qemu/qemu-nbd.c

1186 lines
38 KiB
C
Raw Normal View History

/*
* Copyright (C) 2005 Anthony Liguori <anthony@codemonkey.ws>
*
* Network Block Device
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; under version 2 of the License.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, see <http://www.gnu.org/licenses/>.
*/
#include "qemu/osdep.h"
#include <getopt.h>
#include <libgen.h>
#include <pthread.h>
#include "qemu/help-texts.h"
2016-03-14 11:01:28 +03:00
#include "qapi/error.h"
#include "qemu/cutils.h"
block: New BlockBackend A block device consists of a frontend device model and a backend. A block backend has a tree of block drivers doing the actual work. The tree is managed by the block layer. We currently use a single abstraction BlockDriverState both for tree nodes and the backend as a whole. Drawbacks: * Its API includes both stuff that makes sense only at the block backend level (root of the tree) and stuff that's only for use within the block layer. This makes the API bigger and more complex than necessary. Moreover, it's not obvious which interfaces are meant for device models, and which really aren't. * Since device models keep a reference to their backend, the backend object can't just be destroyed. But for media change, we need to replace the tree. Our solution is to make the BlockDriverState generic, with actual driver state in a separate object, pointed to by member opaque. That lets us replace the tree by deinitializing and reinitializing its root. This special need of the root makes the data structure awkward everywhere in the tree. The general plan is to separate the APIs into "block backend", for use by device models, monitor and whatever other code dealing with block backends, and "block driver", for use by the block layer and whatever other code (if any) dealing with trees and tree nodes. Code dealing with block backends, device models in particular, should become completely oblivious of BlockDriverState. This should let us clean up both APIs, and the tree data structures. This commit is a first step. It creates a minimal "block backend" API: type BlockBackend and functions to create, destroy and find them. BlockBackend objects are created and destroyed exactly when root BlockDriverState objects are created and destroyed. "Root" in the sense of "in bdrv_states". They're not yet used for anything; that'll come shortly. A root BlockDriverState is created with bdrv_new_root(), so where to create a BlockBackend is obvious. Where these roots get destroyed isn't always as obvious. It is obvious in qemu-img.c, qemu-io.c and qemu-nbd.c, and in error paths of blockdev_init(), blk_connect(). That leaves destruction of objects successfully created by blockdev_init() and blk_connect(). blockdev_init() is used only by drive_new() and qmp_blockdev_add(). Objects created by the latter are currently indestructible (see commit 48f364d "blockdev: Refuse to drive_del something added with blockdev-add" and commit 2d246f0 "blockdev: Introduce DriveInfo.enable_auto_del"). Objects created by the former get destroyed by drive_del(). Objects created by blk_connect() get destroyed by blk_disconnect(). BlockBackend is reference-counted. Its reference count never exceeds one so far, but that's going to change. In drive_del(), the BB's reference count is surely one now. The BDS's reference count is greater than one when something else is holding a reference, such as a block job. In this case, the BB is destroyed right away, but the BDS lives on until all extra references get dropped. Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Max Reitz <mreitz@redhat.com> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2014-10-07 15:59:04 +04:00
#include "sysemu/block-backend.h"
#include "sysemu/runstate.h" /* for qemu_system_killed() prototype */
#include "block/block_int.h"
#include "block/nbd.h"
#include "qemu/main-loop.h"
#include "qemu/module.h"
#include "qemu/option.h"
#include "qemu/error-report.h"
#include "qemu/config-file.h"
#include "qemu/bswap.h"
#include "qemu/log.h"
#include "qemu/systemd.h"
#include "block/snapshot.h"
#include "qapi/qmp/qdict.h"
#include "qapi/qmp/qstring.h"
#include "qom/object_interfaces.h"
#include "io/channel-socket.h"
#include "io/net-listener.h"
#include "crypto/init.h"
#include "crypto/tlscreds.h"
#include "trace/control.h"
#include "qemu-version.h"
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
#ifdef CONFIG_SELINUX
#include <selinux/selinux.h>
#endif
#ifdef __linux__
#define HAVE_NBD_DEVICE 1
#else
#define HAVE_NBD_DEVICE 0
#endif
#define SOCKET_PATH "/var/lock/qemu-nbd-%s"
#define QEMU_NBD_OPT_CACHE 256
#define QEMU_NBD_OPT_AIO 257
#define QEMU_NBD_OPT_DISCARD 258
#define QEMU_NBD_OPT_DETECT_ZEROES 259
#define QEMU_NBD_OPT_OBJECT 260
#define QEMU_NBD_OPT_TLSCREDS 261
#define QEMU_NBD_OPT_IMAGE_OPTS 262
#define QEMU_NBD_OPT_FORK 263
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
#define QEMU_NBD_OPT_TLSAUTHZ 264
#define QEMU_NBD_OPT_PID_FILE 265
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
#define QEMU_NBD_OPT_SELINUX_LABEL 266
#define QEMU_NBD_OPT_TLSHOSTNAME 267
#define MBR_SIZE 512
static int verbose;
static char *srcpath;
static SocketAddress *saddr;
static int persistent = 0;
static enum { RUNNING, TERMINATE, TERMINATED } state;
static int shared = 1;
static int nb_fds;
static QIONetListener *server;
static QCryptoTLSCreds *tlscreds;
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
static const char *tlsauthz;
static void usage(const char *name)
{
(printf) (
"Usage: %s [OPTIONS] FILE\n"
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
" or: %s -L [OPTIONS]\n"
"QEMU Disk Network Block Device Utility\n"
"\n"
" -h, --help display this help and exit\n"
" -V, --version output version information and exit\n"
"\n"
"Connection properties:\n"
" -p, --port=PORT port to listen on (default `%d')\n"
" -b, --bind=IFACE interface to bind to (default `0.0.0.0')\n"
" -k, --socket=PATH path to the unix socket\n"
" (default '"SOCKET_PATH"')\n"
" -e, --shared=NUM device can be shared by NUM clients (default '1')\n"
" -t, --persistent don't exit on the last connection\n"
" -v, --verbose display extra debugging information\n"
" -x, --export-name=NAME expose export by name (default is empty string)\n"
" -D, --description=TEXT export a human-readable description\n"
"\n"
"Exposing part of the image:\n"
" -o, --offset=OFFSET offset into the image\n"
" -A, --allocation-depth expose the allocation depth\n"
" -B, --bitmap=NAME expose a persistent dirty bitmap\n"
"\n"
"General purpose options:\n"
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
" -L, --list list exports available from another NBD server\n"
" --object type,id=ID,... define an object such as 'secret' for providing\n"
" passwords and/or encryption keys\n"
" --tls-creds=ID use id of an earlier --object to provide TLS\n"
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
" --tls-authz=ID use id of an earlier --object to provide\n"
" authorization\n"
" -T, --trace [[enable=]<pattern>][,events=<file>][,file=<file>]\n"
" 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"
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
#ifdef CONFIG_SELINUX
" --selinux-label=LABEL set SELinux process label on listening socket\n"
#endif
#if HAVE_NBD_DEVICE
"\n"
"Kernel NBD client support:\n"
" -c, --connect=DEV connect FILE to the local NBD device DEV\n"
" -d, --disconnect disconnect the specified device\n"
#endif
"\n"
"Block device options:\n"
" -f, --format=FORMAT set image format (raw, qcow2, ...)\n"
" -r, --read-only export read-only\n"
" -s, --snapshot use FILE as an external snapshot, create a temporary\n"
" file with backing_file=FILE, redirect the write to\n"
" the temporary one\n"
" -l, --load-snapshot=SNAPSHOT_PARAM\n"
" load an internal snapshot inside FILE and export it\n"
" as an read-only device, SNAPSHOT_PARAM format is\n"
" 'snapshot.id=[ID],snapshot.name=[NAME]', or\n"
" '[ID_OR_NAME]'\n"
" -n, --nocache disable host cache\n"
qemu-nbd: Change default cache mode to writeback Both qemu and qemu-img use writeback cache mode by default, which is already documented in qemu(1). qemu-nbd uses writethrough cache mode by default, and the default cache mode is not documented. According to the qemu-nbd(8): --cache=CACHE The cache mode to be used with the file. See the documentation of the emulator's -drive cache=... option for allowed values. qemu(1) says: The default mode is cache=writeback. So users have no reason to assume that qemu-nbd is using writethough cache mode. The only hint is the painfully slow writing when using the defaults. Looking in git history, it seems that qemu used writethrough in the past to support broken guests that did not flush data properly, or could not flush due to limitations in qemu. But qemu-nbd clients can use NBD_CMD_FLUSH to flush data, so using writethrough does not help anyone. Change the default cache mode to writback, and document the default and available values properly in the online help and manual. With this change converting image via qemu-nbd is 3.5 times faster. $ qemu-img create dst.img 50g $ qemu-nbd -t -f raw -k /tmp/nbd.sock dst.img Before this change: $ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock" Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock Time (mean ± σ): 83.639 s ± 5.970 s [User: 2.733 s, System: 6.112 s] Range (min … max): 76.749 s … 87.245 s 3 runs After this change: $ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock" Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock Time (mean ± σ): 23.522 s ± 0.433 s [User: 2.083 s, System: 5.475 s] Range (min … max): 23.234 s … 24.019 s 3 runs Users can avoid the issue by using --cache=writeback[1] but the defaults should give good performance for the common use case. [1] https://bugzilla.redhat.com/1990656 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20210813205519.50518-1-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> CC: qemu-stable@nongnu.org Signed-off-by: Eric Blake <eblake@redhat.com>
2021-08-13 23:55:19 +03:00
" --cache=MODE set cache mode used to access the disk image, the\n"
" valid options are: 'none', 'writeback' (default),\n"
" 'writethrough', 'directsync' and 'unsafe'\n"
" --aio=MODE set AIO mode (native, io_uring or threads)\n"
" --discard=MODE set discard mode (ignore, unmap)\n"
" --detect-zeroes=MODE set detect-zeroes mode (off, on, unmap)\n"
" --image-opts treat FILE as a full set of image options\n"
"\n"
QEMU_HELP_BOTTOM "\n"
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
, name, name, NBD_DEFAULT_PORT, "DEVICE");
}
static void version(const char *name)
{
printf(
"%s " QEMU_FULL_VERSION "\n"
"Written by Anthony Liguori.\n"
"\n"
QEMU_COPYRIGHT "\n"
"This is free software; see the source for copying conditions. There is NO\n"
"warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n"
, name);
}
#ifdef CONFIG_POSIX
/*
* The client thread uses SIGTERM to interrupt the server. A signal
* handler ensures that "qemu-nbd -v -c" exits with a nice status code.
*/
void qemu_system_killed(int signum, pid_t pid)
{
qatomic_cmpxchg(&state, RUNNING, TERMINATE);
qemu_notify_event();
}
#endif /* CONFIG_POSIX */
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
static int qemu_nbd_client_list(SocketAddress *saddr, QCryptoTLSCreds *tls,
const char *hostname)
{
int ret = EXIT_FAILURE;
int rc;
Error *err = NULL;
QIOChannelSocket *sioc;
NBDExportInfo *list;
int i, j;
sioc = qio_channel_socket_new();
if (qio_channel_socket_connect_sync(sioc, saddr, &err) < 0) {
error_report_err(err);
goto out;
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
}
rc = nbd_receive_export_list(QIO_CHANNEL(sioc), tls, hostname, &list,
&err);
if (rc < 0) {
if (err) {
error_report_err(err);
}
goto out;
}
printf("exports available: %d\n", rc);
for (i = 0; i < rc; i++) {
printf(" export: '%s'\n", list[i].name);
if (list[i].description && *list[i].description) {
printf(" description: %s\n", list[i].description);
}
if (list[i].flags & NBD_FLAG_HAS_FLAGS) {
static const char *const flag_names[] = {
[NBD_FLAG_READ_ONLY_BIT] = "readonly",
[NBD_FLAG_SEND_FLUSH_BIT] = "flush",
[NBD_FLAG_SEND_FUA_BIT] = "fua",
[NBD_FLAG_ROTATIONAL_BIT] = "rotational",
[NBD_FLAG_SEND_TRIM_BIT] = "trim",
[NBD_FLAG_SEND_WRITE_ZEROES_BIT] = "zeroes",
[NBD_FLAG_SEND_DF_BIT] = "df",
[NBD_FLAG_CAN_MULTI_CONN_BIT] = "multi",
[NBD_FLAG_SEND_RESIZE_BIT] = "resize",
[NBD_FLAG_SEND_CACHE_BIT] = "cache",
nbd: Prepare for NBD_CMD_FLAG_FAST_ZERO Commit fe0480d6 and friends added BDRV_REQ_NO_FALLBACK as a way to avoid wasting time on a preliminary write-zero request that will later be rewritten by actual data, if it is known that the write-zero request will use a slow fallback; but in doing so, could not optimize for NBD. The NBD specification is now considering an extension that will allow passing on those semantics; this patch updates the new protocol bits and 'qemu-nbd --list' output to recognize the bit, as well as the new errno value possible when using the new flag; while upcoming patches will improve the client to use the feature when present, and the server to advertise support for it. The NBD spec recommends (but not requires) that ENOTSUP be avoided for all but failures of a fast zero (the only time it is mandatory to avoid an ENOTSUP failure is when fast zero is supported but not requested during write zeroes; the questionable use is for ENOTSUP to other actions like a normal write request). However, clients that get an unexpected ENOTSUP will either already be treating it the same as EINVAL, or may appreciate the extra bit of information. We were equally loose for returning EOVERFLOW in more situations than recommended by the spec, so if it turns out to be a problem in practice, a later patch can tighten handling for both error codes. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20190823143726.27062-3-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: tweak commit message, also handle EOPNOTSUPP]
2019-08-23 17:37:23 +03:00
[NBD_FLAG_SEND_FAST_ZERO_BIT] = "fast-zero",
};
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
printf(" size: %" PRIu64 "\n", list[i].size);
printf(" flags: 0x%x (", list[i].flags);
for (size_t bit = 0; bit < ARRAY_SIZE(flag_names); bit++) {
if (flag_names[bit] && (list[i].flags & (1 << bit))) {
printf(" %s", flag_names[bit]);
}
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
}
printf(" )\n");
}
if (list[i].min_block) {
printf(" min block: %u\n", list[i].min_block);
printf(" opt block: %u\n", list[i].opt_block);
printf(" max block: %u\n", list[i].max_block);
}
if (list[i].n_contexts) {
printf(" available meta contexts: %d\n", list[i].n_contexts);
for (j = 0; j < list[i].n_contexts; j++) {
printf(" %s\n", list[i].contexts[j]);
}
}
}
nbd_free_export_list(list, rc);
ret = EXIT_SUCCESS;
out:
object_unref(OBJECT(sioc));
return ret;
}
#if HAVE_NBD_DEVICE
static void *show_parts(void *arg)
{
char *device = arg;
int nbd;
/* linux just needs an open() to trigger
* the partition table update
* but remember to load the module with max_part != 0 :
* modprobe nbd max_part=63
*/
nbd = open(device, O_RDWR);
if (nbd >= 0) {
close(nbd);
}
return NULL;
}
struct NbdClientOpts {
char *device;
2023-07-17 17:55:41 +03:00
bool fork_process;
};
static void *nbd_client_thread(void *arg)
{
struct NbdClientOpts *opts = arg;
NBDExportInfo info = { .request_sizes = false, .name = g_strdup("") };
QIOChannelSocket *sioc;
int fd = -1;
int ret = EXIT_FAILURE;
pthread_t show_parts_thread;
Error *local_error = NULL;
sioc = qio_channel_socket_new();
if (qio_channel_socket_connect_sync(sioc,
saddr,
&local_error) < 0) {
error_report_err(local_error);
goto out;
}
if (nbd_receive_negotiate(NULL, QIO_CHANNEL(sioc),
NULL, NULL, NULL, &info, &local_error) < 0) {
if (local_error) {
error_report_err(local_error);
}
goto out;
}
fd = open(opts->device, O_RDWR);
if (fd < 0) {
/* Linux-only, we can use %m in printf. */
error_report("Failed to open %s: %m", opts->device);
goto out;
}
if (nbd_init(fd, sioc, &info, &local_error) < 0) {
error_report_err(local_error);
goto out;
}
/* update partition table */
pthread_create(&show_parts_thread, NULL, show_parts, opts->device);
2023-07-17 17:55:41 +03:00
if (verbose && !opts->fork_process) {
fprintf(stderr, "NBD device %s is now connected to %s\n",
opts->device, srcpath);
} else {
/* Close stderr so that the qemu-nbd process exits. */
dup2(STDOUT_FILENO, STDERR_FILENO);
}
if (nbd_client(fd) < 0) {
goto out;
}
ret = EXIT_SUCCESS;
out:
if (fd >= 0) {
close(fd);
}
object_unref(OBJECT(sioc));
g_free(info.name);
kill(getpid(), SIGTERM);
return (void *) (intptr_t) ret;
}
#endif /* HAVE_NBD_DEVICE */
static int nbd_can_accept(void)
{
return state == RUNNING && (shared == 0 || nb_fds < shared);
}
static void nbd_update_server_watch(void);
nbd: Fix regression on resiliency to port scan Back in qemu 2.5, qemu-nbd was immune to port probes (a transient server would not quit, regardless of how many probe connections came and went, until a connection actually negotiated). But we broke that in commit ee7d7aa when removing the return value to nbd_client_new(), although that patch also introduced a bug causing an assertion failure on a client that fails negotiation. We then made it worse during refactoring in commit 1a6245a (a segfault before we could even assert); the (masked) assertion was cleaned up in d3780c2 (still in 2.6), and just recently we finally fixed the segfault ("nbd: Fully intialize client in case of failed negotiation"). But that still means that ever since we added TLS support to qemu-nbd, we have been vulnerable to an ill-timed port-scan being able to cause a denial of service by taking down qemu-nbd before a real client has a chance to connect. Since negotiation is now handled asynchronously via coroutines, we no longer have a synchronous point of return by re-adding a return value to nbd_client_new(). So this patch instead wires things up to pass the negotiation status through the close_fn callback function. Simple test across two terminals: $ qemu-nbd -f raw -p 30001 file $ nmap 127.0.0.1 -p 30001 && \ qemu-io -c 'r 0 512' -f raw nbd://localhost:30001 Note that this patch does not change what constitutes successful negotiation (thus, a client must enter transmission phase before that client can be considered as a reason to terminate the server when the connection ends). Perhaps we may want to tweak things in a later patch to also treat a client that uses NBD_OPT_ABORT as being a 'successful' negotiation (the client correctly talked the NBD protocol, and informed us it was not going to use our export after all), but that's a discussion for another day. Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1451614 Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20170608222617.20376-1-eblake@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2017-06-09 01:26:17 +03:00
static void nbd_client_closed(NBDClient *client, bool negotiated)
{
nb_fds--;
nbd: Fix regression on resiliency to port scan Back in qemu 2.5, qemu-nbd was immune to port probes (a transient server would not quit, regardless of how many probe connections came and went, until a connection actually negotiated). But we broke that in commit ee7d7aa when removing the return value to nbd_client_new(), although that patch also introduced a bug causing an assertion failure on a client that fails negotiation. We then made it worse during refactoring in commit 1a6245a (a segfault before we could even assert); the (masked) assertion was cleaned up in d3780c2 (still in 2.6), and just recently we finally fixed the segfault ("nbd: Fully intialize client in case of failed negotiation"). But that still means that ever since we added TLS support to qemu-nbd, we have been vulnerable to an ill-timed port-scan being able to cause a denial of service by taking down qemu-nbd before a real client has a chance to connect. Since negotiation is now handled asynchronously via coroutines, we no longer have a synchronous point of return by re-adding a return value to nbd_client_new(). So this patch instead wires things up to pass the negotiation status through the close_fn callback function. Simple test across two terminals: $ qemu-nbd -f raw -p 30001 file $ nmap 127.0.0.1 -p 30001 && \ qemu-io -c 'r 0 512' -f raw nbd://localhost:30001 Note that this patch does not change what constitutes successful negotiation (thus, a client must enter transmission phase before that client can be considered as a reason to terminate the server when the connection ends). Perhaps we may want to tweak things in a later patch to also treat a client that uses NBD_OPT_ABORT as being a 'successful' negotiation (the client correctly talked the NBD protocol, and informed us it was not going to use our export after all), but that's a discussion for another day. Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1451614 Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20170608222617.20376-1-eblake@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2017-06-09 01:26:17 +03:00
if (negotiated && nb_fds == 0 && !persistent && state == RUNNING) {
state = TERMINATE;
}
nbd_update_server_watch();
nbd_client_put(client);
}
static void nbd_accept(QIONetListener *listener, QIOChannelSocket *cioc,
gpointer opaque)
{
if (state >= TERMINATE) {
return;
}
nb_fds++;
nbd_update_server_watch();
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
nbd_client_new(cioc, tlscreds, tlsauthz, nbd_client_closed);
}
static void nbd_update_server_watch(void)
{
if (nbd_can_accept()) {
qio_net_listener_set_client_func(server, nbd_accept, NULL, NULL);
} else {
qio_net_listener_set_client_func(server, NULL, NULL, NULL);
}
}
static SocketAddress *nbd_build_socket_address(const char *sockpath,
const char *bindto,
const char *port)
{
SocketAddress *saddr;
saddr = g_new0(SocketAddress, 1);
if (sockpath) {
saddr->type = SOCKET_ADDRESS_TYPE_UNIX;
saddr->u.q_unix.path = g_strdup(sockpath);
} else {
InetSocketAddress *inet;
saddr->type = SOCKET_ADDRESS_TYPE_INET;
inet = &saddr->u.inet;
inet->host = g_strdup(bindto);
if (port) {
inet->port = g_strdup(port);
} else {
inet->port = g_strdup_printf("%d", NBD_DEFAULT_PORT);
}
}
return saddr;
}
static QemuOptsList file_opts = {
.name = "file",
.implied_opt_name = "file",
.head = QTAILQ_HEAD_INITIALIZER(file_opts.head),
.desc = {
/* no elements => accept any params */
{ /* end of list */ }
},
};
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, bool list,
Error **errp)
{
Object *obj;
QCryptoTLSCreds *creds;
obj = object_resolve_path_component(
object_get_objects_root(), id);
if (!obj) {
error_setg(errp, "No TLS credentials with id '%s'",
id);
return NULL;
}
creds = (QCryptoTLSCreds *)
object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
if (!creds) {
error_setg(errp, "Object with id '%s' is not TLS credentials",
id);
return NULL;
}
if (!qcrypto_tls_creds_check_endpoint(creds,
list
? QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT
: QCRYPTO_TLS_CREDS_ENDPOINT_SERVER,
errp)) {
return NULL;
}
object_ref(obj);
return creds;
}
static void setup_address_and_port(const char **address, const char **port)
{
if (*address == NULL) {
*address = "0.0.0.0";
}
if (*port == NULL) {
*port = stringify(NBD_DEFAULT_PORT);
}
}
/*
* Check socket parameters compatibility when socket activation is used.
*/
static const char *socket_activation_validate_opts(const char *device,
const char *sockpath,
const char *address,
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
const char *port,
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
const char *selinux,
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
bool list)
{
if (device != NULL) {
return "NBD device can't be set when using socket activation";
}
if (sockpath != NULL) {
return "Unix socket can't be set when using socket activation";
}
if (address != NULL) {
return "The interface can't be set when using socket activation";
}
if (port != NULL) {
return "TCP port number can't be set when using socket activation";
}
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
if (selinux != NULL) {
return "SELinux label can't be set when using socket activation";
}
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
if (list) {
return "List mode is incompatible with socket activation";
}
return NULL;
}
static void qemu_nbd_shutdown(void)
{
job_cancel_sync_all();
blk_exp_close_all();
bdrv_close_all();
}
int main(int argc, char **argv)
{
block: New BlockBackend A block device consists of a frontend device model and a backend. A block backend has a tree of block drivers doing the actual work. The tree is managed by the block layer. We currently use a single abstraction BlockDriverState both for tree nodes and the backend as a whole. Drawbacks: * Its API includes both stuff that makes sense only at the block backend level (root of the tree) and stuff that's only for use within the block layer. This makes the API bigger and more complex than necessary. Moreover, it's not obvious which interfaces are meant for device models, and which really aren't. * Since device models keep a reference to their backend, the backend object can't just be destroyed. But for media change, we need to replace the tree. Our solution is to make the BlockDriverState generic, with actual driver state in a separate object, pointed to by member opaque. That lets us replace the tree by deinitializing and reinitializing its root. This special need of the root makes the data structure awkward everywhere in the tree. The general plan is to separate the APIs into "block backend", for use by device models, monitor and whatever other code dealing with block backends, and "block driver", for use by the block layer and whatever other code (if any) dealing with trees and tree nodes. Code dealing with block backends, device models in particular, should become completely oblivious of BlockDriverState. This should let us clean up both APIs, and the tree data structures. This commit is a first step. It creates a minimal "block backend" API: type BlockBackend and functions to create, destroy and find them. BlockBackend objects are created and destroyed exactly when root BlockDriverState objects are created and destroyed. "Root" in the sense of "in bdrv_states". They're not yet used for anything; that'll come shortly. A root BlockDriverState is created with bdrv_new_root(), so where to create a BlockBackend is obvious. Where these roots get destroyed isn't always as obvious. It is obvious in qemu-img.c, qemu-io.c and qemu-nbd.c, and in error paths of blockdev_init(), blk_connect(). That leaves destruction of objects successfully created by blockdev_init() and blk_connect(). blockdev_init() is used only by drive_new() and qmp_blockdev_add(). Objects created by the latter are currently indestructible (see commit 48f364d "blockdev: Refuse to drive_del something added with blockdev-add" and commit 2d246f0 "blockdev: Introduce DriveInfo.enable_auto_del"). Objects created by the former get destroyed by drive_del(). Objects created by blk_connect() get destroyed by blk_disconnect(). BlockBackend is reference-counted. Its reference count never exceeds one so far, but that's going to change. In drive_del(), the BB's reference count is surely one now. The BDS's reference count is greater than one when something else is holding a reference, such as a block job. In this case, the BB is destroyed right away, but the BDS lives on until all extra references get dropped. Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Max Reitz <mreitz@redhat.com> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2014-10-07 15:59:04 +04:00
BlockBackend *blk;
BlockDriverState *bs;
uint64_t dev_offset = 0;
nbd: Improve per-export flag handling in server When creating a read-only image, we are still advertising support for TRIM and WRITE_ZEROES to the client, even though the client should not be issuing those commands. But seeing this requires looking across multiple functions: All callers to nbd_export_new() passed a single flag based solely on whether the export allows writes. Later, we then pass a constant set of flags to nbd_negotiate_options() (namely, the set of flags which we always support, at least for writable images), which is then further dynamically modified with NBD_FLAG_SEND_DF based on client requests for structured options. Finally, when processing NBD_OPT_EXPORT_NAME or NBD_OPT_EXPORT_GO we bitwise-or the original caller's flag with the runtime set of flags we've built up over several functions. Let's refactor things to instead compute a baseline of flags as soon as possible which gets shared between multiple clients, in nbd_export_new(), and changing the signature for the callers to pass in a simpler bool rather than having to figure out flags. We can then get rid of the 'myflags' parameter to various functions, and instead refer to client for everything we need (we still have to perform a bitwise-OR for NBD_FLAG_SEND_DF during NBD_OPT_EXPORT_NAME and NBD_OPT_EXPORT_GO, but it's easier to see what is being computed). This lets us quit advertising senseless flags for read-only images, as well as making the next patch for exposing FAST_ZERO support easier to write. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20190823143726.27062-2-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: improve commit message, update iotest 223]
2019-08-23 17:37:22 +03:00
bool readonly = false;
bool disconnect = false;
const char *bindto = NULL;
const char *port = NULL;
char *sockpath = NULL;
char *device = NULL;
QemuOpts *sn_opts = NULL;
const char *sn_id_or_name = NULL;
const char *sopt = "hVb:o:p:rsnc:dvk:e:f:tl:x:T:D:AB:L";
struct option lopt[] = {
{ "help", no_argument, NULL, 'h' },
{ "version", no_argument, NULL, 'V' },
{ "bind", required_argument, NULL, 'b' },
{ "port", required_argument, NULL, 'p' },
{ "socket", required_argument, NULL, 'k' },
{ "offset", required_argument, NULL, 'o' },
{ "read-only", no_argument, NULL, 'r' },
{ "allocation-depth", no_argument, NULL, 'A' },
{ "bitmap", required_argument, NULL, 'B' },
{ "connect", required_argument, NULL, 'c' },
{ "disconnect", no_argument, NULL, 'd' },
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
{ "list", no_argument, NULL, 'L' },
{ "snapshot", no_argument, NULL, 's' },
{ "load-snapshot", required_argument, NULL, 'l' },
{ "nocache", no_argument, NULL, 'n' },
{ "cache", required_argument, NULL, QEMU_NBD_OPT_CACHE },
{ "aio", required_argument, NULL, QEMU_NBD_OPT_AIO },
{ "discard", required_argument, NULL, QEMU_NBD_OPT_DISCARD },
{ "detect-zeroes", required_argument, NULL,
QEMU_NBD_OPT_DETECT_ZEROES },
{ "shared", required_argument, NULL, 'e' },
{ "format", required_argument, NULL, 'f' },
{ "persistent", no_argument, NULL, 't' },
{ "verbose", no_argument, NULL, 'v' },
{ "object", required_argument, NULL, QEMU_NBD_OPT_OBJECT },
{ "export-name", required_argument, NULL, 'x' },
{ "description", required_argument, NULL, 'D' },
{ "tls-creds", required_argument, NULL, QEMU_NBD_OPT_TLSCREDS },
{ "tls-hostname", required_argument, NULL, QEMU_NBD_OPT_TLSHOSTNAME },
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
{ "tls-authz", required_argument, NULL, QEMU_NBD_OPT_TLSAUTHZ },
{ "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 },
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
{ "selinux-label", required_argument, NULL,
QEMU_NBD_OPT_SELINUX_LABEL },
{ NULL, 0, NULL, 0 }
};
int ch;
int opt_ind = 0;
int flags = BDRV_O_RDWR;
int ret = 0;
bool seen_cache = false;
bool seen_discard = false;
bool seen_aio = false;
pthread_t client_thread;
const char *fmt = NULL;
Error *local_err = NULL;
BlockdevDetectZeroesOptions detect_zeroes = BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF;
QDict *options = NULL;
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
const char *export_name = NULL; /* defaults to "" later for server mode */
const char *export_description = NULL;
BlockDirtyBitmapOrStrList *bitmaps = NULL;
bool alloc_depth = false;
const char *tlscredsid = NULL;
const char *tlshostname = NULL;
bool imageOpts = false;
qemu-nbd: Change default cache mode to writeback Both qemu and qemu-img use writeback cache mode by default, which is already documented in qemu(1). qemu-nbd uses writethrough cache mode by default, and the default cache mode is not documented. According to the qemu-nbd(8): --cache=CACHE The cache mode to be used with the file. See the documentation of the emulator's -drive cache=... option for allowed values. qemu(1) says: The default mode is cache=writeback. So users have no reason to assume that qemu-nbd is using writethough cache mode. The only hint is the painfully slow writing when using the defaults. Looking in git history, it seems that qemu used writethrough in the past to support broken guests that did not flush data properly, or could not flush due to limitations in qemu. But qemu-nbd clients can use NBD_CMD_FLUSH to flush data, so using writethrough does not help anyone. Change the default cache mode to writback, and document the default and available values properly in the online help and manual. With this change converting image via qemu-nbd is 3.5 times faster. $ qemu-img create dst.img 50g $ qemu-nbd -t -f raw -k /tmp/nbd.sock dst.img Before this change: $ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock" Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock Time (mean ± σ): 83.639 s ± 5.970 s [User: 2.733 s, System: 6.112 s] Range (min … max): 76.749 s … 87.245 s 3 runs After this change: $ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock" Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock Time (mean ± σ): 23.522 s ± 0.433 s [User: 2.083 s, System: 5.475 s] Range (min … max): 23.234 s … 24.019 s 3 runs Users can avoid the issue by using --cache=writeback[1] but the defaults should give good performance for the common use case. [1] https://bugzilla.redhat.com/1990656 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20210813205519.50518-1-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> CC: qemu-stable@nongnu.org Signed-off-by: Eric Blake <eblake@redhat.com>
2021-08-13 23:55:19 +03:00
bool writethrough = false; /* Client will flush as needed. */
bool fork_process = false;
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
bool list = false;
unsigned socket_activation;
const char *pid_file_name = NULL;
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
const char *selinux_label = NULL;
BlockExportOptions *export_opts;
#if HAVE_NBD_DEVICE
struct NbdClientOpts opts;
#endif
#ifdef CONFIG_POSIX
os_setup_early_signal_handling();
os_setup_signal_handling();
#endif
socket_init();
error_init(argv[0]);
module_call_init(MODULE_INIT_TRACE);
qcrypto_init(&error_fatal);
module_call_init(MODULE_INIT_QOM);
qemu_add_opts(&qemu_trace_opts);
qemu_init_exec_dir(argv[0]);
while ((ch = getopt_long(argc, argv, sopt, lopt, &opt_ind)) != -1) {
switch (ch) {
case 's':
flags |= BDRV_O_SNAPSHOT;
break;
case 'n':
optarg = (char *) "none";
/* fallthrough */
case QEMU_NBD_OPT_CACHE:
if (seen_cache) {
error_report("-n and --cache can only be specified once");
exit(EXIT_FAILURE);
}
seen_cache = true;
if (bdrv_parse_cache_mode(optarg, &flags, &writethrough) == -1) {
error_report("Invalid cache mode `%s'", optarg);
exit(EXIT_FAILURE);
}
break;
case QEMU_NBD_OPT_AIO:
if (seen_aio) {
error_report("--aio can only be specified once");
exit(EXIT_FAILURE);
}
seen_aio = true;
if (bdrv_parse_aio(optarg, &flags) < 0) {
error_report("Invalid aio mode '%s'", optarg);
exit(EXIT_FAILURE);
}
break;
case QEMU_NBD_OPT_DISCARD:
if (seen_discard) {
error_report("--discard can only be specified once");
exit(EXIT_FAILURE);
}
seen_discard = true;
if (bdrv_parse_discard_flags(optarg, &flags) == -1) {
error_report("Invalid discard mode `%s'", optarg);
exit(EXIT_FAILURE);
}
break;
case QEMU_NBD_OPT_DETECT_ZEROES:
detect_zeroes =
qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup,
optarg,
BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF,
&local_err);
if (local_err) {
error_reportf_err(local_err,
"Failed to parse detect_zeroes mode: ");
exit(EXIT_FAILURE);
}
if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP &&
!(flags & BDRV_O_UNMAP)) {
error_report("setting detect-zeroes to unmap is not allowed "
"without setting discard operation to unmap");
exit(EXIT_FAILURE);
}
break;
case 'b':
bindto = optarg;
break;
case 'p':
port = optarg;
break;
case 'o':
if (qemu_strtou64(optarg, NULL, 0, &dev_offset) < 0) {
error_report("Invalid offset '%s'", optarg);
exit(EXIT_FAILURE);
}
break;
case 'l':
if (strstart(optarg, SNAPSHOT_OPT_BASE, NULL)) {
QemuOpts: Wean off qerror_report_err() qerror_report_err() is a transitional interface to help with converting existing monitor commands to QMP. It should not be used elsewhere. The only remaining user in qemu-option.c is qemu_opts_parse(). Is it used in QMP context? If not, we can simply replace qerror_report_err() by error_report_err(). The uses in qemu-img.c, qemu-io.c, qemu-nbd.c and under tests/ are clearly not in QMP context. The uses in vl.c aren't either, because the only QMP command handlers there are qmp_query_status() and qmp_query_machines(), and they don't call it. Remaining uses: * drive_def(): Command line -drive and such, HMP drive_add and pci_add * hmp_chardev_add(): HMP chardev-add * monitor_parse_command(): HMP core * tmp_config_parse(): Command line -tpmdev * net_host_device_add(): HMP host_net_add * net_client_parse(): Command line -net and -netdev * qemu_global_option(): Command line -global * vnc_parse_func(): Command line -display, -vnc, default display, HMP change, QMP change. Bummer. * qemu_pci_hot_add_nic(): HMP pci_add * usb_net_init(): Command line -usbdevice, HMP usb_add Propagate errors through qemu_opts_parse(). Create a convenience function qemu_opts_parse_noisily() that passes errors to error_report_err(). Switch all non-QMP users outside tests to it. That leaves vnc_parse_func(). Propagate errors through it. Since I'm touching it anyway, rename it to vnc_parse(). Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> Reviewed-by: Luiz Capitulino <lcapitulino@redhat.com>
2015-02-13 14:50:26 +03:00
sn_opts = qemu_opts_parse_noisily(&internal_snapshot_opts,
optarg, false);
if (!sn_opts) {
error_report("Failed in parsing snapshot param `%s'",
optarg);
exit(EXIT_FAILURE);
}
} else {
sn_id_or_name = optarg;
}
/* fall through */
case 'r':
nbd: Improve per-export flag handling in server When creating a read-only image, we are still advertising support for TRIM and WRITE_ZEROES to the client, even though the client should not be issuing those commands. But seeing this requires looking across multiple functions: All callers to nbd_export_new() passed a single flag based solely on whether the export allows writes. Later, we then pass a constant set of flags to nbd_negotiate_options() (namely, the set of flags which we always support, at least for writable images), which is then further dynamically modified with NBD_FLAG_SEND_DF based on client requests for structured options. Finally, when processing NBD_OPT_EXPORT_NAME or NBD_OPT_EXPORT_GO we bitwise-or the original caller's flag with the runtime set of flags we've built up over several functions. Let's refactor things to instead compute a baseline of flags as soon as possible which gets shared between multiple clients, in nbd_export_new(), and changing the signature for the callers to pass in a simpler bool rather than having to figure out flags. We can then get rid of the 'myflags' parameter to various functions, and instead refer to client for everything we need (we still have to perform a bitwise-OR for NBD_FLAG_SEND_DF during NBD_OPT_EXPORT_NAME and NBD_OPT_EXPORT_GO, but it's easier to see what is being computed). This lets us quit advertising senseless flags for read-only images, as well as making the next patch for exposing FAST_ZERO support easier to write. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20190823143726.27062-2-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: improve commit message, update iotest 223]
2019-08-23 17:37:22 +03:00
readonly = true;
flags &= ~BDRV_O_RDWR;
break;
case 'A':
alloc_depth = true;
break;
case 'B':
{
BlockDirtyBitmapOrStr *el = g_new(BlockDirtyBitmapOrStr, 1);
*el = (BlockDirtyBitmapOrStr) {
.type = QTYPE_QSTRING,
.u.local = g_strdup(optarg),
};
QAPI_LIST_PREPEND(bitmaps, el);
}
break;
case 'k':
sockpath = optarg;
if (sockpath[0] != '/') {
error_report("socket path must be absolute");
exit(EXIT_FAILURE);
}
break;
case 'd':
disconnect = true;
break;
case 'c':
device = optarg;
break;
case 'e':
if (qemu_strtoi(optarg, NULL, 0, &shared) < 0 ||
shared < 0) {
error_report("Invalid shared device number '%s'", optarg);
exit(EXIT_FAILURE);
}
break;
case 'f':
fmt = optarg;
break;
case 't':
persistent = 1;
break;
case 'x':
export_name = optarg;
if (strlen(export_name) > NBD_MAX_STRING_SIZE) {
error_report("export name '%s' too long", export_name);
exit(EXIT_FAILURE);
}
break;
case 'D':
export_description = optarg;
if (strlen(export_description) > NBD_MAX_STRING_SIZE) {
error_report("export description '%s' too long",
export_description);
exit(EXIT_FAILURE);
}
break;
case 'v':
verbose = 1;
break;
case 'V':
version(argv[0]);
exit(0);
break;
case 'h':
usage(argv[0]);
exit(0);
break;
case '?':
error_report("Try `%s --help' for more information.", argv[0]);
exit(EXIT_FAILURE);
case QEMU_NBD_OPT_OBJECT:
user_creatable_process_cmdline(optarg);
break;
case QEMU_NBD_OPT_TLSCREDS:
tlscredsid = optarg;
break;
case QEMU_NBD_OPT_TLSHOSTNAME:
tlshostname = optarg;
break;
case QEMU_NBD_OPT_IMAGE_OPTS:
imageOpts = true;
break;
case 'T':
trace_opt_parse(optarg);
break;
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
case QEMU_NBD_OPT_TLSAUTHZ:
tlsauthz = optarg;
break;
case QEMU_NBD_OPT_FORK:
fork_process = true;
break;
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
case 'L':
list = true;
break;
case QEMU_NBD_OPT_PID_FILE:
pid_file_name = optarg;
break;
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
case QEMU_NBD_OPT_SELINUX_LABEL:
selinux_label = optarg;
break;
}
}
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
if (list) {
if (argc != optind) {
error_report("List mode is incompatible with a file name");
exit(EXIT_FAILURE);
}
if (export_name || export_description || dev_offset ||
device || disconnect || fmt || sn_id_or_name || bitmaps ||
alloc_depth || seen_aio || seen_discard || seen_cache) {
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
error_report("List mode is incompatible with per-device settings");
exit(EXIT_FAILURE);
}
if (fork_process) {
error_report("List mode is incompatible with forking");
exit(EXIT_FAILURE);
}
} else if ((argc - optind) != 1) {
error_report("Invalid number of arguments");
error_printf("Try `%s --help' for more information.\n", argv[0]);
exit(EXIT_FAILURE);
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
} else if (!export_name) {
export_name = "";
}
if (!trace_init_backends()) {
exit(1);
}
trace_init_file();
qemu_set_log(LOG_TRACE, &error_fatal);
socket_activation = check_socket_activation();
if (socket_activation == 0) {
if (!sockpath) {
setup_address_and_port(&bindto, &port);
}
} else {
/* Using socket activation - check user didn't use -p etc. */
const char *err_msg = socket_activation_validate_opts(device, sockpath,
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
bindto, port,
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
selinux_label,
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
list);
if (err_msg != NULL) {
error_report("%s", err_msg);
exit(EXIT_FAILURE);
}
/* qemu-nbd can only listen on a single socket. */
if (socket_activation > 1) {
error_report("qemu-nbd does not support socket activation with %s > 1",
"LISTEN_FDS");
exit(EXIT_FAILURE);
}
}
if (tlscredsid) {
if (device) {
error_report("TLS is not supported with a host device");
exit(EXIT_FAILURE);
}
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
if (tlsauthz && list) {
error_report("TLS authorization is incompatible with export list");
exit(EXIT_FAILURE);
}
if (tlshostname && !list) {
error_report("TLS hostname is only supported with export list");
exit(EXIT_FAILURE);
}
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
tlscreds = nbd_get_tls_creds(tlscredsid, list, &local_err);
if (local_err) {
error_reportf_err(local_err, "Failed to get TLS creds: ");
exit(EXIT_FAILURE);
}
qemu-nbd: add support for authorization of TLS clients Currently any client which can complete the TLS handshake is able to use the NBD server. The server admin can turn on the 'verify-peer' option for the x509 creds to require the client to provide a x509 certificate. This means the client will have to acquire a certificate from the CA before they are permitted to use the NBD server. This is still a fairly low bar to cross. This adds a '--tls-authz OBJECT-ID' option to the qemu-nbd command which takes the ID of a previously added 'QAuthZ' object instance. This will be used to validate the client's x509 distinguished name. Clients failing the authorization check will not be permitted to use the NBD server. For example to setup authorization that only allows connection from a client whose x509 certificate distinguished name is CN=laptop.example.com,O=Example Org,L=London,ST=London,C=GB escape the commas in the name and use: qemu-nbd --object tls-creds-x509,id=tls0,dir=/home/berrange/qemutls,\ endpoint=server,verify-peer=yes \ --object 'authz-simple,id=auth0,identity=CN=laptop.example.com,,\ O=Example Org,,L=London,,ST=London,,C=GB' \ --tls-creds tls0 \ --tls-authz authz0 \ ....other qemu-nbd args... NB: a real shell command line would not have leading whitespace after the line continuation, it is just included here for clarity. Reviewed-by: Juan Quintela <quintela@redhat.com> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> Message-Id: <20190227162035.18543-2-berrange@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: split long line in --help text, tweak 233 to show that whitespace after ,, in identity= portion is actually okay] Signed-off-by: Eric Blake <eblake@redhat.com>
2019-02-27 19:20:33 +03:00
} else {
if (tlsauthz) {
error_report("--tls-authz is not permitted without --tls-creds");
exit(EXIT_FAILURE);
}
if (tlshostname) {
error_report("--tls-hostname is not permitted without --tls-creds");
exit(EXIT_FAILURE);
}
}
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
if (selinux_label) {
#ifdef CONFIG_SELINUX
if (sockpath == NULL && device == NULL) {
error_report("--selinux-label is not permitted without --socket");
exit(EXIT_FAILURE);
}
#else
error_report("SELinux support not enabled in this binary");
exit(EXIT_FAILURE);
#endif
}
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
if (list) {
saddr = nbd_build_socket_address(sockpath, bindto, port);
return qemu_nbd_client_list(saddr, tlscreds,
tlshostname ? tlshostname : bindto);
qemu-nbd: Add --list option We want to be able to detect whether a given qemu NBD server is exposing the right export(s) and dirty bitmaps, at least for regression testing. We could use 'nbd-client -l' from the upstream NBD project to list exports, but it's annoying to rely on out-of-tree binaries; furthermore, nbd-client doesn't necessarily know about all of the qemu NBD extensions. Thus, it is time to add a new mode to qemu-nbd that merely sniffs all possible information from the server during handshake phase, then disconnects and dumps the information. This patch actually implements --list/-L, while reusing other options such as --tls-creds for now designating how to connect as the client (rather than their non-list usage of how to operate as the server). I debated about adding this functionality to something akin to 'qemu-img info' - but that tool does not readily lend itself to connecting to an arbitrary NBD server without also tying to a specific export (I may, however, still add ImageInfoSpecificNBD for reporting the bitmaps available when connecting to a single export). And, while it may feel a bit odd that normally qemu-nbd is a server but 'qemu-nbd -L' is a client, we are not really making the qemu-nbd binary that much larger, because 'qemu-nbd -c' has to operate as both server and client simultaneously across two threads when feeding the kernel module for /dev/nbdN access. Sample output: $ qemu-nbd -L exports available: 1 export: '' size: 65536 flags: 0x4ed ( flush fua trim zeroes df cache ) min block: 512 opt block: 4096 max block: 33554432 available meta contexts: 1 base:allocation Note that the output only lists sizes if the server sent NBD_FLAG_HAS_FLAGS, because a newstyle server does not give the size otherwise. It has the side effect that for really old servers that did not send any flags, the size is not output even though it was available. However, I'm not too concerned about that - oldstyle servers are (rightfully) getting less common to encounter (qemu 3.0 was the last version where we even serve it), and most existing servers that still even offer oldstyle negotiation (such as nbdkit) still send flags (since that was added to the NBD protocol in 2007 to permit read-only connections). Not done here, but maybe worth future experiments: capture the meat of NBDExportInfo into a QAPI struct, and use the generated QAPI pretty-printers instead of hand-rolling our output loop. It would also permit us to add a JSON output mode for machine parsing. Signed-off-by: Eric Blake <eblake@redhat.com> Reviewed-by: Richard W.M. Jones <rjones@redhat.com> Message-Id: <20190117193658.16413-20-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
2019-01-17 22:36:56 +03:00
}
#if !HAVE_NBD_DEVICE
if (disconnect || device) {
error_report("Kernel /dev/nbdN support not available");
exit(EXIT_FAILURE);
}
#else /* HAVE_NBD_DEVICE */
if (disconnect) {
int nbdfd = open(argv[optind], O_RDWR);
if (nbdfd < 0) {
error_report("Cannot open %s: %s", argv[optind],
strerror(errno));
exit(EXIT_FAILURE);
}
nbd_disconnect(nbdfd);
close(nbdfd);
printf("%s disconnected\n", argv[optind]);
return 0;
}
#endif
if ((device && !verbose) || fork_process) {
#ifndef WIN32
g_autoptr(GError) err = NULL;
int stderr_fd[2];
pid_t pid;
int ret;
if (!g_unix_open_pipe(stderr_fd, FD_CLOEXEC, &err)) {
error_report("Error setting up communication pipe: %s",
err->message);
exit(EXIT_FAILURE);
}
/* Now daemonize, but keep a communication channel open to
* print errors and exit with the proper status code.
*/
pid = fork();
if (pid < 0) {
error_report("Failed to fork: %s", strerror(errno));
exit(EXIT_FAILURE);
} else if (pid == 0) {
close(stderr_fd[0]);
ret = qemu_daemon(1, 0);
/* Temporarily redirect stderr to the parent's pipe... */
dup2(stderr_fd[1], STDERR_FILENO);
if (ret < 0) {
error_report("Failed to daemonize: %s", strerror(errno));
exit(EXIT_FAILURE);
}
/* ... close the descriptor we inherited and go on. */
close(stderr_fd[1]);
} else {
bool errors = false;
char *buf;
/* In the parent. Print error messages from the child until
* it closes the pipe.
*/
close(stderr_fd[1]);
buf = g_malloc(1024);
while ((ret = read(stderr_fd[0], buf, 1024)) > 0) {
errors = true;
ret = qemu_write_full(STDERR_FILENO, buf, ret);
if (ret < 0) {
exit(EXIT_FAILURE);
}
}
if (ret < 0) {
error_report("Cannot read from daemon: %s",
strerror(errno));
exit(EXIT_FAILURE);
}
/* Usually the daemon should not print any message.
* Exit with zero status in that case.
*/
exit(errors);
}
#else /* WIN32 */
error_report("Unable to fork into background on Windows hosts");
exit(EXIT_FAILURE);
#endif /* WIN32 */
}
if (device != NULL && sockpath == NULL) {
sockpath = g_malloc(128);
snprintf(sockpath, 128, SOCKET_PATH, basename(device));
}
server = qio_net_listener_new();
if (socket_activation == 0) {
qemu-nbd: Use SOMAXCONN for socket listen() backlog Our default of a backlog of 1 connection is rather puny; it gets in the way when we are explicitly allowing multiple clients (such as qemu-nbd -e N [--shared], or nbd-server-start with its default "max-connections":0 for unlimited), but is even a problem when we stick to qemu-nbd's default of only 1 active client but use -t [--persistent] where a second client can start using the server once the first finishes. While the effects are less noticeable on TCP sockets (since the client can poll() to learn when the server is ready again), it is definitely observable on Unix sockets, where on Linux, a client will fail with EAGAIN and no recourse but to sleep an arbitrary amount of time before retrying if the server backlog is already full. Since QMP nbd-server-start is always persistent, it now always requests a backlog of SOMAXCONN; meanwhile, qemu-nbd will request SOMAXCONN if persistent, otherwise its backlog should be based on the expected number of clients. See https://bugzilla.redhat.com/1925045 for a demonstration of where our low backlog prevents libnbd from connecting as many parallel clients as it wants. Reported-by: Richard W.M. Jones <rjones@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com> CC: qemu-stable@nongnu.org Message-Id: <20210209152759.209074-2-eblake@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-02-09 18:27:58 +03:00
int backlog;
if (persistent || shared == 0) {
qemu-nbd: Use SOMAXCONN for socket listen() backlog Our default of a backlog of 1 connection is rather puny; it gets in the way when we are explicitly allowing multiple clients (such as qemu-nbd -e N [--shared], or nbd-server-start with its default "max-connections":0 for unlimited), but is even a problem when we stick to qemu-nbd's default of only 1 active client but use -t [--persistent] where a second client can start using the server once the first finishes. While the effects are less noticeable on TCP sockets (since the client can poll() to learn when the server is ready again), it is definitely observable on Unix sockets, where on Linux, a client will fail with EAGAIN and no recourse but to sleep an arbitrary amount of time before retrying if the server backlog is already full. Since QMP nbd-server-start is always persistent, it now always requests a backlog of SOMAXCONN; meanwhile, qemu-nbd will request SOMAXCONN if persistent, otherwise its backlog should be based on the expected number of clients. See https://bugzilla.redhat.com/1925045 for a demonstration of where our low backlog prevents libnbd from connecting as many parallel clients as it wants. Reported-by: Richard W.M. Jones <rjones@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com> CC: qemu-stable@nongnu.org Message-Id: <20210209152759.209074-2-eblake@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-02-09 18:27:58 +03:00
backlog = SOMAXCONN;
} else {
backlog = MIN(shared, SOMAXCONN);
}
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
#ifdef CONFIG_SELINUX
if (selinux_label && setsockcreatecon_raw(selinux_label) == -1) {
error_report("Cannot set SELinux socket create context to %s: %s",
selinux_label, strerror(errno));
exit(EXIT_FAILURE);
}
#endif
saddr = nbd_build_socket_address(sockpath, bindto, port);
qemu-nbd: Use SOMAXCONN for socket listen() backlog Our default of a backlog of 1 connection is rather puny; it gets in the way when we are explicitly allowing multiple clients (such as qemu-nbd -e N [--shared], or nbd-server-start with its default "max-connections":0 for unlimited), but is even a problem when we stick to qemu-nbd's default of only 1 active client but use -t [--persistent] where a second client can start using the server once the first finishes. While the effects are less noticeable on TCP sockets (since the client can poll() to learn when the server is ready again), it is definitely observable on Unix sockets, where on Linux, a client will fail with EAGAIN and no recourse but to sleep an arbitrary amount of time before retrying if the server backlog is already full. Since QMP nbd-server-start is always persistent, it now always requests a backlog of SOMAXCONN; meanwhile, qemu-nbd will request SOMAXCONN if persistent, otherwise its backlog should be based on the expected number of clients. See https://bugzilla.redhat.com/1925045 for a demonstration of where our low backlog prevents libnbd from connecting as many parallel clients as it wants. Reported-by: Richard W.M. Jones <rjones@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com> CC: qemu-stable@nongnu.org Message-Id: <20210209152759.209074-2-eblake@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-02-09 18:27:58 +03:00
if (qio_net_listener_open_sync(server, saddr, backlog,
&local_err) < 0) {
object_unref(OBJECT(server));
error_report_err(local_err);
exit(EXIT_FAILURE);
}
nbd/server: Add --selinux-label option Under SELinux, Unix domain sockets have two labels. One is on the disk and can be set with commands such as chcon(1). There is a different label stored in memory (called the process label). This can only be set by the process creating the socket. When using SELinux + SVirt and wanting qemu to be able to connect to a qemu-nbd instance, you must set both labels correctly first. For qemu-nbd the options to set the second label are awkward. You can create the socket in a wrapper program and then exec into qemu-nbd. Or you could try something with LD_PRELOAD. This commit adds the ability to set the label straightforwardly on the command line, via the new --selinux-label flag. (The name of the flag is the same as the equivalent nbdkit option.) A worked example showing how to use the new option can be found in this bug: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1984938 Signed-off-by: Richard W.M. Jones <rjones@redhat.com> Reviewed-by: Daniel P. Berrangé <berrange@redhat.com> [eblake: rebase to configure changes, reject --selinux-label if it is not compiled in or not used on a Unix socket] Note that we may relax some of these restrictions at a later date, such as making it possible to label a TCP socket, although it may be smarter to do so as a generic QMP action rather than more one-off command lines in qemu-nbd. Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20211115202944.615966-1-eblake@redhat.com> Reviewed-by: Thomas Huth <thuth@redhat.com> [eblake: adjust meson output as suggested by thuth] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-11-15 23:29:43 +03:00
#ifdef CONFIG_SELINUX
if (selinux_label && setsockcreatecon_raw(NULL) == -1) {
error_report("Cannot clear SELinux socket create context: %s",
strerror(errno));
exit(EXIT_FAILURE);
}
#endif
} else {
size_t i;
/* See comment in check_socket_activation above. */
for (i = 0; i < socket_activation; i++) {
QIOChannelSocket *sioc;
sioc = qio_channel_socket_new_fd(FIRST_SOCKET_ACTIVATION_FD + i,
&local_err);
if (sioc == NULL) {
object_unref(OBJECT(server));
error_reportf_err(local_err,
"Failed to use socket activation: ");
exit(EXIT_FAILURE);
}
qio_net_listener_add(server, sioc);
object_unref(OBJECT(sioc));
}
}
qemu_init_main_loop(&error_fatal);
bdrv_init();
atexit(qemu_nbd_shutdown);
srcpath = argv[optind];
if (imageOpts) {
QemuOpts *opts;
if (fmt) {
error_report("--image-opts and -f are mutually exclusive");
exit(EXIT_FAILURE);
}
opts = qemu_opts_parse_noisily(&file_opts, srcpath, true);
if (!opts) {
qemu_opts_reset(&file_opts);
exit(EXIT_FAILURE);
}
options = qemu_opts_to_qdict(opts, NULL);
qemu_opts_reset(&file_opts);
blk = blk_new_open(NULL, NULL, options, flags, &local_err);
} else {
if (fmt) {
options = qdict_new();
qdict_put_str(options, "driver", fmt);
}
blk = blk_new_open(srcpath, NULL, options, flags, &local_err);
}
if (!blk) {
error_reportf_err(local_err, "Failed to blk_new_open '%s': ",
argv[optind]);
exit(EXIT_FAILURE);
}
bs = blk_bs(blk);
if (dev_offset) {
QDict *raw_opts = qdict_new();
qdict_put_str(raw_opts, "driver", "raw");
qdict_put_str(raw_opts, "file", bs->node_name);
qdict_put_int(raw_opts, "offset", dev_offset);
bs = bdrv_open(NULL, NULL, raw_opts, flags, &error_fatal);
blk_remove_bs(blk);
blk_insert_bs(blk, bs, &error_fatal);
bdrv_unref(bs);
}
blk_set_enable_write_cache(blk, !writethrough);
if (sn_opts) {
ret = bdrv_snapshot_load_tmp(bs,
qemu_opt_get(sn_opts, SNAPSHOT_OPT_ID),
qemu_opt_get(sn_opts, SNAPSHOT_OPT_NAME),
&local_err);
} else if (sn_id_or_name) {
ret = bdrv_snapshot_load_tmp_by_id_or_name(bs, sn_id_or_name,
&local_err);
}
if (ret < 0) {
error_reportf_err(local_err, "Failed to load snapshot: ");
exit(EXIT_FAILURE);
}
bs->detect_zeroes = detect_zeroes;
nbd_server_is_qemu_nbd(shared);
export_opts = g_new(BlockExportOptions, 1);
*export_opts = (BlockExportOptions) {
.type = BLOCK_EXPORT_TYPE_NBD,
.id = g_strdup("qemu-nbd-export"),
.node_name = g_strdup(bdrv_get_node_name(bs)),
.has_writethrough = true,
.writethrough = writethrough,
.has_writable = true,
.writable = !readonly,
.u.nbd = {
.has_name = true,
.name = g_strdup(export_name),
.has_description = !!export_description,
.description = g_strdup(export_description),
.has_bitmaps = !!bitmaps,
.bitmaps = bitmaps,
.has_allocation_depth = alloc_depth,
.allocation_depth = alloc_depth,
},
};
blk_exp_add(export_opts, &error_fatal);
qapi_free_BlockExportOptions(export_opts);
if (device) {
#if HAVE_NBD_DEVICE
int ret;
opts = (struct NbdClientOpts) {
.device = device,
2023-07-17 17:55:41 +03:00
.fork_process = fork_process,
};
ret = pthread_create(&client_thread, NULL, nbd_client_thread, &opts);
if (ret != 0) {
error_report("Failed to create client thread: %s", strerror(ret));
exit(EXIT_FAILURE);
}
#endif
} else {
/* Shut up GCC warnings. */
memset(&client_thread, 0, sizeof(client_thread));
}
nbd_update_server_watch();
if (pid_file_name) {
qemu_write_pidfile(pid_file_name, &error_fatal);
}
/* now when the initialization is (almost) complete, chdir("/")
* to free any busy filesystems */
if (chdir("/") < 0) {
error_report("Could not chdir to root directory: %s",
strerror(errno));
exit(EXIT_FAILURE);
}
if (fork_process) {
2023-07-17 17:55:41 +03:00
dup2(STDOUT_FILENO, STDERR_FILENO);
}
state = RUNNING;
do {
main_loop_wait(false);
if (state == TERMINATE) {
blk_exp_close_all();
state = TERMINATED;
}
} while (state != TERMINATED);
block: New BlockBackend A block device consists of a frontend device model and a backend. A block backend has a tree of block drivers doing the actual work. The tree is managed by the block layer. We currently use a single abstraction BlockDriverState both for tree nodes and the backend as a whole. Drawbacks: * Its API includes both stuff that makes sense only at the block backend level (root of the tree) and stuff that's only for use within the block layer. This makes the API bigger and more complex than necessary. Moreover, it's not obvious which interfaces are meant for device models, and which really aren't. * Since device models keep a reference to their backend, the backend object can't just be destroyed. But for media change, we need to replace the tree. Our solution is to make the BlockDriverState generic, with actual driver state in a separate object, pointed to by member opaque. That lets us replace the tree by deinitializing and reinitializing its root. This special need of the root makes the data structure awkward everywhere in the tree. The general plan is to separate the APIs into "block backend", for use by device models, monitor and whatever other code dealing with block backends, and "block driver", for use by the block layer and whatever other code (if any) dealing with trees and tree nodes. Code dealing with block backends, device models in particular, should become completely oblivious of BlockDriverState. This should let us clean up both APIs, and the tree data structures. This commit is a first step. It creates a minimal "block backend" API: type BlockBackend and functions to create, destroy and find them. BlockBackend objects are created and destroyed exactly when root BlockDriverState objects are created and destroyed. "Root" in the sense of "in bdrv_states". They're not yet used for anything; that'll come shortly. A root BlockDriverState is created with bdrv_new_root(), so where to create a BlockBackend is obvious. Where these roots get destroyed isn't always as obvious. It is obvious in qemu-img.c, qemu-io.c and qemu-nbd.c, and in error paths of blockdev_init(), blk_connect(). That leaves destruction of objects successfully created by blockdev_init() and blk_connect(). blockdev_init() is used only by drive_new() and qmp_blockdev_add(). Objects created by the latter are currently indestructible (see commit 48f364d "blockdev: Refuse to drive_del something added with blockdev-add" and commit 2d246f0 "blockdev: Introduce DriveInfo.enable_auto_del"). Objects created by the former get destroyed by drive_del(). Objects created by blk_connect() get destroyed by blk_disconnect(). BlockBackend is reference-counted. Its reference count never exceeds one so far, but that's going to change. In drive_del(), the BB's reference count is surely one now. The BDS's reference count is greater than one when something else is holding a reference, such as a block job. In this case, the BB is destroyed right away, but the BDS lives on until all extra references get dropped. Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Max Reitz <mreitz@redhat.com> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2014-10-07 15:59:04 +04:00
blk_unref(blk);
if (sockpath) {
unlink(sockpath);
}
qemu_opts_del(sn_opts);
if (device) {
void *ret;
pthread_join(client_thread, &ret);
exit(ret != NULL);
} else {
exit(EXIT_SUCCESS);
}
}