| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-Id: <20190812144144.2646-1-mika.westerberg@linux.intel.com>
Date: Mon, 12 Aug 2019 17:41:44 +0300
From: Mika Westerberg <mika.westerberg@...ux.intel.com>
To: Bjorn Helgaas <bhelgaas@...gle.com>
Cc: Benjamin Herrenschmidt <benh@...nel.crashing.org>,
Nicholas Johnson <nicholas.johnson-opensource@...look.com.au>,
Mika Westerberg <mika.westerberg@...ux.intel.com>,
linux-pci@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: [PATCH] PCI: Use minimum window alignment when calculating memory window size
There is an issue in Linux PCI resource allocation that if we remove an
existing device that was initially configured by the BIOS and then issue
rescan, it will not fit in to the memory space allocated by the BIOS
even if it originally it fit there just fine.
The system in question is just a regular PC with a FPGA card connected
to PCIe slot. The initial BIOS resource allocation right after boot
looks like:
00:01.1 Root port
Memory behind bridge: d0000000-f01fffff (514M)
02:00.0 PCI-Express to PCI/PCI-X Bridge
BAR 0: Memory at d0000000-dfffffff (256M)
BAR 1: Memory at f0100000-f01fffff (1M)
Memory behind bridge: e0000000-f00fffff (257M)
03:0a.0 FPGA
BAR 0: Memory at f0000000-f000ffff (64k)
BAR 2: Memory at e0000000-efffffff (256M)
The FPGA card consists of a PCIe-to-PCI-X bridge (02:00.0) and the FPGA
device itself (03:0a.0) right below the bridge. In order to update the
FPGA image to the card without need for rebooting the system we remove
the device through sysfs:
# echo 1 > /sys/bus/pci/devices/0000:00:01.1/0000:02:00.0/remove
At this point the same image is burned to the FPGA or alternatively it
is just reset to make sure the config space gets cleared and the kernel
needs to do the resource allocation in the next step. Next we issue
rescan to find and re-configure the same device:
# echo 1 > /sys/bus/pci/devices/0000:00:01.1/rescan
But the end result is not the same and in fact the FPGA device cannot
enable its BARs because there is no memory window from bridge 02:00.0 to
the FPGA:
00:01.1 Root port
Memory behind bridge: d0000000-f01fffff (514M)
02:00.0 PCI-Express to PCI/PCI-X Bridge
BAR 0: Memory at d0000000-dfffffff (256M)
BAR 1: Memory at e0000000-e00fffff (1M)
Below are the messages from the rescan with the failure highlighted:
[ 208.396066] pci_bus 0000:02: scanning bus
[ 208.396090] pci 0000:02:00.0: [8086:0bcd] type 01 class 0x060400
[ 208.396119] pci 0000:02:00.0: reg 0x10: [mem 0x00000000-0x0fffffff]
[ 208.396128] pci 0000:02:00.0: reg 0x14: [mem 0x00000000-0x000fffff]
[ 208.396156] pci 0000:02:00.0: enabling Extended Tags
[ 208.396236] pci 0000:02:00.0: PME# supported from D0
[ 208.396241] pci 0000:02:00.0: PME# disabled
[ 208.407979] pci 0000:02:00.0: scanning [bus 00-00] behind bridge, pass 0
[ 208.407984] pci 0000:02:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring
[ 208.407994] pci 0000:02:00.0: scanning [bus 00-00] behind bridge, pass 1
[ 208.408039] pci_bus 0000:03: extended config space not accessible
[ 208.408119] pci_bus 0000:03: scanning bus
[ 208.456396] pci 0000:03:0a.0: [8086:0bce] type 00 class 0x000000
[ 208.456528] pci 0000:03:0a.0: reg 0x10: [mem 0xf0000000-0xf000ffff 64bit]
[ 208.456611] pci 0000:03:0a.0: reg 0x18: [mem 0xe0000000-0xefffffff 64bit]
[ 208.456664] pci 0000:03:0a.0: reg 0x20: [io 0xe000-0xe0ff]
[ 208.938323] pci_bus 0000:03: fixups for bus
[ 208.938324] pci 0000:02:00.0: PCI bridge to [bus 03]
[ 208.938330] pci 0000:02:00.0: bridge window [io 0x0000-0x0fff]
[ 208.938334] pci 0000:02:00.0: bridge window [mem 0x00000000-0x000fffff]
[ 208.938339] pci 0000:02:00.0: bridge window [mem 0x00000000-0x000fffff 64bit pref]
[ 208.938341] pci_bus 0000:03: bus scan returning with max=03
[ 208.938343] pci_bus 0000:03: busn_res: [bus 03] end is updated to 03
[ 208.938346] pci_bus 0000:02: bus scan returning with max=03
[ 208.938555] pci 0000:02:00.0: BAR 0: assigned [mem 0xd0000000-0xdfffffff]
[ 208.938558] pci 0000:02:00.0: BAR 14: no space for [mem size 0x18000000]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[ 208.938559] pci 0000:02:00.0: BAR 14: failed to assign [mem size 0x18000000]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[ 208.938560] pci 0000:02:00.0: BAR 1: assigned [mem 0xe0000000-0xe00fffff]
[ 208.938564] pci 0000:02:00.0: BAR 15: assigned [mem 0xbfb00000-0xbfbfffff 64bit pref]
[ 208.938565] pci 0000:02:00.0: BAR 13: assigned [io 0xe000-0xefff]
[ 208.938566] pci 0000:03:0a.0: BAR 2: no space for [mem size 0x10000000 64bit]
[ 208.938567] pci 0000:03:0a.0: BAR 2: failed to assign [mem size 0x10000000 64bit]
[ 208.938568] pci 0000:03:0a.0: BAR 0: no space for [mem size 0x00010000 64bit]
[ 208.938569] pci 0000:03:0a.0: BAR 0: failed to assign [mem size 0x00010000 64bit]
[ 208.938569] pci 0000:03:0a.0: BAR 4: assigned [io 0xe000-0xe0ff]
It seems that Linux tries to open memory window of 0x18000000 (384M)
from the bridge 02:00.0 towards the FPGA 03:0a.0. This of does not fit
to the memory available for the bridge itself and thus the assignment
fails.
The size 0x18000000 comes from the following code which tries to
calculate size of the devices below the bridge and the required
alignment:
drivers/pci/setup-bus.c::pbus_size_mem(...)
{
// size = 0x10000000 + 0x10000 = 0x10010000 (256M + 64k)
min_align = calculate_mem_align(aligns, max_order);
min_align = max(min_align, window_alignment(bus, b_res->flags));
// min_align = 0x8000000 (128M)
size0 = calculate_memsize(size, min_size, 0, 0, resource_size(b_res), min_align);
// size0 = ALIGN(0x10010000, 0x8000000) = 0x18000000 (384M)
We align the calculated size to the next 128M which increases to 384M
and then later in pci_assign_resource() try to assing the resource which
fails because there is no room available in the parent bridge.
Since the minimum alignment (min_align) is kept as part of the window
resource and taken into account later in pbus_assign_resources_sorted()
and pci_assign_resource(), I think aligning size to min_align is not
necessary.
For this reason make the size calculation to use minimum memory window
alignment (1M) instead. The resulting resource allocation matches the
initial allocation done by the BIOS.
Signed-off-by: Mika Westerberg <mika.westerberg@...ux.intel.com>
---
The previous RFC version can be found here:
https://patchwork.kernel.org/patch/10917249/
There are no changes to that version except that I changed "RFC" to "PATCH"
in the subject and included wider audience in Cc list.
drivers/pci/setup-bus.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/pci/setup-bus.c b/drivers/pci/setup-bus.c
index 79b1fa6519be..c8c947f17675 100644
--- a/drivers/pci/setup-bus.c
+++ b/drivers/pci/setup-bus.c
@@ -1049,9 +1049,10 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
}
}
+ size0 = calculate_memsize(size, min_size, 0, 0, resource_size(b_res),
+ window_alignment(bus, b_res->flags));
min_align = calculate_mem_align(aligns, max_order);
min_align = max(min_align, window_alignment(bus, b_res->flags));
- size0 = calculate_memsize(size, min_size, 0, 0, resource_size(b_res), min_align);
add_align = max(min_align, add_align);
size1 = (!realloc_head || (realloc_head && !add_size && !children_add_size)) ? size0 :
calculate_memsize(size, min_size, add_size, children_add_size,
--
2.20.1
Powered by blists - more mailing lists