From: Michael Kelley <mhklinux@xxxxxxxxxxx>
Code specific to Hyper-V guests currently assumes the cpu_possible_mask
is "dense" -- i.e., all bit positions 0 thru (nr_cpu_ids - 1) are set,
with no "holes". Therefore, num_possible_cpus() is assumed to be equal
to nr_cpu_ids.
Per a separate discussion[1], this assumption is not valid in the
general case. For example, the function setup_nr_cpu_ids() in
kernel/smp.c is coded to assume cpu_possible_mask may be sparse,
and other patches have been made in the past to correctly handle
the sparseness. See bc75e99983df1efd ("rcu: Correctly handle sparse
possible cpu") as noted by Mark Rutland.
The general case notwithstanding, the configurations that Hyper-V
provides to guest VMs on x86 and ARM64 hardware, in combination
with the algorithms currently used by architecture specific code
to assign Linux CPU numbers, *does* always produce a dense
cpu_possible_mask. So the invalid assumption is not currently
causing failures. But in the interest of correctness, and robustness
against future changes in the code that populates cpu_possible_mask,
update the Hyper-V code to no longer assume denseness.
The typical code pattern with the invalid assumption is as follows:
array = kcalloc(num_possible_cpus(), sizeof(<some struct>),
GFP_KERNEL);
....
index into "array" with smp_processor_id()
In such as case, the array might be indexed by a value beyond the size
of the array. The correct approach is to allocate the array with size
"nr_cpu_ids". While this will probably leave unused any array entries
corresponding to holes in cpu_possible_mask, the holes are assumed to
be minimal and hence the amount of memory wasted by unused entries is
minimal.
Removing the assumption in Hyper-V code is done in several patches
because they touch different kernel subsystems:
Patch 1: Hyper-V x86 initialization of hv_vp_assist_page (there's no
hv_vp_assist_page on ARM64)
Patch 2: Hyper-V common init of hv_vp_index
Patch 3: Hyper-V IOMMU driver
Patch 4: storvsc driver
Patch 5: netvsc driver
I tested the changes by hacking the construction of cpu_possible_mask
to include a hole on x86. With a configuration set to demonstrate the
problem, a Hyper-V guest kernel eventually crashes due to memory
corruption. After the patches in this series, the crash does not occur.
[1] https://lore.kernel.org/lkml/SN6PR02MB4157210CC36B2593F8572E5ED4692@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
Michael Kelley (5):
x86/hyperv: Don't assume cpu_possible_mask is dense
Drivers: hv: Don't assume cpu_possible_mask is dense
iommu/hyper-v: Don't assume cpu_possible_mask is dense
scsi: storvsc: Don't assume cpu_possible_mask is dense
hv_netvsc: Don't assume cpu_possible_mask is dense
arch/x86/hyperv/hv_init.c | 2 +-
drivers/hv/hv_common.c | 4 ++--
drivers/iommu/hyperv-iommu.c | 4 ++--
drivers/net/hyperv/netvsc_drv.c | 2 +-
drivers/scsi/storvsc_drv.c | 13 ++++++-------
5 files changed, 12 insertions(+), 13 deletions(-)
--
2.25.1
