ret = -EINVAL;
 249         }
 250
 251         rcu_read_unlock();
 252         return ret;
 253 }