Commit ef2b6d5f authored Jan 22, 2025 by Jim Harris

env_dpdk: don't touch dpdk device handle after DPDK removal



If we get back-to-back removals, the first removal could be pending
removal in the SPDK device list. When callback happens for the second
removal, we can't touch the rte dev_handle for that first device,
because DPDK has already freed it. So use the removed flag
on our spdk_pci_device structure to skip it and avoid dereferencing
freed memory.

While here, remove debug prints that were added specifically to
triage this issue.

Fixes issue #3599.

Signed-off-by: Jim Harris <jim.harris@nvidia.com>
Change-Id: I43f81145470f190a51c235ab512c1b9f419ba29e
Reviewed-on: https://review.spdk.io/c/spdk/spdk/+/25586


Reviewed-by: Changpeng Liu <changpeliu@tencent.com>
Tested-by: SPDK Automated Test System <spdkbot@gmail.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Community-CI: Mellanox Build Bot

parent 9f051b98

lib/env_dpdk/pci.c

+8 −6

Original line number	Diff line number	Diff line
		@@ -226,15 +226,17 @@ pci_device_rte_dev_event(const char *device_name,
		TAILQ_FOREACH(dev, &g_pci_devices, internal.tailq) {
		struct rte_pci_device *rte_dev = dev->dev_handle;

		if (strcmp(dpdk_pci_device_get_name(rte_dev), device_name)) {
		if (dev->internal.removed) {
		/* DPDK already removed this device, we are still pending
		* removal of the device from the SPDK device list. Since
		* DPDK freed the device handle, we must not try to
		* get its device name.
		*/
		continue;
		}

		/* Note: these ERRLOGs are useful for triaging issue #2983. */
		if (dev->internal.pending_removal \|\| dev->internal.removed) {
		SPDK_ERRLOG("Received event for device SPDK already tried to remove\n");
		SPDK_ERRLOG("pending_removal=%d removed=%d\n", dev->internal.pending_removal,
		dev->internal.removed);
		if (strcmp(dpdk_pci_device_get_name(rte_dev), device_name)) {
		continue;
		}

		if (!dev->internal.pending_removal) {