xen/xen-pciback/pciback_ops.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * PCI Backend Operations - respond to PCI requests from Frontend
0004  *
0005  *   Author: Ryan Wilson <hap9@epoch.ncsc.mil>
0006  */
0007
0008 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0009 #define dev_fmt pr_fmt
0010
0011 #include <linux/moduleparam.h>
0012 #include <linux/wait.h>
0013 #include <linux/bitops.h>
0014 #include <xen/events.h>
0015 #include <linux/sched.h>
0016 #include "pciback.h"
0017
0018 static irqreturn_t xen_pcibk_guest_interrupt(int irq, void *dev_id);
0019
0020 /* Ensure a device is has the fake IRQ handler "turned on/off" and is
0021  * ready to be exported. This MUST be run after xen_pcibk_reset_device
0022  * which does the actual PCI device enable/disable.
0023  */
0024 static void xen_pcibk_control_isr(struct pci_dev *dev, int reset)
0025 {
0026     struct xen_pcibk_dev_data *dev_data;
0027     int rc;
0028     int enable = 0;
0029
0030     dev_data = pci_get_drvdata(dev);
0031     if (!dev_data)
0032         return;
0033
0034     /* We don't deal with bridges */
0035     if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
0036         return;
0037
0038     if (reset) {
0039         dev_data->enable_intx = 0;
0040         dev_data->ack_intr = 0;
0041     }
0042     enable =  dev_data->enable_intx;
0043
0044     /* Asked to disable, but ISR isn't runnig */
0045     if (!enable && !dev_data->isr_on)
0046         return;
0047
0048     /* Squirrel away the IRQs in the dev_data. We need this
0049      * b/c when device transitions to MSI, the dev->irq is
0050      * overwritten with the MSI vector.
0051      */
0052     if (enable)
0053         dev_data->irq = dev->irq;
0054
0055     /*
0056      * SR-IOV devices in all use MSI-X and have no legacy
0057      * interrupts, so inhibit creating a fake IRQ handler for them.
0058      */
0059     if (dev_data->irq == 0)
0060         goto out;
0061
0062     dev_dbg(&dev->dev, "%s: #%d %s %s%s %s-> %s\n",
0063         dev_data->irq_name,
0064         dev_data->irq,
0065         pci_is_enabled(dev) ? "on" : "off",
0066         dev->msi_enabled ? "MSI" : "",
0067         dev->msix_enabled ? "MSI/X" : "",
0068         dev_data->isr_on ? "enable" : "disable",
0069         enable ? "enable" : "disable");
0070
0071     if (enable) {
0072         /*
0073          * The MSI or MSI-X should not have an IRQ handler. Otherwise
0074          * if the guest terminates we BUG_ON in free_msi_irqs.
0075          */
0076         if (dev->msi_enabled || dev->msix_enabled)
0077             goto out;
0078
0079         rc = request_irq(dev_data->irq,
0080                 xen_pcibk_guest_interrupt, IRQF_SHARED,
0081                 dev_data->irq_name, dev);
0082         if (rc) {
0083             dev_err(&dev->dev, "%s: failed to install fake IRQ " \
0084                 "handler for IRQ %d! (rc:%d)\n",
0085                 dev_data->irq_name, dev_data->irq, rc);
0086             goto out;
0087         }
0088     } else {
0089         free_irq(dev_data->irq, dev);
0090         dev_data->irq = 0;
0091     }
0092     dev_data->isr_on = enable;
0093     dev_data->ack_intr = enable;
0094 out:
0095     dev_dbg(&dev->dev, "%s: #%d %s %s%s %s\n",
0096         dev_data->irq_name,
0097         dev_data->irq,
0098         pci_is_enabled(dev) ? "on" : "off",
0099         dev->msi_enabled ? "MSI" : "",
0100         dev->msix_enabled ? "MSI/X" : "",
0101         enable ? (dev_data->isr_on ? "enabled" : "failed to enable") :
0102             (dev_data->isr_on ? "failed to disable" : "disabled"));
0103 }
0104
0105 /* Ensure a device is "turned off" and ready to be exported.
0106  * (Also see xen_pcibk_config_reset to ensure virtual configuration space is
0107  * ready to be re-exported)
0108  */
0109 void xen_pcibk_reset_device(struct pci_dev *dev)
0110 {
0111     u16 cmd;
0112
0113     xen_pcibk_control_isr(dev, 1 /* reset device */);
0114
0115     /* Disable devices (but not bridges) */
0116     if (dev->hdr_type == PCI_HEADER_TYPE_NORMAL) {
0117 #ifdef CONFIG_PCI_MSI
0118         /* The guest could have been abruptly killed without
0119          * disabling MSI/MSI-X interrupts.*/
0120         if (dev->msix_enabled)
0121             pci_disable_msix(dev);
0122         if (dev->msi_enabled)
0123             pci_disable_msi(dev);
0124 #endif
0125         if (pci_is_enabled(dev))
0126             pci_disable_device(dev);
0127
0128         dev->is_busmaster = 0;
0129     } else {
0130         pci_read_config_word(dev, PCI_COMMAND, &cmd);
0131         if (cmd & (PCI_COMMAND_INVALIDATE)) {
0132             cmd &= ~(PCI_COMMAND_INVALIDATE);
0133             pci_write_config_word(dev, PCI_COMMAND, cmd);
0134
0135             dev->is_busmaster = 0;
0136         }
0137     }
0138 }
0139
0140 #ifdef CONFIG_PCI_MSI
0141 static
0142 int xen_pcibk_enable_msi(struct xen_pcibk_device *pdev,
0143              struct pci_dev *dev, struct xen_pci_op *op)
0144 {
0145     struct xen_pcibk_dev_data *dev_data;
0146     int status;
0147
0148     if (dev->msi_enabled)
0149         status = -EALREADY;
0150     else if (dev->msix_enabled)
0151         status = -ENXIO;
0152     else
0153         status = pci_enable_msi(dev);
0154
0155     if (status) {
0156         dev_warn_ratelimited(&dev->dev, "error enabling MSI for guest %u: err %d\n",
0157                      pdev->xdev->otherend_id, status);
0158         op->value = 0;
0159         return XEN_PCI_ERR_op_failed;
0160     }
0161
0162     /* The value the guest needs is actually the IDT vector, not
0163      * the local domain's IRQ number. */
0164
0165     op->value = dev->irq ? xen_pirq_from_irq(dev->irq) : 0;
0166
0167     dev_dbg(&dev->dev, "MSI: %d\n", op->value);
0168
0169     dev_data = pci_get_drvdata(dev);
0170     if (dev_data)
0171         dev_data->ack_intr = 0;
0172
0173     return 0;
0174 }
0175
0176 static
0177 int xen_pcibk_disable_msi(struct xen_pcibk_device *pdev,
0178               struct pci_dev *dev, struct xen_pci_op *op)
0179 {
0180     if (dev->msi_enabled) {
0181         struct xen_pcibk_dev_data *dev_data;
0182
0183         pci_disable_msi(dev);
0184
0185         dev_data = pci_get_drvdata(dev);
0186         if (dev_data)
0187             dev_data->ack_intr = 1;
0188     }
0189     op->value = dev->irq ? xen_pirq_from_irq(dev->irq) : 0;
0190
0191     dev_dbg(&dev->dev, "MSI: %d\n", op->value);
0192
0193     return 0;
0194 }
0195
0196 static
0197 int xen_pcibk_enable_msix(struct xen_pcibk_device *pdev,
0198               struct pci_dev *dev, struct xen_pci_op *op)
0199 {
0200     struct xen_pcibk_dev_data *dev_data;
0201     int i, result;
0202     struct msix_entry *entries;
0203     u16 cmd;
0204
0205     dev_dbg(&dev->dev, "enable MSI-X\n");
0206
0207     if (op->value > SH_INFO_MAX_VEC)
0208         return -EINVAL;
0209
0210     if (dev->msix_enabled)
0211         return -EALREADY;
0212
0213     /*
0214      * PCI_COMMAND_MEMORY must be enabled, otherwise we may not be able
0215      * to access the BARs where the MSI-X entries reside.
0216      * But VF devices are unique in which the PF needs to be checked.
0217      */
0218     pci_read_config_word(pci_physfn(dev), PCI_COMMAND, &cmd);
0219     if (dev->msi_enabled || !(cmd & PCI_COMMAND_MEMORY))
0220         return -ENXIO;
0221
0222     entries = kmalloc_array(op->value, sizeof(*entries), GFP_KERNEL);
0223     if (entries == NULL)
0224         return -ENOMEM;
0225
0226     for (i = 0; i < op->value; i++) {
0227         entries[i].entry = op->msix_entries[i].entry;
0228         entries[i].vector = op->msix_entries[i].vector;
0229     }
0230
0231     result = pci_enable_msix_exact(dev, entries, op->value);
0232     if (result == 0) {
0233         for (i = 0; i < op->value; i++) {
0234             op->msix_entries[i].entry = entries[i].entry;
0235             if (entries[i].vector) {
0236                 op->msix_entries[i].vector =
0237                     xen_pirq_from_irq(entries[i].vector);
0238                 dev_dbg(&dev->dev, "MSI-X[%d]: %d\n", i,
0239                     op->msix_entries[i].vector);
0240             }
0241         }
0242     } else
0243         dev_warn_ratelimited(&dev->dev, "error enabling MSI-X for guest %u: err %d!\n",
0244                      pdev->xdev->otherend_id, result);
0245     kfree(entries);
0246
0247     op->value = result;
0248     dev_data = pci_get_drvdata(dev);
0249     if (dev_data)
0250         dev_data->ack_intr = 0;
0251
0252     return result > 0 ? 0 : result;
0253 }
0254
0255 static
0256 int xen_pcibk_disable_msix(struct xen_pcibk_device *pdev,
0257                struct pci_dev *dev, struct xen_pci_op *op)
0258 {
0259     if (dev->msix_enabled) {
0260         struct xen_pcibk_dev_data *dev_data;
0261
0262         pci_disable_msix(dev);
0263
0264         dev_data = pci_get_drvdata(dev);
0265         if (dev_data)
0266             dev_data->ack_intr = 1;
0267     }
0268     /*
0269      * SR-IOV devices (which don't have any legacy IRQ) have
0270      * an undefined IRQ value of zero.
0271      */
0272     op->value = dev->irq ? xen_pirq_from_irq(dev->irq) : 0;
0273
0274     dev_dbg(&dev->dev, "MSI-X: %d\n", op->value);
0275
0276     return 0;
0277 }
0278 #endif
0279
0280 static inline bool xen_pcibk_test_op_pending(struct xen_pcibk_device *pdev)
0281 {
0282     return test_bit(_XEN_PCIF_active,
0283             (unsigned long *)&pdev->sh_info->flags) &&
0284            !test_and_set_bit(_PDEVF_op_active, &pdev->flags);
0285 }
0286
0287 /*
0288 * Now the same evtchn is used for both pcifront conf_read_write request
0289 * as well as pcie aer front end ack. We use a new work_queue to schedule
0290 * xen_pcibk conf_read_write service for avoiding confict with aer_core
0291 * do_recovery job which also use the system default work_queue
0292 */
0293 static void xen_pcibk_test_and_schedule_op(struct xen_pcibk_device *pdev)
0294 {
0295     bool eoi = true;
0296
0297     /* Check that frontend is requesting an operation and that we are not
0298      * already processing a request */
0299     if (xen_pcibk_test_op_pending(pdev)) {
0300         schedule_work(&pdev->op_work);
0301         eoi = false;
0302     }
0303     /*_XEN_PCIB_active should have been cleared by pcifront. And also make
0304     sure xen_pcibk is waiting for ack by checking _PCIB_op_pending*/
0305     if (!test_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags)
0306         && test_bit(_PCIB_op_pending, &pdev->flags)) {
0307         wake_up(&xen_pcibk_aer_wait_queue);
0308         eoi = false;
0309     }
0310
0311     /* EOI if there was nothing to do. */
0312     if (eoi)
0313         xen_pcibk_lateeoi(pdev, XEN_EOI_FLAG_SPURIOUS);
0314 }
0315
0316 /* Performing the configuration space reads/writes must not be done in atomic
0317  * context because some of the pci_* functions can sleep (mostly due to ACPI
0318  * use of semaphores). This function is intended to be called from a work
0319  * queue in process context taking a struct xen_pcibk_device as a parameter */
0320
0321 static void xen_pcibk_do_one_op(struct xen_pcibk_device *pdev)
0322 {
0323     struct pci_dev *dev;
0324     struct xen_pcibk_dev_data *dev_data = NULL;
0325     struct xen_pci_op *op = &pdev->op;
0326     int test_intx = 0;
0327 #ifdef CONFIG_PCI_MSI
0328     unsigned int nr = 0;
0329 #endif
0330
0331     *op = pdev->sh_info->op;
0332     barrier();
0333     dev = xen_pcibk_get_pci_dev(pdev, op->domain, op->bus, op->devfn);
0334
0335     if (dev == NULL)
0336         op->err = XEN_PCI_ERR_dev_not_found;
0337     else {
0338         dev_data = pci_get_drvdata(dev);
0339         if (dev_data)
0340             test_intx = dev_data->enable_intx;
0341         switch (op->cmd) {
0342         case XEN_PCI_OP_conf_read:
0343             op->err = xen_pcibk_config_read(dev,
0344                   op->offset, op->size, &op->value);
0345             break;
0346         case XEN_PCI_OP_conf_write:
0347             op->err = xen_pcibk_config_write(dev,
0348                   op->offset, op->size, op->value);
0349             break;
0350 #ifdef CONFIG_PCI_MSI
0351         case XEN_PCI_OP_enable_msi:
0352             op->err = xen_pcibk_enable_msi(pdev, dev, op);
0353             break;
0354         case XEN_PCI_OP_disable_msi:
0355             op->err = xen_pcibk_disable_msi(pdev, dev, op);
0356             break;
0357         case XEN_PCI_OP_enable_msix:
0358             nr = op->value;
0359             op->err = xen_pcibk_enable_msix(pdev, dev, op);
0360             break;
0361         case XEN_PCI_OP_disable_msix:
0362             op->err = xen_pcibk_disable_msix(pdev, dev, op);
0363             break;
0364 #endif
0365         default:
0366             op->err = XEN_PCI_ERR_not_implemented;
0367             break;
0368         }
0369     }
0370     if (!op->err && dev && dev_data) {
0371         /* Transition detected */
0372         if ((dev_data->enable_intx != test_intx))
0373             xen_pcibk_control_isr(dev, 0 /* no reset */);
0374     }
0375     pdev->sh_info->op.err = op->err;
0376     pdev->sh_info->op.value = op->value;
0377 #ifdef CONFIG_PCI_MSI
0378     if (op->cmd == XEN_PCI_OP_enable_msix && op->err == 0) {
0379         unsigned int i;
0380
0381         for (i = 0; i < nr; i++)
0382             pdev->sh_info->op.msix_entries[i].vector =
0383                 op->msix_entries[i].vector;
0384     }
0385 #endif
0386     /* Tell the driver domain that we're done. */
0387     wmb();
0388     clear_bit(_XEN_PCIF_active, (unsigned long *)&pdev->sh_info->flags);
0389     notify_remote_via_irq(pdev->evtchn_irq);
0390
0391     /* Mark that we're done. */
0392     smp_mb__before_atomic(); /* /after/ clearing PCIF_active */
0393     clear_bit(_PDEVF_op_active, &pdev->flags);
0394     smp_mb__after_atomic(); /* /before/ final check for work */
0395 }
0396
0397 void xen_pcibk_do_op(struct work_struct *data)
0398 {
0399     struct xen_pcibk_device *pdev =
0400         container_of(data, struct xen_pcibk_device, op_work);
0401
0402     do {
0403         xen_pcibk_do_one_op(pdev);
0404     } while (xen_pcibk_test_op_pending(pdev));
0405
0406     xen_pcibk_lateeoi(pdev, 0);
0407 }
0408
0409 irqreturn_t xen_pcibk_handle_event(int irq, void *dev_id)
0410 {
0411     struct xen_pcibk_device *pdev = dev_id;
0412     bool eoi;
0413
0414     /* IRQs might come in before pdev->evtchn_irq is written. */
0415     if (unlikely(pdev->evtchn_irq != irq))
0416         pdev->evtchn_irq = irq;
0417
0418     eoi = test_and_set_bit(_EOI_pending, &pdev->flags);
0419     WARN(eoi, "IRQ while EOI pending\n");
0420
0421     xen_pcibk_test_and_schedule_op(pdev);
0422
0423     return IRQ_HANDLED;
0424 }
0425 static irqreturn_t xen_pcibk_guest_interrupt(int irq, void *dev_id)
0426 {
0427     struct pci_dev *dev = (struct pci_dev *)dev_id;
0428     struct xen_pcibk_dev_data *dev_data = pci_get_drvdata(dev);
0429
0430     if (dev_data->isr_on && dev_data->ack_intr) {
0431         dev_data->handled++;
0432         if ((dev_data->handled % 1000) == 0) {
0433             if (xen_test_irq_shared(irq)) {
0434                 dev_info(&dev->dev, "%s IRQ line is not shared "
0435                     "with other domains. Turning ISR off\n",
0436                      dev_data->irq_name);
0437                 dev_data->ack_intr = 0;
0438             }
0439         }
0440         return IRQ_HANDLED;
0441     }
0442     return IRQ_NONE;
0443 }