Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 1 | #include <linux/atomic.h> |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 2 | #include <linux/rwsem.h> |
| 3 | #include <linux/percpu.h> |
| 4 | #include <linux/wait.h> |
| 5 | #include <linux/percpu-rwsem.h> |
| 6 | #include <linux/rcupdate.h> |
| 7 | #include <linux/sched.h> |
| 8 | #include <linux/errno.h> |
| 9 | |
| 10 | int percpu_init_rwsem(struct percpu_rw_semaphore *brw) |
| 11 | { |
| 12 | brw->fast_read_ctr = alloc_percpu(int); |
| 13 | if (unlikely(!brw->fast_read_ctr)) |
| 14 | return -ENOMEM; |
| 15 | |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 16 | init_rwsem(&brw->rw_sem); |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 17 | atomic_set(&brw->write_ctr, 0); |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 18 | atomic_set(&brw->slow_read_ctr, 0); |
| 19 | init_waitqueue_head(&brw->write_waitq); |
| 20 | return 0; |
| 21 | } |
| 22 | |
| 23 | void percpu_free_rwsem(struct percpu_rw_semaphore *brw) |
| 24 | { |
| 25 | free_percpu(brw->fast_read_ctr); |
| 26 | brw->fast_read_ctr = NULL; /* catch use after free bugs */ |
| 27 | } |
| 28 | |
| 29 | /* |
| 30 | * This is the fast-path for down_read/up_read, it only needs to ensure |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 31 | * there is no pending writer (atomic_read(write_ctr) == 0) and inc/dec the |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 32 | * fast per-cpu counter. The writer uses synchronize_sched_expedited() to |
| 33 | * serialize with the preempt-disabled section below. |
| 34 | * |
| 35 | * The nontrivial part is that we should guarantee acquire/release semantics |
| 36 | * in case when |
| 37 | * |
| 38 | * R_W: down_write() comes after up_read(), the writer should see all |
| 39 | * changes done by the reader |
| 40 | * or |
| 41 | * W_R: down_read() comes after up_write(), the reader should see all |
| 42 | * changes done by the writer |
| 43 | * |
| 44 | * If this helper fails the callers rely on the normal rw_semaphore and |
| 45 | * atomic_dec_and_test(), so in this case we have the necessary barriers. |
| 46 | * |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 47 | * But if it succeeds we do not have any barriers, atomic_read(write_ctr) or |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 48 | * __this_cpu_add() below can be reordered with any LOAD/STORE done by the |
| 49 | * reader inside the critical section. See the comments in down_write and |
| 50 | * up_write below. |
| 51 | */ |
| 52 | static bool update_fast_ctr(struct percpu_rw_semaphore *brw, unsigned int val) |
| 53 | { |
| 54 | bool success = false; |
| 55 | |
| 56 | preempt_disable(); |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 57 | if (likely(!atomic_read(&brw->write_ctr))) { |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 58 | __this_cpu_add(*brw->fast_read_ctr, val); |
| 59 | success = true; |
| 60 | } |
| 61 | preempt_enable(); |
| 62 | |
| 63 | return success; |
| 64 | } |
| 65 | |
| 66 | /* |
| 67 | * Like the normal down_read() this is not recursive, the writer can |
| 68 | * come after the first percpu_down_read() and create the deadlock. |
| 69 | */ |
| 70 | void percpu_down_read(struct percpu_rw_semaphore *brw) |
| 71 | { |
| 72 | if (likely(update_fast_ctr(brw, +1))) |
| 73 | return; |
| 74 | |
| 75 | down_read(&brw->rw_sem); |
| 76 | atomic_inc(&brw->slow_read_ctr); |
| 77 | up_read(&brw->rw_sem); |
| 78 | } |
| 79 | |
| 80 | void percpu_up_read(struct percpu_rw_semaphore *brw) |
| 81 | { |
| 82 | if (likely(update_fast_ctr(brw, -1))) |
| 83 | return; |
| 84 | |
| 85 | /* false-positive is possible but harmless */ |
| 86 | if (atomic_dec_and_test(&brw->slow_read_ctr)) |
| 87 | wake_up_all(&brw->write_waitq); |
| 88 | } |
| 89 | |
| 90 | static int clear_fast_ctr(struct percpu_rw_semaphore *brw) |
| 91 | { |
| 92 | unsigned int sum = 0; |
| 93 | int cpu; |
| 94 | |
| 95 | for_each_possible_cpu(cpu) { |
| 96 | sum += per_cpu(*brw->fast_read_ctr, cpu); |
| 97 | per_cpu(*brw->fast_read_ctr, cpu) = 0; |
| 98 | } |
| 99 | |
| 100 | return sum; |
| 101 | } |
| 102 | |
| 103 | /* |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 104 | * A writer increments ->write_ctr to force the readers to switch to the |
| 105 | * slow mode, note the atomic_read() check in update_fast_ctr(). |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 106 | * |
| 107 | * After that the readers can only inc/dec the slow ->slow_read_ctr counter, |
| 108 | * ->fast_read_ctr is stable. Once the writer moves its sum into the slow |
| 109 | * counter it represents the number of active readers. |
| 110 | * |
| 111 | * Finally the writer takes ->rw_sem for writing and blocks the new readers, |
| 112 | * then waits until the slow counter becomes zero. |
| 113 | */ |
| 114 | void percpu_down_write(struct percpu_rw_semaphore *brw) |
| 115 | { |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 116 | /* tell update_fast_ctr() there is a pending writer */ |
| 117 | atomic_inc(&brw->write_ctr); |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 118 | /* |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 119 | * 1. Ensures that write_ctr != 0 is visible to any down_read/up_read |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 120 | * so that update_fast_ctr() can't succeed. |
| 121 | * |
| 122 | * 2. Ensures we see the result of every previous this_cpu_add() in |
| 123 | * update_fast_ctr(). |
| 124 | * |
| 125 | * 3. Ensures that if any reader has exited its critical section via |
| 126 | * fast-path, it executes a full memory barrier before we return. |
| 127 | * See R_W case in the comment above update_fast_ctr(). |
| 128 | */ |
| 129 | synchronize_sched_expedited(); |
| 130 | |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 131 | /* exclude other writers, and block the new readers completely */ |
| 132 | down_write(&brw->rw_sem); |
| 133 | |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 134 | /* nobody can use fast_read_ctr, move its sum into slow_read_ctr */ |
| 135 | atomic_add(clear_fast_ctr(brw), &brw->slow_read_ctr); |
| 136 | |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 137 | /* wait for all readers to complete their percpu_up_read() */ |
| 138 | wait_event(brw->write_waitq, !atomic_read(&brw->slow_read_ctr)); |
| 139 | } |
| 140 | |
| 141 | void percpu_up_write(struct percpu_rw_semaphore *brw) |
| 142 | { |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 143 | /* release the lock, but the readers can't use the fast-path */ |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 144 | up_write(&brw->rw_sem); |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 145 | /* |
| 146 | * Insert the barrier before the next fast-path in down_read, |
| 147 | * see W_R case in the comment above update_fast_ctr(). |
| 148 | */ |
| 149 | synchronize_sched_expedited(); |
Oleg Nesterov | 9390ef0 | 2012-12-17 16:01:36 -0800 | [diff] [blame^] | 150 | /* the last writer unblocks update_fast_ctr() */ |
| 151 | atomic_dec(&brw->write_ctr); |
Oleg Nesterov | a1fd3e2 | 2012-12-17 16:01:32 -0800 | [diff] [blame] | 152 | } |