[Qemu-devel] [PATCH v4 5/8] nbd/server: Include human-readable message in structured errors

Eric Blake posted 8 patches 8 years, 3 months ago
[Qemu-devel] [PATCH v4 5/8] nbd/server: Include human-readable message in structured errors
Posted by Eric Blake 8 years, 3 months ago
The NBD spec permits including a human-readable error string if
structured replies are in force, so we might as well send the
client the message that we logged on any error.

Signed-off-by: Eric Blake <eblake@redhat.com>
---
 nbd/server.c     | 22 ++++++++++++++++------
 nbd/trace-events |  2 +-
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/nbd/server.c b/nbd/server.c
index 23dc6be708..8085d79076 100644
--- a/nbd/server.c
+++ b/nbd/server.c
@@ -1289,23 +1289,25 @@ static int coroutine_fn nbd_co_send_structured_read(NBDClient *client,
 static int coroutine_fn nbd_co_send_structured_error(NBDClient *client,
                                                      uint64_t handle,
                                                      uint32_t error,
+                                                     char *msg,
                                                      Error **errp)
 {
     NBDStructuredError chunk;
     int nbd_err = system_errno_to_nbd_errno(error);
     struct iovec iov[] = {
         {.iov_base = &chunk, .iov_len = sizeof(chunk)},
-        /* FIXME: Support human-readable error message */
+        {.iov_base = msg, .iov_len = strlen(msg)},
     };

     assert(nbd_err);
-    trace_nbd_co_send_structured_error(handle, nbd_err);
+    trace_nbd_co_send_structured_error(handle, nbd_err,
+                                       nbd_err_lookup(nbd_err), msg);
     set_be_chunk(&chunk.h, NBD_REPLY_FLAG_DONE, NBD_REPLY_TYPE_ERROR, handle,
                  sizeof(chunk) - sizeof(chunk.h));
     stl_be_p(&chunk.error, nbd_err);
-    stw_be_p(&chunk.message_length, 0);
+    stw_be_p(&chunk.message_length, iov[1].iov_len);

-    return nbd_co_send_iov(client, iov, 1, errp);
+    return nbd_co_send_iov(client, iov, 1 + !!iov[1].iov_len, errp);
 }

 /* nbd_co_receive_request
@@ -1406,6 +1408,7 @@ static coroutine_fn void nbd_trip(void *opaque)
     int flags;
     int reply_data_len = 0;
     Error *local_err = NULL;
+    char *msg = NULL;

     trace_nbd_trip();
     if (client->closing) {
@@ -1521,14 +1524,20 @@ reply:
     if (local_err) {
         /* If we get here, local_err was not a fatal error, and should be sent
          * to the client. */
+        assert(ret < 0);
+        msg = g_strdup(error_get_pretty(local_err));
         error_report_err(local_err);
         local_err = NULL;
     }

-    if (client->structured_reply && request.type == NBD_CMD_READ) {
+    if (client->structured_reply &&
+        (ret < 0 || request.type == NBD_CMD_READ)) {
         if (ret < 0) {
+            if (!msg) {
+                msg = g_strdup("");
+            }
             ret = nbd_co_send_structured_error(req->client, request.handle,
-                                               -ret, &local_err);
+                                               -ret, msg, &local_err);
         } else {
             ret = nbd_co_send_structured_read(req->client, request.handle,
                                               request.from, req->data,
@@ -1539,6 +1548,7 @@ reply:
                                        ret < 0 ? -ret : 0,
                                        req->data, reply_data_len, &local_err);
     }
+    g_free(msg);
     if (ret < 0) {
         error_prepend(&local_err, "Failed to send reply: ");
         goto disconnect;
diff --git a/nbd/trace-events b/nbd/trace-events
index 15a9294445..880f211c07 100644
--- a/nbd/trace-events
+++ b/nbd/trace-events
@@ -57,7 +57,7 @@ nbd_blk_aio_attached(const char *name, void *ctx) "Export %s: Attaching clients
 nbd_blk_aio_detach(const char *name, void *ctx) "Export %s: Detaching clients from AIO context %p\n"
 nbd_co_send_simple_reply(uint64_t handle, uint32_t error, const char *errname, int len) "Send simple reply: handle = %" PRIu64 ", error = %" PRIu32 " (%s), len = %d"
 nbd_co_send_structured_read(uint64_t handle, uint64_t offset, void *data, size_t size) "Send structured read data reply: handle = %" PRIu64 ", offset = %" PRIu64 ", data = %p, len = %zu"
-nbd_co_send_structured_error(uint64_t handle, int err) "Send structured error reply: handle = %" PRIu64 ", error = %d"
+nbd_co_send_structured_error(uint64_t handle, int err, const char *errname, const char *msg) "Send structured error reply: handle = %" PRIu64 ", error = %d (%s), msg = '%s'"
 nbd_co_receive_request_decode_type(uint64_t handle, uint16_t type, const char *name) "Decoding type: handle = %" PRIu64 ", type = %" PRIu16 " (%s)"
 nbd_co_receive_request_payload_received(uint64_t handle, uint32_t len) "Payload received: handle = %" PRIu64 ", len = %" PRIu32
 nbd_co_receive_request_cmd_write(uint32_t len) "Reading %" PRIu32 " byte(s)"
-- 
2.13.6


Re: [Qemu-devel] [PATCH v4 5/8] nbd/server: Include human-readable message in structured errors
Posted by Vladimir Sementsov-Ogievskiy 8 years, 3 months ago
15.10.2017 04:01, Eric Blake wrote:
> The NBD spec permits including a human-readable error string if
> structured replies are in force, so we might as well send the
> client the message that we logged on any error.
>
> Signed-off-by: Eric Blake <eblake@redhat.com>
> ---
>   nbd/server.c     | 22 ++++++++++++++++------
>   nbd/trace-events |  2 +-
>   2 files changed, 17 insertions(+), 7 deletions(-)
>
> diff --git a/nbd/server.c b/nbd/server.c
> index 23dc6be708..8085d79076 100644
> --- a/nbd/server.c
> +++ b/nbd/server.c
> @@ -1289,23 +1289,25 @@ static int coroutine_fn nbd_co_send_structured_read(NBDClient *client,
>   static int coroutine_fn nbd_co_send_structured_error(NBDClient *client,
>                                                        uint64_t handle,
>                                                        uint32_t error,
> +                                                     char *msg,

it's not const because we want to put it into iov..
may be it is better to make it const and convert to (char *) like in 
qio_channel_write_all, to
1: more native message parameter type
2: allow constat strings like nbd_co_send_structured_error(..., "some 
error")

>                                                        Error **errp)
>   {
>       NBDStructuredError chunk;
>       int nbd_err = system_errno_to_nbd_errno(error);
>       struct iovec iov[] = {
>           {.iov_base = &chunk, .iov_len = sizeof(chunk)},
> -        /* FIXME: Support human-readable error message */
> +        {.iov_base = msg, .iov_len = strlen(msg)},
>       };

msg is always non-zero? so we don't want to send errors without 
message.. (1)

>
>       assert(nbd_err);
> -    trace_nbd_co_send_structured_error(handle, nbd_err);
> +    trace_nbd_co_send_structured_error(handle, nbd_err,
> +                                       nbd_err_lookup(nbd_err), msg);

it doesn't really matter, but "nbd_err_lookup(nbd_err)" is a bit 
unrelated and looks like part of previous patch

>       set_be_chunk(&chunk.h, NBD_REPLY_FLAG_DONE, NBD_REPLY_TYPE_ERROR, handle,
>                    sizeof(chunk) - sizeof(chunk.h));
>       stl_be_p(&chunk.error, nbd_err);
> -    stw_be_p(&chunk.message_length, 0);
> +    stw_be_p(&chunk.message_length, iov[1].iov_len);
>
> -    return nbd_co_send_iov(client, iov, 1, errp);
> +    return nbd_co_send_iov(client, iov, 1 + !!iov[1].iov_len, errp);

but you allow messages of zero length..
it's ok, but looks a bit strange in connection with (1)

>   }
>
>   /* nbd_co_receive_request
> @@ -1406,6 +1408,7 @@ static coroutine_fn void nbd_trip(void *opaque)
>       int flags;
>       int reply_data_len = 0;
>       Error *local_err = NULL;
> +    char *msg = NULL;
>
>       trace_nbd_trip();
>       if (client->closing) {
> @@ -1521,14 +1524,20 @@ reply:
>       if (local_err) {
>           /* If we get here, local_err was not a fatal error, and should be sent
>            * to the client. */
> +        assert(ret < 0);
> +        msg = g_strdup(error_get_pretty(local_err));
>           error_report_err(local_err);
>           local_err = NULL;
>       }
>
> -    if (client->structured_reply && request.type == NBD_CMD_READ) {
> +    if (client->structured_reply &&
> +        (ret < 0 || request.type == NBD_CMD_READ)) {
>           if (ret < 0) {
> +            if (!msg) {
> +                msg = g_strdup("");

I'd prefer to handle msg=NULL in nbd_co_send_structured_error instead of 
this.

> +            }
>               ret = nbd_co_send_structured_error(req->client, request.handle,
> -                                               -ret, &local_err);
> +                                               -ret, msg, &local_err);
>           } else {
>               ret = nbd_co_send_structured_read(req->client, request.handle,
>                                                 request.from, req->data,
> @@ -1539,6 +1548,7 @@ reply:
>                                          ret < 0 ? -ret : 0,
>                                          req->data, reply_data_len, &local_err);
>       }
> +    g_free(msg);
>       if (ret < 0) {
>           error_prepend(&local_err, "Failed to send reply: ");
>           goto disconnect;
> diff --git a/nbd/trace-events b/nbd/trace-events
> index 15a9294445..880f211c07 100644
> --- a/nbd/trace-events
> +++ b/nbd/trace-events
> @@ -57,7 +57,7 @@ nbd_blk_aio_attached(const char *name, void *ctx) "Export %s: Attaching clients
>   nbd_blk_aio_detach(const char *name, void *ctx) "Export %s: Detaching clients from AIO context %p\n"
>   nbd_co_send_simple_reply(uint64_t handle, uint32_t error, const char *errname, int len) "Send simple reply: handle = %" PRIu64 ", error = %" PRIu32 " (%s), len = %d"
>   nbd_co_send_structured_read(uint64_t handle, uint64_t offset, void *data, size_t size) "Send structured read data reply: handle = %" PRIu64 ", offset = %" PRIu64 ", data = %p, len = %zu"
> -nbd_co_send_structured_error(uint64_t handle, int err) "Send structured error reply: handle = %" PRIu64 ", error = %d"
> +nbd_co_send_structured_error(uint64_t handle, int err, const char *errname, const char *msg) "Send structured error reply: handle = %" PRIu64 ", error = %d (%s), msg = '%s'"
>   nbd_co_receive_request_decode_type(uint64_t handle, uint16_t type, const char *name) "Decoding type: handle = %" PRIu64 ", type = %" PRIu16 " (%s)"
>   nbd_co_receive_request_payload_received(uint64_t handle, uint32_t len) "Payload received: handle = %" PRIu64 ", len = %" PRIu32
>   nbd_co_receive_request_cmd_write(uint32_t len) "Reading %" PRIu32 " byte(s)"

anyway it should work, so, with or without my suggestions:

Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>

-- 
Best regards,
Vladimir


Re: [Qemu-devel] [PATCH v4 5/8] nbd/server: Include human-readable message in structured errors
Posted by Eric Blake 8 years, 3 months ago
On 10/16/2017 05:59 AM, Vladimir Sementsov-Ogievskiy wrote:
> 15.10.2017 04:01, Eric Blake wrote:
>> The NBD spec permits including a human-readable error string if
>> structured replies are in force, so we might as well send the
>> client the message that we logged on any error.
>>
>> Signed-off-by: Eric Blake <eblake@redhat.com>
>> ---
>>   nbd/server.c     | 22 ++++++++++++++++------
>>   nbd/trace-events |  2 +-
>>   2 files changed, 17 insertions(+), 7 deletions(-)
>>
>> diff --git a/nbd/server.c b/nbd/server.c
>> index 23dc6be708..8085d79076 100644
>> --- a/nbd/server.c
>> +++ b/nbd/server.c
>> @@ -1289,23 +1289,25 @@ static int coroutine_fn
>> nbd_co_send_structured_read(NBDClient *client,
>>   static int coroutine_fn nbd_co_send_structured_error(NBDClient *client,
>>                                                        uint64_t handle,
>>                                                        uint32_t error,
>> +                                                     char *msg,
> 
> it's not const because we want to put it into iov..
> may be it is better to make it const and convert to (char *) like in
> qio_channel_write_all, to
> 1: more native message parameter type
> 2: allow constat strings like nbd_co_send_structured_error(..., "some
> error")

Yes, casting away const would be a more convenient interface (it's a
bummer that iov can't be const-correct for write, because it is also
used by read which cannot be const).

> 
>>                                                        Error **errp)
>>   {
>>       NBDStructuredError chunk;
>>       int nbd_err = system_errno_to_nbd_errno(error);
>>       struct iovec iov[] = {
>>           {.iov_base = &chunk, .iov_len = sizeof(chunk)},
>> -        /* FIXME: Support human-readable error message */
>> +        {.iov_base = msg, .iov_len = strlen(msg)},
>>       };
> 
> msg is always non-zero? so we don't want to send errors without
> message.. (1)

I played with the idea of allowing NULL, and don't know whether it was
easier to require non-NULL message (which may require NULL check at all
callers) or to allow NULL (requiring more code here).

> 
>>
>>       assert(nbd_err);
>> -    trace_nbd_co_send_structured_error(handle, nbd_err);
>> +    trace_nbd_co_send_structured_error(handle, nbd_err,
>> +                                       nbd_err_lookup(nbd_err), msg);
> 
> it doesn't really matter, but "nbd_err_lookup(nbd_err)" is a bit
> unrelated and looks like part of previous patch

The previous patch was yours; I tried to minimize the changes I made to
your patches, so I stuck this one in mine instead. But if you think we
should trace accurately from the get-go, I'm just fine rebasing the
series to make your patch trace the error name at its introduction,
rather than in my followup that supports error messages.

> 
>>       set_be_chunk(&chunk.h, NBD_REPLY_FLAG_DONE,
>> NBD_REPLY_TYPE_ERROR, handle,
>>                    sizeof(chunk) - sizeof(chunk.h));
>>       stl_be_p(&chunk.error, nbd_err);
>> -    stw_be_p(&chunk.message_length, 0);
>> +    stw_be_p(&chunk.message_length, iov[1].iov_len);
>>
>> -    return nbd_co_send_iov(client, iov, 1, errp);
>> +    return nbd_co_send_iov(client, iov, 1 + !!iov[1].iov_len, errp);
> 
> but you allow messages of zero length..
> it's ok, but looks a bit strange in connection with (1)

Passing "" is the easiest thing to do for callers that can't pass NULL.
My first attempt had:

.iov_len = msg ? strlen(msg) : 0

and

trace_...(msg ? msg : "")

For this patch, there is only a single caller, so it's really hard to
judge which style (required non-NULL parameter, vs. doing more work in
the common function) will be more useful until we add further callers.


>> -    if (client->structured_reply && request.type == NBD_CMD_READ) {
>> +    if (client->structured_reply &&
>> +        (ret < 0 || request.type == NBD_CMD_READ)) {
>>           if (ret < 0) {
>> +            if (!msg) {
>> +                msg = g_strdup("");
> 
> I'd prefer to handle msg=NULL in nbd_co_send_structured_error instead of
> this.

Okay, then it sounds like accepting a NULL parameter is the way to go!

It's also worth considering that casting away const would mean I don't
have to g_strdup("").


> anyway it should work, so, with or without my suggestions:
> 
> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>

Thanks for your reviews!

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org

Re: [Qemu-devel] [PATCH v4 5/8] nbd/server: Include human-readable message in structured errors
Posted by Eric Blake 8 years, 3 months ago
On 10/14/2017 08:01 PM, Eric Blake wrote:
> The NBD spec permits including a human-readable error string if
> structured replies are in force, so we might as well send the
> client the message that we logged on any error.
> 
> Signed-off-by: Eric Blake <eblake@redhat.com>
> ---
>  nbd/server.c     | 22 ++++++++++++++++------
>  nbd/trace-events |  2 +-
>  2 files changed, 17 insertions(+), 7 deletions(-)
> 

>      assert(nbd_err);
> -    trace_nbd_co_send_structured_error(handle, nbd_err);
> +    trace_nbd_co_send_structured_error(handle, nbd_err,
> +                                       nbd_err_lookup(nbd_err), msg);
>      set_be_chunk(&chunk.h, NBD_REPLY_FLAG_DONE, NBD_REPLY_TYPE_ERROR, handle,
>                   sizeof(chunk) - sizeof(chunk.h));

Bug - it's a bad idea to not include the message length in the overall
length, because the client then gets out of sync with the server (it
reads only 6 bytes instead of 6+strlen(msg) bytes, and expects the
message to start with the magic number for the next reply).

>      stl_be_p(&chunk.error, nbd_err);
> -    stw_be_p(&chunk.message_length, 0);
> +    stw_be_p(&chunk.message_length, iov[1].iov_len);

But this also highlights a bug in 9/8, where we have:

> +static int nbd_parse_error_payload(NBDStructuredReplyChunk *chunk,
> +                                   uint8_t *payload, int *request_ret,
> +                                   Error **errp)
> +{
> +    uint32_t error;
> +    uint16_t message_size;
> +
> +    assert(chunk->type & (1 << 15));
> +
> +    if (chunk->length < sizeof(error) + sizeof(message_size)) {
> +        error_setg(errp,
> +                   "Protocol error: invalid payload for structured error");
> +        return -EINVAL;
> +    }
> +
> +    error = nbd_errno_to_system_errno(payload_advance32(&payload));
> +    if (error == 0) {
> +        error_setg(errp, "Protocol error: server sent structured error chunk"
> +                         "with error = 0");
> +        return -EINVAL;
> +    }
> +
> +    *request_ret = error;
> +    message_size = payload_advance16(&payload);
> +    error_setg_errno(errp, error, "%.*s", message_size, payload);

Whoops - no sanity check that message_size fits within chunk->length.
So when we read message_length 33 (when the server sends a message 33
bytes long), we are then dereferencing up to 33 bytes of garbage beyond
the end of payload.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org