open-gpu-kernel-modules/kernel-open/nvidia/nv-mmap.c
Andy Ritger fe0728787f
515.76
2022-09-20 13:54:59 -07:00

795 lines
22 KiB
C

/*
* SPDX-FileCopyrightText: Copyright (c) 1999-2021 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
* DEALINGS IN THE SOFTWARE.
*/
#define __NO_VERSION__
#include "os-interface.h"
#include "nv-linux.h"
#include "nv_speculation_barrier.h"
/*
* The 'struct vm_operations' open() callback is called by the Linux
* kernel when the parent VMA is split or copied, close() when the
* current VMA is about to be deleted.
*
* We implement these callbacks to keep track of the number of user
* mappings of system memory allocations. This was motivated by a
* subtle interaction problem between the driver and the kernel with
* respect to the bookkeeping of pages marked reserved and later
* mapped with mmap().
*
* Traditionally, the Linux kernel ignored reserved pages, such that
* when they were mapped via mmap(), the integrity of their usage
* counts depended on the reserved bit being set for as long as user
* mappings existed.
*
* Since we mark system memory pages allocated for DMA reserved and
* typically map them with mmap(), we need to ensure they remain
* reserved until the last mapping has been torn down. This worked
* correctly in most cases, but in a few, the RM API called into the
* RM to free memory before calling munmap() to unmap it.
*
* In the past, we allowed nv_free_pages() to remove the 'at' from
* the parent device's allocation list in this case, but didn't
* release the underlying pages until the last user mapping had been
* destroyed:
*
* In nvidia_vma_release(), we freed any resources associated with
* the allocation (IOMMU mappings, etc.) and cleared the
* underlying pages' reserved bits, but didn't free them. The kernel
* was expected to do this.
*
* This worked in practise, but made dangerous assumptions about the
* kernel's behavior and could fail in some cases. We now handle
* this case differently (see below).
*/
static void
nvidia_vma_open(struct vm_area_struct *vma)
{
nv_alloc_t *at = NV_VMA_PRIVATE(vma);
NV_PRINT_VMA(NV_DBG_MEMINFO, vma);
if (at != NULL)
{
NV_ATOMIC_INC(at->usage_count);
NV_PRINT_AT(NV_DBG_MEMINFO, at);
}
}
/*
* (see above for additional information)
*
* If the 'at' usage count drops to zero with the updated logic, the
* the allocation is recorded in the free list of the private
* data associated with the file pointer; nvidia_close() uses this
* list to perform deferred free operations when the parent file
* descriptor is closed. This will typically happen when the process
* exits.
*
* Since this is technically a workaround to handle possible fallout
* from misbehaving clients, we additionally print a warning.
*/
static void
nvidia_vma_release(struct vm_area_struct *vma)
{
nv_alloc_t *at = NV_VMA_PRIVATE(vma);
nv_linux_file_private_t *nvlfp = NV_GET_LINUX_FILE_PRIVATE(NV_VMA_FILE(vma));
static int count = 0;
NV_PRINT_VMA(NV_DBG_MEMINFO, vma);
if (at != NULL && nv_alloc_release(nvlfp, at))
{
if ((at->pid == os_get_current_process()) &&
(count++ < NV_MAX_RECURRING_WARNING_MESSAGES))
{
nv_printf(NV_DBG_MEMINFO,
"NVRM: VM: %s: late unmap, comm: %s, 0x%p\n",
__FUNCTION__, current->comm, at);
}
}
}
static int
nvidia_vma_access(
struct vm_area_struct *vma,
unsigned long addr,
void *buffer,
int length,
int write
)
{
nv_alloc_t *at = NULL;
nv_linux_file_private_t *nvlfp = NV_GET_LINUX_FILE_PRIVATE(NV_VMA_FILE(vma));
nv_state_t *nv = NV_STATE_PTR(nvlfp->nvptr);
NvU32 pageIndex, pageOffset;
void *kernel_mapping;
const nv_alloc_mapping_context_t *mmap_context = &nvlfp->mmap_context;
NvU64 offset;
pageIndex = ((addr - vma->vm_start) >> PAGE_SHIFT);
pageOffset = (addr & ~PAGE_MASK);
if (!mmap_context->valid)
{
nv_printf(NV_DBG_ERRORS, "NVRM: VM: invalid mmap context\n");
return -EINVAL;
}
offset = mmap_context->mmap_start;
if (nv->flags & NV_FLAG_CONTROL)
{
at = NV_VMA_PRIVATE(vma);
/*
* at can be NULL for peer IO mem.
*/
if (!at)
return -EINVAL;
if (pageIndex >= at->num_pages)
return -EINVAL;
/*
* For PPC64LE build, nv_array_index_no_speculate() is not defined
* therefore call nv_speculation_barrier().
* When this definition is added, this platform check should be removed.
*/
#if !defined(NVCPU_PPC64LE)
pageIndex = nv_array_index_no_speculate(pageIndex, at->num_pages);
#else
nv_speculation_barrier();
#endif
kernel_mapping = (void *)(at->page_table[pageIndex]->virt_addr + pageOffset);
}
else if (IS_FB_OFFSET(nv, offset, length))
{
addr = (offset & PAGE_MASK);
kernel_mapping = os_map_kernel_space(addr, PAGE_SIZE, NV_MEMORY_UNCACHED);
if (kernel_mapping == NULL)
return -ENOMEM;
kernel_mapping = ((char *)kernel_mapping + pageOffset);
}
else
return -EINVAL;
length = NV_MIN(length, (int)(PAGE_SIZE - pageOffset));
if (write)
memcpy(kernel_mapping, buffer, length);
else
memcpy(buffer, kernel_mapping, length);
if (at == NULL)
{
kernel_mapping = ((char *)kernel_mapping - pageOffset);
os_unmap_kernel_space(kernel_mapping, PAGE_SIZE);
}
return length;
}
static vm_fault_t nvidia_fault(
#if !defined(NV_VM_OPS_FAULT_REMOVED_VMA_ARG)
struct vm_area_struct *vma,
#endif
struct vm_fault *vmf
)
{
#if defined(NV_VM_OPS_FAULT_REMOVED_VMA_ARG)
struct vm_area_struct *vma = vmf->vma;
#endif
nv_linux_file_private_t *nvlfp = NV_GET_LINUX_FILE_PRIVATE(NV_VMA_FILE(vma));
nv_linux_state_t *nvl = nvlfp->nvptr;
nv_state_t *nv = NV_STATE_PTR(nvl);
vm_fault_t ret = VM_FAULT_NOPAGE;
NvU64 page;
NvU64 num_pages = NV_VMA_SIZE(vma) >> PAGE_SHIFT;
NvU64 pfn_start =
(nvlfp->mmap_context.mmap_start >> PAGE_SHIFT) + vma->vm_pgoff;
// Mapping revocation is only supported for GPU mappings.
if (NV_IS_CTL_DEVICE(nv))
{
return VM_FAULT_SIGBUS;
}
// Wake up GPU and reinstate mappings only if we are not in S3/S4 entry
if (!down_read_trylock(&nv_system_pm_lock))
{
return VM_FAULT_NOPAGE;
}
down(&nvl->mmap_lock);
// Wake up the GPU if it is not currently safe to mmap.
if (!nvl->safe_to_mmap)
{
NV_STATUS status;
if (!nvl->gpu_wakeup_callback_needed)
{
// GPU wakeup callback already scheduled.
up(&nvl->mmap_lock);
up_read(&nv_system_pm_lock);
return VM_FAULT_NOPAGE;
}
/*
* GPU wakeup cannot be completed directly in the fault handler due to the
* inability to take the GPU lock while mmap_lock is held.
*/
status = rm_schedule_gpu_wakeup(nvl->sp[NV_DEV_STACK_GPU_WAKEUP], nv);
if (status != NV_OK)
{
nv_printf(NV_DBG_ERRORS,
"NVRM: VM: rm_schedule_gpu_wakeup failed: %x\n", status);
up(&nvl->mmap_lock);
up_read(&nv_system_pm_lock);
return VM_FAULT_SIGBUS;
}
// Ensure that we do not schedule duplicate GPU wakeup callbacks.
nvl->gpu_wakeup_callback_needed = NV_FALSE;
up(&nvl->mmap_lock);
up_read(&nv_system_pm_lock);
return VM_FAULT_NOPAGE;
}
// Safe to mmap, map all pages in this VMA.
for (page = 0; page < num_pages; page++)
{
NvU64 virt_addr = vma->vm_start + (page << PAGE_SHIFT);
NvU64 pfn = pfn_start + page;
ret = nv_insert_pfn(vma, virt_addr, pfn,
nvlfp->mmap_context.remap_prot_extra);
if (ret != VM_FAULT_NOPAGE)
{
nv_printf(NV_DBG_ERRORS,
"NVRM: VM: nv_insert_pfn failed: %x\n", ret);
break;
}
nvl->all_mappings_revoked = NV_FALSE;
}
up(&nvl->mmap_lock);
up_read(&nv_system_pm_lock);
return ret;
}
static struct vm_operations_struct nv_vm_ops = {
.open = nvidia_vma_open,
.close = nvidia_vma_release,
.fault = nvidia_fault,
.access = nvidia_vma_access,
};
int nv_encode_caching(
pgprot_t *prot,
NvU32 cache_type,
nv_memory_type_t memory_type
)
{
pgprot_t tmp;
if (prot == NULL)
{
tmp = __pgprot(0);
prot = &tmp;
}
switch (cache_type)
{
case NV_MEMORY_UNCACHED_WEAK:
#if defined(NV_PGPROT_UNCACHED_WEAK)
*prot = NV_PGPROT_UNCACHED_WEAK(*prot);
break;
#endif
case NV_MEMORY_UNCACHED:
*prot = (memory_type == NV_MEMORY_TYPE_SYSTEM) ?
NV_PGPROT_UNCACHED(*prot) :
NV_PGPROT_UNCACHED_DEVICE(*prot);
break;
#if defined(NV_PGPROT_WRITE_COMBINED) && \
defined(NV_PGPROT_WRITE_COMBINED_DEVICE)
case NV_MEMORY_WRITECOMBINED:
if (NV_ALLOW_WRITE_COMBINING(memory_type))
{
*prot = (memory_type == NV_MEMORY_TYPE_FRAMEBUFFER) ?
NV_PGPROT_WRITE_COMBINED_DEVICE(*prot) :
NV_PGPROT_WRITE_COMBINED(*prot);
break;
}
/*
* If WC support is unavailable, we need to return an error
* code to the caller, but need not print a warning.
*
* For frame buffer memory, callers are expected to use the
* UC- memory type if we report WC as unsupported, which
* translates to the effective memory type WC if a WC MTRR
* exists or else UC.
*/
return 1;
#endif
case NV_MEMORY_CACHED:
if (NV_ALLOW_CACHING(memory_type))
break;
// Intentional fallthrough.
default:
nv_printf(NV_DBG_ERRORS,
"NVRM: VM: cache type %d not supported for memory type %d!\n",
cache_type, memory_type);
return 1;
}
return 0;
}
int static nvidia_mmap_peer_io(
struct vm_area_struct *vma,
nv_alloc_t *at,
NvU64 page_index,
NvU64 pages
)
{
int ret;
NvU64 start;
NvU64 size;
BUG_ON(!at->flags.contig);
start = at->page_table[page_index]->phys_addr;
size = pages * PAGE_SIZE;
ret = nv_io_remap_page_range(vma, start, size, 0);
return ret;
}
int static nvidia_mmap_sysmem(
struct vm_area_struct *vma,
nv_alloc_t *at,
NvU64 page_index,
NvU64 pages
)
{
NvU64 j;
int ret = 0;
unsigned long start = 0;
NV_ATOMIC_INC(at->usage_count);
start = vma->vm_start;
for (j = page_index; j < (page_index + pages); j++)
{
/*
* For PPC64LE build, nv_array_index_no_speculate() is not defined
* therefore call nv_speculation_barrier().
* When this definition is added, this platform check should be removed.
*/
#if !defined(NVCPU_PPC64LE)
j = nv_array_index_no_speculate(j, (page_index + pages));
#else
nv_speculation_barrier();
#endif
#if defined(NV_VGPU_KVM_BUILD)
if (at->flags.guest)
{
ret = nv_remap_page_range(vma, start, at->page_table[j]->phys_addr,
PAGE_SIZE, vma->vm_page_prot);
}
else
#endif
{
vma->vm_page_prot = nv_adjust_pgprot(vma->vm_page_prot, 0);
ret = vm_insert_page(vma, start,
NV_GET_PAGE_STRUCT(at->page_table[j]->phys_addr));
}
if (ret)
{
NV_ATOMIC_DEC(at->usage_count);
return -EAGAIN;
}
start += PAGE_SIZE;
}
return ret;
}
static int nvidia_mmap_numa(
struct vm_area_struct *vma,
const nv_alloc_mapping_context_t *mmap_context)
{
NvU64 start, addr;
NvU64 pages;
NvU64 i;
pages = NV_VMA_SIZE(vma) >> PAGE_SHIFT;
start = vma->vm_start;
if (mmap_context->num_pages < pages)
{
return -EINVAL;
}
// Needed for the linux kernel for mapping compound pages
vma->vm_flags |= VM_MIXEDMAP;
for (i = 0, addr = mmap_context->page_array[0]; i < pages;
addr = mmap_context->page_array[++i], start += PAGE_SIZE)
{
if (vm_insert_page(vma, start, NV_GET_PAGE_STRUCT(addr)) != 0)
{
return -EAGAIN;
}
}
return 0;
}
int nvidia_mmap_helper(
nv_state_t *nv,
nv_linux_file_private_t *nvlfp,
nvidia_stack_t *sp,
struct vm_area_struct *vma,
void *vm_priv
)
{
NvU32 prot = 0;
int ret;
const nv_alloc_mapping_context_t *mmap_context = &nvlfp->mmap_context;
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
NV_STATUS status;
if (nvlfp == NULL)
return NV_ERR_INVALID_ARGUMENT;
/*
* If mmap context is not valid on this file descriptor, this mapping wasn't
* previously validated with the RM so it must be rejected.
*/
if (!mmap_context->valid)
{
nv_printf(NV_DBG_ERRORS, "NVRM: VM: invalid mmap\n");
return -EINVAL;
}
NV_PRINT_VMA(NV_DBG_MEMINFO, vma);
status = nv_check_gpu_state(nv);
if (status != NV_OK)
{
NV_DEV_PRINTF(NV_DBG_INFO, nv,
"GPU is lost, skipping nvidia_mmap_helper\n");
return status;
}
NV_VMA_PRIVATE(vma) = vm_priv;
prot = mmap_context->prot;
/*
* Nvidia device node(nvidia#) maps device's BAR memory,
* Nvidia control node(nvidiactrl) maps system memory.
*/
if (!NV_IS_CTL_DEVICE(nv))
{
NvU32 remap_prot_extra = mmap_context->remap_prot_extra;
NvU64 mmap_start = mmap_context->mmap_start;
NvU64 mmap_length = mmap_context->mmap_size;
NvU64 access_start = mmap_context->access_start;
NvU64 access_len = mmap_context->access_size;
if (IS_REG_OFFSET(nv, access_start, access_len))
{
if (nv_encode_caching(&vma->vm_page_prot, NV_MEMORY_UNCACHED,
NV_MEMORY_TYPE_REGISTERS))
{
return -ENXIO;
}
}
else if (IS_FB_OFFSET(nv, access_start, access_len))
{
if (IS_UD_OFFSET(nv, access_start, access_len))
{
if (nv_encode_caching(&vma->vm_page_prot, NV_MEMORY_UNCACHED,
NV_MEMORY_TYPE_FRAMEBUFFER))
{
return -ENXIO;
}
}
else
{
if (nv_encode_caching(&vma->vm_page_prot,
rm_disable_iomap_wc() ? NV_MEMORY_UNCACHED : NV_MEMORY_WRITECOMBINED,
NV_MEMORY_TYPE_FRAMEBUFFER))
{
if (nv_encode_caching(&vma->vm_page_prot,
NV_MEMORY_UNCACHED_WEAK, NV_MEMORY_TYPE_FRAMEBUFFER))
{
return -ENXIO;
}
}
}
}
down(&nvl->mmap_lock);
if (nvl->safe_to_mmap)
{
nvl->all_mappings_revoked = NV_FALSE;
//
// This path is similar to the sysmem mapping code.
// TODO: Refactor is needed as part of bug#2001704.
// Use pfn_valid to determine whether the physical address has
// backing struct page. This is used to isolate P8 from P9.
//
if ((nv_get_numa_status(nvl) == NV_NUMA_STATUS_ONLINE) &&
!IS_REG_OFFSET(nv, access_start, access_len) &&
(pfn_valid(PFN_DOWN(mmap_start))))
{
ret = nvidia_mmap_numa(vma, mmap_context);
if (ret)
{
up(&nvl->mmap_lock);
return ret;
}
}
else
{
if (nv_io_remap_page_range(vma, mmap_start, mmap_length,
remap_prot_extra) != 0)
{
up(&nvl->mmap_lock);
return -EAGAIN;
}
}
}
up(&nvl->mmap_lock);
vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND;
}
else
{
nv_alloc_t *at;
NvU64 page_index;
NvU64 pages;
NvU64 mmap_size;
at = (nv_alloc_t *)mmap_context->alloc;
page_index = mmap_context->page_index;
mmap_size = NV_VMA_SIZE(vma);
pages = mmap_size >> PAGE_SHIFT;
if ((page_index + pages) > at->num_pages)
{
return -ERANGE;
}
/*
* Callers that pass in non-NULL VMA private data must never reach this
* code. They should be mapping on a non-control node.
*/
BUG_ON(NV_VMA_PRIVATE(vma));
if (at->flags.peer_io)
{
if (nv_encode_caching(&vma->vm_page_prot,
at->cache_type,
NV_MEMORY_TYPE_DEVICE_MMIO))
{
return -ENXIO;
}
/*
* There is no need to keep 'peer IO at' alive till vma_release like
* 'sysmem at' because there are no security concerns where a client
* could free RM allocated sysmem before unmapping it. Hence, vm_ops
* are NOP, and at->usage_count is never being used.
*/
NV_VMA_PRIVATE(vma) = NULL;
ret = nvidia_mmap_peer_io(vma, at, page_index, pages);
BUG_ON(NV_VMA_PRIVATE(vma));
}
else
{
if (nv_encode_caching(&vma->vm_page_prot,
at->cache_type,
NV_MEMORY_TYPE_SYSTEM))
{
return -ENXIO;
}
NV_VMA_PRIVATE(vma) = at;
ret = nvidia_mmap_sysmem(vma, at, page_index, pages);
}
if (ret)
{
return ret;
}
NV_PRINT_AT(NV_DBG_MEMINFO, at);
vma->vm_flags |= (VM_IO | VM_LOCKED | VM_RESERVED);
vma->vm_flags |= (VM_DONTEXPAND | VM_DONTDUMP);
}
if ((prot & NV_PROTECT_WRITEABLE) == 0)
{
vma->vm_page_prot = NV_PGPROT_READ_ONLY(vma->vm_page_prot);
vma->vm_flags &= ~VM_WRITE;
vma->vm_flags &= ~VM_MAYWRITE;
}
vma->vm_ops = &nv_vm_ops;
return 0;
}
int nvidia_mmap(
struct file *file,
struct vm_area_struct *vma
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_FILEP(file);
nv_state_t *nv = NV_STATE_PTR(nvl);
nv_linux_file_private_t *nvlfp = NV_GET_LINUX_FILE_PRIVATE(file);
nvidia_stack_t *sp = NULL;
int status;
//
// Do not allow mmap operation if this is a fd into
// which rm objects have been exported.
//
if (nvlfp->nvfp.handles != NULL)
{
return -EINVAL;
}
down(&nvlfp->fops_sp_lock[NV_FOPS_STACK_INDEX_MMAP]);
sp = nvlfp->fops_sp[NV_FOPS_STACK_INDEX_MMAP];
status = nvidia_mmap_helper(nv, nvlfp, sp, vma, NULL);
up(&nvlfp->fops_sp_lock[NV_FOPS_STACK_INDEX_MMAP]);
return status;
}
void
nv_revoke_gpu_mappings_locked(
nv_state_t *nv
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
nv_linux_file_private_t *nvlfp;
/* Revoke all mappings for every open file */
list_for_each_entry (nvlfp, &nvl->open_files, entry)
{
unmap_mapping_range(&nvlfp->mapping, 0, ~0, 1);
}
nvl->all_mappings_revoked = NV_TRUE;
}
NV_STATUS NV_API_CALL nv_revoke_gpu_mappings(
nv_state_t *nv
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
// Mapping revocation is only supported for GPU mappings.
if (NV_IS_CTL_DEVICE(nv))
{
return NV_ERR_NOT_SUPPORTED;
}
down(&nvl->mmap_lock);
nv_revoke_gpu_mappings_locked(nv);
up(&nvl->mmap_lock);
return NV_OK;
}
void NV_API_CALL nv_acquire_mmap_lock(
nv_state_t *nv
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
down(&nvl->mmap_lock);
}
void NV_API_CALL nv_release_mmap_lock(
nv_state_t *nv
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
up(&nvl->mmap_lock);
}
NvBool NV_API_CALL nv_get_all_mappings_revoked_locked(
nv_state_t *nv
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
// Caller must hold nvl->mmap_lock for all decisions based on this
return nvl->all_mappings_revoked;
}
void NV_API_CALL nv_set_safe_to_mmap_locked(
nv_state_t *nv,
NvBool safe_to_mmap
)
{
nv_linux_state_t *nvl = NV_GET_NVL_FROM_NV_STATE(nv);
// Caller must hold nvl->mmap_lock
/*
* If nvl->safe_to_mmap is transitioning from TRUE to FALSE, we expect to
* need to schedule a GPU wakeup callback when we fault.
*
* nvl->gpu_wakeup_callback_needed will be set to FALSE in nvidia_fault()
* after scheduling the GPU wakeup callback, preventing us from scheduling
* duplicates.
*/
if (!safe_to_mmap && nvl->safe_to_mmap)
{
nvl->gpu_wakeup_callback_needed = NV_TRUE;
}
nvl->safe_to_mmap = safe_to_mmap;
}