8def14e24c3758ed95571dd89c05d3d58586c768
[~shefty/rdma-dev.git] / 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/ioprio.h>
25 #include <linux/blkdev.h>
26 #include <linux/capability.h>
27 #include <linux/syscalls.h>
28 #include <linux/security.h>
29 #include <linux/pid_namespace.h>
30
31 int set_task_ioprio(struct task_struct *task, int ioprio)
32 {
33         int err;
34         struct io_context *ioc;
35         const struct cred *cred = current_cred(), *tcred;
36
37         rcu_read_lock();
38         tcred = __task_cred(task);
39         if (tcred->uid != cred->euid &&
40             tcred->uid != cred->uid && !capable(CAP_SYS_NICE)) {
41                 rcu_read_unlock();
42                 return -EPERM;
43         }
44         rcu_read_unlock();
45
46         err = security_task_setioprio(task, ioprio);
47         if (err)
48                 return err;
49
50         task_lock(task);
51         do {
52                 ioc = task->io_context;
53                 /* see wmb() in current_io_context() */
54                 smp_read_barrier_depends();
55                 if (ioc)
56                         break;
57
58                 ioc = alloc_io_context(GFP_ATOMIC, -1);
59                 if (!ioc) {
60                         err = -ENOMEM;
61                         break;
62                 }
63                 task->io_context = ioc;
64         } while (1);
65
66         if (!err) {
67                 ioc->ioprio = ioprio;
68                 ioc->ioprio_changed = 1;
69         }
70
71         task_unlock(task);
72         return err;
73 }
74 EXPORT_SYMBOL_GPL(set_task_ioprio);
75
76 SYSCALL_DEFINE3(ioprio_set, int, which, int, who, int, ioprio)
77 {
78         int class = IOPRIO_PRIO_CLASS(ioprio);
79         int data = IOPRIO_PRIO_DATA(ioprio);
80         struct task_struct *p, *g;
81         struct user_struct *user;
82         struct pid *pgrp;
83         int ret;
84
85         switch (class) {
86                 case IOPRIO_CLASS_RT:
87                         if (!capable(CAP_SYS_ADMIN))
88                                 return -EPERM;
89                         /* fall through, rt has prio field too */
90                 case IOPRIO_CLASS_BE:
91                         if (data >= IOPRIO_BE_NR || data < 0)
92                                 return -EINVAL;
93
94                         break;
95                 case IOPRIO_CLASS_IDLE:
96                         break;
97                 case IOPRIO_CLASS_NONE:
98                         if (data)
99                                 return -EINVAL;
100                         break;
101                 default:
102                         return -EINVAL;
103         }
104
105         ret = -ESRCH;
106         /*
107          * We want IOPRIO_WHO_PGRP/IOPRIO_WHO_USER to be "atomic",
108          * so we can't use rcu_read_lock(). See re-copy of ->ioprio
109          * in copy_process().
110          */
111         read_lock(&tasklist_lock);
112         switch (which) {
113                 case IOPRIO_WHO_PROCESS:
114                         if (!who)
115                                 p = current;
116                         else
117                                 p = find_task_by_vpid(who);
118                         if (p)
119                                 ret = set_task_ioprio(p, ioprio);
120                         break;
121                 case IOPRIO_WHO_PGRP:
122                         if (!who)
123                                 pgrp = task_pgrp(current);
124                         else
125                                 pgrp = find_vpid(who);
126                         do_each_pid_thread(pgrp, PIDTYPE_PGID, p) {
127                                 ret = set_task_ioprio(p, ioprio);
128                                 if (ret)
129                                         break;
130                         } while_each_pid_thread(pgrp, PIDTYPE_PGID, p);
131                         break;
132                 case IOPRIO_WHO_USER:
133                         if (!who)
134                                 user = current_user();
135                         else
136                                 user = find_user(who);
137
138                         if (!user)
139                                 break;
140
141                         do_each_thread(g, p) {
142                                 int match;
143
144                                 rcu_read_lock();
145                                 match = __task_cred(p)->uid == who;
146                                 rcu_read_unlock();
147                                 if (!match)
148                                         continue;
149                                 ret = set_task_ioprio(p, ioprio);
150                                 if (ret)
151                                         goto free_uid;
152                         } while_each_thread(g, p);
153 free_uid:
154                         if (who)
155                                 free_uid(user);
156                         break;
157                 default:
158                         ret = -EINVAL;
159         }
160
161         read_unlock(&tasklist_lock);
162         return ret;
163 }
164
165 static int get_task_ioprio(struct task_struct *p)
166 {
167         int ret;
168
169         ret = security_task_getioprio(p);
170         if (ret)
171                 goto out;
172         ret = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, IOPRIO_NORM);
173         if (p->io_context)
174                 ret = p->io_context->ioprio;
175 out:
176         return ret;
177 }
178
179 int ioprio_best(unsigned short aprio, unsigned short bprio)
180 {
181         unsigned short aclass = IOPRIO_PRIO_CLASS(aprio);
182         unsigned short bclass = IOPRIO_PRIO_CLASS(bprio);
183
184         if (aclass == IOPRIO_CLASS_NONE)
185                 aclass = IOPRIO_CLASS_BE;
186         if (bclass == IOPRIO_CLASS_NONE)
187                 bclass = IOPRIO_CLASS_BE;
188
189         if (aclass == bclass)
190                 return min(aprio, bprio);
191         if (aclass > bclass)
192                 return bprio;
193         else
194                 return aprio;
195 }
196
197 SYSCALL_DEFINE2(ioprio_get, int, which, int, who)
198 {
199         struct task_struct *g, *p;
200         struct user_struct *user;
201         struct pid *pgrp;
202         int ret = -ESRCH;
203         int tmpio;
204
205         read_lock(&tasklist_lock);
206         switch (which) {
207                 case IOPRIO_WHO_PROCESS:
208                         if (!who)
209                                 p = current;
210                         else
211                                 p = find_task_by_vpid(who);
212                         if (p)
213                                 ret = get_task_ioprio(p);
214                         break;
215                 case IOPRIO_WHO_PGRP:
216                         if (!who)
217                                 pgrp = task_pgrp(current);
218                         else
219                                 pgrp = find_vpid(who);
220                         do_each_pid_thread(pgrp, PIDTYPE_PGID, p) {
221                                 tmpio = get_task_ioprio(p);
222                                 if (tmpio < 0)
223                                         continue;
224                                 if (ret == -ESRCH)
225                                         ret = tmpio;
226                                 else
227                                         ret = ioprio_best(ret, tmpio);
228                         } while_each_pid_thread(pgrp, PIDTYPE_PGID, p);
229                         break;
230                 case IOPRIO_WHO_USER:
231                         if (!who)
232                                 user = current_user();
233                         else
234                                 user = find_user(who);
235
236                         if (!user)
237                                 break;
238
239                         do_each_thread(g, p) {
240                                 int match;
241
242                                 rcu_read_lock();
243                                 match = __task_cred(p)->uid == user->uid;
244                                 rcu_read_unlock();
245                                 if (!match)
246                                         continue;
247                                 tmpio = get_task_ioprio(p);
248                                 if (tmpio < 0)
249                                         continue;
250                                 if (ret == -ESRCH)
251                                         ret = tmpio;
252                                 else
253                                         ret = ioprio_best(ret, tmpio);
254                         } while_each_thread(g, p);
255
256                         if (who)
257                                 free_uid(user);
258                         break;
259                 default:
260                         ret = -EINVAL;
261         }
262
263         read_unlock(&tasklist_lock);
264         return ret;
265 }