[pve-devel] [PATCH qemu] PVE: add zero block handling to PBS dump callback

[pve-devel] [PATCH qemu] PVE: add zero block handling to PBS dump callback

[Stefan Reiter]

Both the PBS and VMA dump callbacks assume that a NULL pointer can be
passed as *pbuf, but that never happens, as backup-dump.c calls this
function with contents of an iovec.

So first, remove that assumption and add an 'assert' to verify.

Secondly, while the vma-writer already does the buffer_is_zero check
internally, for PBS we relied on that non-existant behaviour for zero
chunks, so do the buffer_is_zero check manually and pass NULL to the
rust lib in case it is true.

Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
---

This increases backup speed to PBS for VMs with free space quite dramatically.

 pve-backup.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/pve-backup.c b/pve-backup.c
index 40d8136f1a..7c99554514 100644
--- a/pve-backup.c
+++ b/pve-backup.c
@@ -8,6 +8,7 @@
 #include "block/blockjob.h"
 #include "qapi/qapi-commands-block.h"
 #include "qapi/qmp/qerror.h"
+#include "qemu/cutils.h"
 
 /* PVE backup state and related function */
 
@@ -136,10 +137,13 @@ pvebackup_co_dump_pbs_cb(
     PVEBackupDevInfo *di = opaque;
 
     assert(backup_state.pbs);
+    assert(buf);
 
     Error *local_err = NULL;
     int pbs_res = -1;
 
+    bool is_zero_block = size == di->block_size && buffer_is_zero(buf, size);
+
     qemu_co_mutex_lock(&backup_state.dump_callback_mutex);
 
     // avoid deadlock if job is cancelled
@@ -155,7 +159,8 @@ pvebackup_co_dump_pbs_cb(
         uint64_t to_transfer = left < di->block_size ? left : di->block_size;
 
         pbs_res = proxmox_backup_co_write_data(backup_state.pbs, di->dev_id,
-            buf ? buf + transferred : NULL, start + transferred, to_transfer, &local_err);
+            is_zero_block ? NULL : buf + transferred, start + transferred,
+            to_transfer, &local_err);
         transferred += to_transfer;
 
         if (pbs_res < 0) {
@@ -168,7 +173,7 @@ pvebackup_co_dump_pbs_cb(
     }
 
     qemu_co_mutex_unlock(&backup_state.dump_callback_mutex);
-    pvebackup_add_transfered_bytes(size, !buf ? size : 0, reused);
+    pvebackup_add_transfered_bytes(size, is_zero_block ? size : 0, reused);
 
     return size;
 }
@@ -190,6 +195,7 @@ pvebackup_co_dump_vma_cb(
     int ret = -1;
 
     assert(backup_state.vmaw);
+    assert(buf);
 
     uint64_t remaining = size;
 
@@ -216,9 +222,7 @@ pvebackup_co_dump_vma_cb(
         qemu_co_mutex_unlock(&backup_state.dump_callback_mutex);
 
         ++cluster_num;
-        if (buf) {
-            buf += VMA_CLUSTER_SIZE;
-        }
+        buf += VMA_CLUSTER_SIZE;
         if (ret < 0) {
             Error *local_err = NULL;
             vma_writer_error_propagate(backup_state.vmaw, &local_err);
-- 
2.20.1

[pve-devel] applied: [PATCH qemu] PVE: add zero block handling to PBS dump callback

[Thomas Lamprecht]

On 13.08.20 13:50, Stefan Reiter wrote:
> Both the PBS and VMA dump callbacks assume that a NULL pointer can be
> passed as *pbuf, but that never happens, as backup-dump.c calls this
> function with contents of an iovec.
> 
> So first, remove that assumption and add an 'assert' to verify.
> 
> Secondly, while the vma-writer already does the buffer_is_zero check
> internally, for PBS we relied on that non-existant behaviour for zero
> chunks, so do the buffer_is_zero check manually and pass NULL to the
> rust lib in case it is true.
> 
> Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
> ---
> 
> This increases backup speed to PBS for VMs with free space quite dramatically.
> 
>  pve-backup.c | 14 +++++++++-----
>  1 file changed, 9 insertions(+), 5 deletions(-)
> 
>

applied, thanks!