nvme pci: two fixes on nvme_setup_irqs

Hi,

The 1st one fixes the case that -ENOSPC is returned from pci_alloc_irq_vectors_affinity().

The 2nd one fixes the case that -EINVAL is returned from pci_alloc_irq_vectors_affinity().

Ming Lei (2):
nvme pci: fix nvme_setup_irqs()
nvme pci: try to allocate multiple irq vectors again in case of
-EINVAL

 drivers/nvme/host/pci.c | 26 ++++++++++++++------------
 1 file changed, 14 insertions(+), 12 deletions(-)

2.9.5

When -ENOSPC is returned from pci_alloc_irq_vectors_affinity(), we still try to allocate multiple irq vectors again, so irq queues covers the admin queue acturally. But we don't consider that, then number of the allocated irq vector may be same with sum of io_queues[HCTX_TYPE_DEFAULT] and io_queues[HCTX_TYPE_READ], this way is wrong, and finally breaks nvme_pci_map_queues(), and triggeres warning from pci_irq_get_affinity().

Irq queues should cover admin queues, this patch makes this point explicitely in nvme_calc_io_queues().

---
 drivers/nvme/host/pci.c | 21 +++++++++++++--------
 1 file changed, 13 insertions(+), 8 deletions(-)

index 5a0bf6a24d50..584ea7a57122 100644
@@ -2028,14 +2028,18 @@ static int nvme_setup_host_mem(struct nvme_dev *dev)
 	return ret;
 }
 
+/* irq_queues covers admin queue */
 static void nvme_calc_io_queues(struct nvme_dev *dev, unsigned int irq_queues)
 {
 	unsigned int this_w_queues = write_queues;
 
+	WARN_ON(!irq_queues);
+
 	/*
-	 * Setup read/write queue split
+	 * Setup read/write queue split, assign admin queue one independent
+	 * irq vector if irq_queues is > 1.
 	 */
-	if (irq_queues == 1) {
+	if (irq_queues <= 2) {
 		dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
 		dev->io_queues[HCTX_TYPE_READ] = 0;
 		return;

@@ -2043,21 +2047,21 @@ static void nvme_calc_io_queues(struct nvme_dev *dev, unsigned int irq_queues)
 
 	/*
 	 * If 'write_queues' is set, ensure it leaves room for at least
-	 * one read queue
+	 * one read queue and one admin queue
 	 */
 	if (this_w_queues >= irq_queues)
-		this_w_queues = irq_queues - 1;
+		this_w_queues = irq_queues - 2;
 
 	/*
 	 * If 'write_queues' is set to zero, reads and writes will share
 	 * a queue set.
 	 */
 	if (!this_w_queues) {
-		dev->io_queues[HCTX_TYPE_DEFAULT] = irq_queues;
+		dev->io_queues[HCTX_TYPE_DEFAULT] = irq_queues - 1;
 		dev->io_queues[HCTX_TYPE_READ] = 0;
 	} else {
 		dev->io_queues[HCTX_TYPE_DEFAULT] = this_w_queues;
-		dev->io_queues[HCTX_TYPE_READ] = irq_queues - this_w_queues;
+		dev->io_queues[HCTX_TYPE_READ] = irq_queues - this_w_queues - 1;
 	}
 }
 

@@ -2082,7 +2086,7 @@ static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
 		this_p_queues = nr_io_queues - 1;
 		irq_queues = 1;
 	} else {
-		irq_queues = nr_io_queues - this_p_queues;
+		irq_queues = nr_io_queues - this_p_queues + 1;
 	}
 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
 

@@ -2102,8 +2106,9 @@ static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
 		 * If we got a failure and we're down to asking for just
 		 * 1 + 1 queues, just ask for a single vector. We'll share
 		 * that between the single IO queue and the admin queue.
+		 * Otherwise, we assign one independent vector to admin queue.
 		 */
-		if (result >= 0 && irq_queues > 1)
+		if (irq_queues > 1)
 			irq_queues = irq_sets[0] + irq_sets[1] + 1;
 
 		result = pci_alloc_irq_vectors_affinity(pdev, irq_queues,

2.9.5

It is observed on QEMU that pci_alloc_irq_vectors_affinity() may returns -EINVAL when the requested number is too big(such as 64).

However, the allocation can be done successfully if we ask for a smaller number, such as 63.

So reduce queue number and reallocate again in case of -EINVAL.

---
 drivers/nvme/host/pci.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

index 584ea7a57122..d839bbe408c3 100644
@@ -2121,14 +2121,11 @@ static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
 		 * to decrease our ask. If we get EINVAL, the platform
 		 * likely does not. Back down to ask for just one vector.
 		 */
-		if (result == -ENOSPC) {
+		if (result == -ENOSPC || result == -EINVAL) {
 			irq_queues--;
 			if (!irq_queues)
 				return result;
 			continue;
-		} else if (result == -EINVAL) {
-			irq_queues = 1;
-			continue;
 		} else if (result <= 0)
 			return -EIO;
 		break;

2.9.5

Which is not how this API is supposed to work and documented to work.

We need to fix pci_alloc_irq_vectors_affinity to not return a spurious error and just return the allocated number of vectors instead of hacking around that in drivers.

Yeah, you are right.

The issue is that QEMU nvme-pci is MSIX-capable only, and hasn't MSI capability.

__pci_enable_msix_range() actually returns -ENOSPC, but __pci_enable_msi_range() returns -EINVAL because dev->msi_cap is zero.

Maybe we need the following fix?

index 265ed3e4c920..b0bf260dc154 100644
@@ -1186,7 +1186,7 @@ int pci_alloc_irq_vectors_affinity(struct pci_dev *dev, unsigned int min_vecs,
                        return vecs;
        }
 
-       if (flags & PCI_IRQ_MSI) {
+       if ((flags & PCI_IRQ_MSI) && dev->msi_cap) {
                vecs = __pci_enable_msi_range(dev, min_vecs, max_vecs, affd);
                if (vecs > 0)
                        return vecs;

Thanks,
Ming

Should it matter? We still get a negative vecs back, and still fall back to the next option. Unless ther are no irqs available at all for the selected types pci_alloc_irq_vectors_affinity should never return an error.

The patch in last email does fix this issue.

In this case, the number of NVMe PCI's MSI-X table entries is 64, so __pci_enable_msix_range() return -ENOSPC when we ask for 65.

However, the following __pci_enable_msi_range() returns -EINVAL because the NVMe PCI isn't capable of MSI, then this error is returned from pci_alloc_irq_vectors_affinity() finally to NVMe driver.

Of course, -EINVAL makes a difference because the current code only tries to assign one irq vector in this case, and it shouldn't be returned from pci_alloc_irq_vectors_affinity(), given there is enough msix entries for fallback, right?

Thanks,
Ming

I'm not sure how it matters either, since pci_alloc_irq_vectors_affinity() will fail either way. It *would* be nice to return the correct error in case the caller uses it to emit a message.

But if the caller wants to use -ENOSPC to reduce @min_vecs and try again, that sounds like an incorrect use of the interface -- the caller should have just used the smaller @min_vecs the first time around.

I don't quite understand this last sentence. If @min_vecs == 5 and the device only supports 4 MSI-X and 4 MSI vectors, the function comment says we should fail with -ENOSPC.

Bjorn