Skip to contentSkip to conversation

iotests: Let 233 run concurrently

PATCH Superseded Max Reitz 3 participants 19 messages May 7, 2019 – May 8, 2019

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Max Reitz

author
#0 lore raw

cover iotests: Let 233 run concurrently

Currently, 233 cannot reliably run concurrently to other NBD TCP tests. When it starts, it looks for a free port and then attempts to use that for the whole duration of the test run. This is a TOCTTOU race
condition: It does not reserve that port, so another NBD TCP test that runs in parallel can grab it.

To fix this, we must not use the same port all the time, but always choose a new one when qemu-nbd is started. We cannot check whether it is free, but must let qemu-nbd do so and take it atomically. We can achieve this by using the existing --fork option.

There are two problems with --fork, however. First, it does not give us a chance to reliably get the server’s PID, which we need. We can change that by letting qemu-nbd (optionally) write a PID file, though. (Which makes sense if we have a daemon mode.)

Second, it currently discards all output after the server has been started. That looks like an accident to me, because we clearly try to restore the old stderr channel after having redirected all startup messages to the parent process. If it is a bug, we can fix it.

Max Reitz (5):
qemu-nbd: Add --pid-file option
iotests.py: Add qemu_nbd_early_pipe()
qemu-nbd: Do not close stderr
iotests: Use qemu-nbd's --pid-file
iotests: Let 233 run concurrently

 qemu-nbd.c                    | 32 +++++++++++-
 qemu-nbd.texi                 |  2 +
 tests/qemu-iotests/147        |  4 +-
 tests/qemu-iotests/233        |  1 -
 tests/qemu-iotests/common.nbd | 93 ++++++++++++++++-------------------
 tests/qemu-iotests/common.rc  |  4 +-
 tests/qemu-iotests/iotests.py |  9 ++--
 7 files changed, 85 insertions(+), 60 deletions(-)
Signature · 1 line
2.20.1
author
#1 lore raw

PATCH 1/5 qemu-nbd: Add --pid-file option

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
qemu-nbd.c +29 -0
index dca9e72cee..7e48154f44 100644
@@ -59,6 +59,7 @@
 #define QEMU_NBD_OPT_IMAGE_OPTS    262
 #define QEMU_NBD_OPT_FORK          263
 #define QEMU_NBD_OPT_TLSAUTHZ      264
+#define QEMU_NBD_OPT_PID_FILE      265
 
 #define MBR_SIZE 512
 
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"
 #if HAVE_NBD_DEVICE
 "\n"
 "Kernel NBD client support:\n"
@@ -651,6 +653,7 @@ int main(int argc, char **argv)
         { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
         { "trace", required_argument, NULL, 'T' },
         { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
+        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
         { NULL, 0, NULL, 0 }
     };
     int ch;
@@ -677,6 +680,8 @@ int main(int argc, char **argv)
     bool list = false;
     int old_stderr = -1;
     unsigned socket_activation;
+    const char *pid_path = NULL;
+    FILE *pid_file;
 
     /* The client thread uses SIGTERM to interrupt the server.  A signal
      * handler ensures that "qemu-nbd -v -c" exits with a nice status code.
@@ -876,6 +881,9 @@ int main(int argc, char **argv)
         case 'L':
             list = true;
             break;
+        case QEMU_NBD_OPT_PID_FILE:
+            pid_path = optarg;
+            break;
         }
     }
 
@@ -1196,6 +1204,27 @@ int main(int argc, char **argv)
 
     nbd_update_server_watch();
 
+    if (pid_path) {
+        pid_file = fopen(pid_path, "w");
+        if (!pid_file) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(errno));
+            exit(EXIT_FAILURE);
+        }
+
+        ret = fprintf(pid_file, "%ld", (long)getpid());
+        if (ret < 0) {
+            ret = -errno;
+        }
+        fclose(pid_file);
+
+        if (ret < 0) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(-ret));
+            exit(EXIT_FAILURE);
+        }
+    }
+
     /* now when the initialization is (almost) complete, chdir("/")
      * to free any busy filesystems */
     if (chdir("/") < 0) {
qemu-nbd.texi +2 -0
index de342c76b8..7f55657722 100644
@@ -117,6 +117,8 @@ option; or provide the credentials needed for connecting as a client
 in list mode.
 @item --fork
 Fork off the server process and exit the parent once the server is running.
+@item --pid-file=PATH
+Store the server's process ID in the given file.
 @item --tls-authz=ID
 Specify the ID of a qauthz object previously created with the
 --object option. This will be used to authorize connecting users
Signature · 1 line
2.20.1
Reviewed-by on v1 1/5
#2 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

Quoted 11 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)

qemu-nbd.c static void usage(const char *name) · line 115

Quoted 4 lines
qemu-nbd.c +1 -0
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"

Should --pid-file imply --fork, or be an error if --fork was not supplied? As coded, it writes a pid file regardless of --fork, even though it is less obvious that it is useful in that case. I don't have a strong preference (there doesn't seem to be a useful consensus on what forking daemons should do), but it would at least be worth documenting the intended action (even if that implies a tweak to the patch to match the intent).

Quoted 12 lines
qemu-nbd.c +1 -0
 #if HAVE_NBD_DEVICE
 "\n"
 "Kernel NBD client support:\n"
@@ -651,6 +653,7 @@ int main(int argc, char **argv)
         { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
         { "trace", required_argument, NULL, 'T' },
         { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
+        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
         { NULL, 0, NULL, 0 }
     };
     int ch;

qemu-nbd.c int main(int argc, char **argv) · line 683

Quoted 4 lines
qemu-nbd.c +1 -0
@@ -677,6 +680,8 @@ int main(int argc, char **argv)
     bool list = false;
     int old_stderr = -1;
     unsigned socket_activation;
+    const char *pid_path = NULL;

Bikeshedding: pid_name is nicer (path makes me think of $PATH and other colon-separated lists, which this is not).

Otherwise, I agree that this is long overdue. Thanks! If you can justify the behavior without --fork,

Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#3 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

Quoted 11 lines · Eric Blake
Quoted 9 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
Quoted 13 lines

qemu-nbd.c static void usage(const char *name) · line 115

Quoted 5 lines
qemu-nbd.c +1 -0
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"

Should --pid-file imply --fork, or be an error if --fork was not supplied? As coded, it writes a pid file regardless of --fork, even though it is less obvious that it is useful in that case. I don't have a strong preference (there doesn't seem to be a useful consensus on what forking daemons should do), but it would at least be worth documenting the intended action (even if that implies a tweak to the patch to match the intent).

I think the documentation is pretty clear. It stores the server's PID, whether it has been forked or not.

I don't think we would gain anything from forbidding --pid-file without --fork, would we?

Quoted 11 lines
Quoted 11 lines
qemu-nbd.c +1 -0
 #if HAVE_NBD_DEVICE
 "\n"
 "Kernel NBD client support:\n"
@@ -651,6 +653,7 @@ int main(int argc, char **argv)
         { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
         { "trace", required_argument, NULL, 'T' },
         { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
+        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
         { NULL, 0, NULL, 0 }
     };
     int ch;
Quoted 8 lines

qemu-nbd.c int main(int argc, char **argv) · line 683

Quoted 5 lines
qemu-nbd.c +1 -0
@@ -677,6 +680,8 @@ int main(int argc, char **argv)
     bool list = false;
     int old_stderr = -1;
     unsigned socket_activation;
+    const char *pid_path = NULL;

Bikeshedding: pid_name is nicer (path makes me think of $PATH and other colon-separated lists, which this is not).

I'd prefer pid_filename myself, then, because pid_name sounds like a weird way to say "process name". O:-)

Quoted 2 lines

Otherwise, I agree that this is long overdue. Thanks! If you can justify the behavior without --fork,

I just can’t think of a reason not to allow it without --fork. Maybe a user doesn’t need --fork because they just start the server in the background and that’s good enough, but they still want a PID file. So basically like common.rc’s _qemu_nbd_wrapper() before this series.

Max

#4 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

Quoted 26 lines · Max Reitz
Quoted 11 lines · Eric Blake
Quoted 9 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
Quoted 13 lines

qemu-nbd.c static void usage(const char *name) · line 115

Quoted 5 lines
qemu-nbd.c +1 -0
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"

Should --pid-file imply --fork, or be an error if --fork was not supplied? As coded, it writes a pid file regardless of --fork, even though it is less obvious that it is useful in that case. I don't have a strong preference (there doesn't seem to be a useful consensus on what forking daemons should do), but it would at least be worth documenting the intended action (even if that implies a tweak to the patch to match the intent).

Quoted 5 lines

I think the documentation is pretty clear. It stores the server's PID, whether it has been forked or not.

I don't think we would gain anything from forbidding --pid-file without --fork, would we?

I can't think of any reason to forbid it. So it sounds like we are intentional, this writes the pid into --pid-file regardless of whether that pid can be learned by other means as well.

Quoted 7 lines
Quoted 4 lines

qemu-nbd.c int main(int argc, char **argv) · line 683

Quoted 1 line
qemu-nbd.c +1 -0
+    const char *pid_path = NULL;

Bikeshedding: pid_name is nicer (path makes me think of $PATH and other colon-separated lists, which this is not).

I'd prefer pid_filename myself, then, because pid_name sounds like a weird way to say "process name". O:-)

Works for me, even if it is longer. Do you want to respin, or just have me touch it up when folding it into my NBD tree?

Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#5 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

Quoted 40 lines · Eric Blake
Quoted 26 lines · Max Reitz
Quoted 11 lines · Eric Blake
Quoted 9 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
Quoted 13 lines

qemu-nbd.c static void usage(const char *name) · line 115

Quoted 5 lines
qemu-nbd.c +1 -0
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"

Should --pid-file imply --fork, or be an error if --fork was not supplied? As coded, it writes a pid file regardless of --fork, even though it is less obvious that it is useful in that case. I don't have a strong preference (there doesn't seem to be a useful consensus on what forking daemons should do), but it would at least be worth documenting the intended action (even if that implies a tweak to the patch to match the intent).

Quoted 5 lines

I think the documentation is pretty clear. It stores the server's PID, whether it has been forked or not.

I don't think we would gain anything from forbidding --pid-file without --fork, would we?

I can't think of any reason to forbid it. So it sounds like we are intentional, this writes the pid into --pid-file regardless of whether that pid can be learned by other means as well.

Quoted 2 lines
Quoted 2 lines
Quoted 1 line
qemu-nbd.c +1 -0
+    const char *pid_path = NULL;
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

Bikeshedding: pid_name is nicer (path makes me think of $PATH and other colon-separated lists, which this is not).

I'd prefer pid_filename myself, then, because pid_name sounds like a weird way to say "process name". O:-)

Works for me, even if it is longer. Do you want to respin, or just have me touch it up when folding it into my NBD tree?

I suppose I’d prefer a respin, independently of what you make of patches 4 and 5.

Max

#6 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

In reply to Max Reitz
Quoted 26 lines · Max Reitz
Quoted 11 lines · Eric Blake
Quoted 9 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
Quoted 13 lines

qemu-nbd.c static void usage(const char *name) · line 115

Quoted 5 lines
qemu-nbd.c +1 -0
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"

Should --pid-file imply --fork, or be an error if --fork was not supplied? As coded, it writes a pid file regardless of --fork, even though it is less obvious that it is useful in that case. I don't have a strong preference (there doesn't seem to be a useful consensus on what forking daemons should do), but it would at least be worth documenting the intended action (even if that implies a tweak to the patch to match the intent).

Quoted 5 lines

I think the documentation is pretty clear. It stores the server's PID, whether it has been forked or not.

I don't think we would gain anything from forbidding --pid-file without --fork, would we?

Indeed, use of --pid-file should be independant of --fork, as a mgmt app may have already forked it into the background, and merely want to get the pidfile

Regards,
Daniel

Signature · 3 lines
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
#7 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

In reply to Max Reitz
Quoted 70 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
qemu-nbd.c +18 -0
index dca9e72cee..7e48154f44 100644
@@ -59,6 +59,7 @@
 #define QEMU_NBD_OPT_IMAGE_OPTS    262
 #define QEMU_NBD_OPT_FORK          263
 #define QEMU_NBD_OPT_TLSAUTHZ      264
+#define QEMU_NBD_OPT_PID_FILE      265
 
 #define MBR_SIZE 512
 
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"
 #if HAVE_NBD_DEVICE
 "\n"
 "Kernel NBD client support:\n"
@@ -651,6 +653,7 @@ int main(int argc, char **argv)
         { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
         { "trace", required_argument, NULL, 'T' },
         { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
+        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
         { NULL, 0, NULL, 0 }
     };
     int ch;
@@ -677,6 +680,8 @@ int main(int argc, char **argv)
     bool list = false;
     int old_stderr = -1;
     unsigned socket_activation;
+    const char *pid_path = NULL;
+    FILE *pid_file;
 
     /* The client thread uses SIGTERM to interrupt the server.  A signal
      * handler ensures that "qemu-nbd -v -c" exits with a nice status code.
@@ -876,6 +881,9 @@ int main(int argc, char **argv)
         case 'L':
             list = true;
             break;
+        case QEMU_NBD_OPT_PID_FILE:
+            pid_path = optarg;
+            break;
         }
     }
 
@@ -1196,6 +1204,27 @@ int main(int argc, char **argv)
 
     nbd_update_server_watch();
 
+    if (pid_path) {
+        pid_file = fopen(pid_path, "w");
+        if (!pid_file) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(errno));
+            exit(EXIT_FAILURE);
+        }
+
+        ret = fprintf(pid_file, "%ld", (long)getpid());
+        if (ret < 0) {

qemu-nbd.c int main(int argc, char **argv) · line 1226

Quoted 10 lines
qemu-nbd.c +10 -0
+            ret = -errno;
+        }
+        fclose(pid_file);
+
+        if (ret < 0) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(-ret));
+            exit(EXIT_FAILURE);
+        }
+    }

This is racy because multiple qemu-nbd's can be started pointing to the same pidfile and one will blindly overwrite the other.

Please use qemu_write_pidfile instead which acquires locks on the pidfile in a race free manner.

Regards,
Daniel

Signature · 3 lines
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
author
#8 lore raw

Re: [Qemu-devel] [PATCH 1/5] qemu-nbd: Add --pid-file option

Quoted 85 lines · Daniel P. Berrangé
Quoted 70 lines · Max Reitz

--fork is a bit boring if there is no way to get the child's PID. This option helps.

---
 qemu-nbd.c    | 29 +++++++++++++++++++++++++++++
 qemu-nbd.texi |  2 ++
 2 files changed, 31 insertions(+)
qemu-nbd.c +18 -0
index dca9e72cee..7e48154f44 100644
@@ -59,6 +59,7 @@
 #define QEMU_NBD_OPT_IMAGE_OPTS    262
 #define QEMU_NBD_OPT_FORK          263
 #define QEMU_NBD_OPT_TLSAUTHZ      264
+#define QEMU_NBD_OPT_PID_FILE      265
 
 #define MBR_SIZE 512
 
@@ -111,6 +112,7 @@ static void usage(const char *name)
 "                            specify tracing options\n"
 "  --fork                    fork off the server process and exit the parent\n"
 "                            once the server is running\n"
+"  --pid-file=PATH           store the server's process ID in the given file\n"
 #if HAVE_NBD_DEVICE
 "\n"
 "Kernel NBD client support:\n"
@@ -651,6 +653,7 @@ int main(int argc, char **argv)
         { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
         { "trace", required_argument, NULL, 'T' },
         { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
+        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
         { NULL, 0, NULL, 0 }
     };
     int ch;
@@ -677,6 +680,8 @@ int main(int argc, char **argv)
     bool list = false;
     int old_stderr = -1;
     unsigned socket_activation;
+    const char *pid_path = NULL;
+    FILE *pid_file;
 
     /* The client thread uses SIGTERM to interrupt the server.  A signal
      * handler ensures that "qemu-nbd -v -c" exits with a nice status code.
@@ -876,6 +881,9 @@ int main(int argc, char **argv)
         case 'L':
             list = true;
             break;
+        case QEMU_NBD_OPT_PID_FILE:
+            pid_path = optarg;
+            break;
         }
     }
 
@@ -1196,6 +1204,27 @@ int main(int argc, char **argv)
 
     nbd_update_server_watch();
 
+    if (pid_path) {
+        pid_file = fopen(pid_path, "w");
+        if (!pid_file) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(errno));
+            exit(EXIT_FAILURE);
+        }
+
+        ret = fprintf(pid_file, "%ld", (long)getpid());
+        if (ret < 0) {

qemu-nbd.c int main(int argc, char **argv) · line 1226

Quoted 10 lines
qemu-nbd.c +10 -0
+            ret = -errno;
+        }
+        fclose(pid_file);
+
+        if (ret < 0) {
+            error_report("Failed to store PID in %s: %s",
+                         pid_path, strerror(-ret));
+            exit(EXIT_FAILURE);
+        }
+    }

This is racy because multiple qemu-nbd's can be started pointing to the same pidfile and one will blindly overwrite the other.

Quoted 2 lines

Please use qemu_write_pidfile instead which acquires locks on the pidfile in a race free manner.

Ah, nice, that makes things better and easier. :-)

Max

author
#9 lore raw

PATCH 2/5 iotests.py: Add qemu_nbd_early_pipe()

In reply to Max Reitz

qemu_nbd_pipe() currently unconditionally reads qemu-nbd's output. That is not ideal because qemu-nbd may keep stderr open after the parent process has exited.

Currently, the only user of qemu_nbd_pipe() is 147, which discards the whole output if the parent process returned success and only evaluates it on error. Therefore, we can replace qemu_nbd_pipe() by qemu_nbd_early_pipe() that does the same: Discard the output on success, and return it on error.

---
 tests/qemu-iotests/147        | 4 ++--
 tests/qemu-iotests/iotests.py | 9 ++++++---
 2 files changed, 8 insertions(+), 5 deletions(-)
tests/qemu-iotests/147 +2 -2
index 82513279b0..2d84fddb01 100755
@@ -24,7 +24,7 @@ import socket
 import stat
 import time
 import iotests
-from iotests import cachemode, imgfmt, qemu_img, qemu_nbd, qemu_nbd_pipe
+from iotests import cachemode, imgfmt, qemu_img, qemu_nbd, qemu_nbd_early_pipe
 
 NBD_PORT_START      = 32768
 NBD_PORT_END        = NBD_PORT_START + 1024
@@ -93,7 +93,7 @@ class QemuNBD(NBDBlockdevAddBase):
             pass
 
     def _try_server_up(self, *args):
-        status, msg = qemu_nbd_pipe('-f', imgfmt, test_img, *args)
+        status, msg = qemu_nbd_early_pipe('-f', imgfmt, test_img, *args)
         if status == 0:
             return True
         if 'Address already in use' in msg:
tests/qemu-iotests/iotests.py +6 -3
index f811f69135..ce21d83182 100644
@@ -204,9 +204,9 @@ def qemu_nbd(*args):
     '''Run qemu-nbd in daemon mode and return the parent's exit code'''
     return subprocess.call(qemu_nbd_args + ['--fork'] + list(args))
 
-def qemu_nbd_pipe(*args):
+def qemu_nbd_early_pipe(*args):
     '''Run qemu-nbd in daemon mode and return both the parent's exit code
-       and its output'''
+       and its output in case of an error'''
     subp = subprocess.Popen(qemu_nbd_args + ['--fork'] + list(args),
                             stdout=subprocess.PIPE,
                             stderr=subprocess.STDOUT,
@@ -216,7 +216,10 @@ def qemu_nbd_pipe(*args):
         sys.stderr.write('qemu-nbd received signal %i: %s\n' %
                          (-exitcode,
                           ' '.join(qemu_nbd_args + ['--fork'] + list(args))))
-    return exitcode, subp.communicate()[0]
+    if exitcode == 0:
+        return exitcode, ''
+    else:
+        return exitcode, subp.communicate()[0]
 
 def compare_images(img1, img2, fmt1=imgfmt, fmt2=imgfmt):
     '''Return True if two image files are identical'''
Signature · 1 line
2.20.1
Reviewed-by on v1 2/5
#10 lore raw

Re: [Qemu-devel] [PATCH 2/5] iotests.py: Add qemu_nbd_early_pipe()

Quoted 18 lines · Max Reitz

qemu_nbd_pipe() currently unconditionally reads qemu-nbd's output. That is not ideal because qemu-nbd may keep stderr open after the parent process has exited.

Currently, the only user of qemu_nbd_pipe() is 147, which discards the whole output if the parent process returned success and only evaluates it on error. Therefore, we can replace qemu_nbd_pipe() by qemu_nbd_early_pipe() that does the same: Discard the output on success, and return it on error.

---
 tests/qemu-iotests/147        | 4 ++--
 tests/qemu-iotests/iotests.py | 9 ++++++---
 2 files changed, 8 insertions(+), 5 deletions(-)

tests/qemu-iotests/iotests.py def qemu_nbd_pipe(*args): · line 223

Quoted 9 lines
tests/qemu-iotests/iotests.py +4 -1
@@ -216,7 +216,10 @@ def qemu_nbd_pipe(*args):
         sys.stderr.write('qemu-nbd received signal %i: %s\n' %
                          (-exitcode,
                           ' '.join(qemu_nbd_args + ['--fork'] + list(args))))
-    return exitcode, subp.communicate()[0]
+    if exitcode == 0:
+        return exitcode, ''
+    else:
+        return exitcode, subp.communicate()[0]
 
Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#11 lore raw

PATCH 3/5 qemu-nbd: Do not close stderr

In reply to Max Reitz

We kept old_stderr specifically so we could keep emitting error message on stderr. However, qemu_daemon() closes stderr. Therefore, we need to dup() stderr to old_stderr before invoking qemu_daemon().

---
As I hinted at in the cover letter, I am not sure whether this is truly a bug or whether the current behavior is intentional. So if you disagree with me on this patch, you are welcome to suggest an alternative.

I personally can see two:
(1) Add a --log option for a file to store the server's messages in. Seems a bit cumbersome to me.

(2) Add a --keep-stderr option, which specifically enables this behavior here. Without this option we keep the old behavior.

---
 qemu-nbd.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
qemu-nbd.c +2 -1
index 7e48154f44..3805324153 100644
@@ -1015,10 +1015,11 @@ int main(int argc, char **argv)
             exit(EXIT_FAILURE);
         } else if (pid == 0) {
             close(stderr_fd[0]);
+
+            old_stderr = dup(STDERR_FILENO);
             ret = qemu_daemon(1, 0);
 
             /* Temporarily redirect stderr to the parent's pipe...  */
-            old_stderr = dup(STDERR_FILENO);
             dup2(stderr_fd[1], STDERR_FILENO);
             if (ret < 0) {
                 error_report("Failed to daemonize: %s", strerror(errno));
Signature · 1 line
2.20.1
Reviewed-by on v1 3/5
#12 lore raw

Re: [Qemu-devel] [PATCH 3/5] qemu-nbd: Do not close stderr

Quoted 10 lines · Max Reitz

We kept old_stderr specifically so we could keep emitting error message on stderr. However, qemu_daemon() closes stderr. Therefore, we need to dup() stderr to old_stderr before invoking qemu_daemon().

---
As I hinted at in the cover letter, I am not sure whether this is truly a bug or whether the current behavior is intentional. So if you disagree with me on this patch, you are welcome to suggest an alternative.

Looks to me like a bug fix, for a problem present since commit c1f8fdc3 added old_stderr in 2011.

Quoted 7 lines

I personally can see two:
(1) Add a --log option for a file to store the server's messages in. Seems a bit cumbersome to me.

(2) Add a --keep-stderr option, which specifically enables this behavior here. Without this option we keep the old behavior.

The approach done here is simplest, I wouldn't worry about your two alternatives.

Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#13 lore raw

PATCH 4/5 iotests: Use qemu-nbd's --pid-file

In reply to Max Reitz
---
 tests/qemu-iotests/common.rc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tests/qemu-iotests/common.rc +2 -2
index 93f87389b6..217cf3874d 100644
@@ -106,8 +106,8 @@ _qemu_io_wrapper()
 _qemu_nbd_wrapper()
 {
     (
-        echo $BASHPID > "${QEMU_TEST_DIR}/qemu-nbd.pid"
-        exec "$QEMU_NBD_PROG" $QEMU_NBD_OPTIONS "$@"
+        exec "$QEMU_NBD_PROG" --pid-file="${QEMU_TEST_DIR}/qemu-nbd.pid" \
+                              $QEMU_NBD_OPTIONS "$@"
     )
 }
 
Signature · 1 line
2.20.1
#14 lore raw

Re: [Qemu-devel] [PATCH 4/5] iotests: Use qemu-nbd's --pid-file

Quoted 10 lines · Max Reitz
---
 tests/qemu-iotests/common.rc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tests/qemu-iotests/common.rc +0 -0
index 93f87389b6..217cf3874d 100644

tests/qemu-iotests/common.rc _qemu_io_wrapper() · line 111

Quoted 8 lines
tests/qemu-iotests/common.rc +2 -2
@@ -106,8 +106,8 @@ _qemu_io_wrapper()
 _qemu_nbd_wrapper()
 {
     (
-        echo $BASHPID > "${QEMU_TEST_DIR}/qemu-nbd.pid"
-        exec "$QEMU_NBD_PROG" $QEMU_NBD_OPTIONS "$@"
+        exec "$QEMU_NBD_PROG" --pid-file="${QEMU_TEST_DIR}/qemu-nbd.pid" \
+                              $QEMU_NBD_OPTIONS "$@"
     )

Beforehand, we needed the subshell + exec to guarantee that the pid we were writing was that of the subshell. Now, we don't need either; this could be simplified to:

_qemu_nbd_wrapper()

{
    "$QEMU_NBD_PROG" --pid-file... "$@"
}
Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#15 lore raw

Re: [Qemu-devel] [PATCH 4/5] iotests: Use qemu-nbd's --pid-file

Quoted 24 lines · Eric Blake
Quoted 10 lines · Max Reitz
---
 tests/qemu-iotests/common.rc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tests/qemu-iotests/common.rc +0 -0
index 93f87389b6..217cf3874d 100644

tests/qemu-iotests/common.rc _qemu_io_wrapper() · line 111

Quoted 8 lines
tests/qemu-iotests/common.rc +2 -2
@@ -106,8 +106,8 @@ _qemu_io_wrapper()
 _qemu_nbd_wrapper()
 {
     (
-        echo $BASHPID > "${QEMU_TEST_DIR}/qemu-nbd.pid"
-        exec "$QEMU_NBD_PROG" $QEMU_NBD_OPTIONS "$@"
+        exec "$QEMU_NBD_PROG" --pid-file="${QEMU_TEST_DIR}/qemu-nbd.pid" \
+                              $QEMU_NBD_OPTIONS "$@"
     )

Beforehand, we needed the subshell + exec to guarantee that the pid we were writing was that of the subshell. Now, we don't need either; this could be simplified to:

Quoted 4 lines

_qemu_nbd_wrapper()

{
    "$QEMU_NBD_PROG" --pid-file... "$@"
}

True, but I just followed _qemu_img_wrapper()’s example. I could change both, of course...

Max

author
#16 lore raw

PATCH 5/5 iotests: Let 233 run concurrently

In reply to Max Reitz

common.nbd's nbd_server_set_tcp_port() tries to find a free port, and then uses it for the whole test run. However, this is racey because even if the port was free at the beginning, there is no guarantee it will continue to be available. Therefore, 233 currently cannot reliably be run concurrently with other NBD TCP tests.

This patch addresses the problem by dropping nbd_server_set_tcp_port(), and instead finding a new port every time nbd_server_start_tcp_socket() is invoked. For this, we run qemu-nbd with --fork and on error evaluate the output to see whether it contains "Address already in use". If so, we try the next port.

On success, we still want to continually redirect the output from qemu-nbd to stderr. To achieve both, we redirect qemu-nbd's stderr to a FIFO that we then open in bash. If the parent process exits with status 0 (which means that the server has started successfully), we launch a background cat process that copies the FIFO to stderr. On failure, we read the whole content into a variable and then evaluate it.

While at it, use --fork in nbd_server_start_unix_socket(), too. Doing so allows us to drop nbd_server_wait_for_*_socket().

Note that the reason common.nbd did not use --fork before is that qemu-nbd did not have --pid-file.

---
 tests/qemu-iotests/233        |  1 -
 tests/qemu-iotests/common.nbd | 93 ++++++++++++++++-------------------
 2 files changed, 42 insertions(+), 52 deletions(-)
tests/qemu-iotests/233 +0 -1
index b8b6c8cc4c..8682ea277c 100755
@@ -50,7 +50,6 @@ _supported_proto file
 _supported_os Linux
 _require_command QEMU_NBD
 
-nbd_server_set_tcp_port
 tls_x509_init
 
 echo
tests/qemu-iotests/common.nbd +42 -51

124 lines not on this page · raw message

Signature · 1 line
2.20.1
Reviewed-by on v1 5/5
#17 lore raw
Quoted 2 lines · Max Reitz

common.nbd's nbd_server_set_tcp_port() tries to find a free port, and then uses it for the whole test run. However, this is racey because

racy

Quoted 38 lines

even if the port was free at the beginning, there is no guarantee it will continue to be available. Therefore, 233 currently cannot reliably be run concurrently with other NBD TCP tests.

This patch addresses the problem by dropping nbd_server_set_tcp_port(), and instead finding a new port every time nbd_server_start_tcp_socket() is invoked. For this, we run qemu-nbd with --fork and on error evaluate the output to see whether it contains "Address already in use". If so, we try the next port.

On success, we still want to continually redirect the output from qemu-nbd to stderr. To achieve both, we redirect qemu-nbd's stderr to a FIFO that we then open in bash. If the parent process exits with status 0 (which means that the server has started successfully), we launch a background cat process that copies the FIFO to stderr. On failure, we read the whole content into a variable and then evaluate it.

While at it, use --fork in nbd_server_start_unix_socket(), too. Doing so allows us to drop nbd_server_wait_for_*_socket().

Note that the reason common.nbd did not use --fork before is that qemu-nbd did not have --pid-file.

---
 tests/qemu-iotests/233        |  1 -
 tests/qemu-iotests/common.nbd | 93 ++++++++++++++++-------------------
 2 files changed, 42 insertions(+), 52 deletions(-)
tests/qemu-iotests/common.nbd +0 -4
@@ -34,76 +39,62 @@ nbd_server_stop()
         fi
     fi
     rm -f "$nbd_unix_socket"
-}
-
-nbd_server_wait_for_unix_socket()
-{

...

tests/qemu-iotests/common.nbd nbd_server_stop() · line 42

Quoted 3 lines
tests/qemu-iotests/common.nbd +1 -2
-    echo "Failed in check of unix socket created by qemu-nbd"
-    exit 1
+    rm -f "$nbd_stderr_fifo"

You could use a single 'rm -f "$nbd_unix_socket" "$nbd_stderr_fifo"'. That's cosmetic, though.

Are we sure that even on failure, our fifo will not fill up and cause deadlock? If the failing qemu-nbd has so much output as to be non-atomic so that it blocks waiting for a reader, but we don't read anything until after qemu-nbd exits after forking the daemon, then we have deadlock. But in the common case, I don't think qemu-nbd ever spits out that much in errors, even when it fails to start whether due to a socket in use or for other reasons. And even if it does hang, it is our testsuite (and our CI tools will probably notice it), rather than our main code.

Otherwise, it's a lot of shell code with quite a few bash-isms, but we already require bash, and I didn't spot anything blatantly wrong.

Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
author
#18 lore raw
Quoted 49 lines · Eric Blake
Quoted 2 lines · Max Reitz

common.nbd's nbd_server_set_tcp_port() tries to find a free port, and then uses it for the whole test run. However, this is racey because

racy

Quoted 38 lines

even if the port was free at the beginning, there is no guarantee it will continue to be available. Therefore, 233 currently cannot reliably be run concurrently with other NBD TCP tests.

This patch addresses the problem by dropping nbd_server_set_tcp_port(), and instead finding a new port every time nbd_server_start_tcp_socket() is invoked. For this, we run qemu-nbd with --fork and on error evaluate the output to see whether it contains "Address already in use". If so, we try the next port.

On success, we still want to continually redirect the output from qemu-nbd to stderr. To achieve both, we redirect qemu-nbd's stderr to a FIFO that we then open in bash. If the parent process exits with status 0 (which means that the server has started successfully), we launch a background cat process that copies the FIFO to stderr. On failure, we read the whole content into a variable and then evaluate it.

While at it, use --fork in nbd_server_start_unix_socket(), too. Doing so allows us to drop nbd_server_wait_for_*_socket().

Note that the reason common.nbd did not use --fork before is that qemu-nbd did not have --pid-file.

---
 tests/qemu-iotests/233        |  1 -
 tests/qemu-iotests/common.nbd | 93 ++++++++++++++++-------------------
 2 files changed, 42 insertions(+), 52 deletions(-)
tests/qemu-iotests/common.nbd +0 -4
@@ -34,76 +39,62 @@ nbd_server_stop()
         fi
     fi
     rm -f "$nbd_unix_socket"
-}
-
-nbd_server_wait_for_unix_socket()
-{

...

tests/qemu-iotests/common.nbd nbd_server_stop() · line 42

Quoted 3 lines
tests/qemu-iotests/common.nbd +1 -2
-    echo "Failed in check of unix socket created by qemu-nbd"
-    exit 1
+    rm -f "$nbd_stderr_fifo"
Quoted 7 lines

You could use a single 'rm -f "$nbd_unix_socket" "$nbd_stderr_fifo"'. That's cosmetic, though.

Are we sure that even on failure, our fifo will not fill up and cause deadlock? If the failing qemu-nbd has so much output as to be non-atomic so that it blocks waiting for a reader, but we don't read anything until after qemu-nbd exits after forking the daemon, then we have deadlock.

Hm, right. I don’t think it will happen, but if it does, it won’t be because of an “Address already in use”. So if it did happen, the test should fail anyway.

Of course, a hang is not the nicest way to fail a test, but I think as long as we don’t think it will be a problem, it should be fine.

(The alternative I can think of would be to start a background cat that copies data over to a log file, and then kill it after the qemu-nbd parent process has exited. On error, we read the log; on success, we print it to stderr and then start the cat from nbd_stderr_fifo to stderr.)

Quoted 9 lines

But in the common case, I don't think qemu-nbd ever spits out that much in errors, even when it fails to start whether due to a socket in use or for other reasons. And even if it does hang, it is our testsuite (and our CI tools will probably notice it), rather than our main code.

Otherwise, it's a lot of shell code with quite a few bash-isms, but we already require bash, and I didn't spot anything blatantly wrong.

Thanks again!

I’ll prepare the v2.

Max