This patch applies on top of 9P2000.L patches that we have on the list.
I took a look at how 9P server is handling open() flags in 9P2000.L path.
I think we can do away with the valid_flags() function and simplify the
code. The reasoning is as follows:
O_NOCTTY: (If the file is a terminal, don't make it the controlling
terminal of the process even though the process does not have a controlling
terminal) By the time the control reaches 9P client it is clear that what
we have is not a terminal device. Hence it does not matter what we do with
this flag. In any case 9P server can filter this flag out before making the
syscall.
O_NONBLOCK: (Don't block if i) Can't read/write to the file ii) Can't get
locks) This has an impact on FIFOs, but also on file locks. Hence we can
pass it down to the system call.
O_ASYNC: From the manpage:
O_ASYNC
Enable signal-driven I/O: generate a signal (SIGIO by default, but
this can be changed via fcntl(2)) when input or output becomes pos-
sible on this file descriptor. This feature is only available for
terminals, pseudo-terminals, sockets, and (since Linux 2.6) pipes
and FIFOs. See fcntl(2) for further details.
Again, this does not make any impact on regular files handled by 9P. Also,
we don't want 9P server to receive SIGIO. Hence I think 9P server can
filter this flag out before making the syscall.
O_CLOEXEC: This flag makes sense only on the client. If guest user space
sets this flag the guest VFS will take care of calling close() on the fd if
an exec() happens. Hence 9P client need not be bothered with this flag.
Also I think QEMU will not do an exec, but if it does, it makes sense to
close these fds. Hence we can pass this flag down to the syscall.
O_CREAT: Since we are in open() path it means we have confirmed that the file
exists. Hence there is no need to pass O_CREAT flag down to the system. In fact
on some versions of glibc this causes problems, because we pass O_CREAT flag,
but don't have permission bits. Hence we can just mask this flag out.
So in summary:
Mask out:
O_NOCTTY
O_ASYNC
O_CREAT
Pass-through:
O_NONBLOCK
O_CLOEXEC
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
This is equivalent to SM_PASSTHROUGH security model.
The only exception is, failure of privilige operation like chown
are ignored. This makes a passthrough like security model usable
for people who runs kvm as non root
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
With mapped security mode we use "user.virtfs" namespace is used
to store the virtFs related attributes. So hide it from user.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
TXATTRCREATE: Prepare a fid for setting xattr value on a file system object.
size[4] TXATTRCREATE tag[2] fid[4] name[s] attr_size[8] flags[4]
size[4] RXATTRWALK tag[2]
txattrcreate gets a fid pointing to xattr. This fid can later be
used to get set the xattr value.
flag value is derived from set Linux setxattr. The manpage says
"The flags parameter can be used to refine the semantics of the operation.
XATTR_CREATE specifies a pure create, which fails if the named attribute
exists already. XATTR_REPLACE specifies a pure replace operation, which
fails if the named attribute does not already exist. By default (no flags),
the extended attribute will be created if need be, or will simply replace
the value if the attribute exists."
The actual setxattr operation happens when the fid is clunked. At that point
the written byte count and the attr_size specified in TXATTRCREATE should be
same otherwise an error will be returned.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
TXATTRWALK: Descend a ATTR namespace
size[4] TXATTRWALK tag[2] fid[4] newfid[4] name[s]
size[4] RXATTRWALK tag[2] size[8]
txattrwalk gets a fid pointing to xattr. This fid can later be
used to get read the xattr value. If name is NULL the fid returned
can be used to get the list of extended attribute associated to
the file system object.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Implement 9p2000.L version of open(LOPEN) interface in qemu 9p server.
For LOPEN, no need to convert the flags to and from 9p mode to VFS mode.
Synopsis:
size[4] Tlopen tag[2] fid[4] mode[4]
size[4] Rlopen tag[2] qid[13] iounit[4]
Current qemu 9p server does not support following flags:
O_NOCTTY, O_NONBLOCK, O_ASYNC & O_CLOEXEC
[Fix mode format - jvrao@linux.vnet.ibm.com]
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
size[4] Trename tag[2] fid[4] newdirfid[4] name[s]
size[4] Rrename tag[2]
Implement the 2000.L rename operation. A new function
v9fs_complete_rename is introduced that acts as a common entry point
for 2000.L rename operation and 2000.U rename opearation (via wstat).
As part of this change the field 'nname' (used only for rename) is
removed from the structure V9fsWstatState. Instead a new structure
V9fsRenameState is used for rename operations both by 2000.U and 2000.L
code paths. Both 2000.U and 2000.L rename code paths construct the
V9fsRenameState structure and passes that to v9fs_complete_rename
function.
Changes from previous version:
Use qemu_mallocz to initialize
Use strcpy,strcat functions instead of memcpy
Changed the variable name to newdirfid
Introduced post rename function
Error checking
Removed nname field from V9fsWstatState
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Synopsis
size[4] Tmkdir tag[2] fid[4] name[s] mode[4] gid[4]
size[4] Rmkdir tag[2] qid[13]
Description
mkdir asks the file server to create a directory with given name,
mode and gid. The qid for the new directory is returned with
the mkdir reply message.
Note: 72 is selected as the opcode for TMKDIR from the reserved list.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
[jvrao@linux.vnet.ibm.com: Fix perm handling when creating directory]
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Implement TMKNOD as part of 2000.L Work
Synopsis
size[4] Tmknod tag[2] fid[4] name[s] mode[4] major[4] minor[4] gid[4]
size[4] Rmknod tag[2] qid[13]
Description
mknod asks the file server to create a device node with given device
type, mode and gid. The qid for the new device node is returned with
the mknod reply message.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
SYNOPSIS
size[4] Tlcreate tag[2] fid[4] name[s] flags[4] mode[4] gid[4]
size[4] Rlcreate tag[2] qid[13] iounit[4]
DESCRIPTION
The Tlreate request asks the file server to create a new regular file with the
name supplied, in the directory (dir) represented by fid.
The mode argument specifies the permissions to use. New file is created with
the uid if the fid and with supplied gid.
The flags argument represent Linux access mode flags with which the caller
is requesting to open the file with. Protocol allows all the Linux access
modes but it is upto the server to allow/disallow any of these acess modes.
If the server doesn't support any of the access mode, it is expected to
return error.
To start with we will not restricit/limit any Linux flags on this server.
If needed, We can start restricting as we move forward with various use cases.
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
This patch implements creating a symlink for TSYMLINK request
and responds with RSYMLINK. In the case of error, we return RERROR.
SYNOPSIS
size[4] Tsymlink tag[2] fid[4] name[s] symtgt[s] gid[4]
size[4] Rsymlink tag[2] qid[13]
DESCRIPTION
Create a symbolic link named 'name' pointing to 'symtgt'.
gid represents the effective group id of the caller.
The permissions of a symbolic link are irrelevant hence it is omitted
from the protocol.
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
SYNOPSIS
size[4] Tsetattr tag[2] attr[n]
size[4] Rsetattr tag[2]
DESCRIPTION
The setattr command changes some of the file status information.
attr resembles the iattr structure used in Linux kernel. It
specifies which status parameter is to be changed and to what
value. It is laid out as follows:
valid[4]
specifies which status information is to be changed. Possible
values are:
ATTR_MODE (1 << 0)
ATTR_UID (1 << 1)
ATTR_GID (1 << 2)
ATTR_SIZE (1 << 3)
ATTR_ATIME (1 << 4)
ATTR_MTIME (1 << 5)
ATTR_CTIME (1 << 5)
ATTR_ATIME_SET (1 << 7)
ATTR_MTIME_SET (1 << 8)
The last two bits represent whether the time information
is being sent by the client's user space. In the absense
of these bits the server always uses server's time.
mode[4]
File permission bits
uid[4]
Owner id of file
gid[4]
Group id of the file
size[8]
File size
atime_sec[8]
Time of last file access, seconds
atime_nsec[8]
Time of last file access, nanoseconds
mtime_sec[8]
Time of last file modification, seconds
mtime_nsec[8]
Time of last file modification, nanoseconds
Explanation of the patches:
--------------------------
*) The kernel just copies relevent contents of iattr structure to p9_iattr_dotl
structure and passes it down to the client. The only check it has is calling
inode_change_ok()
*) The p9_iattr_dotl structure does not have ctime and ia_file parameters because
I don't think these are needed in our case. The client user space can request
updating just ctime by calling chown(fd, -1, -1). This is handled on server
side without a need for putting ctime on the wire.
*) The server currently supports changing mode, time, ownership and size of the
file.
*) 9P RFC says "Either all the changes in wstat request happen, or none of them
does: if the request succeeds, all changes were made; if it fails, none were."
I have not done anything to implement this specifically because I don't see
a reason.
[jvrao@linux.vnet.ibm.com: Parts of code for handling chown(-1,-1)
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Currently v9fs_do_utimensat takes a V9fsStat argument and builds
timespec structures. It sets tv_nsec values to 0 by default. Instead
of this it should take struct timespec[2] and pass it down to the
system directly. This will make it more generic and useful
elsewhere.
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Current code resets file's atime to 0 when there is a change in mtime.
This results in resetting the atime to "1970-01-01 05:30:00". For
example, truncate -s 0 filename results in changing the mtime to the
truncate time, but resets the atime to "1970-01-01 05:30:00". utime
system call does not have any provision to set only mtime or atime. So
change v9fs_wstat_post_chmod function to use utimensat function to change
the atime and mtime fields. If tv_nsec field is set to the special value
"UTIME_OMIT", corresponding file time stamp is not updated.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
SYNOPSIS
size[4] Tgetattr tag[2] fid[4] request_mask[8]
size[4] Rgetattr tag[2] lstat[n]
DESCRIPTION
The getattr transaction inquires about the file identified by fid.
request_mask is a bit mask that specifies which fields of the
stat structure is the client interested in.
The reply will contain a machine-independent directory entry,
laid out as follows:
st_result_mask[8]
Bit mask that indicates which fields in the stat structure
have been populated by the server
qid.type[1]
the type of the file (directory, etc.), represented as a bit
vector corresponding to the high 8 bits of the file's mode
word.
qid.vers[4]
version number for given path
qid.path[8]
the file server's unique identification for the file
st_mode[4]
Permission and flags
st_uid[4]
User id of owner
st_gid[4]
Group ID of owner
st_nlink[8]
Number of hard links
st_rdev[8]
Device ID (if special file)
st_size[8]
Size, in bytes
st_blksize[8]
Block size for file system IO
st_blocks[8]
Number of file system blocks allocated
st_atime_sec[8]
Time of last access, seconds
st_atime_nsec[8]
Time of last access, nanoseconds
st_mtime_sec[8]
Time of last modification, seconds
st_mtime_nsec[8]
Time of last modification, nanoseconds
st_ctime_sec[8]
Time of last status change, seconds
st_ctime_nsec[8]
Time of last status change, nanoseconds
st_btime_sec[8]
Time of creation (birth) of file, seconds
st_btime_nsec[8]
Time of creation (birth) of file, nanoseconds
st_gen[8]
Inode generation
st_data_version[8]
Data version number
request_mask and result_mask bit masks contain the following bits
#define P9_STATS_MODE 0x00000001ULL
#define P9_STATS_NLINK 0x00000002ULL
#define P9_STATS_UID 0x00000004ULL
#define P9_STATS_GID 0x00000008ULL
#define P9_STATS_RDEV 0x00000010ULL
#define P9_STATS_ATIME 0x00000020ULL
#define P9_STATS_MTIME 0x00000040ULL
#define P9_STATS_CTIME 0x00000080ULL
#define P9_STATS_INO 0x00000100ULL
#define P9_STATS_SIZE 0x00000200ULL
#define P9_STATS_BLOCKS 0x00000400ULL
#define P9_STATS_BTIME 0x00000800ULL
#define P9_STATS_GEN 0x00001000ULL
#define P9_STATS_DATA_VERSION 0x00002000ULL
#define P9_STATS_BASIC 0x000007ffULL
#define P9_STATS_ALL 0x00003fffULL
This patch implements the client side of getattr implementation for 9P2000.L.
It introduces a new structure p9_stat_dotl for getting Linux stat information
along with QID. The data layout is similar to stat structure in Linux user
space with the following major differences:
inode (st_ino) is not part of data. Instead qid is.
device (st_dev) is not part of data because this doesn't make sense on the
client.
All time variables are 64 bit wide on the wire. The kernel seems to use
32 bit variables for these variables. However, some of the architectures
have used 64 bit variables and glibc exposes 64 bit variables to user
space on some architectures. Hence to be on the safer side we have made
these 64 bit in the protocol. Refer to the comments in
include/asm-generic/stat.h
There are some additional fields: st_btime_sec, st_btime_nsec, st_gen,
st_data_version apart from the bitmask, st_result_mask. The bit mask
is filled by the server to indicate which stat fields have been
populated by the server. Currently there is no clean way for the
server to obtain these additional fields, so it sends back just the
basic fields.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Compute iounit based on the host filesystem block size and pass it to
client with open/create response. Also return iounit as statfs's f_bsize
for optimal block size transfers.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Reviewd-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
This patch implements the server part of readdir() implementation for
9p2000.L
SYNOPSIS
size[4] Treaddir tag[2] fid[4] offset[8] count[4]
size[4] Rreaddir tag[2] count[4] data[count]
DESCRIPTION
The readdir request asks the server to read the directory specified by 'fid'
at an offset specified by 'offset' and return as many dirent structures as
possible that fit into count bytes. Each dirent structure is laid out as
follows.
qid.type[1]
the type of the file (directory, etc.), represented as a bit
vector corresponding to the high 8 bits of the file's mode
word.
qid.vers[4]
version number for given path
qid.path[8]
the file server's unique identification for the file
offset[8]
offset into the next dirent.
type[1]
type of this directory entry.
name[256]
name of this directory entry.
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Reviewed-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
In v9fs_remove_post_remove() we currently ignore the error returned by
the previous call to remove() and return an error only if freeing the
fid fails. However, the client expects to see the error from remove().
Currently the client falsely thinks that the remove call has always
succeeded. For example, doing rmdir on a non-empty directory does
not return ENOTEMPTY.
With this patch we ignore the error from free_fid(). The client cannot
use this error value anyway.
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Implement statfs support in qemu server based on Sripathi's
initial statfs patch.
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Make 9P server recognize 9P2000.L protocol version
Signed-off-by: M. Mohan Kumar <mohan@in.ibm.com>
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
When a new cluster was allocated, we only need a flush after the write to the
L2 table if it was a COW and we need to decrease the refcounts of the old
clusters.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
BDRV_O_CACHE_MASK should have been extended when cache=unsafe introduced a new
flag BDRV_O_NO_FLUSH. There are currently no users that would change their
behaviour because of this, but let's clean it up before things break.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
If qemu-img crashes during the conversion, the user will throw away the broken
output file anyway and start over. So no need to be too cautious.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Allow symbolic links which point to /dev/sgX devices.
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
I use a legacy OS which depends on some optional SCSI commands.
In fact this implementation does nothing special, but provides minimum
support for the following commands:
REZERO UNIT
WRITE AND VERIFY(10)
WRITE AND VERIFY(12)
WRITE AND VERIFY(16)
MODE SELECT(6)
MODE SELECT(10)
SEEK(6)
SEEK(10)
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Fill in word 64 of IDENTIFY data to indicate support for PIO modes 3 and 4.
This allows NetBSD guests to use UltraDMA modes instead of just PIO mode 0.
Signed-off-by: Jonathan A. Kollasch <jakllsch@kollasch.net>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
On Linux, we have code to detect CD-ROMs using an ioctl. We shouldn't lose
anything but false positives by removing the check for a /dev/cd* path.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There is no indication whether or not the sector is allocated when
nb_sectors=1:
sector allocated at offset 64 KiB
This message is produced whether or not the sector is allocated.
Simply use the same message as the plural case, I don't think the
English is so broken that we need special case output here:
0/1 sectors allocated at offset 64 KiB
This change does not affect qemu-iotests since nb_sectors=1 is not used
there.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The DBD bit does not work as expected.
SCSI-Spec:
http://ldkelley.com/SCSI2/SCSI2/SCSI2-08.html#8.2.10
"A disable block descriptors (DBD) bit of zero indicates that the target
may return zero or more block descriptors in the returned MODE SENSE
data (see 8.3.3), at the target's discretion. A DBD bit of one
specifies that the target shall not return any block descriptors in the
returned MODE SENSE data."
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
SCSI-Spec:
http://ldkelley.com/SCSI2/SCSI2/SCSI2-08.html#8.2.10
"An initiator may request any one or all of the supported mode pages
from a target. If an initiator issues a MODE SENSE command with a
page code value not implemented by the target, the target shall return
CHECK CONDITION status and shall set the sense key to ILLEGAL REQUEST
and the additional sense code to INVALID FIELD IN CDB."
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The block descriptor contains the number of blocks, not the highest LBA.
Real hard disks return 0 if the number of blocks exceed the maximum 0xFFFFFF.
SCSI-Spec:
http://ldkelley.com/SCSI2/SCSI2/SCSI2-08.html#8.3.3
"The number of blocks field specifies the number of logical blocks on the
medium to which the density code and block length fields apply. A value
of zero indicates that all of the remaining logical blocks of the logical
unit shall have the medium characteristics specified."
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The page control (PC) field defines the type of mode parameter values
to be returned in the mode pages:
PC=0 : Current values
PC=1 : Changeable values
PC=2 : Default values
PC=3 : Saved values
The current implementation always returns the same type of parameters.
This is OK for Current and Default values as we don't support changes
to be done by the MODE SELECT command.
For Saved values the following applies (implemented by this patch):
"A PC field value of 3h requests that the target return the saved
values of the mode parameters. Implementation of saved page parameters
is optional. Mode parameters not supported by the target shall be set
to zero. If saved values are not implemented, the command shall be
terminated with CHECK CONDITION status, the sense key set to
ILLEGAL REQUEST and the additional sense code set to
SAVING PARAMETERS NOT SUPPORTED."
For Changeable values the following applies (implemented by this patch):
"A PC field value of 1h requests that the target return a mask denoting
those mode parameters that are changeable. In the mask, the fields of
the mode parameters that are changeable shall be set to all one bits and
the fields of the mode parameters that are non-changeable (i.e. defined
by the target) shall be set to all zero bits."
In newer versions of the SCSI-2 spec the following clause was added.
"If the logical unit does not implement changeable parameters mode pages
and the device server receives a MODE SENSE command with 01b in the PC
field, then the command shall be terminated with CHECK CONDITION status,
with the sense key set to ILLEGAL REQUEST, and the additional sense code
set to INVALID FIELD IN CDB."
This was not yet included in the SCSI-2 Working Drafts from 1986-1993.
I assume that the variant to return CHECK CONDITION for PC=1 is not
widely implemented by real devices. I have a legacy OS which fails,
if MODE_SENSE returns non GOOD for PC=1. So for highest compatibility I
implemented the former variant with this patch.
The last Working Draft X3T9.2 Rev. 10L 7-SEP-93 can be found here:
http://ldkelley.com/SCSI2/SCSI2/SCSI2-08.html#8.2.10
In mode_sense_page() this patch also avoids multiple hard coded
definitions of the same mode page length. Instead I use the varable
p[1]. In fact the returned length of the mode pages 4 and 5 were wrong
(2 bytes less).
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The header for the MODE SENSE(10) command is 8 bytes long.
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The MODE DATA LENGTH field indicates the length in bytes of the following
data that is available to be transferred. The mode data length does not include
the number of bytes in the MODE DATA LENGTH field.
Signed-off-by: Bernhard Kohl <bernhard.kohl@nsn.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Based on a patch from Mark McLoughlin, this patch introduces a new
bottom half packet transmitter that avoids the latency imposed by
the tx_timer approach. Rather than scheduling a timer when a TX
packet comes in, schedule a bottom half to be run from the iothread.
The bottom half handler first attempts to flush the queue with
notification disabled (this is where we could race with a guest
without txburst). If we flush a full burst, reschedule immediately.
If we send short of a full burst, try to re-enable notification.
To avoid a race with TXs that may have occurred, we must then
flush again. If we find some packets to send, the guest it probably
active, so we can reschedule again.
tx_timer and tx_bh are mutually exclusive, so we can re-use the
tx_waiting flag to indicate one or the other needs to be setup.
This allows us to seamlessly migrate between timer and bh TX
handling.
The bottom half handler becomes the new default and we add a new
tx= option to virtio-net-pci. Usage:
-device virtio-net-pci,tx=timer # select timer mitigation vs "bh"
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
De-couple this from the timer since we might want to use
different backends to send the packet.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
If virtio_net_flush_tx() is called with notification disabled, we can
race with the guest, processing packets at the same rate as they
get produced. The trouble is that this means we have no guaranteed
exit condition from the function and can spend minutes in there.
Currently flush_tx is only called with notification on, which seems
to limit us to one pass through the queue per call. An upcoming
patch changes this.
Also add an option to set this value on the command line as different
workloads may wish to use different values. We can't necessarily
support any random value, so this is a developer option: x-txburst=
Usage:
-device virtio-net-pci,x-txburst=64 # 64 packets per tx flush
One pass through the queue (256) seems to be a good default value
for this, balancing latency with throughput. We use a signed int
for x-txburst because 2^31 packets in a burst would take many, many
minutes to process and it allows us to easily return a negative
value value from virtio_net_flush_tx() to indicate a back-off
or error condition.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Add an option to make the TX mitigation timer adjustable as a device
option. The 150us hard coded default used currently is reasonable,
but may not be suitable for all workloads, this gives us a way to
adjust it using a single binary. We can't support any random option
though, so use the "x-" prefix to indicate this is a developer
option. Usage:
-device virtio-net-pci,x-txtimer=500000,... # .5ms timeout
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
