On Wed, 2009-02-25 at 15:54 +0000, Daniel P. Berrange wrote:
On Tue, Feb 24, 2009 at 10:21:47PM +0000, Mark McLoughlin wrote:
Add implementations of dettach, reattach and reset for PCI devices.
Background to this code can be found here:
http://marc.info/?l=kvm&m=123454366317045
Some notes:
* pci-stub was first introduced in 2.6.29; if it's not available, then dettach can only unbind the device from the host driver
* remove_id is queued up for 2.6.30; in it's absence reattach cannot re-bind the device to the original driver
* If the device supports Function Level Reset, we just don't bother doing the reset - KVM will do this before and after a guest runs
* Currently, if a reset would affect another device on the same bus, or another function on the same multi-function device, then we just refuse to do the reset. In future, we could allow the reset as long as it doesn't affect a device in use by the host or assigned to another guest
* The device reset code is similar in concept to some of Xen's code
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
I can't claim to understand all the gory PCI code in here, but a few API level suggestions.
Also, would it be worth (or possible) using libpciaccess.so for this instead of poking Linux sysfs files directly ?
Ah, yes - I knew I forgot another note! Originally the code used libpci, but since that doesn't handle malloc failure (well, you can handle it, but only by exiting) I implemented it directly. libpciaccess looks much saner, so we should use that. It doesn't take much code to implement it, though, so I don't consider switching to the library to be very high priority.
+static int +pciRead(pciDevice *dev, unsigned pos, uint8_t *buf, unsigned buflen) +{ + memset(buf, 0, buflen); + + if (pciOpenConfig(dev) < 0) + return -1; + + if (pread(dev->fd, buf, buflen, pos) < 0) { + char ebuf[1024]; + VIR_WARN(_("Failed to read from '%s' : %s"), dev->path, + virStrerror(errno, ebuf, sizeof(ebuf))); + return -1; + } + return 0; +}
I think this needs to check for a possible short read too. Perhaps better off doing a seek and then saferead(), or defining a safepread() function.
+static int +pciWrite(pciDevice *dev, unsigned pos, uint8_t *buf, unsigned buflen) +{ + if (pciOpenConfig(dev) < 0) + return -1; + + if (pwrite(dev->fd, buf, buflen, pos) < 0) { + char ebuf[1024]; + VIR_WARN(_("Failed to write to '%s' : %s"), dev->path, + virStrerror(errno, ebuf, sizeof(ebuf))); + return -1; + } + return 0; +}
Likewise here needs to validate actual num of bytes written
Yep, except we shouldn't see such failures with sysfs - e.g. the likes of libpci doesn't handle such failures. Suggest it's not a big issue and fix it by switching to libpciaccess
+int +pciDettachDevice(pciDevice *dev) +{ + char path[PATH_MAX]; + + /* Try loading the pci-stub module if it isn't already loaded; + * return an error if there is no pci-stub available. + */ + if (!virFileExists(PCI_SYSFS "drivers/pci-stub")) { + const char *const modprobeargv[] = { MODPROBE, "pci-stub", NULL }; + + if (virRun(dev->conn, modprobeargv, NULL) < 0) { + char ebuf[1024]; + VIR_WARN(_("modprobe pci-stub failed: %s"), + virStrerror(errno, ebuf, sizeof ebuf)); + } + } + + if (!virFileExists(PCI_SYSFS "drivers/pci-stub")) { + VIR_WARN(_("pci-stub driver not available, cannot bind device %s to it"), + dev->name); + } else { + /* Add the PCI device ID to pci-stub's dynamic ID table; + * this is needed to allow us to bind the device to pci-stub. + * Note: if the device is not currently bound to any driver, + * pci-stub will immediately be bound to the device. Also, note + * that if a new device with this ID is hotplugged, or if a probe + * is triggered for such a device, it will also be immediately + * bound by pci-stub. + */ + if (virFileWriteStr(PCI_SYSFS "drivers/pci-stub/new_id", dev->id) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to add PCI device ID '%s' to pci-stub/new_id"), + dev->id); + return -1; + } + } + + /* If the device is already bound to a driver, unbind it. + * Note, this will have rather unpleasant side effects if this + * PCI device happens to be IDE controller for the disk hosting + * your root filesystem. + */ + snprintf(path, sizeof(path), + PCI_SYSFS "devices/%s/driver/unbind", dev->name); + if (virFileExists(path) && virFileWriteStr(path, dev->name) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to unbind PCI device '%s'"), dev->name); + return -1; + } + + if (virFileExists(PCI_SYSFS "drivers/pci-stub")) { + /* If the device isn't already bound to pci-stub, try binding it now. + */ + snprintf(path, sizeof(path), PCI_SYSFS "devices/%s/driver", dev->name); + if (!virFileLinkPointsTo(path, PCI_SYSFS "drivers/pci-stub") && + virFileWriteStr(PCI_SYSFS "drivers/pci-stub/bind", dev->name) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to bind PCI device '%s' to pci-stub"), + dev->name); + return -1; + } + + /* If 'remove_id' exists, remove the device id from pci-stub's dynamic + * ID table so that 'drivers_probe' works below. + */ + if (virFileExists(PCI_SYSFS "drivers/pci-stub/remove_id") && + virFileWriteStr(PCI_SYSFS "drivers/pci-stub/remove_id", dev->id) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to remove PCI ID '%s' with pci-stub/remove_id"), + dev->id); + return -1; + } + } + + return 0; +} + +int +pciReAttachDevice(pciDevice *dev) +{ + if (virFileExists(PCI_SYSFS "drivers/pci-stub")) { + char path[PATH_MAX]; + + /* If the device is bound to pci-stub, unbind it. + */ + snprintf(path, sizeof(path), PCI_SYSFS "devices/%s/driver", dev->name); + if (virFileLinkPointsTo(path, PCI_SYSFS "drivers/pci-stub") && + virFileWriteStr(PCI_SYSFS "drivers/pci-stub/unbind", dev->name) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to bind PCI device '%s' to pci-stub"), + dev->name); + return -1; + } + } + + /* Trigger a re-probe of the device is not in pci-stub's dynamic + * ID table. If pci-stub is available, but 'remove_id' isn't + * available, then re-probing would just cause the device to be + * re-bound to pci-stub. + */ + if (!virFileExists(PCI_SYSFS "drivers/pci-stub") || + virFileExists(PCI_SYSFS "drivers/pci-stub/remove_id")) { + if (virFileWriteStr(PCI_SYSFS "drivers_probe", dev->name) < 0) { + virReportSystemError(dev->conn, errno, + _("Failed to trigger a re-probe for PCI device '%s'"), + dev->name); + return -1; + } + } + + return 0; +}
These functions really need to take the driver name as an argument, so we can pass 'pci-stub' for KVM (or modern pv_ops Xen dom0), or 'pciback' for old pre-pv_ops Xen dom0.
Okay.
+pciDevice *pciGetDevice (virConnectPtr conn, + unsigned vendor, + unsigned product, + unsigned domain, + unsigned bus, + unsigned slot, + unsigned function); +void pciFreeDevice (pciDevice *dev); +int pciDettachDevice (pciDevice *dev); +int pciReAttachDevice (pciDevice *dev); +int pciResetDevice (pciDevice *dev);
I prefer it to pass the virConnectPtr conn into each of these calls rather than storing it in the pciDevice struct.
Okay.
It'd also be good to remove the vendor/product ID here, and have this code lookup them up from te address info, since the domain XML format only provides the caller with the domain/bus/slot/function address info.
Okay. I'll follow up with a new series with those changes. Cheers, Mark.