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

[Paraschiv, Andra-Irina]

On 25/04/2020 17:52, Liran Alon wrote:
> 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@xxxxxxxxxx>
> > Signed-off-by: Andra Paraschiv <andraprs@xxxxxxxxxx>
> > ---
> > Â .../virt/amazon/nitro_enclaves/ne_pci_dev.cÂÂ | 264 ++++++++++++++++++
> > Â 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[] = {
> > Â Â MODULE_DEVICE_TABLE(pci, ne_pci_ids);
> > Â +/**
> > + * 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,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ enum ne_pci_dev_cmd_type cmd_type,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void *cmd_request, size_t cmd_request_size)
> > +{
> > +ÂÂÂ struct ne_pci_dev *ne_pci_dev = NULL;
> These local vars are unnecessarily initialized.

I would keep this initialized overall.

> > +
> > +ÂÂÂ BUG_ON(!pdev);
> > +
> > +ÂÂÂ ne_pci_dev = pci_get_drvdata(pdev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev->iomem_base);
> You should remove these defensive BUG_ON() calls.

Done.

> > +
> > +ÂÂÂ if (WARN_ON(cmd_type <= INVALID_CMD || cmd_type >= MAX_CMD)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev, "Invalid cmd type=%d\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_type);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ if (WARN_ON(!cmd_request))
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +
> > +ÂÂÂ if (WARN_ON(cmd_request_size > NE_SEND_DATA_SIZE)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Invalid req size=%ld for cmd type=%d\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_request_size, cmd_type);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> 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?

Updated to not use WARN_ON. Right, they were checked before, but I kept 
them here just for checking the parameters.

> > +
> > +ÂÂÂ memcpy_toio(ne_pci_dev->iomem_base + NE_SEND_DATA, cmd_request,
> > +ÂÂÂÂÂÂÂÂÂÂÂ cmd_request_size);
> > +
> > +ÂÂÂ iowrite32(cmd_type, ne_pci_dev->iomem_base + NE_COMMAND);
> > +
> > +ÂÂÂ 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,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ struct ne_pci_dev_cmd_reply *cmd_reply,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ size_t cmd_reply_size)
> > +{
> > +ÂÂÂ struct ne_pci_dev *ne_pci_dev = NULL;
> These local vars are unnecessarily initialized.
> > +
> > +ÂÂÂ BUG_ON(!pdev);
> > +
> > +ÂÂÂ ne_pci_dev = pci_get_drvdata(pdev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev->iomem_base);
> You should remove these defensive BUG_ON() calls.
> > +
> > +ÂÂÂ if (WARN_ON(!cmd_reply))
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +
> > +ÂÂÂ if (WARN_ON(cmd_reply_size > NE_RECV_DATA_SIZE)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev, "Invalid reply size=%ld\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_reply_size);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> 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?
>
> > +
> > +ÂÂÂ memcpy_fromio(cmd_reply, ne_pci_dev->iomem_base + NE_RECV_DATA,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_reply_size);
> > +
> > +ÂÂÂ 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)
> > +{
> > +ÂÂÂ struct ne_pci_dev *ne_pci_dev = NULL;
> > +ÂÂÂ int rc = -EINVAL;
> These local vars are unnecessarily initialized.
> > +
> > +ÂÂÂ BUG_ON(!pdev);
> > +
> > +ÂÂÂ ne_pci_dev = pci_get_drvdata(pdev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev);
> You should remove these defensive BUG_ON() calls.
> > +
> > +ÂÂÂ /*
> > +ÂÂÂÂ * TODO: Update to _interruptible and handle interrupted wait event
> > +ÂÂÂÂ * e.g. -ERESTARTSYS, incoming signals + add / update timeout.
> > +ÂÂÂÂ */
> > +ÂÂÂ rc = wait_event_timeout(ne_pci_dev->cmd_reply_wait_q,
> > + atomic_read(&ne_pci_dev->cmd_reply_avail) != 0,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ msecs_to_jiffies(DEFAULT_TIMEOUT_MSECS));
> > +ÂÂÂ if (!rc) {
> > +ÂÂÂÂÂÂÂ pr_err("Wait event timed out when waiting for PCI cmd 
> > reply\n");
> > +
> > +ÂÂÂÂÂÂÂ return -ETIMEDOUT;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ return 0;
> > +}
> > +
> > +int ne_do_request(struct pci_dev *pdev, enum ne_pci_dev_cmd_type 
> > cmd_type,
> > +ÂÂÂÂÂÂÂÂÂ void *cmd_request, size_t cmd_request_size,
> > +ÂÂÂÂÂÂÂÂÂ 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.

This function is externally available, via the ne_pci_dev header, so it 
shouldn't raise warnings.

> > +{
> > +ÂÂÂ struct ne_pci_dev *ne_pci_dev = NULL;
> > +ÂÂÂ int rc = -EINVAL;
> These local vars are unnecessarily initialized.
> > +
> > +ÂÂÂ BUG_ON(!pdev);
> > +
> > +ÂÂÂ ne_pci_dev = pci_get_drvdata(pdev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev);
> > +ÂÂÂ BUG_ON(!ne_pci_dev->iomem_base);
> You should remove these defensive BUG_ON() calls.
> > +
> > +ÂÂÂ if (WARN_ON(cmd_type <= INVALID_CMD || cmd_type >= MAX_CMD)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev, "Invalid cmd type=%d\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_type);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ if (WARN_ON(!cmd_request))
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +
> > +ÂÂÂ if (WARN_ON(cmd_request_size > NE_SEND_DATA_SIZE)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Invalid req size=%ld for cmd type=%d\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_request_size, cmd_type);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ if (WARN_ON(!cmd_reply))
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +
> > +ÂÂÂ if (WARN_ON(cmd_reply_size > NE_RECV_DATA_SIZE)) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev, "Invalid reply size=%ld\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_reply_size);
> > +
> > +ÂÂÂÂÂÂÂ return -EINVAL;
> > +ÂÂÂ }
> I would consider specifying all these conditions in function 
> documentation instead of enforcing them at runtime on every function 
> call.

I think that both PCI dev logic checks and documentation would be 
helpful in this case. :)

> > +
> > +ÂÂÂ /*
> > +ÂÂÂÂ * Use this mutex so that the PCI device handles one command 
> > request at
> > +ÂÂÂÂ * a time.
> > +ÂÂÂÂ */
> > +ÂÂÂ mutex_lock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂ atomic_set(&ne_pci_dev->cmd_reply_avail, 0);
> > +
> > +ÂÂÂ rc = ne_submit_request(pdev, cmd_type, cmd_request, 
> > cmd_request_size);
> > +ÂÂÂ if (rc < 0) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Failure in submit cmd request [rc=%d]\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ rc);
> > +
> > +ÂÂÂÂÂÂÂ mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂÂÂÂÂ return rc;
> Consider leaving function with a goto to a label that unlocks mutex 
> and then return.

Done, I added a goto for mutex unlock and return. In this patch and in a 
following one, that was having a similar cleanup structure.

> > +ÂÂÂ }
> > +
> > +ÂÂÂ rc = ne_wait_for_reply(pdev);
> > +ÂÂÂ if (rc < 0) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Failure in wait cmd reply [rc=%d]\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ rc);
> > +
> > +ÂÂÂÂÂÂÂ mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂÂÂÂÂ return rc;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ rc = ne_retrieve_reply(pdev, cmd_reply, cmd_reply_size);
> > +ÂÂÂ if (rc < 0) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Failure in retrieve cmd reply [rc=%d]\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ rc);
> > +
> > +ÂÂÂÂÂÂÂ mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂÂÂÂÂ return rc;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ atomic_set(&ne_pci_dev->cmd_reply_avail, 0);
> > +
> > +ÂÂÂ if (cmd_reply->rc < 0) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Failure in cmd process logic [rc=%d]\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ cmd_reply->rc);
> > +
> > +ÂÂÂÂÂÂÂ mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂÂÂÂÂ return cmd_reply->rc;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ mutex_unlock(&ne_pci_dev->pci_dev_mutex);
> > +
> > +ÂÂÂ 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)
> > +{
> > +ÂÂÂ struct ne_pci_dev *ne_pci_dev = (struct ne_pci_dev *)args;
> > +
> > +ÂÂÂ atomic_set(&ne_pci_dev->cmd_reply_avail, 1);
> > +
> > +ÂÂÂ /* TODO: Update to _interruptible. */
> > +ÂÂÂ wake_up(&ne_pci_dev->cmd_reply_wait_q);
> > +
> > +ÂÂÂ return IRQ_HANDLED;
> > +}
> > +
> > Â /**
> > ÂÂ * ne_setup_msix - Setup MSI-X vectors for the PCI device.
> > ÂÂ *
> > @@ -75,8 +320,25 @@ static int ne_setup_msix(struct pci_dev *pdev, 
> > struct ne_pci_dev *ne_pci_dev)
> > ÂÂÂÂÂÂÂÂÂ goto err_alloc_irq_vecs;
> > ÂÂÂÂÂ }
> > Â +ÂÂÂ /*
> > +ÂÂÂÂ * This IRQ gets triggered every time the PCI device responds to a
> > +ÂÂÂÂ * command request. The reply is then retrieved, reading from 
> > the MMIO
> > +ÂÂÂÂ * space of the PCI device.
> > +ÂÂÂÂ */
> > +ÂÂÂ rc = request_irq(pci_irq_vector(pdev, NE_VEC_REPLY),
> > +ÂÂÂÂÂÂÂÂÂÂÂÂ ne_reply_handler, 0, "enclave_cmd", ne_pci_dev);
> > +ÂÂÂ if (rc < 0) {
> > +ÂÂÂÂÂÂÂ dev_err_ratelimited(&pdev->dev,
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ "Failure in allocating irq reply [rc=%d]\n",
> > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ rc);
> > +
> > +ÂÂÂÂÂÂÂ goto err_req_irq_reply;
> > +ÂÂÂ }
> > +
> > ÂÂÂÂÂ return 0;
> > Â +err_req_irq_reply:
> > +ÂÂÂ pci_free_irq_vectors(pdev);
> > Â err_alloc_irq_vecs:
> > ÂÂÂÂÂ return rc;
> > Â }
> > @@ -232,6 +494,7 @@ static int ne_probe(struct pci_dev *pdev, const 
> > struct pci_device_id *id)
> > Â Â err_ne_pci_dev_enable:
> > Â err_ne_pci_dev_disable:
> > +ÂÂÂ free_irq(pci_irq_vector(pdev, NE_VEC_REPLY), ne_pci_dev);
> > ÂÂÂÂÂ pci_free_irq_vectors(pdev);
> I suggest to introduce a ne_teardown_msix() utility. That is aimed to 
> cleanup after ne_setup_msix().

I added this functionality in a new function, then I used it for cleanup 
in this function and teardown in pci remove function.

Thank you.

Andra

> > Â err_setup_msix:
> > ÂÂÂÂÂ pci_iounmap(pdev, ne_pci_dev->iomem_base);
> > @@ -255,6 +518,7 @@ static void ne_remove(struct pci_dev *pdev)
> > Â ÂÂÂÂÂ pci_set_drvdata(pdev, NULL);
> > Â +ÂÂÂ free_irq(pci_irq_vector(pdev, NE_VEC_REPLY), ne_pci_dev);
> > ÂÂÂÂÂ pci_free_irq_vectors(pdev);
> > Â ÂÂÂÂÂ pci_iounmap(pdev, ne_pci_dev->iomem_base);



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.