Re: [PATCH v1 04/15] nitro_enclaves: Init PCI device driver

["Paraschiv, Andra-Irina" <andraprs@amazon.com>]

On 25/04/2020 17:25, Liran Alon wrote:
>
> On 21/04/2020 21:41, Andra Paraschiv wrote:
>> +
>> +/**
>> + * ne_setup_msix - Setup MSI-X vectors for the PCI device.
>> + *
>> + * @pdev: PCI device to setup the MSI-X for.
>> + * @ne_pci_dev: PCI device private data structure.
>> + *
>> + * @returns: 0 on success, negative return value on failure.
>> + */
>> +static int ne_setup_msix(struct pci_dev *pdev, struct ne_pci_dev 
>> *ne_pci_dev)
>> +{
>> +    int nr_vecs = 0;
>> +    int rc = -EINVAL;
>> +
>> +    BUG_ON(!ne_pci_dev);
> This kind of defensive programming does not align with Linux coding 
> convention.
> I think these BUG_ON() conditions should be removed.

I replaced with WARN_ON here and in the other places in the codebase.

>> +
>> +    nr_vecs = pci_msix_vec_count(pdev);
>> +    if (nr_vecs < 0) {
>> +        rc = nr_vecs;
>> +
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in getting vec count [rc=%d]\n",
>> +                    rc);
>> +
>> +        return rc;
>> +    }
>> +
>> +    rc = pci_alloc_irq_vectors(pdev, nr_vecs, nr_vecs, PCI_IRQ_MSIX);
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in alloc MSI-X vecs [rc=%d]\n",
>> +                    rc);
>> +
>> +        goto err_alloc_irq_vecs;
> You should just replace this with "return rc;" as no cleanup is 
> required here.

Done.

>> +    }
>> +
>> +    return 0;
>> +
>> +err_alloc_irq_vecs:
>> +    return rc;
>> +}
>> +
>> +/**
>> + * ne_pci_dev_enable - Select PCI device version and enable it.
>> + *
>> + * @pdev: PCI device to select version for and then enable.
>> + * @ne_pci_dev: PCI device private data structure.
>> + *
>> + * @returns: 0 on success, negative return value on failure.
>> + */
>> +static int ne_pci_dev_enable(struct pci_dev *pdev,
>> +                 struct ne_pci_dev *ne_pci_dev)
>> +{
>> +    u8 dev_enable_reply = 0;
>> +    u16 dev_version_reply = 0;
>> +
>> +    BUG_ON(!pdev);
>> +    BUG_ON(!ne_pci_dev);
>> +    BUG_ON(!ne_pci_dev->iomem_base);
> Same.
>> +
>> +    iowrite16(NE_VERSION_MAX, ne_pci_dev->iomem_base + NE_VERSION);
>> +
>> +    dev_version_reply = ioread16(ne_pci_dev->iomem_base + NE_VERSION);
>> +    if (dev_version_reply != NE_VERSION_MAX) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci dev version cmd\n");
>> +
>> +        return -EIO;
>> +    }
>> +
>> +    iowrite8(NE_ENABLE_ON, ne_pci_dev->iomem_base + NE_ENABLE);
>> +
>> +    dev_enable_reply = ioread8(ne_pci_dev->iomem_base + NE_ENABLE);
>> +    if (dev_enable_reply != NE_ENABLE_ON) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci dev enable cmd\n");
>> +
>> +        return -EIO;
>> +    }
>> +
>> +    return 0;
>> +}
>> +
>> +/**
>> + * ne_pci_dev_disable - Disable PCI device.
>> + *
>> + * @pdev: PCI device to disable.
>> + * @ne_pci_dev: PCI device private data structure.
>> + *
>> + * @returns: 0 on success, negative return value on failure.
>> + */
>> +static int ne_pci_dev_disable(struct pci_dev *pdev,
>> +                  struct ne_pci_dev *ne_pci_dev)
>> +{
>> +    u8 dev_disable_reply = 0;
>> +
>> +    BUG_ON(!pdev);
>> +    BUG_ON(!ne_pci_dev);
>> +    BUG_ON(!ne_pci_dev->iomem_base);
> Same.
>> +
>> +    iowrite8(NE_ENABLE_OFF, ne_pci_dev->iomem_base + NE_ENABLE);
>> +
>> +    /*
>> +     * TODO: Check for NE_ENABLE_OFF in a loop, to handle cases when 
>> the
>> +     * device state is not immediately set to disabled and going 
>> through a
>> +     * transitory state of disabling.
>> +     */
>> +    dev_disable_reply = ioread8(ne_pci_dev->iomem_base + NE_ENABLE);
>> +    if (dev_disable_reply != NE_ENABLE_OFF) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci dev disable cmd\n");
>> +
>> +        return -EIO;
>> +    }
>> +
>> +    return 0;
>> +}
>> +
>> +static int ne_probe(struct pci_dev *pdev, const struct pci_device_id 
>> *id)
>> +{
>> +    struct ne_pci_dev *ne_pci_dev = NULL;
>> +    int rc = -EINVAL;
> Unnecessary variable initialization.
> ne_pci_dev and rc are initialized below always before they are used.

I would rather keep the initialization in place overall, to not have a 
mix of init and uninit vars, when needed.

>> +
>> +    ne_pci_dev = kzalloc(sizeof(*ne_pci_dev), GFP_KERNEL);
>> +    if (!ne_pci_dev)
>> +        return -ENOMEM;
>> +
>> +    rc = pci_enable_device(pdev);
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci dev enable [rc=%d]\n", rc);
>> +
> Why is this dev_err_ratelimited() instead of dev_err()?
> Same for the rest of error printing in this probe() method and other 
> places in this patch.

Just to avoid any misbehaving scenario, where would be way too many logs 
in a short timeframe. Here may not happen, but while handling PCI dev 
requests could be.

>> +        goto err_pci_enable_dev;
> I find it confusing that the error labels are named based on the 
> failure-case they are used,
> instead of the action they do (i.e. Unwind previous successful 
> operation that requires unwinding).
> This doesn't seem to match Linux kernel coding convention.
> It also created an unnecessary 2 labels pointing to the same place in 
> cleanup code.

Yep, that's better this way wrt the naming of the labels. I updated the 
gotos in the patch series.

>> +    }
>> +
>> +    rc = pci_request_regions_exclusive(pdev, "ne_pci_dev");
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci request regions [rc=%d]\n",
>> +                    rc);
>> +
>> +        goto err_req_regions;
>> +    }
>> +
>> +    ne_pci_dev->iomem_base = pci_iomap(pdev, PCI_BAR_NE, 0);
>> +    if (!ne_pci_dev->iomem_base) {
>> +        rc = -ENOMEM;
>> +
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci bar mapping [rc=%d]\n", rc);
>> +
>> +        goto err_iomap;
>> +    }
>> +
>> +    rc = ne_setup_msix(pdev, ne_pci_dev);
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in pci dev msix setup [rc=%d]\n",
>> +                    rc);
>> +
>> +        goto err_setup_msix;
>> +    }
>> +
>> +    rc = ne_pci_dev_disable(pdev, ne_pci_dev);
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in ne_pci_dev disable [rc=%d]\n",
>> +                    rc);
>> +
>> +        goto err_ne_pci_dev_disable;
>> +    }
> It seems weird that we need to disable the device before enabling it 
> on the probe() method.
> Why can't we just enable the device without disabling it?

The pci dev disable call cleanups the internal state of the device and 
terminates any running / "dangling" enclaves; here it is included just 
in case any remaining state from a previous PCI device use. The below 
enable call would fail in that case, though.

>> +
>> +    rc = ne_pci_dev_enable(pdev, ne_pci_dev);
>> +    if (rc < 0) {
>> +        dev_err_ratelimited(&pdev->dev,
>> +                    "Failure in ne_pci_dev enable [rc=%d]\n",
>> +                    rc);
>> +
>> +        goto err_ne_pci_dev_enable;
>> +    }
>> +
>> +    atomic_set(&ne_pci_dev->cmd_reply_avail, 0);
>> +    init_waitqueue_head(&ne_pci_dev->cmd_reply_wait_q);
>> +    INIT_LIST_HEAD(&ne_pci_dev->enclaves_list);
>> +    mutex_init(&ne_pci_dev->enclaves_list_mutex);
>> +    mutex_init(&ne_pci_dev->pci_dev_mutex);
>> +
>> +    pci_set_drvdata(pdev, ne_pci_dev);
> If you would have pci_set_drvdata() as one of the first operations in 
> ne_probe(), then you could have avoided
> passing both struct pci_dev  and struct ne_pci_dev parameters to 
> ne_setup_msix(), ne_pci_dev_enable() and ne_pci_dev_disable().
> Which would have been a bit more elegant.

Fair point. I moved pci_set_drvdata() upper in the logic and updated the 
signature of the functions to only include the pci_dev parameter.

>> +
>> +    return 0;
>> +
>> +err_ne_pci_dev_enable:
>> +err_ne_pci_dev_disable:
>> +    pci_free_irq_vectors(pdev);
>> +err_setup_msix:
>> +    pci_iounmap(pdev, ne_pci_dev->iomem_base);
>> +err_iomap:
>> +    pci_release_regions(pdev);
>> +err_req_regions:
>> +    pci_disable_device(pdev);
>> +err_pci_enable_dev:
>> +    kzfree(ne_pci_dev);
> An empty new-line is appropriate here.
> To separate the return statement from the cleanup logic.

Done.

>> +    return rc;
>> +}
>> +
>> +static void ne_remove(struct pci_dev *pdev)
>> +{
>> +    struct ne_pci_dev *ne_pci_dev = pci_get_drvdata(pdev);
>> +
>> +    if (!ne_pci_dev || !ne_pci_dev->iomem_base)
>> +        return;
> Why is this condition necessary?
> The ne_remove() function should be called only in case ne_probe() 
> succeeded.
> In that case, both ne_pci_dev and ne_pci_dev->iomem_base should be 
> non-NULL.

Correct, that shouldn't happen.

Just for early exit in case of bad behavior.

>> +
>> +    ne_pci_dev_disable(pdev, ne_pci_dev);
>> +
>> +    pci_set_drvdata(pdev, NULL);
>> +
>> +    pci_free_irq_vectors(pdev);
>> +
>> +    pci_iounmap(pdev, ne_pci_dev->iomem_base);
>> +
>> +    kzfree(ne_pci_dev);
>> +
>> +    pci_release_regions(pdev);
>> +
>> +    pci_disable_device(pdev);
> You should aspire to keep ne_remove() order of operations to be the 
> reverse order of operations done in ne_probe().
> Which would also nicely match the order of operations done in 
> ne_probe() cleanup.
> i.e. The following order:
>
> pci_set_drvdata();
> ne_pci_dev_disable();
> pci_free_irq_vectors();
> pci_iounmap();
> pci_release_regions();
> pci_disable_device()
> kzfree();

I updated the order of operations.


Thanks for review, Liran.

Andra





Amazon Development Center (Romania) S.R.L. registered office: 27A Sf. Lazar Street, UBC5, floor 2, Iasi, Iasi County, 700045, Romania. Registered in Romania. Registration number J22/2621/2005.