| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * FUSE inode io modes. |
| 4 | * |
| 5 | * Copyright (c) 2024 CTERA Networks. |
| 6 | */ |
| 7 | |
| 8 | #include "fuse_i.h" |
| 9 | |
| 10 | #include <linux/kernel.h> |
| 11 | #include <linux/sched.h> |
| 12 | #include <linux/file.h> |
| 13 | #include <linux/fs.h> |
| 14 | |
| 15 | /* |
| 16 | * Return true if need to wait for new opens in caching mode. |
| 17 | */ |
| 18 | static inline bool fuse_is_io_cache_wait(struct fuse_inode *fi) |
| 19 | { |
| 20 | return READ_ONCE(fi->iocachectr) < 0 && !fuse_inode_backing(fi); |
| 21 | } |
| 22 | |
| 23 | /* |
| 24 | * Called on cached file open() and on first mmap() of direct_io file. |
| 25 | * Takes cached_io inode mode reference to be dropped on file release. |
| 26 | * |
| 27 | * Blocks new parallel dio writes and waits for the in-progress parallel dio |
| 28 | * writes to complete. |
| 29 | */ |
| 30 | int fuse_file_cached_io_open(struct inode *inode, struct fuse_file *ff) |
| 31 | { |
| 32 | struct fuse_inode *fi = get_fuse_inode(inode); |
| 33 | |
| 34 | /* There are no io modes if server does not implement open */ |
| 35 | if (!ff->args) |
| 36 | return 0; |
| 37 | |
| 38 | spin_lock(&fi->lock); |
| 39 | /* |
| 40 | * Setting the bit advises new direct-io writes to use an exclusive |
| 41 | * lock - without it the wait below might be forever. |
| 42 | */ |
| 43 | while (fuse_is_io_cache_wait(fi)) { |
| 44 | set_bit(FUSE_I_CACHE_IO_MODE, &fi->state); |
| 45 | spin_unlock(&fi->lock); |
| 46 | wait_event(fi->direct_io_waitq, !fuse_is_io_cache_wait(fi)); |
| 47 | spin_lock(&fi->lock); |
| 48 | } |
| 49 | |
| 50 | /* |
| 51 | * Check if inode entered passthrough io mode while waiting for parallel |
| 52 | * dio write completion. |
| 53 | */ |
| 54 | if (fuse_inode_backing(fi)) { |
| 55 | clear_bit(FUSE_I_CACHE_IO_MODE, &fi->state); |
| 56 | spin_unlock(&fi->lock); |
| 57 | return -ETXTBSY; |
| 58 | } |
| 59 | |
| 60 | WARN_ON(ff->iomode == IOM_UNCACHED); |
| 61 | if (ff->iomode == IOM_NONE) { |
| 62 | ff->iomode = IOM_CACHED; |
| 63 | if (fi->iocachectr == 0) |
| 64 | set_bit(FUSE_I_CACHE_IO_MODE, &fi->state); |
| 65 | fi->iocachectr++; |
| 66 | } |
| 67 | spin_unlock(&fi->lock); |
| 68 | return 0; |
| 69 | } |
| 70 | |
| 71 | static void fuse_file_cached_io_release(struct fuse_file *ff, |
| 72 | struct fuse_inode *fi) |
| 73 | { |
| 74 | spin_lock(&fi->lock); |
| 75 | WARN_ON(fi->iocachectr <= 0); |
| 76 | WARN_ON(ff->iomode != IOM_CACHED); |
| 77 | ff->iomode = IOM_NONE; |
| 78 | fi->iocachectr--; |
| 79 | if (fi->iocachectr == 0) |
| 80 | clear_bit(FUSE_I_CACHE_IO_MODE, &fi->state); |
| 81 | spin_unlock(&fi->lock); |
| 82 | } |
| 83 | |
| 84 | /* Start strictly uncached io mode where cache access is not allowed */ |
| 85 | int fuse_inode_uncached_io_start(struct fuse_inode *fi, struct fuse_backing *fb) |
| 86 | { |
| 87 | struct fuse_backing *oldfb; |
| 88 | int err = 0; |
| 89 | |
| 90 | spin_lock(&fi->lock); |
| 91 | /* deny conflicting backing files on same fuse inode */ |
| 92 | oldfb = fuse_inode_backing(fi); |
| 93 | if (fb && oldfb && oldfb != fb) { |
| 94 | err = -EBUSY; |
| 95 | goto unlock; |
| 96 | } |
| 97 | if (fi->iocachectr > 0) { |
| 98 | err = -ETXTBSY; |
| 99 | goto unlock; |
| 100 | } |
| 101 | fi->iocachectr--; |
| 102 | |
| 103 | /* fuse inode holds a single refcount of backing file */ |
| 104 | if (fb && !oldfb) { |
| 105 | oldfb = fuse_inode_backing_set(fi, fb); |
| 106 | WARN_ON_ONCE(oldfb != NULL); |
| 107 | } else { |
| 108 | fuse_backing_put(fb); |
| 109 | } |
| 110 | unlock: |
| 111 | spin_unlock(&fi->lock); |
| 112 | return err; |
| 113 | } |
| 114 | |
| 115 | /* Takes uncached_io inode mode reference to be dropped on file release */ |
| 116 | static int fuse_file_uncached_io_open(struct inode *inode, |
| 117 | struct fuse_file *ff, |
| 118 | struct fuse_backing *fb) |
| 119 | { |
| 120 | struct fuse_inode *fi = get_fuse_inode(inode); |
| 121 | int err; |
| 122 | |
| 123 | err = fuse_inode_uncached_io_start(fi, fb); |
| 124 | if (err) |
| 125 | return err; |
| 126 | |
| 127 | WARN_ON(ff->iomode != IOM_NONE); |
| 128 | ff->iomode = IOM_UNCACHED; |
| 129 | return 0; |
| 130 | } |
| 131 | |
| 132 | void fuse_inode_uncached_io_end(struct fuse_inode *fi) |
| 133 | { |
| 134 | struct fuse_backing *oldfb = NULL; |
| 135 | |
| 136 | spin_lock(&fi->lock); |
| 137 | WARN_ON(fi->iocachectr >= 0); |
| 138 | fi->iocachectr++; |
| 139 | if (!fi->iocachectr) { |
| 140 | wake_up(&fi->direct_io_waitq); |
| 141 | oldfb = fuse_inode_backing_set(fi, NULL); |
| 142 | } |
| 143 | spin_unlock(&fi->lock); |
| 144 | if (oldfb) |
| 145 | fuse_backing_put(oldfb); |
| 146 | } |
| 147 | |
| 148 | /* Drop uncached_io reference from passthrough open */ |
| 149 | static void fuse_file_uncached_io_release(struct fuse_file *ff, |
| 150 | struct fuse_inode *fi) |
| 151 | { |
| 152 | WARN_ON(ff->iomode != IOM_UNCACHED); |
| 153 | ff->iomode = IOM_NONE; |
| 154 | fuse_inode_uncached_io_end(fi); |
| 155 | } |
| 156 | |
| 157 | /* |
| 158 | * Open flags that are allowed in combination with FOPEN_PASSTHROUGH. |
| 159 | * A combination of FOPEN_PASSTHROUGH and FOPEN_DIRECT_IO means that read/write |
| 160 | * operations go directly to the server, but mmap is done on the backing file. |
| 161 | * FOPEN_PASSTHROUGH mode should not co-exist with any users of the fuse inode |
| 162 | * page cache, so FOPEN_KEEP_CACHE is a strange and undesired combination. |
| 163 | */ |
| 164 | #define FOPEN_PASSTHROUGH_MASK \ |
| 165 | (FOPEN_PASSTHROUGH | FOPEN_DIRECT_IO | FOPEN_PARALLEL_DIRECT_WRITES | \ |
| 166 | FOPEN_NOFLUSH) |
| 167 | |
| 168 | static int fuse_file_passthrough_open(struct inode *inode, struct file *file) |
| 169 | { |
| 170 | struct fuse_file *ff = file->private_data; |
| 171 | struct fuse_conn *fc = get_fuse_conn(inode); |
| 172 | struct fuse_backing *fb; |
| 173 | int err; |
| 174 | |
| 175 | /* Check allowed conditions for file open in passthrough mode */ |
| 176 | if (!IS_ENABLED(CONFIG_FUSE_PASSTHROUGH) || !fc->passthrough || |
| 177 | (ff->open_flags & ~FOPEN_PASSTHROUGH_MASK)) |
| 178 | return -EINVAL; |
| 179 | |
| 180 | fb = fuse_passthrough_open(file, inode, |
| 181 | ff->args->open_outarg.backing_id); |
| 182 | if (IS_ERR(fb)) |
| 183 | return PTR_ERR(fb); |
| 184 | |
| 185 | /* First passthrough file open denies caching inode io mode */ |
| 186 | err = fuse_file_uncached_io_open(inode, ff, fb); |
| 187 | if (!err) |
| 188 | return 0; |
| 189 | |
| 190 | fuse_passthrough_release(ff, fb); |
| 191 | fuse_backing_put(fb); |
| 192 | |
| 193 | return err; |
| 194 | } |
| 195 | |
| 196 | /* Request access to submit new io to inode via open file */ |
| 197 | int fuse_file_io_open(struct file *file, struct inode *inode) |
| 198 | { |
| 199 | struct fuse_file *ff = file->private_data; |
| 200 | struct fuse_inode *fi = get_fuse_inode(inode); |
| 201 | int err; |
| 202 | |
| 203 | /* |
| 204 | * io modes are not relevant with DAX and with server that does not |
| 205 | * implement open. |
| 206 | */ |
| 207 | if (FUSE_IS_DAX(inode) || !ff->args) |
| 208 | return 0; |
| 209 | |
| 210 | /* |
| 211 | * Server is expected to use FOPEN_PASSTHROUGH for all opens of an inode |
| 212 | * which is already open for passthrough. |
| 213 | */ |
| 214 | err = -EINVAL; |
| 215 | if (fuse_inode_backing(fi) && !(ff->open_flags & FOPEN_PASSTHROUGH)) |
| 216 | goto fail; |
| 217 | |
| 218 | /* |
| 219 | * FOPEN_PARALLEL_DIRECT_WRITES requires FOPEN_DIRECT_IO. |
| 220 | */ |
| 221 | if (!(ff->open_flags & FOPEN_DIRECT_IO)) |
| 222 | ff->open_flags &= ~FOPEN_PARALLEL_DIRECT_WRITES; |
| 223 | |
| 224 | /* |
| 225 | * First passthrough file open denies caching inode io mode. |
| 226 | * First caching file open enters caching inode io mode. |
| 227 | * |
| 228 | * Note that if user opens a file open with O_DIRECT, but server did |
| 229 | * not specify FOPEN_DIRECT_IO, a later fcntl() could remove O_DIRECT, |
| 230 | * so we put the inode in caching mode to prevent parallel dio. |
| 231 | */ |
| 232 | if ((ff->open_flags & FOPEN_DIRECT_IO) && |
| 233 | !(ff->open_flags & FOPEN_PASSTHROUGH)) |
| 234 | return 0; |
| 235 | |
| 236 | if (ff->open_flags & FOPEN_PASSTHROUGH) |
| 237 | err = fuse_file_passthrough_open(inode, file); |
| 238 | else |
| 239 | err = fuse_file_cached_io_open(inode, ff); |
| 240 | if (err) |
| 241 | goto fail; |
| 242 | |
| 243 | return 0; |
| 244 | |
| 245 | fail: |
| 246 | pr_debug("failed to open file in requested io mode (open_flags=0x%x, err=%i).\n", |
| 247 | ff->open_flags, err); |
| 248 | /* |
| 249 | * The file open mode determines the inode io mode. |
| 250 | * Using incorrect open mode is a server mistake, which results in |
| 251 | * user visible failure of open() with EIO error. |
| 252 | */ |
| 253 | return -EIO; |
| 254 | } |
| 255 | |
| 256 | /* No more pending io and no new io possible to inode via open/mmapped file */ |
| 257 | void fuse_file_io_release(struct fuse_file *ff, struct inode *inode) |
| 258 | { |
| 259 | struct fuse_inode *fi = get_fuse_inode(inode); |
| 260 | |
| 261 | /* |
| 262 | * Last passthrough file close allows caching inode io mode. |
| 263 | * Last caching file close exits caching inode io mode. |
| 264 | */ |
| 265 | switch (ff->iomode) { |
| 266 | case IOM_NONE: |
| 267 | /* Nothing to do */ |
| 268 | break; |
| 269 | case IOM_UNCACHED: |
| 270 | fuse_file_uncached_io_release(ff, fi); |
| 271 | break; |
| 272 | case IOM_CACHED: |
| 273 | fuse_file_cached_io_release(ff, fi); |
| 274 | break; |
| 275 | } |
| 276 | } |