X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ian/git?a=blobdiff_plain;f=cdb%2Fwriteable.c;h=0074a6126181e2d76378c1ca21099346223ccbaa;hb=fad73e0340b64f1a7da862d8f511ce9094e14375;hp=15afdf5f5ba287eaeeedfe96f8fe44221f20076b;hpb=01601728a492b537c49fae629b5688d74650c4fb;p=chiark-tcl.git

diff --git a/cdb/writeable.c b/cdb/writeable.c
index 15afdf5..0074a61 100644
--- a/cdb/writeable.c
+++ b/cdb/writeable.c
@@ -2,9 +2,26 @@
 
 #include "chiark_tcl_cdb.h"
 
+#define ftello ftell
+#define fseeko fseek
+
+/*---------- Forward declarations ----------*/
+
+struct ht_forall_ctx;
+
+/*---------- Useful routines ----------*/
+
+static void maybe_close(int fd) {
+  if (fd>=0) close(fd);
+}
+
+#define PE(m) do{						\
+    rc= cht_posixerr(ip, errno, "failed to " m); goto x_rc;	\
+  }while(0)
+
 /*---------- Pathbuf ----------*/
 
-typedef struct {
+typedef struct Pathbuf {
   char *buf, *sfx;
 } Pathbuf;
 
@@ -27,39 +44,554 @@ static void pathbuf_free(Pathbuf *pb) {
   pb->buf= 0;
 }
 
+/*---------- Our hash table ----------*/
+
+typedef struct HashTable {
+  Tcl_HashTable t;
+  Byte padding[128]; /* allow for expansion by Tcl, urgh */
+  Byte confound[16];
+} HashTable;
+
+typedef struct HashValue {
+  int len;
+  Byte data[1];
+} HashValue;
+
+static HashValue *htv_prep(int len) {
+  HashValue *hd;
+  hd= TALLOC((hd->data - (Byte*)hd) + len);
+  hd->len= len;
+}  
+static Byte *htv_fillptr(HashValue *hd) {
+  return hd->data;
+}
+static void htv_fill(HashValue *hd, const Byte *data) {
+  memcpy(hd->data, data, hd->len);
+}
+
+static void ht_setup(HashTable *ht) {
+  Tcl_InitHashTable(&ht->t, TCL_STRING_KEYS);
+}
+static void ht_update(HashTable *ht, const char *key, HashValue *val_eat) {
+  Tcl_HashEntry *he;
+  int new;
+
+  he= Tcl_CreateHashEntry(&ht->t, (char*)key, &new);
+  if (!new) TFREE(Tcl_GetHashValue(he));
+  Tcl_SetHashValue(he, val_eat);
+    /* eats the value since the data structure owns the memory */
+}
+static void ht_maybeupdate(HashTable *ht, const char *key,
+			   HashValue *val_eat) {
+  /* like ht_update except does not overwrite existing values */
+  Tcl_HashEntry *he;
+  int new;
+
+  he= Tcl_CreateHashEntry(&ht->t, (char*)key, &new);
+  if (!new) { TFREE(val_eat); return; }
+  Tcl_SetHashValue(he, val_eat);
+}
+
+static int ht_forall(HashTable *ht,
+		     int (*fn)(const char *key, const HashValue *val,
+			       struct ht_forall_ctx *ctx),
+		     struct ht_forall_ctx *ctx) {
+  /* Returns first positive value returned by any call to fn, or 0.
+   * If fn returns -1, key is deleted. */
+  Tcl_HashSearch sp;
+  Tcl_HashEntry *he;
+  const char *key;
+  HashValue *val;
+  int r;
+  
+  for (he= Tcl_FirstHashEntry(&ht->t, &sp);
+       he;
+       he= Tcl_NextHashEntry(&sp)) {
+    val= Tcl_GetHashValue(he);
+    if (!val->len) continue;
+
+    key= Tcl_GetHashKey(&ht->t, he);
+    
+    r= fn(key, val, ctx);
+    if (r) return r;
+  }
+  return 0;
+}
+
+static void ht_destroy(HashTable *ht) {
+  Tcl_HashSearch sp;
+  Tcl_HashEntry *he;
+  
+  for (he= Tcl_FirstHashEntry(&ht->t, &sp);
+       he;
+       he= Tcl_NextHashEntry(&sp)) {
+    TFREE(Tcl_GetHashValue(he));
+  }
+  Tcl_DeleteHashTable(&ht->t);
+}
+
 /*---------- Rw data structure ----------*/
 
 typedef struct {
-  Pathbuf pbsome, pbtmp;
+  int ix, autocompact, maxage;
+fixme implement maxage during compact;
+  int cdb_fd, lock_fd;
+  struct cdb cdb; /* valid iff cdb_fd >= 0 */
+  off_t cdb_bytes; /* valid iff cdb_fd >= 0 */
+  FILE *logfile;
+  HashTable logincore;
+  Pathbuf pbsome, pbother;
+  off_t mainsz;
 } Rw;
 
-static void destroy_cdbrw_idtabcb(Tcl_Interp *ip, void *val) { abort(); }
+static int rw_close(Tcl_Interp *ip, Rw *rw) {
+  int rc, r;
+
+  rc= TCL_OK;
+  ht_destroy(&rw->logincore);
+  maybe_close(rw->cdb_fd);
+  maybe_close(rw->lock_fd);
 
+  if (rw->logfile) {
+    r= fclose(rw->logfile);
+    if (r && ip) { rc= cht_posixerr(ip, errno, "probable data loss! failed to"
+				    " fclose logfile during untidy close"); }
+  }
+
+  pathbuf_free(&rw->pbsome); pathbuf_free(&rw->pbother);
+  TFREE(rw);
+  return rc;
+}
+
+static void destroy_cdbrw_idtabcb(Tcl_Interp *ip, void *rw) { rw_close(0,rw); }
 const IdDataSpec cdbtcl_rwdatabases= {
   "cdb-rwdb", "cdb-openrwdatabases-table", destroy_cdbrw_idtabcb
 };
 
+/*---------- File handling ----------*/
+
+static int acquire_lock(Tcl_Interp *ip, Pathbuf *pb, int *lockfd_r) {
+  /* *lockfd_r must be -1 on entry.  If may be set to >=0 even
+   * on error, and must be closed by the caller. */
+  mode_t um, lockmode;
+  struct flock fl;
+  int r;
+
+  um= umask(~(mode_t)0);
+  umask(um);
+
+  lockmode= 0666 & ~((um & 0444)>>1);
+  /* Remove r where umask would remove w;
+   * eg umask intending 0664 here gives 0660 */
+  
+  *lockfd_r= open(pathbuf_sfx(pb,".lock"), O_RDONLY|O_CREAT, lockmode);
+  if (*lockfd_r < 0)
+    return cht_posixerr(ip, errno, "could not open/create lockfile");
+
+  fl.l_type= F_WRLCK;
+  fl.l_whence= SEEK_SET;
+  fl.l_start= 0;
+  fl.l_len= 0;
+  fl.l_pid= getpid();
+
+  r= fcntl(*lockfd_r, F_SETLK, &fl);
+  if (r == -1) {
+    if (errno == EACCES || errno == EAGAIN)
+      return cht_staticerr(ip, "lock held by another process", "CDB LOCKED");
+    else return cht_posixerr(ip, errno, "unexpected error from fcntl while"
+			     " acquiring lock");
+  }
+}
+
+/*---------- Log reading ----------*/
+
+static int readlognum(FILE *f, int delim, int *num_r) {
+  int c;
+  char numbuf[20], *p, *ep;
+  unsigned long ul;
+
+  p= numbuf;
+  for (;;) {
+    c= getc(f);  if (c==EOF) return -2;
+    if (c == delim) break;
+    if (!isdigit((unsigned char)c)) return -2;
+    *p++= c;
+    if (p == numbuf+sizeof(numbuf)) return -2;
+  }
+  if (p == numbuf) return -2;
+  *p= 0;
+
+  errno=0; ul= strtoul(numbuf, &ep, 10);
+  if (*ep || errno || ul >= INT_MAX/2) return -2;
+  *num_r= ul;
+  return 0;
+}
+
+static int readstorelogrecord(FILE *f, HashTable *ht,
+			      void (*updatefn)(HashTable*, const char*,
+					       HashValue*)) {
+  /* returns:
+   *     0 for OK
+   *     -1 eof
+   *     -2 corrupt or error
+   *     -3 got newline indicating end
+   */
+  int keylen, vallen;
+  char *key;
+  HashValue *val;
+  int c, rc, r;
+
+  c= getc(f);
+  if (c==EOF) { if (feof(f)) return -1; return -2; }
+  if (c=='\n') return -3;
+  if (c!='+') return -2;
+
+  rc= readlognum(f, ',', &keylen);  if (rc) return rc;
+  rc= readlognum(f, ':', &vallen);  if (rc) return rc;
+
+  key= TALLOC(keylen+1);
+  val= htv_prep(vallen);
+
+  r= fread(key, 1,keylen, f);
+  if (r!=keylen) goto x2_free_keyval;
+  if (memchr(key,0,keylen)) goto x2_free_keyval;
+  key[keylen]= 0;
+
+  c= getc(f);  if (c!='-') goto x2_free_keyval;
+  c= getc(f);  if (c!='>') goto x2_free_keyval;
+  
+  r= fread(htv_fillptr(val), 1,vallen, f);
+  if (r!=vallen) goto x2_free_keyval;
+
+  updatefn(ht, key, val);
+  TFREE(key);
+  return TCL_OK;
 
+ x2_free_keyval:
+  TFREE(val);
+  TFREE(key);
+  return -2;
+}
 
-/*---------- Misc functionality ----------*/
+/*---------- Creating ----------*/
 
 int cht_do_cdbwr_create_empty(ClientData cd, Tcl_Interp *ip,
 			      const char *pathb) {
+  static const char *const toremoves[]= {
+    ".main", ".cdb", ".log", ".tmp", 0
+  };
+
   Pathbuf pb;
-  int lock_fd=-1, fd=-1;
+  int lock_fd=-1, fd=-1, rc, r;
+  const char *const *toremove;
 
   pathbuf_init(&pb, pathb);
   rc= acquire_lock(ip, &pb, &lock_fd);  if (rc) goto x_rc;
   
-  fd= open(pathbuf_sfx(".lock"), O_RDONLY
+  fd= open(pathbuf_sfx(&pb, ".main"), O_RDWR|O_CREAT|O_EXCL, 0666);
+  if (fd <= 0) PE("create new database file");
+
+  for (toremove=toremoves; *toremove; toremove++) {
+    r= remove(*toremove);
+    if (r && errno != ENOENT)
+      PE("delete possible spurious file during creation");
+  }
+  
+  rc= TCL_OK;
+
+ x_rc:
+  maybe_close(fd);
+  maybe_close(lock_fd);
+  pathbuf_free(&pb);
+  return rc;
+}
+
+/*---------- Info callbacks ----------*/
+
+static int infocbv3(Tcl_Interp *ip, Rw *rw, const char *arg1,
+		    const char *arg2fmt, const char *arg3, va_list al) {
+  abort();
+}
+  
+static int infocb3(Tcl_Interp *ip, Rw *rw, const char *arg1,
+		   const char *arg2fmt, const char *arg3, ...) {
+  int rc;
+  va_list al;
+  va_start(al, arg3);
+  rc= infocbv3(ip,rw,arg1,arg2fmt,arg3,al);
+  va_end(al);
+  return rc;
+}
+  
+static int infocb(Tcl_Interp *ip, Rw *rw, const char *arg1,
+		  const char *arg2fmt, ...) {
+  int rc;
+  va_list al;
+  va_start(al, arg2fmt);
+  rc= infocbv3(ip,rw,arg1,arg2fmt,0,al);
+  va_end(al);
+  return rc;
+}
+  
+/*---------- Opening ----------*/
+
+int cht_do_cdbwr_open(ClientData cd, Tcl_Interp *ip, const char *pathb,
+		      Tcl_Obj *on_info, int maxage, void **result) {
+  const Cdbwr_SubCommand *subcmd= cd;
+  int r, rc, mainfd=-1;
+  Rw *rw;
+  struct stat stab;
+  off_t logrecstart, logjunkpos;
+
+  rw= TALLOC(sizeof(*rw));
+  ht_setup(&rw->logincore);
+  rw->cdb_fd= rw->lock_fd= -1;  rw->logfile= 0;
+  rw->maxage= maxage;
+  pathbuf_init(&rw->pbsome, pathb);
+  pathbuf_init(&rw->pbother, pathb);
+  rw->autocompact= 1;
+
+  mainfd= open(pathbuf_sfx(&rw->pbsome,".main"), O_RDONLY);
+  if (mainfd<0) PE("open exist3ing database file .main");
+  rc= acquire_lock(ip, &rw->pbsome, &rw->lock_fd);  if (rc) goto x_rc;
+
+  r= fstat(mainfd, &stab);  if (r) PE("fstat .main");
+  rw->mainsz= stab.st_size;
+
+  rw->cdb_fd= open(pathbuf_sfx(&rw->pbsome,".cdb"), O_RDONLY);
+  if (rw->cdb_fd >=0) {
+    r= cdb_init(&rw->cdb, rw->cdb_fd);
+    if (r) {
+      rc= cht_posixerr(ip, errno, "failed to initialise cdb reader");
+      close(rw->cdb_fd);  rw->cdb_fd= -1;  goto x_rc;
+    }
+  } else if (errno == ENOENT) {
+    if (rw->mainsz) {
+      rc= cht_staticerr(ip, ".cdb does not exist but .main is nonempty -"
+			" .cdb must have been accidentally deleted!",
+			"CDB CDBMISSING");
+      goto x_rc;
+    }
+    /* fine */
+  } else {
+    PE("open .cdb");
+  }
+
+  rw->logfile= fopen(pathbuf_sfx(&rw->pbsome,".log"), "r+");
+  if (!rw->logfile) {
+    if (errno != ENOENT) PE("failed to open .log during open");
+    rw->logfile= fopen(rw->pbsome.buf, "w");
+    if (!rw->logfile) PE("create .log during (clean) open");
+  } else { /* rw->logfile */
+    r= fstat(fileno(rw->logfile), &stab);
+    if (r==-1) PE("fstat .log during open");
+    rc= infocb(ip, rw, "open-dirty-start", "log=%luby",
+	       (unsigned long)stab.st_size);
+    if (rc) goto x_rc;
+
+    for (;;) {
+      logrecstart= ftello(rw->logfile);
+      if (logrecstart < 0) PE("ftello .log during (dirty) open");
+      r= readstorelogrecord(rw->logfile, &rw->logincore, ht_update);
+      if (ferror(rw->logfile)) {
+	rc= cht_posixerr(ip, errno, "error reading .log during (dirty) open");
+	goto x_rc;
+      }
+      if (r==-1) {
+	break;
+      } else if (r==-2 || r==-3) {
+	char buf[100];
+	logjunkpos= ftello(rw->logfile);
+	if(logjunkpos<0) PE("ftello .log during report of junk in dirty open");
+
+	snprintf(buf,sizeof(buf), "CDB SYNTAX LOG %lu %lu",
+		 (unsigned long)logjunkpos, (unsigned long)logrecstart);
+
+	if (!(subcmd->flags & RWSCF_OKJUNK)) {
+	  Tcl_SetObjErrorCode(ip, Tcl_NewStringObj(buf,-1));
+	  snprintf(buf,sizeof(buf),"%lu",(unsigned long)logjunkpos);
+	  Tcl_ResetResult(ip);
+	  Tcl_AppendResult(ip, "syntax error (junk) in .log during"
+			   " (dirty) open, at file position ", buf, (char*)0);
+	  rc= TCL_ERROR;
+	  goto x_rc;
+	}
+	rc= infocb3(ip, rw, "open-dirty-junk", "errorfpos=%luby", buf,
+		    (unsigned long)logjunkpos);
+	if (rc) goto x_rc;
+
+	r= fseeko(rw->logfile, logrecstart, SEEK_SET);
+	if (r) PE("failed to fseeko .log before junk during dirty open");
+
+	r= ftruncate(fileno(rw->logfile), logrecstart);
+	if (r) PE("ftruncate .log to chop junk during dirty open");
+      } else {
+	assert(!r);
+      }
+    }
+  }
+  /* now log is positioned for appending and everything is read */
+
+  *result= rw;
+  maybe_close(mainfd);
+  return TCL_OK;
+
+ x_rc:
+  rw_close(0,rw);
+  maybe_close(mainfd);
+  return rc;
+}
+
+/*---------- Compacting ----------*/
+
+static int compact_core(Tcl_Interp *ip, Rw *rw, unsigned long logsize) {
+  /* creates new .cdb and .main
+   * closes logfile
+   * leaves .log with old data
+   * leaves cdb fd open onto old db
+   * leaves logincore full of crap
+   */
+  int r, rc;
+  int cdbfd, cdbmaking;
+  off_t errpos;
+  char buf[100];
+  
+  struct ht_forall_ctx {
+    struct cdb_make cdbm;
+    FILE *mainfile;
+    int count;
+  } a;
+
+  a.mainfile= 0;
+  cdbfd= -1;
+  cdbmaking= 0;
+
+  r= fclose(rw->logfile);
+  if (r) { rc= cht_posixerr(ip, errno, "probable data loss!  failed to fclose"
+			    " logfile during compact");  goto x_rc; }
+  rw->logfile= 0;
+  
+  rc= infocb(ip, rw, "compact-start", "log=%luby main=%luby",
+	     logsize, (unsigned long)rw->mainsz);
+  if (rc) goto x_rc;
+
+  /* merge unsuperseded records from main into hash table */
+
+  a.mainfile= fopen(pathbuf_sfx(&rw->pbsome,".main"), "r");
+  if (!a.mainfile) PE("failed to open .main for reading during compact");
+
+  for (;;) {
+    r= readstorelogrecord(a.mainfile, &rw->logincore, ht_maybeupdate);
+    if (ferror(a.mainfile)) { rc= cht_posixerr(ip, errno, "error reading"
+                         " .main during compact"); goto x_rc;
+    }
+    if (r==-3) {
+      break;
+    } else if (r==-1 || r==-2) {
+      errpos= ftello(a.mainfile);
+      if (errpos<0) PE("ftello .main during report of syntax error");
+      snprintf(buf,sizeof(buf), "CDB SYNTAX MAIN %lu", (unsigned long)errpos);
+      Tcl_SetObjErrorCode(ip, Tcl_NewStringObj(buf,-1));
+      snprintf(buf,sizeof(buf), "%lu", (unsigned long)errpos);
+      Tcl_ResetResult(ip);
+      Tcl_AppendResult(ip, "syntax error in .main during"
+		       " compact, at file position ", buf, (char*)0);
+      rc= TCL_ERROR;
+      goto x_rc;
+    } else {
+      assert(!rc);
+    }
+  }
+  fclose(a.mainfile);
+  a.mainfile= 0;
+
+  /* create new cdb */
+
+  cdbfd= open(pathbuf_sfx(&rw->pbsome,".tmp"), O_WRONLY|O_CREAT|O_TRUNC, 0666);
+  if (cdbfd<0) PE("create .tmp for new cdb during compact");
+
+  r= cdb_make_start(&a.cdbm, cdbfd);
+  if (r) PE("cdb_make_start during compact");
+  cdbmaking= 1;
+
+  r= ht_forall(&rw->logincore, addto_cdb, &addctx);
+  if (r) PE("cdb_make_add during compact");
+
+  r= cdb_make_finish(&a.cdbm, cdbfd);
+  if(r) PE("cdb_make_finish during compact");
+  cdbmaking= 0;
+
+  r= fdatasync(cdbfd);  if (r) PE("fdatasync new cdb during compact");
+  r= close(cdbfd);  if (r) PE("close new cdb during compact");
+  cdbfd= -1;
+
+  r= rename(rw->pbsome.buf, pathbuf_sfx(&rw->pbother,".cdb"));
+  if (r) PE("install new .cdb during compact");
+
+  /* create new main */
+
+  a.mainfile= fopen(pathbuf_sfx(&rw->pbsome,".tmp"), "w");
+  if (!a.mainfile) PE("create .tmp for new main during compact");
+
+  a.count= 0;
+  r= ht_forall(&rw->logincore, addto_main, a.mainfile);
+  if (r) { rc= cht_posixerr(ip, r, "error writing to new .main"
+			    " during compact");  goto x_rc; }
+  
+  r= fflush(a.mainfile);  if (r) PE("fflush new main during compact");
+  r= fdatasync(fileno(a.mainfile));
+  if (r) PE("fdatasync new main during compact");
+  
+  r= fclose(a.mainfile);  if (r) PE("fclose new main during compact");
+  a.mainfile= 0;
+
+  r= rename(rw->pbsome.buf, pathbuf_sfx(&rw->pbother,".main"));
+  if (r) PE("install new .main during compact");
+
+  /* done! */
+  
+  rc= infocb(ip, rw, "compact-end", "log=%luby main=%luby",
+	     logsize, (unsigned long)rw->mainsz);
+  if (rc) goto x_rc;
+
+  rc= TCL_OK;
+x_rc:
+  if (mainfile) fclose(mainfile);
+  if (cdbmaking) cdb_make_finish(&a.cdbm, cdbfd);
+  maybe_close(cdbfd);
+  remove(pathbuf_sfx(&rw->pbsome,".tmp")); /* for tidyness */
+}
   
+static void compact_forclose(Tcl_Interp *ip, Rw *rw) {
+  off_t logsz;
+  int rc;
 
+  logsz= ftello(rw->logfile);
+  if (logsz < 0) PE("ftello logfile (during tidy close)");
 
-int cht_do_cdbwr_open(ClientData cd, Tcl_Interp *ip, const char *pathb, Tcl_Obj *on_info, void **result);
+  rc= compact_core(ip, rw, logsz);  if (rc) goto x_rc;
 
+  r= remove(pathbuf_sfx(&rw->pbsome,".log"));
+  if (r) PE("remove .log (during tidy close)");
+}
+  
+int cht_do_cdbwr_close(ClientData cd, Tcl_Interp *ip, void *rw_v) {
+  Rw *rw= rw_v;
+  int rc, compact_rc, infocb_rc;
 
-int cht_do_cdbwr_close(ClientData cd, Tcl_Interp *ip, void *db);
-int cht_do_cdbwr_close_quick(ClientData cd, Tcl_Interp *ip, void *db);
+  if (rw->autocompact) compact_rc= compact_forclose(ip, rw);
+  else compact_rc= TCL_OK;
+
+  rc= rw_close(ip,rw);
+  infocb_rc= infocb_close(rw);
+  
+  cht_tabledataid_disposing(ip, rw_v, &cdbtcl_rwdatabases);
+  if (!rc) rc= compact_rc;
+  if (!rc) rc= infocb_rc;
+  return rc;
+}
+
+  
 int cht_do_cdbwr_lookup(ClientData cd, Tcl_Interp *ip, void *db, Tcl_Obj *key, Tcl_Obj **result);
 int cht_do_cdbwr_lookup_hb(ClientData cd, Tcl_Interp *ip, void *db, HBytes_Value key, HBytes_Value *result);
 int cht_do_cdbwr_update(ClientData cd, Tcl_Interp *ip, void *db, Tcl_Obj *key, Tcl_Obj *value);