Re: [PATCH v1 05/15] nitro_enclaves: Handle PCI device command requests

[Liran Alon <liran.alon@oracle.com>]

On 21/04/2020 21:41, Andra Paraschiv wrote:
> The Nitro Enclaves PCI device exposes a MMIO space that this driver
> uses to submit command requests and to receive command replies e.g. for
> enclave creation / termination or setting enclave resources.
>
> Add logic for handling PCI device command requests based on the given
> command type.
>
> Register an MSI-X interrupt vector for command reply notifications to
> handle this type of communication events.
>
> Signed-off-by: Alexandru-Catalin Vasile <lexnv@amazon.com>
> Signed-off-by: Andra Paraschiv <andraprs@amazon.com>
> ---
>   .../virt/amazon/nitro_enclaves/ne_pci_dev.c   | 264 ++++++++++++++++++
>   1 file changed, 264 insertions(+)
>
> diff --git a/drivers/virt/amazon/nitro_enclaves/ne_pci_dev.c b/drivers/virt/amazon/nitro_enclaves/ne_pci_dev.c
> index 8fbee95ea291..7453d129689a 100644
> --- a/drivers/virt/amazon/nitro_enclaves/ne_pci_dev.c
> +++ b/drivers/virt/amazon/nitro_enclaves/ne_pci_dev.c
> @@ -40,6 +40,251 @@ static const struct pci_device_id ne_pci_ids[] = {
>   
>   MODULE_DEVICE_TABLE(pci, ne_pci_ids);
>   
> +/**
> + * ne_submit_request - Submit command request to the PCI device based on the
> + * command type.
> + *
> + * This function gets called with the ne_pci_dev mutex held.
> + *
> + * @pdev: PCI device to send the command to.
> + * @cmd_type: command type of the request sent to the PCI device.
> + * @cmd_request: command request payload.
> + * @cmd_request_size: size of the command request payload.
> + *
> + * @returns: 0 on success, negative return value on failure.
> + */
> +static int ne_submit_request(struct pci_dev *pdev,
> +			     enum ne_pci_dev_cmd_type cmd_type,
> +			     void *cmd_request, size_t cmd_request_size)
> +{
> +	struct ne_pci_dev *ne_pci_dev = NULL;
These local vars are unnecessarily initialized.
> +
> +	BUG_ON(!pdev);
> +
> +	ne_pci_dev = pci_get_drvdata(pdev);
> +	BUG_ON(!ne_pci_dev);
> +	BUG_ON(!ne_pci_dev->iomem_base);
You should remove these defensive BUG_ON() calls.
> +
> +	if (WARN_ON(cmd_type <= INVALID_CMD || cmd_type >= MAX_CMD)) {
> +		dev_err_ratelimited(&pdev->dev, "Invalid cmd type=%d\n",
> +				    cmd_type);
> +
> +		return -EINVAL;
> +	}
> +
> +	if (WARN_ON(!cmd_request))
> +		return -EINVAL;
> +
> +	if (WARN_ON(cmd_request_size > NE_SEND_DATA_SIZE)) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Invalid req size=%ld for cmd type=%d\n",
> +				    cmd_request_size, cmd_type);
> +
> +		return -EINVAL;
> +	}
It doesn't make sense to have WARN_ON() print error to dmesg on every 
evaluation to true,
together with using dev_err_ratelimited() which attempts to rate-limit 
prints.

Anyway, these conditions were already checked by ne_do_request(). Why 
also check them here?

> +
> +	memcpy_toio(ne_pci_dev->iomem_base + NE_SEND_DATA, cmd_request,
> +		    cmd_request_size);
> +
> +	iowrite32(cmd_type, ne_pci_dev->iomem_base + NE_COMMAND);
> +
> +	return 0;
> +}
> +
> +/**
> + * ne_retrieve_reply - Retrieve reply from the PCI device.
> + *
> + * This function gets called with the ne_pci_dev mutex held.
> + *
> + * @pdev: PCI device to receive the reply from.
> + * @cmd_reply: command reply payload.
> + * @cmd_reply_size: size of the command reply payload.
> + *
> + * @returns: 0 on success, negative return value on failure.
> + */
> +static int ne_retrieve_reply(struct pci_dev *pdev,
> +			     struct ne_pci_dev_cmd_reply *cmd_reply,
> +			     size_t cmd_reply_size)
> +{
> +	struct ne_pci_dev *ne_pci_dev = NULL;
These local vars are unnecessarily initialized.
> +
> +	BUG_ON(!pdev);
> +
> +	ne_pci_dev = pci_get_drvdata(pdev);
> +	BUG_ON(!ne_pci_dev);
> +	BUG_ON(!ne_pci_dev->iomem_base);
You should remove these defensive BUG_ON() calls.
> +
> +	if (WARN_ON(!cmd_reply))
> +		return -EINVAL;
> +
> +	if (WARN_ON(cmd_reply_size > NE_RECV_DATA_SIZE)) {
> +		dev_err_ratelimited(&pdev->dev, "Invalid reply size=%ld\n",
> +				    cmd_reply_size);
> +
> +		return -EINVAL;
> +	}
It doesn't make sense to have WARN_ON() print error to dmesg on every 
evaluation to true,
together with using dev_err_ratelimited() which attempts to rate-limit 
prints.

Anyway, these conditions were already checked by ne_do_request(). Why 
also check them here?

> +
> +	memcpy_fromio(cmd_reply, ne_pci_dev->iomem_base + NE_RECV_DATA,
> +		      cmd_reply_size);
> +
> +	return 0;
> +}
> +
> +/**
> + * ne_wait_for_reply - Wait for a reply of a PCI command.
> + *
> + * This function gets called with the ne_pci_dev mutex held.
> + *
> + * @pdev: PCI device for which a reply is waited.
> + *
> + * @returns: 0 on success, negative return value on failure.
> + */
> +static int ne_wait_for_reply(struct pci_dev *pdev)
> +{
> +	struct ne_pci_dev *ne_pci_dev = NULL;
> +	int rc = -EINVAL;
These local vars are unnecessarily initialized.
> +
> +	BUG_ON(!pdev);
> +
> +	ne_pci_dev = pci_get_drvdata(pdev);
> +	BUG_ON(!ne_pci_dev);
You should remove these defensive BUG_ON() calls.
> +
> +	/*
> +	 * TODO: Update to _interruptible and handle interrupted wait event
> +	 * e.g. -ERESTARTSYS, incoming signals + add / update timeout.
> +	 */
> +	rc = wait_event_timeout(ne_pci_dev->cmd_reply_wait_q,
> +				atomic_read(&ne_pci_dev->cmd_reply_avail) != 0,
> +				msecs_to_jiffies(DEFAULT_TIMEOUT_MSECS));
> +	if (!rc) {
> +		pr_err("Wait event timed out when waiting for PCI cmd reply\n");
> +
> +		return -ETIMEDOUT;
> +	}
> +
> +	return 0;
> +}
> +
> +int ne_do_request(struct pci_dev *pdev, enum ne_pci_dev_cmd_type cmd_type,
> +		  void *cmd_request, size_t cmd_request_size,
> +		  struct ne_pci_dev_cmd_reply *cmd_reply, size_t cmd_reply_size)
This function is introduced in this patch but it is not used.
It will cause compiling the kernel on this commit to raise 
warnings/errors on unused functions.
You should introduce functions on the patch that they are used.
> +{
> +	struct ne_pci_dev *ne_pci_dev = NULL;
> +	int rc = -EINVAL;
These local vars are unnecessarily initialized.
> +
> +	BUG_ON(!pdev);
> +
> +	ne_pci_dev = pci_get_drvdata(pdev);
> +	BUG_ON(!ne_pci_dev);
> +	BUG_ON(!ne_pci_dev->iomem_base);
You should remove these defensive BUG_ON() calls.
> +
> +	if (WARN_ON(cmd_type <= INVALID_CMD || cmd_type >= MAX_CMD)) {
> +		dev_err_ratelimited(&pdev->dev, "Invalid cmd type=%d\n",
> +				    cmd_type);
> +
> +		return -EINVAL;
> +	}
> +
> +	if (WARN_ON(!cmd_request))
> +		return -EINVAL;
> +
> +	if (WARN_ON(cmd_request_size > NE_SEND_DATA_SIZE)) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Invalid req size=%ld for cmd type=%d\n",
> +				    cmd_request_size, cmd_type);
> +
> +		return -EINVAL;
> +	}
> +
> +	if (WARN_ON(!cmd_reply))
> +		return -EINVAL;
> +
> +	if (WARN_ON(cmd_reply_size > NE_RECV_DATA_SIZE)) {
> +		dev_err_ratelimited(&pdev->dev, "Invalid reply size=%ld\n",
> +				    cmd_reply_size);
> +
> +		return -EINVAL;
> +	}
I would consider specifying all these conditions in function 
documentation instead of enforcing them at runtime on every function call.
> +
> +	/*
> +	 * Use this mutex so that the PCI device handles one command request at
> +	 * a time.
> +	 */
> +	mutex_lock(&ne_pci_dev->pci_dev_mutex);
> +
> +	atomic_set(&ne_pci_dev->cmd_reply_avail, 0);
> +
> +	rc = ne_submit_request(pdev, cmd_type, cmd_request, cmd_request_size);
> +	if (rc < 0) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Failure in submit cmd request [rc=%d]\n",
> +				    rc);
> +
> +		mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> +
> +		return rc;
Consider leaving function with a goto to a label that unlocks mutex and 
then return.
> +	}
> +
> +	rc = ne_wait_for_reply(pdev);
> +	if (rc < 0) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Failure in wait cmd reply [rc=%d]\n",
> +				    rc);
> +
> +		mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> +
> +		return rc;
> +	}
> +
> +	rc = ne_retrieve_reply(pdev, cmd_reply, cmd_reply_size);
> +	if (rc < 0) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Failure in retrieve cmd reply [rc=%d]\n",
> +				    rc);
> +
> +		mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> +
> +		return rc;
> +	}
> +
> +	atomic_set(&ne_pci_dev->cmd_reply_avail, 0);
> +
> +	if (cmd_reply->rc < 0) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Failure in cmd process logic [rc=%d]\n",
> +				    cmd_reply->rc);
> +
> +		mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> +
> +		return cmd_reply->rc;
> +	}
> +
> +	mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> +
> +	return 0;
> +}
> +
> +/**
> + * ne_reply_handler - Interrupt handler for retrieving a reply matching
> + * a request sent to the PCI device for enclave lifetime management.
> + *
> + * @irq: received interrupt for a reply sent by the PCI device.
> + * @args: PCI device private data structure.
> + *
> + * @returns: IRQ_HANDLED on handled interrupt, IRQ_NONE otherwise.
> + */
> +static irqreturn_t ne_reply_handler(int irq, void *args)
> +{
> +	struct ne_pci_dev *ne_pci_dev = (struct ne_pci_dev *)args;
> +
> +	atomic_set(&ne_pci_dev->cmd_reply_avail, 1);
> +
> +	/* TODO: Update to _interruptible. */
> +	wake_up(&ne_pci_dev->cmd_reply_wait_q);
> +
> +	return IRQ_HANDLED;
> +}
> +
>   /**
>    * ne_setup_msix - Setup MSI-X vectors for the PCI device.
>    *
> @@ -75,8 +320,25 @@ static int ne_setup_msix(struct pci_dev *pdev, struct ne_pci_dev *ne_pci_dev)
>   		goto err_alloc_irq_vecs;
>   	}
>   
> +	/*
> +	 * This IRQ gets triggered every time the PCI device responds to a
> +	 * command request. The reply is then retrieved, reading from the MMIO
> +	 * space of the PCI device.
> +	 */
> +	rc = request_irq(pci_irq_vector(pdev, NE_VEC_REPLY),
> +			 ne_reply_handler, 0, "enclave_cmd", ne_pci_dev);
> +	if (rc < 0) {
> +		dev_err_ratelimited(&pdev->dev,
> +				    "Failure in allocating irq reply [rc=%d]\n",
> +				    rc);
> +
> +		goto err_req_irq_reply;
> +	}
> +
>   	return 0;
>   
> +err_req_irq_reply:
> +	pci_free_irq_vectors(pdev);
>   err_alloc_irq_vecs:
>   	return rc;
>   }
> @@ -232,6 +494,7 @@ static int ne_probe(struct pci_dev *pdev, const struct pci_device_id *id)
>   
>   err_ne_pci_dev_enable:
>   err_ne_pci_dev_disable:
> +	free_irq(pci_irq_vector(pdev, NE_VEC_REPLY), ne_pci_dev);
>   	pci_free_irq_vectors(pdev);
I suggest to introduce a ne_teardown_msix() utility. That is aimed to 
cleanup after ne_setup_msix().
>   err_setup_msix:
>   	pci_iounmap(pdev, ne_pci_dev->iomem_base);
> @@ -255,6 +518,7 @@ static void ne_remove(struct pci_dev *pdev)
>   
>   	pci_set_drvdata(pdev, NULL);
>   
> +	free_irq(pci_irq_vector(pdev, NE_VEC_REPLY), ne_pci_dev);
>   	pci_free_irq_vectors(pdev);
>   
>   	pci_iounmap(pdev, ne_pci_dev->iomem_base);