d/freebsd-nq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

xen/blkback: remove bounce buffering mode

Browse Source 
Remove bounce buffering code for blkback and only attach if Xen
creates IOMMU entries for grant mapped pages.

Such bounce buffering consumed a non trivial amount of memory and CPU
resources to do the memory copy, when it's been a long time since Xen
has been creating IOMMU entries for grant maps.

Refuse to attach blkback if Xen doesn't advertise that IOMMU entries
are created for grant maps.

Sponsored by: Citrix Systems R&D
This commit is contained in:
Roger Pau Monné 2022-04-12 16:17:09 +02:00
parent 137381ca60
commit 1d528f95e8
1 changed files with 22 additions and 178 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

200
sys/dev/xen/blkback/blkback.c
View File

@ -80,6 +80,7 @@ __FBSDID("$FreeBSD$");
#include <xen/gnttab.h>
#include <xen/xen_intr.h>
#include <contrib/xen/arch-x86/cpuid.h>
#include <contrib/xen/event_channel.h>
#include <contrib/xen/grant_table.h>
@ -101,27 +102,6 @@ __FBSDID("$FreeBSD$");
#define XBB_MAX_REQUESTS \
__CONST_RING_SIZE(blkif, PAGE_SIZE * XBB_MAX_RING_PAGES)
/**
* \brief Define to force all I/O to be performed on memory owned by the
* backend device, with a copy-in/out to the remote domain's memory.
*
* \note This option is currently required when this driver's domain is
* operating in HVM mode on a system using an IOMMU.
*
* This driver uses Xen's grant table API to gain access to the memory of
* the remote domains it serves. When our domain is operating in PV mode,
* the grant table mechanism directly updates our domain's page table entries
* to point to the physical pages of the remote domain. This scheme guarantees
* that blkback and the backing devices it uses can safely perform DMA
* operations to satisfy requests. In HVM mode, Xen may use a HW IOMMU to
* insure that our domain cannot DMA to pages owned by another domain. As
* of Xen 4.0, IOMMU mappings for HVM guests are not updated via the grant
* table API. For this reason, in HVM mode, we must bounce all requests into
* memory that is mapped into our domain at domain startup and thus has
* valid IOMMU mappings.
*/
#define XBB_USE_BOUNCE_BUFFERS
/**
* \brief Define to enable rudimentary request logging to the console.
*/
@ -257,14 +237,6 @@ struct xbb_xen_reqlist {
*/
uint64_t gnt_base;
#ifdef XBB_USE_BOUNCE_BUFFERS
/**
* Pre-allocated domain local memory used to proxy remote
* domain memory during I/O operations.
*/
uint8_t *bounce;
#endif
/**
* Array of grant handles (one per page) used to map this request.
*/
@ -500,30 +472,6 @@ struct xbb_file_data {
* so we only need one of these.
*/
struct iovec xiovecs[XBB_MAX_SEGMENTS_PER_REQLIST];
#ifdef XBB_USE_BOUNCE_BUFFERS
/**
* \brief Array of io vectors used to handle bouncing of file reads.
*
* Vnode operations are free to modify uio data during their
* exectuion. In the case of a read with bounce buffering active,
* we need some of the data from the original uio in order to
* bounce-out the read data. This array serves as the temporary
* storage for this saved data.
*/
struct iovec saved_xiovecs[XBB_MAX_SEGMENTS_PER_REQLIST];
/**
* \brief Array of memoized bounce buffer kva offsets used
* in the file based backend.
*
* Due to the way that the mapping of the memory backing an
* I/O transaction is handled by Xen, a second pass through
* the request sg elements is unavoidable. We memoize the computed
* bounce address here to reduce the cost of the second walk.
*/
void *xiovecs_vaddr[XBB_MAX_SEGMENTS_PER_REQLIST];
#endif /* XBB_USE_BOUNCE_BUFFERS */
};
/**
@ -891,25 +839,6 @@ xbb_reqlist_vaddr(struct xbb_xen_reqlist *reqlist, int pagenr, int sector)
return (reqlist->kva + (PAGE_SIZE * pagenr) + (sector << 9));
}
#ifdef XBB_USE_BOUNCE_BUFFERS
/**
* Given a page index and 512b sector offset within that page,
* calculate an offset into a request's local bounce memory region.
*
* \param reqlist The request structure whose bounce region will be accessed.
* \param pagenr The page index used to compute the bounce offset.
* \param sector The 512b sector index used to compute the page relative
* bounce offset.
*
* \return The computed global bounce buffer address.
*/
static inline uint8_t *
xbb_reqlist_bounce_addr(struct xbb_xen_reqlist *reqlist, int pagenr, int sector)
{
return (reqlist->bounce + (PAGE_SIZE * pagenr) + (sector << 9));
}
#endif
/**
* Given a page number and 512b sector offset within that page,
* calculate an offset into the request's memory region that the
@ -929,11 +858,7 @@ xbb_reqlist_bounce_addr(struct xbb_xen_reqlist *reqlist, int pagenr, int sector)
static inline uint8_t *
xbb_reqlist_ioaddr(struct xbb_xen_reqlist *reqlist, int pagenr, int sector)
{
#ifdef XBB_USE_BOUNCE_BUFFERS
return (xbb_reqlist_bounce_addr(reqlist, pagenr, sector));
#else
return (xbb_reqlist_vaddr(reqlist, pagenr, sector));
#endif
}
/**
@ -1508,17 +1433,6 @@ xbb_bio_done(struct bio *bio)
}
}
#ifdef XBB_USE_BOUNCE_BUFFERS
if (bio->bio_cmd == BIO_READ) {
vm_offset_t kva_offset;
kva_offset = (vm_offset_t)bio->bio_data
- (vm_offset_t)reqlist->bounce;
memcpy((uint8_t *)reqlist->kva + kva_offset,
bio->bio_data, bio->bio_bcount);
}
#endif /* XBB_USE_BOUNCE_BUFFERS */
/*
* Decrement the pending count for the request list. When we're
* done with the requests, send status back for all of them.
@ -2180,17 +2094,6 @@ xbb_dispatch_dev(struct xbb_softc *xbb, struct xbb_xen_reqlist *reqlist,
for (bio_idx = 0; bio_idx < nbio; bio_idx++)
{
#ifdef XBB_USE_BOUNCE_BUFFERS
vm_offset_t kva_offset;
kva_offset = (vm_offset_t)bios[bio_idx]->bio_data
- (vm_offset_t)reqlist->bounce;
if (operation == BIO_WRITE) {
memcpy(bios[bio_idx]->bio_data,
(uint8_t *)reqlist->kva + kva_offset,
bios[bio_idx]->bio_bcount);
}
#endif
if (operation == BIO_READ) {
SDT_PROBE3(xbb, kernel, xbb_dispatch_dev, read,
device_get_unit(xbb->dev),
@ -2241,10 +2144,6 @@ xbb_dispatch_file(struct xbb_softc *xbb, struct xbb_xen_reqlist *reqlist,
struct uio xuio;
struct xbb_sg *xbb_sg;
struct iovec *xiovec;
#ifdef XBB_USE_BOUNCE_BUFFERS
void **p_vaddr;
int saved_uio_iovcnt;
#endif /* XBB_USE_BOUNCE_BUFFERS */
int error;
file_data = &xbb->backend.file;
@ -2300,18 +2199,6 @@ xbb_dispatch_file(struct xbb_softc *xbb, struct xbb_xen_reqlist *reqlist,
xiovec = &file_data->xiovecs[xuio.uio_iovcnt];
xiovec->iov_base = xbb_reqlist_ioaddr(reqlist,
seg_idx, xbb_sg->first_sect);
#ifdef XBB_USE_BOUNCE_BUFFERS
/*
* Store the address of the incoming
* buffer at this particular offset
* as well, so we can do the copy
* later without having to do more
* work to recalculate this address.
*/
p_vaddr = &file_data->xiovecs_vaddr[xuio.uio_iovcnt];
*p_vaddr = xbb_reqlist_vaddr(reqlist, seg_idx,
xbb_sg->first_sect);
#endif /* XBB_USE_BOUNCE_BUFFERS */
xiovec->iov_len = 0;
xuio.uio_iovcnt++;
}
@ -2331,28 +2218,6 @@ xbb_dispatch_file(struct xbb_softc *xbb, struct xbb_xen_reqlist *reqlist,
xuio.uio_td = curthread;
#ifdef XBB_USE_BOUNCE_BUFFERS
saved_uio_iovcnt = xuio.uio_iovcnt;
if (operation == BIO_WRITE) {
/* Copy the write data to the local buffer. */
for (seg_idx = 0, p_vaddr = file_data->xiovecs_vaddr,
xiovec = xuio.uio_iov; seg_idx < xuio.uio_iovcnt;
seg_idx++, xiovec++, p_vaddr++) {
memcpy(xiovec->iov_base, *p_vaddr, xiovec->iov_len);
}
} else {
/*
* We only need to save off the iovecs in the case of a
* read, because the copy for the read happens after the
* VOP_READ(). (The uio will get modified in that call
* sequence.)
*/
memcpy(file_data->saved_xiovecs, xuio.uio_iov,
xuio.uio_iovcnt * sizeof(xuio.uio_iov[0]));
}
#endif /* XBB_USE_BOUNCE_BUFFERS */
switch (operation) {
case BIO_READ:
@ -2429,25 +2294,6 @@ xbb_dispatch_file(struct xbb_softc *xbb, struct xbb_xen_reqlist *reqlist,
/* NOTREACHED */
}
#ifdef XBB_USE_BOUNCE_BUFFERS
/* We only need to copy here for read operations */
if (operation == BIO_READ) {
for (seg_idx = 0, p_vaddr = file_data->xiovecs_vaddr,
xiovec = file_data->saved_xiovecs;
seg_idx < saved_uio_iovcnt; seg_idx++,
xiovec++, p_vaddr++) {
/*
* Note that we have to use the copy of the
* io vector we made above. uiomove() modifies
* the uio and its referenced vector as uiomove
* performs the copy, so we can't rely on any
* state from the original uio.
*/
memcpy(*p_vaddr, xiovec->iov_base, xiovec->iov_len);
}
}
#endif /* XBB_USE_BOUNCE_BUFFERS */
bailout_send_response:
if (error != 0)
@ -2826,12 +2672,6 @@ xbb_disconnect(struct xbb_softc *xbb)
/* There is one request list for ever allocated request. */
for (i = 0, reqlist = xbb->request_lists;
i < xbb->max_requests; i++, reqlist++){
#ifdef XBB_USE_BOUNCE_BUFFERS
if (reqlist->bounce != NULL) {
free(reqlist->bounce, M_XENBLOCKBACK);
reqlist->bounce = NULL;
}
#endif
if (reqlist->gnt_handles != NULL) {
free(reqlist->gnt_handles, M_XENBLOCKBACK);
reqlist->gnt_handles = NULL;
@ -3210,17 +3050,6 @@ xbb_alloc_request_lists(struct xbb_softc *xbb)
reqlist->xbb = xbb;
#ifdef XBB_USE_BOUNCE_BUFFERS
reqlist->bounce = malloc(xbb->max_reqlist_size,
M_XENBLOCKBACK, M_NOWAIT);
if (reqlist->bounce == NULL) {
xenbus_dev_fatal(xbb->dev, ENOMEM,
"Unable to allocate request "
"bounce buffers");
return (ENOMEM);
}
#endif /* XBB_USE_BOUNCE_BUFFERS */
reqlist->gnt_handles = malloc(xbb->max_reqlist_segments *
sizeof(*reqlist->gnt_handles),
M_XENBLOCKBACK, M_NOWAIT|M_ZERO);
@ -3489,14 +3318,29 @@ xbb_attach_failed(struct xbb_softc *xbb, int err, const char *fmt, ...)
static int
xbb_probe(device_t dev)
{
uint32_t regs[4];
if (!strcmp(xenbus_get_type(dev), "vbd")) {
device_set_desc(dev, "Backend Virtual Block Device");
device_quiet(dev);
return (0);
}
if (strcmp(xenbus_get_type(dev), "vbd"))
return (ENXIO);
return (ENXIO);
KASSERT(xen_cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
cpuid_count(xen_cpuid_base + 4, 0, regs);
/* Only attach if Xen creates IOMMU entries for grant mapped pages. */
if (!(regs[0] & XEN_HVM_CPUID_IOMMU_MAPPINGS)) {
static bool warned;
if (!warned) {
warned = true;
printf(
"xen-blkback disabled due to grant maps lacking IOMMU entries\n");
}
return (ENXIO);
}
device_set_desc(dev, "Backend Virtual Block Device");
device_quiet(dev);
return (0);
}
/**