On Tue, Jan 26, 2021 at 12:41:05AM +0000, Michael Kelley wrote: > From: Wei Liu <wei.liu@xxxxxxxxxx> Sent: Wednesday, January 20, 2021 4:01 AM > > > > When Linux runs as the root partition, it will need to make hypercalls > > which return data from the hypervisor. > > > > Allocate pages for storing results when Linux runs as the root > > partition. > > > > Signed-off-by: Lillian Grassin-Drake <ligrassi@xxxxxxxxxxxxx> > > Co-Developed-by: Lillian Grassin-Drake <ligrassi@xxxxxxxxxxxxx> > > Signed-off-by: Wei Liu <wei.liu@xxxxxxxxxx> > > --- > > v3: Fix hv_cpu_die to use free_pages. > > v2: Address Vitaly's comments > > --- > > arch/x86/hyperv/hv_init.c | 35 ++++++++++++++++++++++++++++----- > > arch/x86/include/asm/mshyperv.h | 1 + > > 2 files changed, 31 insertions(+), 5 deletions(-) > > > > diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c > > index e04d90af4c27..6f4cb40e53fe 100644 > > --- a/arch/x86/hyperv/hv_init.c > > +++ b/arch/x86/hyperv/hv_init.c > > @@ -41,6 +41,9 @@ EXPORT_SYMBOL_GPL(hv_vp_assist_page); > > void __percpu **hyperv_pcpu_input_arg; > > EXPORT_SYMBOL_GPL(hyperv_pcpu_input_arg); > > > > +void __percpu **hyperv_pcpu_output_arg; > > +EXPORT_SYMBOL_GPL(hyperv_pcpu_output_arg); > > + > > u32 hv_max_vp_index; > > EXPORT_SYMBOL_GPL(hv_max_vp_index); > > > > @@ -73,12 +76,19 @@ static int hv_cpu_init(unsigned int cpu) > > void **input_arg; > > struct page *pg; > > > > - input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg); > > /* hv_cpu_init() can be called with IRQs disabled from hv_resume() */ > > - pg = alloc_page(irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL); > > + pg = alloc_pages(irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL, hv_root_partition ? > > 1 : 0); > > if (unlikely(!pg)) > > return -ENOMEM; > > + > > + input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg); > > *input_arg = page_address(pg); > > + if (hv_root_partition) { > > + void **output_arg; > > + > > + output_arg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg); > > + *output_arg = page_address(pg + 1); > > + } > > > > hv_get_vp_index(msr_vp_index); > > > > @@ -205,14 +215,23 @@ static int hv_cpu_die(unsigned int cpu) > > unsigned int new_cpu; > > unsigned long flags; > > void **input_arg; > > - void *input_pg = NULL; > > + void *pg; > > > > local_irq_save(flags); > > input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg); > > - input_pg = *input_arg; > > + pg = *input_arg; > > *input_arg = NULL; > > + > > + if (hv_root_partition) { > > + void **output_arg; > > + > > + output_arg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg); > > + *output_arg = NULL; > > + } > > + > > local_irq_restore(flags); > > - free_page((unsigned long)input_pg); > > + > > + free_pages((unsigned long)pg, hv_root_partition ? 1 : 0); > > > > if (hv_vp_assist_page && hv_vp_assist_page[cpu]) > > wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, 0); > > @@ -346,6 +365,12 @@ void __init hyperv_init(void) > > > > BUG_ON(hyperv_pcpu_input_arg == NULL); > > > > + /* Allocate the per-CPU state for output arg for root */ > > + if (hv_root_partition) { > > + hyperv_pcpu_output_arg = alloc_percpu(void *); > > + BUG_ON(hyperv_pcpu_output_arg == NULL); > > + } > > + > > /* Allocate percpu VP index */ > > hv_vp_index = kmalloc_array(num_possible_cpus(), sizeof(*hv_vp_index), > > GFP_KERNEL); > > diff --git a/arch/x86/include/asm/mshyperv.h b/arch/x86/include/asm/mshyperv.h > > index ac2b0d110f03..62d9390f1ddf 100644 > > --- a/arch/x86/include/asm/mshyperv.h > > +++ b/arch/x86/include/asm/mshyperv.h > > @@ -76,6 +76,7 @@ static inline void hv_disable_stimer0_percpu_irq(int irq) {} > > #if IS_ENABLED(CONFIG_HYPERV) > > extern void *hv_hypercall_pg; > > extern void __percpu **hyperv_pcpu_input_arg; > > +extern void __percpu **hyperv_pcpu_output_arg; > > > > static inline u64 hv_do_hypercall(u64 control, void *input, void *output) > > { > > -- > > 2.20.1 > > I think this all works OK. But a meta question: Do we need a separate > per-cpu output argument page? From the Hyper-V hypercall standpoint, I > don't think input and output args need to be in separate pages. They both That's correct. They don't have to be in separate pages. > just need to not cross a page boundary. As long as we don't have a hypercall > where the sum of the sizes of the input and output args exceeds a page, > we could just have a single page, and split it up in any manner that works > for the particular hypercall. > There is one more requirement: The pointers must be 8-byte aligned. That means we may need to explicitly pad things a bit. That quickly becomes tedious if we do it in every call site; or we will need to provide a macro to do the calculation correctly. Another consideration is hypercalls that take variable-length input / output. Admittedly I haven't seen one that takes variable-length arguments and needs to do input and output at the same time, I wouldn't want to paint ourselves into the corner now because sizing variable-length input and output at the same time can be non-trivial. Wei. > Thoughts? > > Michael > >