From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from us-smtp-delivery-124.mimecast.com ([63.128.21.124]:35737 "EHLO us-smtp-delivery-124.mimecast.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1728756AbhAUJeX (ORCPT ); Thu, 21 Jan 2021 04:34:23 -0500 Subject: Re: [kvm-unit-tests PATCH v4 2/3] s390x: define UV compatible I/O allocation References: <1611220392-22628-1-git-send-email-pmorel@linux.ibm.com> <1611220392-22628-3-git-send-email-pmorel@linux.ibm.com> From: Thomas Huth Message-ID: Date: Thu, 21 Jan 2021 10:32:29 +0100 MIME-Version: 1.0 In-Reply-To: <1611220392-22628-3-git-send-email-pmorel@linux.ibm.com> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-US Content-Transfer-Encoding: 7bit List-ID: To: Pierre Morel , kvm@vger.kernel.org Cc: linux-s390@vger.kernel.org, frankja@linux.ibm.com, david@redhat.com, cohuck@redhat.com, imbrenda@linux.ibm.com, drjones@redhat.com, pbonzini@redhat.com On 21/01/2021 10.13, Pierre Morel wrote: > To centralize the memory allocation for I/O we define > the alloc_io_page/free_io_page functions which share the I/O > memory with the host in case the guest runs with > protected virtualization. > > Signed-off-by: Pierre Morel > --- > MAINTAINERS | 1 + > lib/s390x/malloc_io.c | 70 +++++++++++++++++++++++++++++++++++++++++++ > lib/s390x/malloc_io.h | 45 ++++++++++++++++++++++++++++ > s390x/Makefile | 1 + > 4 files changed, 117 insertions(+) > create mode 100644 lib/s390x/malloc_io.c > create mode 100644 lib/s390x/malloc_io.h > > diff --git a/MAINTAINERS b/MAINTAINERS > index 54124f6..89cb01e 100644 > --- a/MAINTAINERS > +++ b/MAINTAINERS > @@ -82,6 +82,7 @@ M: Thomas Huth > M: David Hildenbrand > M: Janosch Frank > R: Cornelia Huck > +R: Pierre Morel > L: kvm@vger.kernel.org > L: linux-s390@vger.kernel.org > F: s390x/* > diff --git a/lib/s390x/malloc_io.c b/lib/s390x/malloc_io.c > new file mode 100644 > index 0000000..bfe8c6a > --- /dev/null > +++ b/lib/s390x/malloc_io.c > @@ -0,0 +1,70 @@ > +// SPDX-License-Identifier: GPL-2.0 "GPL-2.0" is deprecated according to https://spdx.org/licenses/ ... please use "GPL-2.0-only" instead. > +/* > + * I/O page allocation > + * > + * Copyright (c) 2021 IBM Corp > + * > + * Authors: > + * Pierre Morel > + * > + * Using this interface provide host access to the allocated pages in > + * case the guest is a secure guest. > + * This is needed for I/O buffers. > + * > + */ > +#include > +#include > +#include > +#include > +#include > +#include > + > +static int share_pages(void *p, int count) > +{ > + int i = 0; > + > + for (i = 0; i < count; i++, p += PAGE_SIZE) > + if (uv_set_shared((unsigned long)p)) > + return i; Just a matter of taste, but you could replace the "return i" here also with a "break" since you're returning i below anyway. > + return i; > +} > + > +static void unshare_pages(void *p, int count) > +{ > + int i; > + > + for (i = count; i > 0; i--, p += PAGE_SIZE) > + uv_remove_shared((unsigned long)p); > +} > + > +void *alloc_io_pages(int size, int flags) I still think the naming or size parameter is confusing here. If I read something like alloc_io_pages(), I'd expect a "num_pages" parameter. So if you want to keep the "size" in bytes, I'd suggest to rename the function to "alloc_io_mem" instead. > +{ > + int order = (size >> PAGE_SHIFT); I think this is wrong. According to the description of alloc_pages_flag, it allocates "1ull << order" pages. So you likely want to do this instead here: int order = get_order(size >> PAGE_SHIFT); > + void *p; > + int n; > + > + assert(size); > + > + p = alloc_pages_flags(order, AREA_DMA31 | flags); > + if (!p || !test_facility(158)) > + return p; > + > + n = share_pages(p, 1 << order); > + if (n == 1 << order) > + return p; > + > + unshare_pages(p, n); > + free_pages(p); > + return NULL; > +} > + > +void free_io_pages(void *p, int size) > +{ > + int order = size >> PAGE_SHIFT; dito? > + assert(IS_ALIGNED((uintptr_t)p, PAGE_SIZE)); > + > + if (test_facility(158)) > + unshare_pages(p, 1 << order); > + free_pages(p); > +} > diff --git a/lib/s390x/malloc_io.h b/lib/s390x/malloc_io.h > new file mode 100644 > index 0000000..494dfe9 > --- /dev/null > +++ b/lib/s390x/malloc_io.h > @@ -0,0 +1,45 @@ > +/* SPDX-License-Identifier: GPL-2.0 */ GPL-2.0-only please. > +/* > + * I/O allocations > + * > + * Copyright (c) 2021 IBM Corp > + * > + * Authors: > + * Pierre Morel > + * > + */ > +#ifndef _S390X_MALLOC_IO_H_ > +#define _S390X_MALLOC_IO_H_ > + > +/* > + * Allocates a page aligned page bound range of contiguous real or > + * absolute memory in the DMA31 region large enough to contain size > + * bytes. > + * If Protected Virtualization facility is present, shares the pages > + * with the host. > + * If all the pages for the specified size cannot be reserved, > + * the function rewinds the partial allocation and a NULL pointer > + * is returned. > + * > + * @size: the minimal size allocated in byte. > + * @flags: the flags used for the underlying page allocator. > + * > + * Errors: > + * The allocation will assert the size parameter, will fail if the > + * underlying page allocator fail or in the case of protected > + * virtualisation if the sharing of the pages fails. I think "virtualization" (with an z) is more common than "virtualisation". > + * > + * Returns a pointer to the first page in case of success, NULL otherwise. > + */ > +void *alloc_io_pages(int size, int flags); > + > +/* > + * Frees a previously memory space allocated by alloc_io_pages. > + * If Protected Virtualization facility is present, unshares the pages > + * with the host. > + * The address must be aligned on a page boundary otherwise an assertion > + * breaks the program. > + */ > +void free_io_pages(void *p, int size); > + > +#endif /* _S390X_MALLOC_IO_H_ */ > diff --git a/s390x/Makefile b/s390x/Makefile > index b079a26..4b6301c 100644 > --- a/s390x/Makefile > +++ b/s390x/Makefile > @@ -63,6 +63,7 @@ cflatobjs += lib/s390x/smp.o > cflatobjs += lib/s390x/vm.o > cflatobjs += lib/s390x/css_dump.o > cflatobjs += lib/s390x/css_lib.o > +cflatobjs += lib/s390x/malloc_io.o > > OBJDIRS += lib/s390x Thomas