FreeBSD source tree https://cgit-dev.freebsd.org/src/atom?h=releng%2F14.3 2025-05-15T22:22:44Z 2025-05-15T22:22:44Z Colin Percival cperciva@FreeBSD.org 2025-05-14T23:36:26Z urn:sha1:29d6968fd36d27b27f953d5c00f9763491b10734 Early versions of Xen, including those used in the early days of EC2, had a bug in their UART emulation whereby the TX FIFO wouldn't send the expected interrupt when emptying; as a result, FreeBSD would write 16 characters to the serial console and then stop because we thought the FIFO was forever full. In 2013 (1c60b24baa50) I added a loader tunable "hw.broken_txfifo" which spinwaits for the FIFO TX rather than relying on the interrupt, and enabled this in loader.conf in EC2 images. A decade later, this workaround is almost certainly no longer needed in EC2 -- most instances don't run Xen, and the bug was long since fixed in Xen anyway -- but we've been holding on to the workaround "just in case". Unfortunately, the spinwait behaviour is causing latency spikes and triggering warnings from the ena(4) driver. This commit removes the hw.broken_txfifo setting from loader.conf in EC2 images, but leaves the loader tunable and associated code, since it has been necessary in some other environments. (It seems that the TX FIFO missing-interrupts bug has been independently written at least three times!) Approved by: re (cperciva) MFC after: 1 minute Sponsored by: Amazon (cherry picked from commit 9a685c09f06a55b18589d75f9307563d84a17fa9) (cherry picked from commit eadda156a50d3487ec1e6fc78f6cfe2df42448fa) 2025-04-30T07:55:41Z Dave Cottlehuber dch@FreeBSD.org 2025-04-27T23:46:47Z urn:sha1:f5418217dbc6c3f9317479183b06fff6be147ab3 The runtime name is taken from the main pkg-base package that this image is built off. Sponsored by: SkunkWerks, GmbH MFC after: 3 days Reviewed by: dfr, emaste Differential Revision: https://reviews.freebsd.org/D50043 (cherry picked from commit a0165254bfeb5e310f92d4e0c88fcb5c6ea802bf) 2025-04-26T19:21:22Z Colin Percival cperciva@FreeBSD.org 2025-04-23T15:17:51Z urn:sha1:66d36f4f1025193a64cf296873ed4a4b34471d91 Having this enabled on Graviton systems prior to Graviton 4 results in a resource leak and a kernel panic after repeated hotplug/unplug. MFC after: 3 days Sponsored by: Amazon (cherry picked from commit ce9a34b1614e37dc3f8763586448063408c7bf16) 2025-04-26T19:17:51Z Colin Percival cperciva@FreeBSD.org 2025-04-20T16:38:48Z urn:sha1:925ec43c152566b28cf00f8ec0d11715263a6c3a Starting in 2015 I have published "AMI Builder AMIs" for FreeBSD/EC2: These boot into a memory disk, extract a "clean" copy of FreeBSD onto the root disk, mount it at /mnt, and allow the user to SSH in to make customizations before creating a new AMI from the "running" instance (in fact, from the FreeBSD installation which is not running but is mounted on /mnt). This provides a much cleaner mechanism for building customized FreeBSD AMIs than the traditional Linux approach of "launch an EC2 instance, SSH in and configure it, then try to wipe logs and credentials before creating an AMI"; and it's easier than building a customized AMI ab initio by modifying the FreeBSD release-building code. This commit brings that functionality into the FreeBSD src tree and into the collection of images built by the release engineering team: The EC2 "BUILDER" flavour AMI is essentially a "SMALL" flavour AMI with a compressed "BASE" flavour disk image, plus an init script which juggles disks around (rerooting into a memory disk and extracting the "BASE" image onto disk). Polished by: bz, emaste MFC after: 1 week Sponsored by: Amazon Differential Revision: https://reviews.freebsd.org/D49930 (cherry picked from commit 58426589030308cd632477d328b9536b1634c54d) 2025-04-16T03:18:03Z Colin Percival cperciva@FreeBSD.org 2025-03-31T04:11:23Z urn:sha1:f852f4f7b6f1666b09b16dba4aaab7acdb49980d Since PCIe device detaching is done via API, there is no opportunity to "press the attention button a second time" and thus the 5 second timeout mandated by PCIe serves no purpose. MFC after: 2 weeks Sponsored by: Amazon (cherry picked from commit 80febaa1714e897d67a8217edc305a0e81498a6f) 2025-04-14T16:09:49Z Colin Percival cperciva@FreeBSD.org 2025-03-06T05:22:33Z urn:sha1:43933d89a3256f33d0eb29473ac56b3e0e7b9318 On some EC2 instances, there is a race between removing a device from the system and making the PCI bus stop reporting the presence of the device. As a result, a PCI BUS_RESCAN performed immediately after the _EJ0 method returns "sees" the device which is being ejected, which then causes problems later (e.g. we won't recognize a new device being plugged into that slot because we never knew it was vacant). On other operating systems the bus is synchronously marked as needing to be rescanned but the rescan does not occur until O(1) seconds later. Create a new ACPI_Q_DELAY_BEFORE_EJECT_RESCAN quirk and set it in EC2 AMIs, and add a 10 ms DELAY between _EJ0 and BUS_RESCAN when tht quirk is set. Reviewed by: jhb MFC after: 1 month Sponsored by: Amazon Differential Revision: https://reviews.freebsd.org/D49252 (cherry picked from commit 55c3348ed78fb1d0891e8bb51a8948f95da3560b) 2025-04-08T18:58:08Z Dave Cottlehuber dch@FreeBSD.org 2025-02-28T22:27:31Z urn:sha1:7eab72fae632115fa93f5b39f1f39453e0630af9 pkg behaviour selects the non-default flavour, this trims 0,5GiB per cloudware image, by being explicit. Approved by: cperciva Differential Revision: https://reviews.freebsd.org/D48599 Sponsored by: SkunkWerks, GmbH PR: 284278 MFC after: 10 days (cherry picked from commit 457d745d90a6b63793c9ca17bcd7ec5ff4cd8b93) 2025-04-01T17:43:58Z Colin Percival cperciva@FreeBSD.org 2025-02-27T00:31:08Z urn:sha1:a64357f31b7c798452e7aca4e6ec34c2842a4fc1 In order to signal to Graviton [123] systems that a device is ready to be "ejected" (after a detach request is made via the EC2 API) we need to set PCIM_PSTAT_PME to 1 and PCIM_PSTAT_PMEENABLE to 0. We are not aware of any rationale for this requirement beyond "another OS kernel happens to do this", i.e. this is effectively bug-for-bug compatibility. Arguably this should be done by the ACPI _EJ0 method on these systems, but it is not. Create a new ACPI_Q_CLEAR_PME_ON_DETACH quirk and set it in EC2 AMIs, and add the PCI register write to acpi_pci_device_notify_handler when that quirk is set. Reviewed by: jhb MFC after: 1 month Sponsored by: Amazon Differential Revision: https://reviews.freebsd.org/D49146 (cherry picked from commit d70bac252d30adec4feba0c866dabe2c16a756d9) 2025-03-04T09:46:03Z Doug Rabson dfr@FreeBSD.org 2024-12-12T16:42:00Z urn:sha1:f60149306ccf784ddaa517132a8be47bb523984d This avoids the need for buildah and skopeo for building releases. Reviewed by: cpersiva MFC after: 1 day Differential Revision: https://reviews.freebsd.org/D48574 (cherry picked from commit e8a5b9fd73f4f437a03c85e7644daa55652e224b) 2025-02-09T21:14:34Z Colin Percival cperciva@FreeBSD.org 2025-02-04T00:10:47Z urn:sha1:eebfb83805c0f9f11ded223e5742a58ab6a5a1d6 pkg used to store copies of upstream repository databases in /var/db/pkg/repo-*.sqlite. About a year ago this was moved to /var/db/pkg/repos/*/, resulting in FreeBSD cloud images no longer having those (unhelpful since they'll be long out of date before the cloud images are launched) databases removed. Remove the correct location, and hope that future pkg updates don't break the base system again. Sponsored by: Amazon (cherry picked from commit 078e8b34b13d6d0663661542eeac9007806fccdc)