2 * Unix SMB/CIFS implementation.
3 * Support for OneFS system interfaces.
5 * Copyright (C) Tim Prouty, 2008
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 3 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, see <http://www.gnu.org/licenses/>.
23 #include "onefs_config.h"
24 #include "oplock_onefs.h"
26 #include <ifs/ifs_syscalls.h>
27 #include <isi_acl/isi_acl_util.h>
28 #include <sys/isi_acl.h>
31 * Initialize the sm_lock struct before passing it to ifs_createfile.
33 static void smlock_init(connection_struct
*conn
, struct sm_lock
*sml
,
34 bool isexe
, uint32_t access_mask
, uint32_t share_access
,
35 uint32_t create_options
)
37 sml
->sm_type
.doc
= false;
38 sml
->sm_type
.isexe
= isexe
;
39 sml
->sm_type
.statonly
= is_stat_open(access_mask
);
40 sml
->sm_type
.access_mask
= access_mask
;
41 sml
->sm_type
.share_access
= share_access
;
44 * private_options was previously used for DENY_DOS/DENY_FCB checks in
45 * the kernel, but are now properly handled by fcb_or_dos_open. In
46 * these cases, ifs_createfile will return a sharing violation, which
47 * gives fcb_or_dos_open the chance to open a duplicate file handle.
49 sml
->sm_type
.private_options
= 0;
51 /* 1 second delay is handled in onefs_open.c by deferring the open */
52 sml
->sm_timeout
= timeval_set(0, 0);
55 static void smlock_dump(int debuglevel
, const struct sm_lock
*sml
)
58 DEBUG(debuglevel
, ("sml == NULL\n"));
63 ("smlock: doc=%s, isexec=%s, statonly=%s, access_mask=0x%x, "
64 "share_access=0x%x, private_options=0x%x timeout=%d/%d\n",
65 sml
->sm_type
.doc
? "True" : "False",
66 sml
->sm_type
.isexe
? "True" : "False",
67 sml
->sm_type
.statonly
? "True" : "False",
68 sml
->sm_type
.access_mask
,
69 sml
->sm_type
.share_access
,
70 sml
->sm_type
.private_options
,
71 (int)sml
->sm_timeout
.tv_sec
,
72 (int)sml
->sm_timeout
.tv_usec
));
76 * External interface to ifs_createfile
78 int onefs_sys_create_file(connection_struct
*conn
,
82 uint32_t open_access_mask
,
83 uint32_t share_access
,
84 uint32_t create_options
,
89 struct security_descriptor
*sd
,
93 struct sm_lock sml
, *psml
= NULL
;
94 enum oplock_type onefs_oplock
;
95 enum oplock_type onefs_granted_oplock
= OPLOCK_NONE
;
96 struct ifs_security_descriptor ifs_sd
= {}, *pifs_sd
= NULL
;
99 uint32_t onefs_dos_attributes
;
100 struct ifs_createfile_flags cf_flags
= CF_FLAGS_NONE
;
102 START_PROFILE(syscall_createfile
);
104 /* Setup security descriptor and get secinfo. */
108 secinfo
= (get_sec_info(sd
) & IFS_SEC_INFO_KNOWN_MASK
);
110 status
= onefs_samba_sd_to_sd(secinfo
, sd
, &ifs_sd
, SNUM(conn
));
112 if (!NT_STATUS_IS_OK(status
)) {
113 DEBUG(1, ("SD initialization failure: %s\n",
122 /* Stripping off private bits will be done for us. */
123 onefs_oplock
= onefs_samba_oplock_to_oplock(oplock_request
);
125 if (!lp_oplocks(SNUM(conn
))) {
126 SMB_ASSERT(onefs_oplock
== OPLOCK_NONE
);
129 /* Convert samba dos flags to UF_DOS_* attributes. */
130 onefs_dos_attributes
= dos_attributes_to_stat_dos_flags(dos_flags
);
133 * Deal with kernel creating Default ACLs. (Isilon bug 47447.)
135 * 1) "nt acl support = no", default_acl = no
136 * 2) "inherit permissions = yes", default_acl = no
138 if (lp_nt_acl_support(SNUM(conn
)) && !lp_inherit_perms(SNUM(conn
)))
139 cf_flags
= cf_flags_or(cf_flags
, CF_FLAGS_DEFAULT_ACL
);
142 * Some customer workflows require the execute bit to be ignored.
144 if (lp_parm_bool(SNUM(conn
), PARM_ONEFS_TYPE
,
145 PARM_ALLOW_EXECUTE_ALWAYS
,
146 PARM_ALLOW_EXECUTE_ALWAYS_DEFAULT
) &&
147 (open_access_mask
& FILE_EXECUTE
)) {
149 DEBUG(3, ("Stripping execute bit from %s: (0x%x)\n", path
,
153 open_access_mask
&= ~FILE_EXECUTE
;
156 * Add READ_DATA, so we're not left with desired_access=0. An
157 * execute call should imply the client will read the data.
159 open_access_mask
|= FILE_READ_DATA
;
161 DEBUGADD(3, ("New stripped access mask: 0x%x\n",
165 DEBUG(10,("onefs_sys_create_file: base_fd = %d, fname = %s"
166 "open_access_mask = 0x%x, flags = 0x%x, mode = 0%o, "
167 "desired_oplock = %s, id = 0x%x, secinfo = 0x%x, sd = %p, "
168 "dos_attributes = 0x%x, path = %s, "
169 "default_acl=%s\n", base_fd
, path
,
170 (unsigned int)open_access_mask
,
173 onefs_oplock_str(onefs_oplock
),
175 (unsigned int)secinfo
, sd
,
176 (unsigned int)onefs_dos_attributes
, path
,
177 cf_flags_and_bool(cf_flags
, CF_FLAGS_DEFAULT_ACL
) ?
180 /* Initialize smlock struct for files/dirs but not internal opens */
181 if (!(oplock_request
& INTERNAL_OPEN_ONLY
)) {
182 smlock_init(conn
, &sml
, is_executable(path
), access_mask
,
183 share_access
, create_options
);
187 smlock_dump(10, psml
);
189 ret_fd
= ifs_createfile(base_fd
, path
,
190 (enum ifs_ace_rights
)open_access_mask
, flags
& ~O_ACCMODE
, mode
,
191 onefs_oplock
, id
, psml
, secinfo
, pifs_sd
, onefs_dos_attributes
,
192 cf_flags
, &onefs_granted_oplock
);
194 DEBUG(10,("onefs_sys_create_file(%s): ret_fd = %d, "
195 "onefs_granted_oplock = %s\n",
196 ret_fd
< 0 ? strerror(errno
) : "success", ret_fd
,
197 onefs_oplock_str(onefs_granted_oplock
)));
199 if (granted_oplock
) {
201 onefs_oplock_to_samba_oplock(onefs_granted_oplock
);
205 END_PROFILE(syscall_createfile
);
206 aclu_free_sd(pifs_sd
, false);
212 * FreeBSD based sendfile implementation that allows for atomic semantics.
214 static ssize_t
onefs_sys_do_sendfile(int tofd
, int fromfd
,
215 const DATA_BLOB
*header
, SMB_OFF_T offset
, size_t count
, bool atomic
)
227 hdr
.headers
= &hdtrl
;
232 /* Set up the header iovec. */
234 hdtrl
.iov_base
= header
->data
;
235 hdtrl
.iov_len
= hdr_len
= header
->length
;
237 hdtrl
.iov_base
= NULL
;
242 while (total
+ hdtrl
.iov_len
) {
247 * FreeBSD sendfile returns 0 on success, -1 on error.
248 * Remember, the tofd and fromfd are reversed..... :-).
249 * nwritten includes the header data sent.
253 ret
= sendfile(fromfd
, tofd
, offset
, total
, &hdr
,
255 } while (ret
== -1 && errno
== EINTR
);
257 /* On error we're done. */
263 * If this was an ATOMIC sendfile, nwritten doesn't
264 * necessarily indicate an error. It could mean count > than
265 * what sendfile can handle atomically (usually 64K) or that
266 * there was a short read due to the file being truncated.
269 return atomic
? 0 : -1;
273 * An atomic sendfile should never send partial data!
275 if (atomic
&& nwritten
!= total
+ hdtrl
.iov_len
) {
276 DEBUG(0,("Atomic sendfile() sent partial data: "
277 "%llu of %d\n", nwritten
,
278 total
+ hdtrl
.iov_len
));
283 * If this was a short (signal interrupted) write we may need
284 * to subtract it from the header data, or null out the header
285 * data altogether if we wrote more than hdtrl.iov_len bytes.
286 * We change nwritten to be the number of file bytes written.
289 if (hdtrl
.iov_base
&& hdtrl
.iov_len
) {
290 if (nwritten
>= hdtrl
.iov_len
) {
291 nwritten
-= hdtrl
.iov_len
;
292 hdtrl
.iov_base
= NULL
;
296 (caddr_t
)hdtrl
.iov_base
+ nwritten
;
297 hdtrl
.iov_len
-= nwritten
;
304 return count
+ hdr_len
;
308 * Handles the subtleties of using sendfile with CIFS.
310 ssize_t
onefs_sys_sendfile(connection_struct
*conn
, int tofd
, int fromfd
,
311 const DATA_BLOB
*header
, SMB_OFF_T offset
,
317 START_PROFILE_BYTES(syscall_sendfile
, count
);
319 if (lp_parm_bool(SNUM(conn
), PARM_ONEFS_TYPE
,
320 PARM_ATOMIC_SENDFILE
,
321 PARM_ATOMIC_SENDFILE_DEFAULT
)) {
325 /* Try the sendfile */
326 ret
= onefs_sys_do_sendfile(tofd
, fromfd
, header
, offset
, count
,
329 /* If the sendfile wasn't atomic, we're done. */
331 DEBUG(10, ("non-atomic sendfile read %ul bytes\n", ret
));
332 END_PROFILE(syscall_sendfile
);
337 * Atomic sendfile takes care to not write anything to the socket
338 * until all of the requested bytes have been read from the file.
339 * There are two atomic cases that need to be handled.
341 * 1. The file was truncated causing less data to be read than was
342 * requested. In this case, we return back to the caller to
343 * indicate 0 bytes were written to the socket. This should
344 * prompt the caller to fallback to the standard read path: read
345 * the data, create a header that indicates how many bytes were
346 * actually read, and send the header/data back to the client.
348 * This saves us from standard sendfile behavior of sending a
349 * header promising more data then will actually be sent. The
350 * only two options are to close the socket and kill the client
351 * connection, or write a bunch of 0s. Closing the client
352 * connection is bad because there could actually be multiple
353 * sessions multiplexed from the same client that are all dropped
354 * because of a truncate. Writing the remaining data as 0s also
355 * isn't good, because the client will have an incorrect version
356 * of the file. If the file is written back to the server, the 0s
357 * will be written back. Fortunately, atomic sendfile allows us
358 * to avoid making this choice in most cases.
360 * 2. One downside of atomic sendfile, is that there is a limit on
361 * the number of bytes that can be sent atomically. The kernel
362 * has a limited amount of mbuf space that it can read file data
363 * into without exhausting the system's mbufs, so a buffer of
364 * length xfsize is used. The xfsize at the time of writing this
365 * is 64K. xfsize bytes are read from the file, and subsequently
366 * written to the socket. This makes it impossible to do the
367 * sendfile atomically for a byte count > xfsize.
369 * To cope with large requests, atomic sendfile returns -1 with
370 * errno set to E2BIG. Since windows maxes out at 64K writes,
371 * this is currently only a concern with non-windows clients.
372 * Posix extensions allow the full 24bit bytecount field to be
373 * used in ReadAndX, and clients such as smbclient and the linux
374 * cifs client can request up to 16MB reads! There are a few
375 * options for handling large sendfile requests.
377 * a. Fall back to the standard read path. This is unacceptable
378 * because it would require prohibitively large mallocs.
380 * b. Fall back to using samba's fake_send_file which emulates
381 * the kernel sendfile in userspace. This still has the same
382 * problem of sending the header before all of the data has
383 * been read, so it doesn't buy us anything, and has worse
384 * performance than the kernel's zero-copy sendfile.
386 * c. Use non-atomic sendfile syscall to attempt a zero copy
387 * read, and hope that there isn't a short read due to
388 * truncation. In the case of a short read, there are two
391 * 1. Kill the client connection
393 * 2. Write zeros to the socket for the remaining bytes
394 * promised in the header.
396 * It is safer from a data corruption perspective to kill the
397 * client connection, so this is our default behavior, but if
398 * this causes problems this can be configured to write zeros
402 /* Handle case 1: short read -> truncated file. */
404 END_PROFILE(syscall_sendfile
);
408 /* Handle case 2: large read. */
409 if (ret
== -1 && errno
== E2BIG
) {
411 if (!lp_parm_bool(SNUM(conn
), PARM_ONEFS_TYPE
,
412 PARM_SENDFILE_LARGE_READS
,
413 PARM_SENDFILE_LARGE_READS_DEFAULT
)) {
414 DEBUG(3, ("Not attempting non-atomic large sendfile: "
415 "%lu bytes\n", count
));
416 END_PROFILE(syscall_sendfile
);
420 if (count
< 0x10000) {
421 DEBUG(0, ("Count < 2^16 and E2BIG was returned! %lu\n",
425 DEBUG(10, ("attempting non-atomic large sendfile: %lu bytes\n",
428 /* Try a non-atomic sendfile. */
429 ret
= onefs_sys_do_sendfile(tofd
, fromfd
, header
, offset
,
431 /* Real error: kill the client connection. */
433 DEBUG(1, ("error on non-atomic large sendfile "
434 "(%lu bytes): %s\n", count
,
436 END_PROFILE(syscall_sendfile
);
440 /* Short read: kill the client connection. */
441 if (ret
!= count
+ header
->length
) {
442 DEBUG(1, ("short read on non-atomic large sendfile "
443 "(%lu of %lu bytes): %s\n", ret
, count
,
447 * Returning ret here would cause us to drop into the
448 * codepath that calls sendfile_short_send, which
449 * sends the client a bunch of zeros instead.
450 * Returning -1 kills the connection.
452 if (lp_parm_bool(SNUM(conn
), PARM_ONEFS_TYPE
,
454 PARM_SENDFILE_SAFE_DEFAULT
)) {
455 END_PROFILE(syscall_sendfile
);
459 END_PROFILE(syscall_sendfile
);
463 DEBUG(10, ("non-atomic large sendfile successful\n"));
466 /* There was error in the atomic sendfile. */
468 DEBUG(1, ("error on %s sendfile (%lu bytes): %s\n",
469 atomic
? "atomic" : "non-atomic",
470 count
, strerror(errno
)));
473 END_PROFILE(syscall_sendfile
);
478 * Only talloc the spill buffer once (reallocing when necessary).
480 static char *get_spill_buffer(size_t new_count
)
482 static int cur_count
= 0;
483 static char *spill_buffer
= NULL
;
485 /* If a sufficiently sized buffer exists, just return. */
486 if (new_count
<= cur_count
) {
487 SMB_ASSERT(spill_buffer
);
491 /* Allocate the first time. */
492 if (cur_count
== 0) {
493 SMB_ASSERT(!spill_buffer
);
494 spill_buffer
= talloc_array(NULL
, char, new_count
);
496 cur_count
= new_count
;
501 /* A buffer exists, but it's not big enough, so realloc. */
502 SMB_ASSERT(spill_buffer
);
503 spill_buffer
= talloc_realloc(NULL
, spill_buffer
, char, new_count
);
505 cur_count
= new_count
;
511 * recvfile does zero-copy writes given an fd to write to, and a socket with
512 * some data to write. If recvfile read more than it was able to write, it
513 * spills the data into a buffer. After first reading any additional data
514 * from the socket into the buffer, the spill buffer is then written with a
517 ssize_t
onefs_sys_recvfile(int fromfd
, int tofd
, SMB_OFF_T offset
,
520 char *spill_buffer
= NULL
;
521 bool socket_drained
= false;
523 off_t total_rbytes
= 0;
524 off_t total_wbytes
= 0;
528 START_PROFILE_BYTES(syscall_recvfile
, count
);
530 DEBUG(10,("onefs_recvfile: from = %d, to = %d, offset=%llu, count = "
531 "%lu\n", fromfd
, tofd
, offset
, count
));
534 END_PROFILE(syscall_recvfile
);
539 * Setup up a buffer for recvfile to spill data that has been read
540 * from the socket but not written.
542 spill_buffer
= get_spill_buffer(count
);
543 if (spill_buffer
== NULL
) {
549 * Keep trying recvfile until:
550 * - There is no data left to read on the socket, or
551 * - bytes read != bytes written, or
552 * - An error is returned that isn't EINTR/EAGAIN
555 /* Keep track of bytes read/written for recvfile */
559 DEBUG(10, ("calling recvfile loop, offset + total_wbytes = "
560 "%llu, count - total_rbytes = %llu\n",
561 offset
+ total_wbytes
, count
- total_rbytes
));
563 ret
= recvfile(tofd
, fromfd
, offset
+ total_wbytes
,
564 count
- total_wbytes
, &rbytes
, &wbytes
, 0,
567 DEBUG(10, ("recvfile ret = %d, errno = %d, rbytes = %llu, "
568 "wbytes = %llu\n", ret
, ret
>= 0 ? 0 : errno
,
571 /* Update our progress so far */
572 total_rbytes
+= rbytes
;
573 total_wbytes
+= wbytes
;
575 } while ((count
- total_rbytes
) && (rbytes
== wbytes
) &&
576 (ret
== -1 && (errno
== EINTR
|| errno
== EAGAIN
)));
578 DEBUG(10, ("total_rbytes = %llu, total_wbytes = %llu\n",
579 total_rbytes
, total_wbytes
));
581 /* Log if recvfile didn't write everything it read. */
582 if (total_rbytes
!= total_wbytes
) {
583 DEBUG(3, ("partial recvfile: total_rbytes=%llu but "
584 "total_wbytes=%llu, diff = %llu\n", total_rbytes
,
585 total_wbytes
, total_rbytes
- total_wbytes
));
586 SMB_ASSERT(total_rbytes
> total_wbytes
);
590 * If there is still data on the socket, read it off.
592 while (total_rbytes
< count
) {
594 DEBUG(3, ("shallow recvfile (%s), reading %llu\n",
595 strerror(errno
), count
- total_rbytes
));
598 * Read the remaining data into the spill buffer. recvfile
599 * may already have some data in the spill buffer, so start
600 * filling the buffer at total_rbytes - total_wbytes.
602 ret
= sys_read(fromfd
,
603 spill_buffer
+ (total_rbytes
- total_wbytes
),
604 count
- total_rbytes
);
608 DEBUG(0, ("shallow recvfile read: EOF\n"));
610 DEBUG(0, ("shallow recvfile read failed: %s\n",
613 /* Socket is dead, so treat as if it were drained. */
614 socket_drained
= true;
618 /* Data was read so update the rbytes */
622 if (total_rbytes
!= count
) {
623 smb_panic("Unread recvfile data still on the socket!");
627 * Now write any spilled data + the extra data read off the socket.
629 while (total_wbytes
< count
) {
631 DEBUG(3, ("partial recvfile, writing %llu\n", count
- total_wbytes
));
633 ret
= sys_pwrite(tofd
, spill_buffer
, count
- total_wbytes
,
634 offset
+ total_wbytes
);
637 DEBUG(0, ("partial recvfile write failed: %s\n",
642 /* Data was written so update the wbytes */
651 END_PROFILE(syscall_recvfile
);
653 /* Make sure we always try to drain the socket. */
654 if (!socket_drained
&& count
- total_rbytes
) {
655 int saved_errno
= errno
;
657 if (drain_socket(fromfd
, count
- total_rbytes
) !=
658 count
- total_rbytes
) {
659 /* Socket is dead! */
660 DEBUG(0, ("drain socket failed: %d\n", errno
));