Final commit from google.

benchmarks/sequentialThroughput.c

@@ -0,0 +1,78 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <stasis/transactional.h>
+#include <stasis/truncation.h>
+
+/*static stasis_handle_t * memory_factory(lsn_t off, lsn_t len, void * ignored) {
+  stasis_handle_t * h = stasis_handle(open_memory)(off);
+  //h = stasis_handle(open_debug)(h);
+  stasis_write_buffer_t * w = h->append_buffer(h, len);
+  w->h->release_write_buffer(w);
+  return h;
+}
+typedef struct sf_args {
+  char * filename;
+  int    openMode;
+  int    filePerm;
+} sf_args;
+static stasis_handle_t * traditional_file_factory(void * argsP) {
+  sf_args * args = (sf_args*) argsP;
+  stasis_handle_t * h =  stasis_handle(open_file)(0, args->filename, args->openMode, args->filePerm);
+  //h = stasis_handle(open_debug)(h);
+  return h;
+}
+*/
+
+static inline long mb_to_page(long mb) {
+  return (mb * 1024 * 1024) / PAGE_SIZE;
+}
+
+int main(int argc, char ** argv) {
+  int direct = 0;
+  int legacyBM = 0;
+  int legacyFH = 0;
+
+  long page_count = mb_to_page(100);
+
+  for(int i = 1; i < argc; i++) {
+    if(!strcmp(argv[i], "--direct")) {
+      direct = 1;
+      bufferManagerO_DIRECT = 1;
+    } else if(!strcmp(argv[i], "--deprecatedBM")) {
+      bufferManagerType = BUFFER_MANAGER_DEPRECATED_HASH;
+      legacyBM = 1;
+    } else if(!strcmp(argv[i], "--deprecatedFH")) {
+      bufferManagerFileHandleType = BUFFER_MANAGER_FILE_HANDLE_DEPRECATED;
+      legacyFH = 1;
+    } else if(!strcmp(argv[i], "--pfile")) {
+      bufferManagerNonBlockingSlowHandleType = IO_HANDLE_PFILE;
+    } else if(!strcmp(argv[i], "--file")) {
+      bufferManagerNonBlockingSlowHandleType = IO_HANDLE_FILE;
+    } else if(!strcmp(argv[i], "--mb")) {
+      i++;
+      page_count = mb_to_page(atoll(argv[i]));
+    } else {
+      printf("Unknown argument: %s\n", argv[i]);
+      return 1;
+    }
+  }
+
+  if(legacyFH && direct) {
+    printf("--direct and --deprecatedFH are incompatible with each other\n");
+    return 1;
+  }
+
+  Tinit();
+
+  for(long i =0; i < page_count; i++) {
+    Page * p = loadPage(-1, i);
+    dirtyPages_add(p);
+    releasePage(p);
+  }
+
+
+  Tdeinit();
+  return 0;
+}

libdfa/rw.h

@@ -22,6 +22,17 @@
 #include <unistd.h>
 #ifndef __LIBDFA_RW_H
 #define __LIBDFA_RW_H
+
+#ifdef __cplusplus
+#  define BEGIN_C_DECLS extern "C" {
+#  define END_C_DECLS   }
+#else /* !__cplusplus */
+#  define BEGIN_C_DECLS
+#  define END_C_DECLS
+#endif /* __cplusplus */
+
+BEGIN_C_DECLS
+
 typedef struct {
 	pthread_mutex_t *mut;
 	int writers;
@@ -55,4 +66,7 @@ rwargs *newRWargs (rwl *l, int i, long d);
 void *reader (void *args);
 void *writer (void *args);
 */
+
+END_C_DECLS
+
 #endif /* rw.h */

src/stasis/bufferManager/bufferHash.c

@@ -108,7 +108,7 @@ inline static Page * writeBackOnePage() {
   assert(old == victim);
 
   //      printf("Write(%ld)\n", (long)victim->id);
-  pageWrite(victim);
+  pageWrite(victim);   /// XXX pageCleanup and pageFlushed might be heavyweight.
   pageCleanup(victim);
   // Make sure that no one mistakenly thinks this is still a live copy.
   victim->id = -1;

src/stasis/io/non_blocking.c

@@ -74,8 +74,8 @@ static inline stasis_read_buffer_t * alloc_read_buffer_error(stasis_handle_t *h,
   return r;
 }
 /** @return a read buffer indicating a write error has occured */
-static inline stasis_write_buffer_t * alloc_write_buffer_error(stasis_handle_t *h,
+static inline stasis_write_buffer_t* alloc_write_buffer_error
-                                                               int error) {
+    (stasis_handle_t *h, int error) {
   assert(error);
   stasis_write_buffer_t * w = malloc(sizeof(stasis_write_buffer_t));
   w->h = h;
@@ -267,15 +267,24 @@ static inline const tree_node * allocFastHandle(nbw_impl * impl, lsn_t off,
   }
 
 #ifdef EAGER_MERGE
-  // check for a mergable range immediately after the point we're interested in.
+  // check for a mergable range immediately after the point we're
+  // interested in.  (DEAD CODE)
   tree_node dummy;
   dummy.start_pos = n->end_pos;
   dummy.end_pos = n->end_pos+1;
 
-  while((np = (tree_node*)RB_ENTRY(find)(&dummy, impl->fast_handles)) && np->dirty && !np->pin_count && np->write_count + n->write_count < MAX_MERGE) {
+  while((np = (tree_node*)RB_ENTRY(find)(&dummy, impl->fast_handles))
-    DEBUG("Did post-merge of page %lld-%lld (%d) and %lld-%lld (%d) outstanding = %d\n", n->start_pos/PAGE_SIZE, -1+n->end_pos/PAGE_SIZE, n->write_count, np->start_pos/PAGE_SIZE, -1+np->end_pos/PAGE_SIZE, np->write_count, impl->fast_handle_count);
+        && np->dirty
+        && !np->pin_count
+        && np->write_count + n->write_count < MAX_MERGE) {
+    DEBUG("Did post-merge of page %lld-%lld (%d) and %lld-%lld (%d) "
+          "outstanding = %d\n", n->start_pos/PAGE_SIZE,
+          -1+n->end_pos/PAGE_SIZE, n->write_count, np->start_pos/PAGE_SIZE,
+          -1+np->end_pos/PAGE_SIZE, np->write_count, impl->fast_handle_count);
     lsn_t appendLen = np->end_pos - np->start_pos;
-    stasis_read_buffer_t * r= np->h->read_buffer(np->h,np->start_pos, appendLen);
+    stasis_read_buffer_t * r= np->h->read_buffer(np->h,np->start_pos,
+                                                 appendLen);
+
     int ret = n->h->write(n->h,np->start_pos,r->buf, appendLen);
     assert(!ret);
     ret = r->h->release_read_buffer(r);
@@ -293,7 +302,8 @@ static inline const tree_node * allocFastHandle(nbw_impl * impl, lsn_t off,
 
   return n;
 }
-static inline const tree_node * findFastHandle(nbw_impl * impl, lsn_t off, lsn_t len) {
+static inline const tree_node * findFastHandle(nbw_impl * impl, lsn_t off,
+                                               lsn_t len) {
   tree_node * np = allocTreeNode(off, len);
 
   pthread_mutex_lock(&impl->mut);
@@ -357,7 +367,8 @@ static int nbw_close(stasis_handle_t * h) {
   free(impl->workers);
 
   DEBUG("nbw had %d slow handles\n", impl->slow_handle_count);
-  DEBUG("fast handles = %d, used buffer = %lld\n", impl->fast_handle_count, impl->used_buffer_size);
+  DEBUG("fast handles = %d, used buffer = %lld\n", impl->fast_handle_count,
+        impl->used_buffer_size);
   if(impl->requested_bytes_written < impl->total_bytes_written) {
     printf("nbw: Problem with write coalescing detected.\n"
            "Client wrote %lld bytes, handle wrote %lld.\n",
@@ -480,7 +491,8 @@ static int nbw_release_read_buffer(stasis_read_buffer_t * r) {
   const tree_node * n = r_impl->n;
   stasis_handle_t * oldHandle = r_impl->r->h;
   r_impl->r->h->release_read_buffer(r_impl->r);
-  // XXX shouldn't need to check for this here; getFastHandle does something similar...
+  // XXX shouldn't need to check for this here; getFastHandle does
+  // something similar...
   if(n) {
     releaseFastHandle(impl, n, 0);
   } else {
@@ -633,7 +645,8 @@ static void * nbw_worker(void * handle) {
           }
 
           stasis_handle_t * fast2 = np->h;
-          stasis_read_buffer_t * r2 = fast2->read_buffer(fast2,np->start_pos, np_len);
+          stasis_read_buffer_t * r2 = fast2->read_buffer(fast2,np->start_pos,
+                                                         np_len);
           memcpy(buf + buf_off, r2->buf, np_len);
           buf_off += np_len;
           r2->h->release_read_buffer(r2);
@@ -647,7 +660,8 @@ static void * nbw_worker(void * handle) {
         pthread_mutex_unlock(&impl->mut);
 
         if(len != PAGE_SIZE) {
-          DEBUG("merged %lld pages at %lld into single write\n", len/PAGE_SIZE, off/PAGE_SIZE);
+          DEBUG("merged %lld pages at %lld into single write\n",
+                len/PAGE_SIZE, off/PAGE_SIZE);
         }
 	slow->write(slow, off, buf, len);
 
@@ -660,7 +674,8 @@ static void * nbw_worker(void * handle) {
 	pthread_mutex_lock(&impl->mut);
 	node->pin_count--;
       }
-      tree_node *new_node = (tree_node*)RB_ENTRY(lookup)(RB_LUGREAT, node, impl->fast_handles);
+      tree_node *new_node = (tree_node*)RB_ENTRY(lookup)(RB_LUGREAT, node,
+                                                         impl->fast_handles);
       if(!node->dirty && !node->pin_count) {
 	impl->fast_handle_count -= node->write_count;
         impl->used_buffer_size -= (node->end_pos - node->start_pos);
@@ -680,12 +695,11 @@ static void * nbw_worker(void * handle) {
   return 0;
 }
 
-stasis_handle_t * stasis_handle(open_non_blocking)(stasis_handle_t * (*slow_factory)(void * arg),
+stasis_handle_t * stasis_handle(open_non_blocking)
-						   void * slow_factory_arg,
+    (stasis_handle_t * (*slow_factory)(void * arg), void * slow_factory_arg,
      stasis_handle_t * (*fast_factory)(lsn_t, lsn_t, void *),
-						   void * fast_factory_arg,
+     void * fast_factory_arg, int worker_thread_count, lsn_t buffer_size,
-						   int worker_thread_count,
+     int max_fast_handles) {
-						   lsn_t buffer_size, int max_fast_handles) {
   nbw_impl * impl = malloc(sizeof(nbw_impl));
   pthread_mutex_init(&impl->mut, 0);
 

src/stasis/io/pfile.c

@@ -5,8 +5,10 @@
 #include <sys/stat.h>
 #include <fcntl.h>
 #include <stdio.h>
+#include <errno.h>
 #include <assert.h>
 
+
 #include <stasis/common.h>
 #include <stasis/io/handle.h>
 #include <pthread.h>
@@ -22,20 +24,52 @@
    @see handle.h
 */
 
+/**
+   Per-handle information for pfile
+*/
 typedef struct pfile_impl {
+  /**
+     This should be held whenever end_pos is accessed.
+  */
   pthread_mutex_t mut;
+  /**
+    The logical offset of the file.  Once the file is open, this will
+    never change, as pfile doesn't support truncation.
+  */
   lsn_t start_pos;
+  /**
+    The logical end of the file.
+  */
   lsn_t end_pos;
+  /**
+    File descriptor
+  */
   int fd;
+  /**
+    Flags passed into open
+  */
   int file_flags;
+  /**
+    File creation mode
+  */
   int file_mode;
+  /**
+    The name of the underlying file.
+  */
   char *filename;
 } pfile_impl;
 
+/**
+   We can pass the caller's buffer directly into pread()/pwrite()
+   without making any copies.
+*/
 static int pfile_num_copies(stasis_handle_t *h) { return 0; }
+/**
+   We have to call malloc(), but not memcpy().  Maybe this should return 1.
+*/
 static int pfile_num_copies_buffer(stasis_handle_t *h) { return 0; }
-static int pfile_close(stasis_handle_t *h) {
 
+static int pfile_close(stasis_handle_t *h) {
   pfile_impl *impl = (pfile_impl*)h->impl;
   DEBUG("Closing pfile: end = %lld\n", impl->end_pos);
   int fd = impl->fd;
@@ -46,10 +80,12 @@ static int pfile_close(stasis_handle_t *h) {
   if (!ret) return 0;
   else     return errno;
 }
+
 static lsn_t pfile_start_position(stasis_handle_t *h) {
   pfile_impl *impl = (pfile_impl*)h->impl;
   return impl->start_pos;
 }
+
 static lsn_t pfile_end_position(stasis_handle_t *h) {
   pfile_impl *impl = (pfile_impl*)h->impl;
   pthread_mutex_lock(&impl->mut);
@@ -57,11 +93,10 @@ static lsn_t pfile_end_position(stasis_handle_t *h) {
   pthread_mutex_unlock(&impl->mut);
   return ret;
 }
+
 inline static int pfile_write_unlocked(int fd, lsn_t off, const byte *dat,
                                        lsn_t len) {
-
   int error = 0;
-
   ssize_t bytes_written = 0;
 
   while (bytes_written < len) {
@@ -92,13 +127,23 @@ inline static int pfile_write_unlocked(int fd, lsn_t off, const byte * dat,
   }
   return error;
 }
+
 static int pfile_read(stasis_handle_t *h, lsn_t off, byte *buf, lsn_t len) {
   pfile_impl *impl = (pfile_impl*)(h->impl);
   int error = 0;
 
+  // reading impl->end_pos is probably atomic on most hardware, but
+  // this is safer.
+  pthread_mutex_lock(&impl->mut);
+  lsn_t end_pos = impl->end_pos;
+  pthread_mutex_unlock(&impl->mut);
+
+  // checking end_pos is adequate, (we assume this is the only handle
+  // touching the file)
+
   if (off < impl->start_pos) {
     error = EDOM;
-  } else if(off + len > impl->end_pos) {
+  } else if (off + len > end_pos) {
     error = EDOM;
   } else {
     ssize_t bytes_read = 0;
@@ -115,6 +160,10 @@ static int pfile_read(stasis_handle_t * h, lsn_t off, byte * buf, lsn_t len) {
           if (errno == EBADF) {
             h->error = EBADF;
           } else {
+            int err = errno;
+            // XXX Why is sys_errlist[] is unavailable here?
+            perror("pfile_read encountered an unknown error code.");
+            fprintf(stderr, "pread() returned -1; errno is %d\n",err); 
             abort(); // XXX other errors?
           }
           error = errno;
@@ -128,10 +177,11 @@ static int pfile_read(stasis_handle_t * h, lsn_t off, byte * buf, lsn_t len) {
     }
     assert(bytes_read == len);
   }
-
   return error;
 }
-static int pfile_write(stasis_handle_t *h, lsn_t off, const byte *dat, lsn_t len) {
+
+static int pfile_write(stasis_handle_t *h, lsn_t off, const byte *dat,
+                       lsn_t len) {
   pfile_impl *impl = (pfile_impl*)(h->impl);
   int error = 0;
   lsn_t phys_off;
@@ -140,6 +190,9 @@ static int pfile_write(stasis_handle_t *h, lsn_t off, const byte *dat, lsn_t len
   } else {
     pthread_mutex_lock(&impl->mut);
     if (impl->end_pos < off+len) {
+      // update end_pos now; the caller is not allowed to look at this
+      // part of the file until we return, so if they notice that the
+      // file hasn't been extended yet, it's a bug on their end.
       impl->end_pos = off+len;
     }
     pthread_mutex_unlock(&impl->mut);
@@ -148,19 +201,20 @@ static int pfile_write(stasis_handle_t *h, lsn_t off, const byte *dat, lsn_t len
   }
   return error;
 }
-static int pfile_append(stasis_handle_t *h, lsn_t *off, const byte *dat, lsn_t len) {
+
+static int pfile_append(stasis_handle_t *h, lsn_t *off, const byte *dat,
+                        lsn_t len) {
   pfile_impl *impl = (pfile_impl*)(h->impl);
   pthread_mutex_lock(&impl->mut);
   *off = impl->end_pos;
   impl->end_pos += len;
   pthread_mutex_unlock(&impl->mut);
   lsn_t phys_off = *off - impl->start_pos;
-  int error = pfile_write_unlocked(impl->fd, phys_off, dat,len);
+  return pfile_write_unlocked(impl->fd, phys_off, dat,len);
-  return error;
 }
+
 static stasis_write_buffer_t * pfile_write_buffer(stasis_handle_t *h,
                                                  lsn_t off, lsn_t len) {
-
   stasis_write_buffer_t *ret = malloc(sizeof(stasis_write_buffer_t));
 
   if (!ret) {
@@ -178,6 +232,7 @@ static stasis_write_buffer_t * pfile_write_buffer(stasis_handle_t * h,
 
   pthread_mutex_lock(&impl->mut);
 
+  // @todo Come up with a reasonable way to avoid sparse files.
   if (off + len > impl->end_pos) {
     impl->end_pos = off+len;
   }
@@ -205,9 +260,9 @@ static stasis_write_buffer_t * pfile_write_buffer(stasis_handle_t * h,
   }
   return ret;
 }
+
 static stasis_write_buffer_t *pfile_append_buffer(stasis_handle_t *h,
 						 lsn_t len) {
-
    // Allocate the handle
   stasis_write_buffer_t *ret = malloc(sizeof(stasis_write_buffer_t));
   if (!ret) { return NULL; }
@@ -244,10 +299,9 @@ static stasis_write_buffer_t *pfile_append_buffer(stasis_handle_t * h,
     ret->error = 0;
   }
   return ret;
-
 }
-static int pfile_release_write_buffer(stasis_write_buffer_t * w) {
 
+static int pfile_release_write_buffer(stasis_write_buffer_t *w) {
   pfile_impl *impl = (pfile_impl*)(w->h->impl);
 
   pthread_mutex_lock(&(impl->mut));
@@ -259,9 +313,12 @@ static int pfile_release_write_buffer(stasis_write_buffer_t * w) {
   pthread_mutex_unlock(&(impl->mut));
 
   if (!error) {
-    error = pfile_write_unlocked(impl->fd, w->off-impl->start_pos, w->buf, w->len);
+    error = pfile_write_unlocked(impl->fd, w->off-impl->start_pos, w->buf,
+                                 w->len);
   }
+  if (w->buf) {
     free(w->buf);
+  }
   free(w);
   return error;
 }
@@ -298,19 +355,25 @@ static stasis_read_buffer_t * pfile_read_buffer(stasis_handle_t * h,
   }
   return ret;
 }
+
 static int pfile_release_read_buffer(stasis_read_buffer_t *r) {
+  if (r->buf) {
     free((void*)r->buf);
+  }
   free(r);
   return 0;
 }
+
 static int pfile_truncate_start(stasis_handle_t *h, lsn_t new_start) {
   static int truncate_warned = 0;
   if (!truncate_warned) {
-    printf("\nWarning: pfile doesn't support truncation; ignoring truncation request\n");
+    printf("\nWarning: pfile doesn't support truncation; "
+           "ignoring truncation request\n");
     truncate_warned = 1;
   }
   return 0;
 }
+
 struct stasis_handle_t pfile_func = {
   .num_copies = pfile_num_copies,
   .num_copies_buffer = pfile_num_copies_buffer,
@@ -328,12 +391,17 @@ struct stasis_handle_t pfile_func = {
   .truncate_start = pfile_truncate_start,
   .error = 0
 };
-stasis_handle_t * stasis_handle(open_pfile)(lsn_t start_offset, char * filename, int flags, int mode) {
+
+stasis_handle_t *stasis_handle(open_pfile)(lsn_t start_offset,
+                                            char *filename,
+                                            int flags, int mode) {
   stasis_handle_t *ret = malloc(sizeof(stasis_handle_t));
   if (!ret) { return NULL; }
   *ret = pfile_func;
 
   pfile_impl *impl = malloc(sizeof(pfile_impl));
+  if (!impl) { free(ret); return NULL; }
+
   ret->impl = impl;
   pthread_mutex_init(&(impl->mut), 0);
   impl->fd = open(filename, flags, mode);
@@ -349,7 +417,8 @@ stasis_handle_t * stasis_handle(open_pfile)(lsn_t start_offset, char * filename,
     ret->error = errno;
   }
   impl->end_pos = file_len + start_offset;
-  DEBUG("file len = %lld, start_off = %lld, end = %lld\n", file_len, start_offset, impl->end_pos);
+  DEBUG("file len = %lld, start_off = %lld, end = %lld\n",
+        file_len, start_offset, impl->end_pos);
 
   impl->filename = strdup(filename);
   impl->file_flags = flags;

src/stasis/operations/lsmTree.c

@@ -450,8 +450,10 @@ recordid TlsmAppendPage(int xid, recordid tree,
 
   if(ret.size == INVALID_SLOT) {
     if(lastLeaf->id != p->id) {
+      assert(s->lastLeaf != tree.page);
       unlock(lastLeaf->rwlatch);
       releasePage(lastLeaf); // don't need that page anymore...
+      lastLeaf = 0;
     }
     // traverse down the root of the tree.
 
@@ -537,6 +539,7 @@ recordid TlsmAppendPage(int xid, recordid tree,
     writeNodeRecord(xid, lastLeaf, ret.slot, key, keySize, val_page);
 
     if(lastLeaf->id != p->id) {
+      assert(s->lastLeaf != tree.page);
       unlock(lastLeaf->rwlatch);
       releasePage(lastLeaf);
     }
@@ -638,6 +641,37 @@ pageid_t TlsmFindPage(int xid, recordid tree, const byte *key) {
 
 }
 
+pageid_t TlsmLastPage(int xid, recordid tree) {
+  Page * root = loadPage(xid, tree.page);
+  readlock(root->rwlatch,0);
+  lsmTreeState *state = root->impl;
+  int keySize = getKeySize(xid,root);
+  if(state->lastLeaf == -1) {
+    const lsmTreeNodeRecord *nr = readNodeRecord(xid,root,DEPTH,
+                                                 keySize);
+    int depth = nr->ptr;
+    state->lastLeaf = findLastLeaf(xid,root,depth);
+  }
+  pageid_t ret = state->lastLeaf;
+  unlock(root->rwlatch);
+
+  // ret points to the last internal node at this point.
+  releasePage(root);
+
+  Page * p = loadPage(xid, ret);
+  readlock(p->rwlatch,0);
+  if(*recordcount_ptr(p) == 2) {
+    ret = -1;
+  } else {
+    const lsmTreeNodeRecord *nr = readNodeRecord(xid,p,(*recordcount_ptr(p))-1,keySize);
+    ret = nr->ptr;
+  }
+  unlock(p->rwlatch);
+  releasePage(p);
+
+  return ret;
+}
+
 /**
     The buffer manager calls this when the lsmTree's root page is
     loaded.  This function allocates some storage for cached values
@@ -683,6 +717,9 @@ lladdIterator_t *lsmTreeIterator_open(int xid, recordid root) {
     releasePage(p);
     p = loadPage(xid,leafid);
     readlock(p->rwlatch,0);
+    assert(depth != 0);
+  } else {
+    assert(depth == 0);
   }
   lsmIteratorImpl *impl = malloc(sizeof(lsmIteratorImpl));
   impl->p = p;
@@ -708,6 +745,24 @@ lladdIterator_t *lsmTreeIterator_open(int xid, recordid root) {
     } */
   return it;
 }
+lladdIterator_t *lsmTreeIterator_copy(int xid, lladdIterator_t* i) {
+  lsmIteratorImpl *it = i->impl;
+  lsmIteratorImpl *mine = malloc(sizeof(lsmIteratorImpl));
+
+  if(it->p) {
+    mine->p = loadPage(xid, it->p->id);
+    readlock(mine->p->rwlatch,0);
+  } else {
+    mine->p = 0;
+  }
+  memcpy(&mine->current, &it->current,sizeof(recordid));
+  mine->t = it->t;
+  mine->justOnePage = it->justOnePage;
+  lladdIterator_t * ret = malloc(sizeof(lladdIterator_t));
+  ret->type = -1; // XXX LSM_TREE_ITERATOR
+  ret->impl = mine;
+  return ret;
+}
 void lsmTreeIterator_close(int xid, lladdIterator_t *it) {
   lsmIteratorImpl *impl = it->impl;
   if(impl->p) {
@@ -738,7 +793,7 @@ int lsmTreeIterator_next(int xid, lladdIterator_t *it) {
       impl->current.size = keySize;
     } else {
       impl->p = 0;
-      impl->current.size = -1;
+      impl->current.size = INVALID_SLOT;
     }
   } else {
     assert(impl->current.size == keySize + sizeof(lsmTreeNodeRecord));

stasis/constants.h

@@ -82,10 +82,10 @@ terms specified in this license.
 
 #ifndef MAX_BUFFER_SIZE
 //#define MAX_BUFFER_SIZE 100003 
-/*#define MAX_BUFFER_SIZE 20029 */
+#define MAX_BUFFER_SIZE 20029
 //#define MAX_BUFFER_SIZE 10007 
 //#define MAX_BUFFER_SIZE 5003
-#define MAX_BUFFER_SIZE 2003 
+//#define MAX_BUFFER_SIZE 2003 
 //#define MAX_BUFFER_SIZE 4006 
 /* #define MAX_BUFFER_SIZE 71  */
 /*#define MAX_BUFFER_SIZE 7 */

stasis/operations/lsmTree.h

@@ -63,7 +63,10 @@ void TlsmSetPageAllocator(pageid_t (*allocer)(int xid, void * ignored),
 */
 pageid_t TlsmFindPage(int xid, recordid tree,
                  const byte *key);
-
+/**
+   @todo TlsmFirstPage for symmetry?
+ */
+pageid_t TlsmLastPage(int xid, recordid tree);
 /// ---------------  Iterator implementation
 
 typedef struct lsmTreeNodeRecord {
@@ -93,7 +96,7 @@ lladdIterator_t * lsmTreeIterator_open(int xid, recordid tree);
 */
 void lsmTreeIterator_close(int xid, lladdIterator_t * it);
 int  lsmTreeIterator_next (int xid, lladdIterator_t * it);
-
+lladdIterator_t *lsmTreeIterator_copy(int xid, lladdIterator_t* i);
 static inline int lsmTreeIterator_key  (int xid, lladdIterator_t *it,
                                         byte **key) {
   lsmIteratorImpl * impl = (lsmIteratorImpl*)it->impl;