diff options
-rw-r--r-- | fs/xfs/xfs_filestream.c | 80 | ||||
-rw-r--r-- | fs/xfs/xfs_filestream.h | 82 |
2 files changed, 77 insertions, 85 deletions
diff --git a/fs/xfs/xfs_filestream.c b/fs/xfs/xfs_filestream.c index d34b9e8d2d37..9b715dce5699 100644 --- a/fs/xfs/xfs_filestream.c +++ b/fs/xfs/xfs_filestream.c | |||
@@ -123,6 +123,82 @@ typedef struct fstrm_item | |||
123 | xfs_inode_t *pip; /* Parent directory inode pointer. */ | 123 | xfs_inode_t *pip; /* Parent directory inode pointer. */ |
124 | } fstrm_item_t; | 124 | } fstrm_item_t; |
125 | 125 | ||
126 | /* | ||
127 | * Allocation group filestream associations are tracked with per-ag atomic | ||
128 | * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a | ||
129 | * particular AG already has active filestreams associated with it. The mount | ||
130 | * point's m_peraglock is used to protect these counters from per-ag array | ||
131 | * re-allocation during a growfs operation. When xfs_growfs_data_private() is | ||
132 | * about to reallocate the array, it calls xfs_filestream_flush() with the | ||
133 | * m_peraglock held in write mode. | ||
134 | * | ||
135 | * Since xfs_mru_cache_flush() guarantees that all the free functions for all | ||
136 | * the cache elements have finished executing before it returns, it's safe for | ||
137 | * the free functions to use the atomic counters without m_peraglock protection. | ||
138 | * This allows the implementation of xfs_fstrm_free_func() to be agnostic about | ||
139 | * whether it was called with the m_peraglock held in read mode, write mode or | ||
140 | * not held at all. The race condition this addresses is the following: | ||
141 | * | ||
142 | * - The work queue scheduler fires and pulls a filestream directory cache | ||
143 | * element off the LRU end of the cache for deletion, then gets pre-empted. | ||
144 | * - A growfs operation grabs the m_peraglock in write mode, flushes all the | ||
145 | * remaining items from the cache and reallocates the mount point's per-ag | ||
146 | * array, resetting all the counters to zero. | ||
147 | * - The work queue thread resumes and calls the free function for the element | ||
148 | * it started cleaning up earlier. In the process it decrements the | ||
149 | * filestreams counter for an AG that now has no references. | ||
150 | * | ||
151 | * With a shrinkfs feature, the above scenario could panic the system. | ||
152 | * | ||
153 | * All other uses of the following macros should be protected by either the | ||
154 | * m_peraglock held in read mode, or the cache's internal locking exposed by the | ||
155 | * interval between a call to xfs_mru_cache_lookup() and a call to | ||
156 | * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode | ||
157 | * when new elements are added to the cache. | ||
158 | * | ||
159 | * Combined, these locking rules ensure that no associations will ever exist in | ||
160 | * the cache that reference per-ag array elements that have since been | ||
161 | * reallocated. | ||
162 | */ | ||
163 | static int | ||
164 | xfs_filestream_peek_ag( | ||
165 | xfs_mount_t *mp, | ||
166 | xfs_agnumber_t agno) | ||
167 | { | ||
168 | struct xfs_perag *pag; | ||
169 | int ret; | ||
170 | |||
171 | pag = xfs_perag_get(mp, agno); | ||
172 | ret = atomic_read(&pag->pagf_fstrms); | ||
173 | xfs_perag_put(pag); | ||
174 | return ret; | ||
175 | } | ||
176 | |||
177 | static int | ||
178 | xfs_filestream_get_ag( | ||
179 | xfs_mount_t *mp, | ||
180 | xfs_agnumber_t agno) | ||
181 | { | ||
182 | struct xfs_perag *pag; | ||
183 | int ret; | ||
184 | |||
185 | pag = xfs_perag_get(mp, agno); | ||
186 | ret = atomic_inc_return(&pag->pagf_fstrms); | ||
187 | xfs_perag_put(pag); | ||
188 | return ret; | ||
189 | } | ||
190 | |||
191 | static void | ||
192 | xfs_filestream_put_ag( | ||
193 | xfs_mount_t *mp, | ||
194 | xfs_agnumber_t agno) | ||
195 | { | ||
196 | struct xfs_perag *pag; | ||
197 | |||
198 | pag = xfs_perag_get(mp, agno); | ||
199 | atomic_dec(&pag->pagf_fstrms); | ||
200 | xfs_perag_put(pag); | ||
201 | } | ||
126 | 202 | ||
127 | /* | 203 | /* |
128 | * Scan the AGs starting at startag looking for an AG that isn't in use and has | 204 | * Scan the AGs starting at startag looking for an AG that isn't in use and has |
@@ -351,16 +427,14 @@ xfs_fstrm_free_func( | |||
351 | { | 427 | { |
352 | fstrm_item_t *item = (fstrm_item_t *)data; | 428 | fstrm_item_t *item = (fstrm_item_t *)data; |
353 | xfs_inode_t *ip = item->ip; | 429 | xfs_inode_t *ip = item->ip; |
354 | int ref; | ||
355 | 430 | ||
356 | ASSERT(ip->i_ino == ino); | 431 | ASSERT(ip->i_ino == ino); |
357 | 432 | ||
358 | xfs_iflags_clear(ip, XFS_IFILESTREAM); | 433 | xfs_iflags_clear(ip, XFS_IFILESTREAM); |
359 | 434 | ||
360 | /* Drop the reference taken on the AG when the item was added. */ | 435 | /* Drop the reference taken on the AG when the item was added. */ |
361 | ref = xfs_filestream_put_ag(ip->i_mount, item->ag); | 436 | xfs_filestream_put_ag(ip->i_mount, item->ag); |
362 | 437 | ||
363 | ASSERT(ref >= 0); | ||
364 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, | 438 | TRACE_FREE(ip->i_mount, ip, item->pip, item->ag, |
365 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); | 439 | xfs_filestream_peek_ag(ip->i_mount, item->ag)); |
366 | 440 | ||
diff --git a/fs/xfs/xfs_filestream.h b/fs/xfs/xfs_filestream.h index 260f757bbc5d..09dd9af45434 100644 --- a/fs/xfs/xfs_filestream.h +++ b/fs/xfs/xfs_filestream.h | |||
@@ -42,88 +42,6 @@ extern ktrace_t *xfs_filestreams_trace_buf; | |||
42 | 42 | ||
43 | #endif | 43 | #endif |
44 | 44 | ||
45 | /* | ||
46 | * Allocation group filestream associations are tracked with per-ag atomic | ||
47 | * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a | ||
48 | * particular AG already has active filestreams associated with it. The mount | ||
49 | * point's m_peraglock is used to protect these counters from per-ag array | ||
50 | * re-allocation during a growfs operation. When xfs_growfs_data_private() is | ||
51 | * about to reallocate the array, it calls xfs_filestream_flush() with the | ||
52 | * m_peraglock held in write mode. | ||
53 | * | ||
54 | * Since xfs_mru_cache_flush() guarantees that all the free functions for all | ||
55 | * the cache elements have finished executing before it returns, it's safe for | ||
56 | * the free functions to use the atomic counters without m_peraglock protection. | ||
57 | * This allows the implementation of xfs_fstrm_free_func() to be agnostic about | ||
58 | * whether it was called with the m_peraglock held in read mode, write mode or | ||
59 | * not held at all. The race condition this addresses is the following: | ||
60 | * | ||
61 | * - The work queue scheduler fires and pulls a filestream directory cache | ||
62 | * element off the LRU end of the cache for deletion, then gets pre-empted. | ||
63 | * - A growfs operation grabs the m_peraglock in write mode, flushes all the | ||
64 | * remaining items from the cache and reallocates the mount point's per-ag | ||
65 | * array, resetting all the counters to zero. | ||
66 | * - The work queue thread resumes and calls the free function for the element | ||
67 | * it started cleaning up earlier. In the process it decrements the | ||
68 | * filestreams counter for an AG that now has no references. | ||
69 | * | ||
70 | * With a shrinkfs feature, the above scenario could panic the system. | ||
71 | * | ||
72 | * All other uses of the following macros should be protected by either the | ||
73 | * m_peraglock held in read mode, or the cache's internal locking exposed by the | ||
74 | * interval between a call to xfs_mru_cache_lookup() and a call to | ||
75 | * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode | ||
76 | * when new elements are added to the cache. | ||
77 | * | ||
78 | * Combined, these locking rules ensure that no associations will ever exist in | ||
79 | * the cache that reference per-ag array elements that have since been | ||
80 | * reallocated. | ||
81 | */ | ||
82 | /* | ||
83 | * xfs_filestream_peek_ag is only used in tracing code | ||
84 | */ | ||
85 | static inline int | ||
86 | xfs_filestream_peek_ag( | ||
87 | xfs_mount_t *mp, | ||
88 | xfs_agnumber_t agno) | ||
89 | { | ||
90 | struct xfs_perag *pag; | ||
91 | int ret; | ||
92 | |||
93 | pag = xfs_perag_get(mp, agno); | ||
94 | ret = atomic_read(&pag->pagf_fstrms); | ||
95 | xfs_perag_put(pag); | ||
96 | return ret; | ||
97 | } | ||
98 | |||
99 | static inline int | ||
100 | xfs_filestream_get_ag( | ||
101 | xfs_mount_t *mp, | ||
102 | xfs_agnumber_t agno) | ||
103 | { | ||
104 | struct xfs_perag *pag; | ||
105 | int ret; | ||
106 | |||
107 | pag = xfs_perag_get(mp, agno); | ||
108 | ret = atomic_inc_return(&pag->pagf_fstrms); | ||
109 | xfs_perag_put(pag); | ||
110 | return ret; | ||
111 | } | ||
112 | |||
113 | static inline int | ||
114 | xfs_filestream_put_ag( | ||
115 | xfs_mount_t *mp, | ||
116 | xfs_agnumber_t agno) | ||
117 | { | ||
118 | struct xfs_perag *pag; | ||
119 | int ret; | ||
120 | |||
121 | pag = xfs_perag_get(mp, agno); | ||
122 | ret = atomic_dec_return(&pag->pagf_fstrms); | ||
123 | xfs_perag_put(pag); | ||
124 | return ret; | ||
125 | } | ||
126 | |||
127 | /* allocation selection flags */ | 45 | /* allocation selection flags */ |
128 | typedef enum xfs_fstrm_alloc { | 46 | typedef enum xfs_fstrm_alloc { |
129 | XFS_PICK_USERDATA = 1, | 47 | XFS_PICK_USERDATA = 1, |