1/*-
2 * See the file LICENSE for redistribution information.
3 *
4 * Copyright (c) 1997,2008 Oracle.  All rights reserved.
5 *
6 * $Id: os_rw.c,v 12.28 2008/02/12 16:08:52 bostic Exp $
7 */
8
9#include "db_config.h"
10
11#include "db_int.h"
12
13/*
14 * __os_io --
15 *	Do an I/O.
16 */
17int
18__os_io(env, op, fhp, pgno, pgsize, relative, io_len, buf, niop)
19	ENV *env;
20	int op;
21	DB_FH *fhp;
22	db_pgno_t pgno;
23	u_int32_t pgsize, relative, io_len;
24	u_int8_t *buf;
25	size_t *niop;
26{
27	int ret;
28
29#ifndef DB_WINCE
30	if (__os_is_winnt()) {
31		DB_ENV *dbenv;
32		DWORD nbytes;
33		OVERLAPPED over;
34		ULONG64 off;
35		dbenv = env == NULL ? NULL : env->dbenv;
36		if ((off = relative) == 0)
37			off = (ULONG64)pgsize * pgno;
38		over.Offset = (DWORD)(off & 0xffffffff);
39		over.OffsetHigh = (DWORD)(off >> 32);
40		over.hEvent = 0; /* we don't want asynchronous notifications */
41
42		if (dbenv != NULL &&
43		    FLD_ISSET(dbenv->verbose, DB_VERB_FILEOPS_ALL))
44			__db_msg(env,
45			    "fileops: %s %s: %lu bytes at offset %lu",
46			    op == DB_IO_READ ? "read" : "write",
47			    fhp->name, (u_long)io_len, (u_long)off);
48
49		LAST_PANIC_CHECK_BEFORE_IO(env);
50
51		switch (op) {
52		case DB_IO_READ:
53#if defined(HAVE_STATISTICS)
54			++fhp->read_count;
55#endif
56			if (!ReadFile(fhp->handle,
57			    buf, (DWORD)io_len, &nbytes, &over))
58				goto slow;
59			break;
60		case DB_IO_WRITE:
61#ifdef HAVE_FILESYSTEM_NOTZERO
62			if (__os_fs_notzero())
63				goto slow;
64#endif
65#if defined(HAVE_STATISTICS)
66			++fhp->write_count;
67#endif
68			if (!WriteFile(fhp->handle,
69			    buf, (DWORD)io_len, &nbytes, &over))
70				goto slow;
71			break;
72		}
73		if (nbytes == io_len) {
74			*niop = (size_t)nbytes;
75			return (0);
76		}
77	}
78
79slow:
80#endif
81	MUTEX_LOCK(env, fhp->mtx_fh);
82
83	if ((ret = __os_seek(env, fhp, pgno, pgsize, relative)) != 0)
84		goto err;
85
86	switch (op) {
87	case DB_IO_READ:
88		ret = __os_read(env, fhp, buf, io_len, niop);
89		break;
90	case DB_IO_WRITE:
91		ret = __os_write(env, fhp, buf, io_len, niop);
92		break;
93	}
94
95err:	MUTEX_UNLOCK(env, fhp->mtx_fh);
96
97	return (ret);
98}
99
100/*
101 * __os_read --
102 *	Read from a file handle.
103 */
104int
105__os_read(env, fhp, addr, len, nrp)
106	ENV *env;
107	DB_FH *fhp;
108	void *addr;
109	size_t len;
110	size_t *nrp;
111{
112	DB_ENV *dbenv;
113	DWORD count;
114	size_t offset, nr;
115	u_int8_t *taddr;
116	int ret;
117
118	dbenv = env == NULL ? NULL : env->dbenv;
119	ret = 0;
120
121#if defined(HAVE_STATISTICS)
122	++fhp->read_count;
123#endif
124	if (dbenv != NULL && FLD_ISSET(dbenv->verbose, DB_VERB_FILEOPS_ALL))
125		__db_msg(env,
126		    "fileops: read %s: %lu bytes", fhp->name, (u_long)len);
127
128	for (taddr = addr,
129	    offset = 0; offset < len; taddr += nr, offset += nr) {
130		LAST_PANIC_CHECK_BEFORE_IO(env);
131		RETRY_CHK((!ReadFile(fhp->handle,
132		    taddr, (DWORD)(len - offset), &count, NULL)), ret);
133		if (count == 0 || ret != 0)
134			break;
135		nr = (size_t)count;
136	}
137	*nrp = taddr - (u_int8_t *)addr;
138	if (ret != 0) {
139		__db_syserr(env, ret, "read: 0x%lx, %lu",
140		    P_TO_ULONG(taddr), (u_long)len - offset);
141		ret = __os_posix_err(ret);
142	}
143	return (ret);
144}
145
146/*
147 * __os_write --
148 *	Write to a file handle.
149 */
150int
151__os_write(env, fhp, addr, len, nwp)
152	ENV *env;
153	DB_FH *fhp;
154	void *addr;
155	size_t len;
156	size_t *nwp;
157{
158	int ret;
159
160#ifdef HAVE_FILESYSTEM_NOTZERO
161	/* Zero-fill as necessary. */
162	if (__os_fs_notzero() &&
163	    (ret = __db_zero_fill(env, fhp)) != 0)
164		return (ret);
165#endif
166	return (__os_physwrite(env, fhp, addr, len, nwp));
167}
168
169/*
170 * __os_physwrite --
171 *	Physical write to a file handle.
172 */
173int
174__os_physwrite(env, fhp, addr, len, nwp)
175	ENV *env;
176	DB_FH *fhp;
177	void *addr;
178	size_t len;
179	size_t *nwp;
180{
181	DB_ENV *dbenv;
182	DWORD count;
183	size_t offset, nw;
184	u_int8_t *taddr;
185	int ret;
186
187	dbenv = env == NULL ? NULL : env->dbenv;
188	ret = 0;
189
190#if defined(HAVE_STATISTICS)
191	++fhp->write_count;
192#endif
193	if (dbenv != NULL && FLD_ISSET(dbenv->verbose, DB_VERB_FILEOPS_ALL))
194		__db_msg(env,
195		    "fileops: write %s: %lu bytes", fhp->name, (u_long)len);
196
197	for (taddr = addr,
198	    offset = 0; offset < len; taddr += nw, offset += nw) {
199		LAST_PANIC_CHECK_BEFORE_IO(env);
200		RETRY_CHK((!WriteFile(fhp->handle,
201		    taddr, (DWORD)(len - offset), &count, NULL)), ret);
202		if (ret != 0)
203			break;
204		nw = (size_t)count;
205	}
206	*nwp = len;
207	if (ret != 0) {
208		__db_syserr(env, ret, "write: %#lx, %lu",
209		    P_TO_ULONG(taddr), (u_long)len - offset);
210		ret = __os_posix_err(ret);
211
212		DB_EVENT(env, DB_EVENT_WRITE_FAILED, NULL);
213	}
214	return (ret);
215}
216