Add a redirect_intent write mode for non-capacitor SSDs

src/blockstore/blockstore_heap.cpp

@@ -1267,6 +1267,34 @@ int blockstore_heap_t::add_big_write(object_id oid, heap_entry_t *old_head, bool
     });
 }
 
+int blockstore_heap_t::add_redirect_intent(object_id oid, heap_entry_t **obj_ptr, uint64_t version,
+    uint32_t offset, uint32_t len, uint64_t location, uint8_t *bitmap, uint8_t *data, uint32_t *modified_block)
+{
+    uint32_t wr_size = get_big_intent_entry_size();
+    // Big-redirect intents, just like regular big writes, are written after writing data so they don't require explicit_complete
+    return add_entry(wr_size, modified_block, false, false, [&](heap_entry_t *wr)
+    {
+        wr->entry_type = BS_HEAP_BIG_INTENT|BS_HEAP_STABLE;
+        wr->inode = oid.inode;
+        wr->stripe = oid.stripe;
+        wr->version = version;
+        wr->set_big_location(this, location);
+        auto & bi = wr->big_intent();
+        bi.offset = offset;
+        bi.len = len;
+        if (bitmap)
+            memcpy(wr->get_ext_bitmap(this), bitmap, dsk->clean_entry_bitmap_size);
+        else
+            memset(wr->get_ext_bitmap(this), 0, dsk->clean_entry_bitmap_size);
+        memset(wr->get_int_bitmap(this), 0, dsk->clean_entry_bitmap_size);
+        bitmap_set(wr->get_int_bitmap(this), offset, len, dsk->bitmap_granularity);
+        if (dsk->data_csum_type)
+            memset(wr->get_checksums(this), 0, get_csum_size(wr));
+        calc_checksums(wr, (uint8_t*)data, true, offset, len);
+        *obj_ptr = wr;
+    });
+}
+
 int blockstore_heap_t::add_big_intent(object_id oid, heap_entry_t **obj_ptr, uint64_t version,
     uint32_t offset, uint32_t len, uint8_t *bitmap, uint8_t *data, uint8_t *checksums, uint32_t *modified_block)
 {

src/blockstore/blockstore_heap.h

@@ -265,6 +265,9 @@ public:
     // adds a big_write (overwrite) entry to an object
     int add_big_write(object_id oid, heap_entry_t *old_head, bool stable, uint64_t version,
         uint32_t offset, uint32_t len, uint64_t location, uint8_t *bitmap, uint8_t *data, uint32_t *modified_block);
+    // adds a "redirecting" big_intent entry to an object (same as big_write, used to avoid fsync on desktop SSDs)
+    int add_redirect_intent(object_id oid, heap_entry_t **obj_ptr, uint64_t version,
+        uint32_t offset, uint32_t len, uint64_t location, uint8_t *bitmap, uint8_t *data, uint32_t *modified_block);
     // adds a big_intent (atomic partial modification) entry to an object
     int add_big_intent(object_id oid, heap_entry_t **obj_ptr, uint64_t version,
         uint32_t offset, uint32_t len, uint8_t *bitmap, uint8_t *data, uint8_t *checksums, uint32_t *modified_block);

src/blockstore/blockstore_write.cpp

@@ -5,6 +5,8 @@
 #include "blockstore_internal.h"
 #include "allocator.h"
 
+#define _REDIRECT_INTENT 0x101
+
 bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
 {
     clock_gettime(CLOCK_REALTIME, &PRIV(op)->tv_begin);
@@ -138,13 +140,12 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         PRIV(op)->op_state = 5;
         write_iodepth++;
     }
-    // FIXME: Add 'big_intent' write mode
     // FIXME: Allow to do initial writes as buffered, not redirected
     // FIXME: Allow to do direct writes over holes
     else if (!obj || obj->type() == BS_HEAP_DELETE || op->offset == 0 && op->len == dsk.data_block_size)
     {
         // Big (redirect) write
-        PRIV(op)->write_type = BS_HEAP_BIG_WRITE;
+        PRIV(op)->write_type = dsk.disable_data_fsync || op->opcode != BS_OP_WRITE_STABLE ? BS_HEAP_BIG_WRITE : _REDIRECT_INTENT;
         BS_SUBMIT_CHECK_SQES(1);
         PRIV(op)->location = heap->find_free_data();
         if (PRIV(op)->location == UINT64_MAX)
@@ -176,9 +177,14 @@ enospc:
         data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
         io_uring_prep_writev(sqe, dsk.data_fd, &data->iov, 1, dsk.data_offset + loc + op->offset);
         PRIV(op)->pending_ops++;
-        PRIV(op)->op_state = 1;
         write_iodepth++;
-        inflight_big++;
+        if (PRIV(op)->write_type == BS_HEAP_BIG_WRITE)
+        {
+            PRIV(op)->op_state = 1;
+            inflight_big++;
+        }
+        else
+            PRIV(op)->op_state = 3;
     }
     else if (intent_write_allowed(op, obj))
     {
@@ -344,8 +350,17 @@ resume_12:
 resume_4:
     {
         auto obj = heap->read_entry(op->oid);
-        int res = heap->add_big_write(op->oid, obj, (op->opcode == BS_OP_WRITE_STABLE), op->version,
+        int res = 0;
-            op->offset, op->len, PRIV(op)->location, op->bitmap, (uint8_t*)op->buf, &PRIV(op)->modified_block);
+        if (PRIV(op)->write_type == _REDIRECT_INTENT)
+        {
+            res = heap->add_redirect_intent(op->oid, &obj, op->version, op->offset, op->len,
+                PRIV(op)->location, op->bitmap, (uint8_t*)op->buf, &PRIV(op)->modified_block);
+        }
+        else
+        {
+            res = heap->add_big_write(op->oid, obj, op->opcode == BS_OP_WRITE_STABLE,
+                op->version, op->offset, op->len, PRIV(op)->location, op->bitmap, (uint8_t*)op->buf, &PRIV(op)->modified_block);
+        }
         if (res == ENOSPC)
         {
             if (!heap->get_to_compact_count())

src/test/test_heap.cpp

@@ -1240,6 +1240,7 @@ void test_intent_write(bool csum)
 }
 
 // FIXME: Add a test for big_intent, incl. explicit_complete with big_intent over big_write over deletion over big_write :)
+// FIXME: Add a test for redirect_intent
 
 int main(int narg, char *args[])
 {