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 <ifs/ifs_syscalls.h>
24 #include <isi_acl/isi_acl_util.h>
27 * Initialize the sm_lock struct before passing it to ifs_createfile.
29 static void smlock_init(connection_struct *conn, struct sm_lock *sml,
30 bool isexe, uint32_t access_mask, uint32_t share_access,
31 uint32_t create_options)
33 sml->sm_type.doc = false;
34 sml->sm_type.isexe = isexe;
35 sml->sm_type.statonly = is_stat_open(access_mask);
36 sml->sm_type.access_mask = access_mask;
37 sml->sm_type.share_access = share_access;
40 * private_options was previously used for DENY_DOS/DENY_FCB checks in
41 * the kernel, but are now properly handled by fcb_or_dos_open. In
42 * these cases, ifs_createfile will return a sharing violation, which
43 * gives fcb_or_dos_open the chance to open a duplicate file handle.
45 sml->sm_type.private_options = 0;
47 /* 1 second delay is handled in onefs_open.c by deferring the open */
48 sml->sm_timeout = timeval_set(0, 0);
51 static void smlock_dump(int debuglevel, const struct sm_lock *sml)
54 DEBUG(debuglevel, ("sml == NULL\n"));
59 ("smlock: doc=%s, isexec=%s, statonly=%s, access_mask=0x%x, "
60 "share_access=0x%x, private_options=0x%x timeout=%d/%d\n",
61 sml->sm_type.doc ? "True" : "False",
62 sml->sm_type.isexe ? "True" : "False",
63 sml->sm_type.statonly ? "True" : "False",
64 sml->sm_type.access_mask,
65 sml->sm_type.share_access,
66 sml->sm_type.private_options,
67 (int)sml->sm_timeout.tv_sec,
68 (int)sml->sm_timeout.tv_usec));
72 * External interface to ifs_createfile
74 int onefs_sys_create_file(connection_struct *conn,
78 uint32_t open_access_mask,
79 uint32_t share_access,
80 uint32_t create_options,
85 struct security_descriptor *sd,
89 struct sm_lock sml, *psml = NULL;
90 enum oplock_type onefs_oplock;
91 enum oplock_type onefs_granted_oplock = OPLOCK_NONE;
92 struct ifs_security_descriptor ifs_sd = {}, *pifs_sd = NULL;
95 uint32_t onefs_dos_attributes;
96 struct ifs_createfile_flags cf_flags = CF_FLAGS_NONE;
98 /* Setup security descriptor and get secinfo. */
102 secinfo = (get_sec_info(sd) & IFS_SEC_INFO_KNOWN_MASK);
104 status = onefs_samba_sd_to_sd(secinfo, sd, &ifs_sd, SNUM(conn));
106 if (!NT_STATUS_IS_OK(status)) {
107 DEBUG(1, ("SD initialization failure: %s",
116 /* Stripping off private bits will be done for us. */
117 onefs_oplock = onefs_samba_oplock_to_oplock(oplock_request);
119 if (!lp_oplocks(SNUM(conn))) {
120 SMB_ASSERT(onefs_oplock == OPLOCK_NONE);
123 /* Convert samba dos flags to UF_DOS_* attributes. */
124 onefs_dos_attributes = dos_attributes_to_stat_dos_flags(dos_flags);
127 * Deal with kernel creating Default ACLs. (Isilon bug 47447.)
129 * 1) "nt acl support = no", default_acl = no
130 * 2) "inherit permissions = yes", default_acl = no
132 if (lp_nt_acl_support(SNUM(conn)) && !lp_inherit_perms(SNUM(conn)))
133 cf_flags = cf_flags_or(cf_flags, CF_FLAGS_DEFAULT_ACL);
136 * Some customer workflows require the execute bit to be ignored.
138 if (lp_parm_bool(SNUM(conn), PARM_ONEFS_TYPE,
139 PARM_ALLOW_EXECUTE_ALWAYS,
140 PARM_ALLOW_EXECUTE_ALWAYS_DEFAULT) &&
141 (open_access_mask & FILE_EXECUTE)) {
143 DEBUG(3, ("Stripping execute bit from %s: (0x%x)\n", path,
147 open_access_mask &= ~FILE_EXECUTE;
150 * Add READ_DATA, so we're not left with desired_access=0. An
151 * execute call should imply the client will read the data.
153 open_access_mask |= FILE_READ_DATA;
155 DEBUGADD(3, ("New stripped access mask: 0x%x\n",
159 DEBUG(10,("onefs_sys_create_file: base_fd = %d, "
160 "open_access_mask = 0x%x, flags = 0x%x, mode = 0%o, "
161 "desired_oplock = %s, id = 0x%x, secinfo = 0x%x, sd = %p, "
162 "dos_attributes = 0x%x, path = %s, "
163 "default_acl=%s\n", base_fd,
164 (unsigned int)open_access_mask,
167 onefs_oplock_str(onefs_oplock),
169 (unsigned int)secinfo, sd,
170 (unsigned int)onefs_dos_attributes, path,
171 cf_flags_and_bool(cf_flags, CF_FLAGS_DEFAULT_ACL) ?
174 /* Initialize smlock struct for files/dirs but not internal opens */
175 if (!(oplock_request & INTERNAL_OPEN_ONLY)) {
176 smlock_init(conn, &sml, is_executable(path), access_mask,
177 share_access, create_options);
181 smlock_dump(10, psml);
183 ret_fd = ifs_createfile(base_fd, path,
184 (enum ifs_ace_rights)open_access_mask, flags & ~O_ACCMODE, mode,
185 onefs_oplock, id, psml, secinfo, pifs_sd, onefs_dos_attributes,
186 cf_flags, &onefs_granted_oplock);
188 DEBUG(10,("onefs_sys_create_file(%s): ret_fd = %d, "
189 "onefs_granted_oplock = %s\n",
190 ret_fd < 0 ? strerror(errno) : "success", ret_fd,
191 onefs_oplock_str(onefs_granted_oplock)));
193 if (granted_oplock) {
195 onefs_oplock_to_samba_oplock(onefs_granted_oplock);
199 aclu_free_sd(pifs_sd, false);
205 * FreeBSD based sendfile implementation that allows for atomic semantics.
207 static ssize_t onefs_sys_do_sendfile(int tofd, int fromfd,
208 const DATA_BLOB *header, SMB_OFF_T offset, size_t count, bool atomic)
220 hdr.headers = &hdtrl;
225 /* Set up the header iovec. */
227 hdtrl.iov_base = header->data;
228 hdtrl.iov_len = hdr_len = header->length;
230 hdtrl.iov_base = NULL;
235 while (total + hdtrl.iov_len) {
240 * FreeBSD sendfile returns 0 on success, -1 on error.
241 * Remember, the tofd and fromfd are reversed..... :-).
242 * nwritten includes the header data sent.
246 ret = sendfile(fromfd, tofd, offset, total, &hdr,
248 } while (ret == -1 && errno == EINTR);
250 /* On error we're done. */
256 * If this was an ATOMIC sendfile, nwritten doesn't
257 * necessarily indicate an error. It could mean count > than
258 * what sendfile can handle atomically (usually 64K) or that
259 * there was a short read due to the file being truncated.
262 return atomic ? 0 : -1;
266 * An atomic sendfile should never send partial data!
268 if (atomic && nwritten != total + hdtrl.iov_len) {
269 DEBUG(0,("Atomic sendfile() sent partial data: "
270 "%llu of %d\n", nwritten,
271 total + hdtrl.iov_len));
276 * If this was a short (signal interrupted) write we may need
277 * to subtract it from the header data, or null out the header
278 * data altogether if we wrote more than hdtrl.iov_len bytes.
279 * We change nwritten to be the number of file bytes written.
282 if (hdtrl.iov_base && hdtrl.iov_len) {
283 if (nwritten >= hdtrl.iov_len) {
284 nwritten -= hdtrl.iov_len;
285 hdtrl.iov_base = NULL;
289 (caddr_t)hdtrl.iov_base + nwritten;
290 hdtrl.iov_len -= nwritten;
297 return count + hdr_len;
301 * Handles the subtleties of using sendfile with CIFS.
303 ssize_t onefs_sys_sendfile(connection_struct *conn, int tofd, int fromfd,
304 const DATA_BLOB *header, SMB_OFF_T offset,
310 if (lp_parm_bool(SNUM(conn), PARM_ONEFS_TYPE,
311 PARM_ATOMIC_SENDFILE,
312 PARM_ATOMIC_SENDFILE_DEFAULT)) {
316 /* Try the sendfile */
317 ret = onefs_sys_do_sendfile(tofd, fromfd, header, offset, count,
320 /* If the sendfile wasn't atomic, we're done. */
322 DEBUG(10, ("non-atomic sendfile read %ul bytes", ret));
327 * Atomic sendfile takes care to not write anything to the socket
328 * until all of the requested bytes have been read from the file.
329 * There are two atomic cases that need to be handled.
331 * 1. The file was truncated causing less data to be read than was
332 * requested. In this case, we return back to the caller to
333 * indicate 0 bytes were written to the socket. This should
334 * prompt the caller to fallback to the standard read path: read
335 * the data, create a header that indicates how many bytes were
336 * actually read, and send the header/data back to the client.
338 * This saves us from standard sendfile behavior of sending a
339 * header promising more data then will actually be sent. The
340 * only two options are to close the socket and kill the client
341 * connection, or write a bunch of 0s. Closing the client
342 * connection is bad because there could actually be multiple
343 * sessions multiplexed from the same client that are all dropped
344 * because of a truncate. Writing the remaining data as 0s also
345 * isn't good, because the client will have an incorrect version
346 * of the file. If the file is written back to the server, the 0s
347 * will be written back. Fortunately, atomic sendfile allows us
348 * to avoid making this choice in most cases.
350 * 2. One downside of atomic sendfile, is that there is a limit on
351 * the number of bytes that can be sent atomically. The kernel
352 * has a limited amount of mbuf space that it can read file data
353 * into without exhausting the system's mbufs, so a buffer of
354 * length xfsize is used. The xfsize at the time of writing this
355 * is 64K. xfsize bytes are read from the file, and subsequently
356 * written to the socket. This makes it impossible to do the
357 * sendfile atomically for a byte count > xfsize.
359 * To cope with large requests, atomic sendfile returns -1 with
360 * errno set to E2BIG. Since windows maxes out at 64K writes,
361 * this is currently only a concern with non-windows clients.
362 * Posix extensions allow the full 24bit bytecount field to be
363 * used in ReadAndX, and clients such as smbclient and the linux
364 * cifs client can request up to 16MB reads! There are a few
365 * options for handling large sendfile requests.
367 * a. Fall back to the standard read path. This is unacceptable
368 * because it would require prohibitively large mallocs.
370 * b. Fall back to using samba's fake_send_file which emulates
371 * the kernel sendfile in userspace. This still has the same
372 * problem of sending the header before all of the data has
373 * been read, so it doesn't buy us anything, and has worse
374 * performance than the kernel's zero-copy sendfile.
376 * c. Use non-atomic sendfile syscall to attempt a zero copy
377 * read, and hope that there isn't a short read due to
378 * truncation. In the case of a short read, there are two
381 * 1. Kill the client connection
383 * 2. Write zeros to the socket for the remaining bytes
384 * promised in the header.
386 * It is safer from a data corruption perspective to kill the
387 * client connection, so this is our default behavior, but if
388 * this causes problems this can be configured to write zeros
392 /* Handle case 1: short read -> truncated file. */
397 /* Handle case 2: large read. */
398 if (ret == -1 && errno == E2BIG) {
400 if (!lp_parm_bool(SNUM(conn), PARM_ONEFS_TYPE,
401 PARM_SENDFILE_LARGE_READS,
402 PARM_SENDFILE_LARGE_READS_DEFAULT)) {
403 DEBUG(3, ("Not attempting non-atomic large sendfile: "
404 "%lu bytes\n", count));
408 if (count < 0x10000) {
409 DEBUG(0, ("Count < 2^16 and E2BIG was returned! %lu",
413 DEBUG(10, ("attempting non-atomic large sendfile: %lu bytes\n",
416 /* Try a non-atomic sendfile. */
417 ret = onefs_sys_do_sendfile(tofd, fromfd, header, offset,
419 /* Real error: kill the client connection. */
421 DEBUG(1, ("error on non-atomic large sendfile "
422 "(%lu bytes): %s\n", count,
427 /* Short read: kill the client connection. */
428 if (ret != count + header->length) {
429 DEBUG(1, ("short read on non-atomic large sendfile "
430 "(%lu of %lu bytes): %s\n", ret, count,
434 * Returning ret here would cause us to drop into the
435 * codepath that calls sendfile_short_send, which
436 * sends the client a bunch of zeros instead.
437 * Returning -1 kills the connection.
439 if (lp_parm_bool(SNUM(conn), PARM_ONEFS_TYPE,
441 PARM_SENDFILE_SAFE_DEFAULT)) {
448 DEBUG(10, ("non-atomic large sendfile successful\n"));
451 /* There was error in the atomic sendfile. */
453 DEBUG(1, ("error on %s sendfile (%lu bytes): %s\n",
454 atomic ? "atomic" : "non-atomic",
455 count, strerror(errno)));
462 * Only talloc the spill buffer once (reallocing when necessary).
464 static char *get_spill_buffer(size_t new_count)
466 static int cur_count = 0;
467 static char *spill_buffer = NULL;
469 /* If a sufficiently sized buffer exists, just return. */
470 if (new_count <= cur_count) {
471 SMB_ASSERT(spill_buffer);
475 /* Allocate the first time. */
476 if (cur_count == 0) {
477 SMB_ASSERT(!spill_buffer);
478 spill_buffer = talloc_array(NULL, char, new_count);
480 cur_count = new_count;
485 /* A buffer exists, but it's not big enough, so realloc. */
486 SMB_ASSERT(spill_buffer);
487 spill_buffer = talloc_realloc(NULL, spill_buffer, char, new_count);
489 cur_count = new_count;
495 * recvfile does zero-copy writes given an fd to write to, and a socket with
496 * some data to write. If recvfile read more than it was able to write, it
497 * spills the data into a buffer. After first reading any additional data
498 * from the socket into the buffer, the spill buffer is then written with a
501 ssize_t onefs_sys_recvfile(int fromfd, int tofd, SMB_OFF_T offset,
504 char *spill_buffer = NULL;
505 bool socket_drained = false;
507 off_t total_rbytes = 0;
508 off_t total_wbytes = 0;
512 DEBUG(10,("onefs_recvfile: from = %d, to = %d, offset=%llu, count = "
513 "%lu\n", fromfd, tofd, offset, count));
520 * Setup up a buffer for recvfile to spill data that has been read
521 * from the socket but not written.
523 spill_buffer = get_spill_buffer(count);
524 if (spill_buffer == NULL) {
530 * Keep trying recvfile until:
531 * - There is no data left to read on the socket, or
532 * - bytes read != bytes written, or
533 * - An error is returned that isn't EINTR/EAGAIN
536 /* Keep track of bytes read/written for recvfile */
540 DEBUG(10, ("calling recvfile loop, offset + total_wbytes = "
541 "%llu, count - total_rbytes = %llu\n",
542 offset + total_wbytes, count - total_rbytes));
544 ret = recvfile(tofd, fromfd, offset + total_wbytes,
545 count - total_wbytes, &rbytes, &wbytes, 0,
548 DEBUG(10, ("recvfile ret = %d, errno = %d, rbytes = %llu, "
549 "wbytes = %llu\n", ret, ret >= 0 ? 0 : errno,
552 /* Update our progress so far */
553 total_rbytes += rbytes;
554 total_wbytes += wbytes;
556 } while ((count - total_rbytes) && (rbytes == wbytes) &&
557 (ret == -1 && (errno == EINTR || errno == EAGAIN)));
559 DEBUG(10, ("total_rbytes = %llu, total_wbytes = %llu\n",
560 total_rbytes, total_wbytes));
562 /* Log if recvfile didn't write everything it read. */
563 if (total_rbytes != total_wbytes) {
564 DEBUG(0, ("partial recvfile: total_rbytes=%llu but "
565 "total_wbytes=%llu, diff = %llu\n", total_rbytes,
566 total_wbytes, total_rbytes - total_wbytes));
567 SMB_ASSERT(total_rbytes > total_wbytes);
571 * If there is still data on the socket, read it off.
573 while (total_rbytes < count) {
575 DEBUG(0, ("shallow recvfile, reading %llu\n",
576 count - total_rbytes));
579 * Read the remaining data into the spill buffer. recvfile
580 * may already have some data in the spill buffer, so start
581 * filling the buffer at total_rbytes - total_wbytes.
583 ret = sys_read(fromfd,
584 spill_buffer + (total_rbytes - total_wbytes),
585 count - total_rbytes);
588 DEBUG(0, ("shallow recvfile read failed: %s\n",
590 /* Socket is dead, so treat as if it were drained. */
591 socket_drained = true;
595 /* Data was read so update the rbytes */
599 if (total_rbytes != count) {
600 smb_panic("Unread recvfile data still on the socket!");
604 * Now write any spilled data + the extra data read off the socket.
606 while (total_wbytes < count) {
608 DEBUG(0, ("partial recvfile, writing %llu\n", count - total_wbytes));
610 ret = sys_pwrite(tofd, spill_buffer, count - total_wbytes,
611 offset + total_wbytes);
614 DEBUG(0, ("partial recvfile write failed: %s\n",
619 /* Data was written so update the wbytes */
627 /* Make sure we always try to drain the socket. */
628 if (!socket_drained && count - total_rbytes) {
629 int saved_errno = errno;
631 if (drain_socket(fromfd, count - total_rbytes) !=
632 count - total_rbytes) {
633 /* Socket is dead! */
634 DEBUG(0, ("drain socket failed: %d\n", errno));