xen/gntdev: Fixes for leaks and VMA splitting
cover xen/gntdev: Fixes for leaks and VMA splitting
Hi all,
The changes in this patch series intend to fix the Xen grant device driver, so that grant mapping leaks caused by partially failed grant mapping operations are avoided with the first patch, and so that the splitting of VMAs does not result in incorrectly unmapped grant pages with the second patch. The second patch also prevents a similar issue in a double-mapping scenario, where mmap() is used with MAP_FIXED to map grants over an existing mapping created with the same grants, and where grant pages are unmapped incorrectly as well.
These commits were tested on top of Linux v5.15.67, but I have also verified that they compile fine on top of the base commit mentioned at the bottom of this cover letter. The base commit in question is tagged as "next-20220909".
Thank you,
Vefa
M. Vefa Bicakci (2):
xen/gntdev: Prevent leaking grants
xen/gntdev: Accommodate VMA splitting
drivers/xen/gntdev-common.h | 3 +- drivers/xen/gntdev.c | 90 +++++++++++++++++++++---------------- 2 files changed, 54 insertions(+), 39 deletions(-)
base-commit: 9a82ccda91ed2b40619cb3c10d446ae1f97bab6e
Signature · 1 line
2.37.3
PATCH 1/2 xen/gntdev: Prevent leaking grants
Prior to this commit, if a grant mapping operation failed partially, some of the entries in the map_ops array would be invalid, whereas all of the entries in the kmap_ops array would be valid. This in turn would cause the following logic in gntdev_map_grant_pages to become invalid:
for (i = 0; i < map->count; i++) {
if (map->map_ops[i].status == GNTST_okay) {
map->unmap_ops[i].handle = map->map_ops[i].handle;
if (!use_ptemod)
alloced++;
}
if (use_ptemod) {
if (map->kmap_ops[i].status == GNTST_okay) {
if (map->map_ops[i].status == GNTST_okay)
alloced++;
map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
}
}
}
...
atomic_add(alloced, &map->live_grants);Assume that use_ptemod is true (i.e., the domain mapping the granted pages is a paravirtualized domain). In the code excerpt above, note that the "alloced" variable is only incremented when both kmap_ops[i].status and map_ops[i].status are set to GNTST_okay (i.e., both mapping operations are successful). However, as also noted above, there are cases where a grant mapping operation fails partially, breaking the assumption of the code excerpt above.
The aforementioned causes map->live_grants to be incorrectly set. In some cases, all of the map_ops mappings fail, but all of the kmap_ops mappings succeed, meaning that live_grants may remain zero. This in turn makes it impossible to unmap the successfully grant-mapped pages pointed to by kmap_ops, because unmap_grant_pages has the following snippet of code at its beginning:
if (atomic_read(&map->live_grants) == 0)
return; /* Nothing to do */
In other cases where only some of the map_ops mappings fail but all kmap_ops mappings succeed, live_grants is made positive, but when the user requests unmapping the grant-mapped pages, __unmap_grant_pages_done will then make map->live_grants negative, because the latter function does not check if all of the pages that were requested to be unmapped were actually unmapped, and the same function unconditionally subtracts "data->count" (i.e., a value that can be greater than map->live_grants) from map->live_grants. The side effects of a negative live_grants value have not been studied.
The net effect of all of this is that grant references are leaked in one of the above conditions. In Qubes OS v4.1 (which uses Xen's grant mechanism extensively for X11 GUI isolation), this issue manifests itself with warning messages like the following to be printed out by the Linux kernel in the VM that had granted pages (that contain X11 GUI window data) to dom0: "g.e. 0x1234 still pending", especially after the user rapidly resizes GUI VM windows (causing some grant-mapping operations to partially or completely fail, due to the fact that the VM unshares some of the pages as part of the window resizing, making the pages impossible to grant-map from dom0).
The fix for this issue involves counting all successful map_ops and kmap_ops mappings separately, and then adding the sum to live_grants. During unmapping, only the number of successfully unmapped grants is subtracted from live_grants. To determine which grants were successfully unmapped, their status fields are set to an arbitrary positive number (1), as was done in commit ebee0eab0859 ("Xen/gntdev: correct error checking in gntdev_map_grant_pages()"). The code is also modified to check for negative live_grants values after the subtraction and warn the user.
--- drivers/xen/gntdev.c | 32 +++++++++++++++++++++++++++----- 1 file changed, 27 insertions(+), 5 deletions(-)
drivers/xen/gntdev.c +27 -5
@@ -367,8 +367,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map)for (i =0 ; i < map->count; i++) {if (map->map_ops[i].status == GNTST_okay) { map->unmap_ops[i].handle = map->map_ops[i].handle; -if (!use_ptemod) - alloced++; + alloced++; }else if (!err) err = -EINVAL ;
@@ -377,8 +376,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map)if (use_ptemod) {if (map->kmap_ops[i].status == GNTST_okay) { -if (map->map_ops[i].status == GNTST_okay) - alloced++; + alloced++; map->kunmap_ops[i].handle = map->kmap_ops[i].handle; }else if (!err) err = -EINVAL ;
@@ -394,8 +392,13 @@ static void __unmap_grant_pages_done(int result,unsigned int i;struct gntdev_grant_map *map = data->data;unsigned int offset = data->unmap_ops - map->unmap_ops; +int successful_unmaps =0 ; +int live_grants;for (i =0 ; i < data->count; i++) { +if (map->unmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++; +WARN_ON (map->unmap_ops[offset + i].status != GNTST_okay && map->unmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE ); pr_debug("unmap handle=%d st=%d\n",
@@ -403,6 +406,9 @@ static void __unmap_grant_pages_done(int result, map->unmap_ops[offset+i].status); map->unmap_ops[offset+i].handle =INVALID_GRANT_HANDLE ;if (use_ptemod) { +if (map->kunmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++; +WARN_ON (map->kunmap_ops[offset + i].status != GNTST_okay && map->kunmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE ); pr_debug("kunmap handle=%u st=%d\n",
@@ -411,11 +417,15 @@ static void __unmap_grant_pages_done(int result, map->kunmap_ops[offset+i].handle =INVALID_GRANT_HANDLE ; } } +/* * Decrease the live-grant counter. This must happen after the loop to * prevent premature reuse of the grants by gnttab_mmap(). */ -atomic_sub (data->count, &map->live_grants); + live_grants =atomic_sub_return (successful_unmaps, &map->live_grants); +if (WARN_ON (live_grants <0 )) +pr_err ("%s: live_grants became negative (%d) after unmapping %d pages!\n" , + __func__, live_grants, successful_unmaps);/* Release reference taken by __unmap_grant_pages */ gntdev_put_map (NULL , map);
@@ -424,6 +434,8 @@ static void __unmap_grant_pages_done(int result,static void __unmap_grant_pages (struct gntdev_grant_map *map,int offset,int pages) { +int idx; +if (map->notify.flags &UNMAP_NOTIFY_CLEAR_BYTE ) {int pgno = (map->notify.addr >>PAGE_SHIFT );
@@ -436,6 +448,16 @@ static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, } } +/* Set all unmap/kunmap status fields to an arbitrary positive value, +* so that it is possible to determine which grants were successfully +* unmapped by inspecting the status fields. +*/ +for (idx = offset; idx < offset + pages; idx++) { + map->unmap_ops[idx].status =1 ; +if (use_ptemod) + map->kunmap_ops[idx].status =1 ; + } + map->unmap_data.unmap_ops = map->unmap_ops + offset; map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset :NULL ; map->unmap_data.pages = map->pages + offset;
Signature · 1 line
2.37.3
Juergen Gross
maintainerRe: [PATCH 1/2] xen/gntdev: Prevent leaking grants
Quoted 104 lines · M. Vefa Bicakci
Prior to this commit, if a grant mapping operation failed partially, some of the entries in the
map_opsarray would be invalid, whereas all of the entries in thekmap_opsarray would be valid. This in turn would cause the following logic ingntdev_map_grant_pagesto become invalid:for (i = 0; i < map->count; i++) { if (map->map_ops[i].status == GNTST_okay) { map->unmap_ops[i].handle = map->map_ops[i].handle; if (!use_ptemod) alloced++; } if (use_ptemod) { if (map->kmap_ops[i].status == GNTST_okay) { if (map->map_ops[i].status == GNTST_okay) alloced++; map->kunmap_ops[i].handle = map->kmap_ops[i].handle; } } } ... atomic_add(alloced, &map->live_grants);Assume that
use_ptemodis true (i.e., the domain mapping the granted pages is a paravirtualized domain). In the code excerpt above, note that the "alloced" variable is only incremented when both kmap_ops[i].status and map_ops[i].status are set toGNTST_okay(i.e., both mapping operations are successful). However, as also noted above, there are cases where a grant mapping operation fails partially, breaking the assumption of the code excerpt above.The aforementioned causes
map->live_grantsto be incorrectly set. In some cases, all of themap_opsmappings fail, but all of thekmap_opsmappings succeed, meaning thatlive_grantsmay remain zero. This in turn makes it impossible to unmap the successfully grant-mapped pages pointed to bykmap_ops, becauseunmap_grant_pageshas the following snippet of code at its beginning:if (
atomic_read(&map->live_grants)== 0)
return; /* Nothing to do */In other cases where only some of the
map_opsmappings fail but allkmap_opsmappings succeed,live_grantsis made positive, but when the user requests unmapping the grant-mapped pages,__unmap_grant_pages_donewill then makemap->live_grantsnegative, because the latter function does not check if all of the pages that were requested to be unmapped were actually unmapped, and the same function unconditionally subtracts "data->count" (i.e., a value that can be greater thanmap->live_grants) frommap->live_grants. The side effects of a negativelive_grantsvalue have not been studied.The net effect of all of this is that grant references are leaked in one of the above conditions. In Qubes OS v4.1 (which uses Xen's grant mechanism extensively for X11 GUI isolation), this issue manifests itself with warning messages like the following to be printed out by the Linux kernel in the VM that had granted pages (that contain X11 GUI window data) to dom0: "g.e.
0x1234still pending", especially after the user rapidly resizes GUI VM windows (causing some grant-mapping operations to partially or completely fail, due to the fact that the VM unshares some of the pages as part of the window resizing, making the pages impossible to grant-map from dom0).The fix for this issue involves counting all successful
map_opsandkmap_opsmappings separately, and then adding the sum tolive_grants. During unmapping, only the number of successfully unmapped grants is subtracted fromlive_grants. To determine which grants were successfully unmapped, their status fields are set to an arbitrary positive number (1), as was done in commitebee0eab0859("Xen/gntdev: correct error checking ingntdev_map_grant_pages()"). The code is also modified to check for negativelive_grantsvalues after the subtraction and warn the user.Fixes dbe97cff7dd9 ("xen/gntdev: Avoid blocking in unmap_grant_pages()")Signed-off-by M. Vefa Bicakci <[email protected]>--- drivers/xen/gntdev.c | 32 +++++++++++++++++++++++++++----- 1 file changed, 27 insertions(+), 5 deletions(-)drivers/xen/gntdev.c +2
-4@@ -367,8 +367,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map)for (i =0 ; i < map->count; i++) {if (map->map_ops[i].status == GNTST_okay) { map->unmap_ops[i].handle = map->map_ops[i].handle; -if (!use_ptemod) - alloced++; + alloced++; }else if (!err) err = -EINVAL ;@@ -377,8 +376,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map)if (use_ptemod) {if (map->kmap_ops[i].status == GNTST_okay) { -if (map->map_ops[i].status == GNTST_okay) - alloced++; + alloced++; map->kunmap_ops[i].handle = map->kmap_ops[i].handle; }else if (!err) err = -EINVAL ;
drivers/xen/gntdev.c static void __unmap_grant_pages_done(int result, · line 400
Quoted 9 lines
drivers/xen/gntdev.c +4
-0@@ -394,8 +392,13 @@ static void __unmap_grant_pages_done(int result,unsigned int i;struct gntdev_grant_map *map = data->data;unsigned int offset = data->unmap_ops - map->unmap_ops; +int successful_unmaps =0 ; +int live_grants;for (i =0 ; i < data->count; i++) { +if (map->unmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++;
Shouldn't this test include "&& handle != INVALID_GRANT_HANDLE" ?
This should enable you to drop setting status to 1 below.
Quoted 47 lines
drivers/xen/gntdev.c +14
-1+WARN_ON (map->unmap_ops[offset + i].status != GNTST_okay && map->unmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE ); pr_debug("unmap handle=%d st=%d\n",@@ -403,6 +406,9 @@ static void __unmap_grant_pages_done(int result, map->unmap_ops[offset+i].status); map->unmap_ops[offset+i].handle =INVALID_GRANT_HANDLE ;if (use_ptemod) { +if (map->kunmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++; +WARN_ON (map->kunmap_ops[offset + i].status != GNTST_okay && map->kunmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE ); pr_debug("kunmap handle=%u st=%d\n",@@ -411,11 +417,15 @@ static void __unmap_grant_pages_done(int result, map->kunmap_ops[offset+i].handle =INVALID_GRANT_HANDLE ; } } +/* * Decrease the live-grant counter. This must happen after the loop to * prevent premature reuse of the grants by gnttab_mmap(). */ -atomic_sub (data->count, &map->live_grants); + live_grants =atomic_sub_return (successful_unmaps, &map->live_grants); +if (WARN_ON (live_grants <0 )) +pr_err ("%s: live_grants became negative (%d) after unmapping %d pages!\n" , + __func__, live_grants, successful_unmaps);/* Release reference taken by __unmap_grant_pages */ gntdev_put_map (NULL , map);@@ -424,6 +434,8 @@ static void __unmap_grant_pages_done(int result,static void __unmap_grant_pages (struct gntdev_grant_map *map,int offset,int pages) { +int idx; +if (map->notify.flags &UNMAP_NOTIFY_CLEAR_BYTE ) {int pgno = (map->notify.addr >>PAGE_SHIFT );@@ -436,6 +448,16 @@ static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, } } +/* Set all unmap/kunmap status fields to an arbitrary positive value, +* so that it is possible to determine which grants were successfully +* unmapped by inspecting the status fields.
drivers/xen/gntdev.c static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, · line 463
Quoted 10 lines
drivers/xen/gntdev.c +7
-0+*/ +for (idx = offset; idx < offset + pages; idx++) { + map->unmap_ops[idx].status =1 ; +if (use_ptemod) + map->kunmap_ops[idx].status =1 ; + } + map->unmap_data.unmap_ops = map->unmap_ops + offset; map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset :NULL ; map->unmap_data.pages = map->pages + offset;
Juergen
Re: [PATCH 1/2] xen/gntdev: Prevent leaking grants
Quoted 103 lines · Juergen Gross
Quoted 102 lines · M. Vefa Bicakci
Prior to this commit, if a grant mapping operation failed partially, some of the entries in the
map_opsarray would be invalid, whereas all of the entries in thekmap_opsarray would be valid. This in turn would cause the following logic ingntdev_map_grant_pagesto become invalid:for (i = 0; i <
map->count; i++) {
if (map->map_ops[i].status ==GNTST_okay) { map->unmap_ops[i].handle = map->map_ops[i].handle; if (!use_ptemod)
alloced++;
}
if (use_ptemod) {
if (map->kmap_ops[i].status ==GNTST_okay) { if (map->map_ops[i].status ==GNTST_okay) alloced++;
map->kunmap_ops[i].handle = map->kmap_ops[i].handle; }
}
}
...
atomic_add(alloced, &map->live_grants);Assume that
use_ptemodis true (i.e., the domain mapping the granted pages is a paravirtualized domain). In the code excerpt above, note that the "alloced" variable is only incremented when both kmap_ops[i].status and map_ops[i].status are set toGNTST_okay(i.e., both mapping operations are successful). However, as also noted above, there are cases where a grant mapping operation fails partially, breaking the assumption of the code excerpt above.The aforementioned causes
map->live_grantsto be incorrectly set. In some cases, all of themap_opsmappings fail, but all of thekmap_opsmappings succeed, meaning thatlive_grantsmay remain zero. This in turn makes it impossible to unmap the successfully grant-mapped pages pointed to bykmap_ops, becauseunmap_grant_pageshas the following snippet of code at its beginning:if (
atomic_read(&map->live_grants)== 0)
return; /* Nothing to do */In other cases where only some of the
map_opsmappings fail but allkmap_opsmappings succeed,live_grantsis made positive, but when the user requests unmapping the grant-mapped pages,__unmap_grant_pages_donewill then makemap->live_grantsnegative, because the latter function does not check if all of the pages that were requested to be unmapped were actually unmapped, and the same function unconditionally subtracts "data->count" (i.e., a value that can be greater thanmap->live_grants) frommap->live_grants. The side effects of a negativelive_grantsvalue have not been studied.The net effect of all of this is that grant references are leaked in one of the above conditions. In Qubes OS v4.1 (which uses Xen's grant mechanism extensively for X11 GUI isolation), this issue manifests itself with warning messages like the following to be printed out by the Linux kernel in the VM that had granted pages (that contain X11 GUI window data) to dom0: "g.e.
0x1234still pending", especially after the user rapidly resizes GUI VM windows (causing some grant-mapping operations to partially or completely fail, due to the fact that the VM unshares some of the pages as part of the window resizing, making the pages impossible to grant-map from dom0).The fix for this issue involves counting all successful
map_opsandkmap_opsmappings separately, and then adding the sum tolive_grants. During unmapping, only the number of successfully unmapped grants is subtracted fromlive_grants. To determine which grants were successfully unmapped, their status fields are set to an arbitrary positive number (1), as was done in commitebee0eab0859("Xen/gntdev: correct error checking ingntdev_map_grant_pages()"). The code is also modified to check for negativelive_grantsvalues after the subtraction and warn the user.Fixes dbe97cff7dd9 ("xen/gntdev: Avoid blocking in unmap_grant_pages()")Signed-off-by M. Vefa Bicakci <[email protected]>---
drivers/xen/gntdev.c| 32 +++++++++++++++++++++++++++-----1 file changed, 27 insertions(+), 5 deletions(-)drivers/xen/gntdev.c +2
-4@@ -367,8 +367,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map) for (i = 0; i < map->count; i++) { if (map->map_ops[i].status == GNTST_okay) { map->unmap_ops[i].handle = map->map_ops[i].handle; - if (!use_ptemod) - alloced++; + alloced++; } else if (!err) err = -EINVAL;@@ -377,8 +376,7 @@ int gntdev_map_grant_pages(struct gntdev_grant_map *map) if (use_ptemod) { if (map->kmap_ops[i].status == GNTST_okay) { - if (map->map_ops[i].status == GNTST_okay) - alloced++; + alloced++; map->kunmap_ops[i].handle = map->kmap_ops[i].handle; } else if (!err) err = -EINVAL;@@ -394,8 +392,13 @@ static void __unmap_grant_pages_done(int result,
Quoted 9 lines
drivers/xen/gntdev.c static void __unmap_grant_pages_done(int result, · line 400
Quoted 8 lines
drivers/xen/gntdev.c +4
-0unsigned int i; struct gntdev_grant_map *map = data->data; unsigned int offset = data->unmap_ops - map->unmap_ops; + int successful_unmaps = 0; + int live_grants; for (i = 0; i < data->count; i++) { + if (map->unmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++;
Hi,
Sorry for the delay, and thank you for reviewing my patches!
Quoted 3 lines
Shouldn't this test include "&& handle !=
INVALID_GRANT_HANDLE" ?This should enable you to drop setting status to 1 below.
I had not thought of the approach you suggested. Just now, I applied your suggestion to my local kernel tree, and I am building a new kernel for testing now. I hope to publish a newer version of this patch set over the weekend.
Thanks again,
Vefa
Quoted 57 lines
Quoted 45 lines
drivers/xen/gntdev.c +1
-0+WARN_ON(map->unmap_ops[offset + i].status !=
GNTST_okay&& map->unmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE); pr_debug("unmap handle=%d st=%d\n",drivers/xen/gntdev.c +3
-0@@ -403,6 +406,9 @@ static void __unmap_grant_pages_done(int result, map->unmap_ops[offset+i].status); map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE; if (use_ptemod) { + if (map->kunmap_ops[offset + i].status == GNTST_okay) + successful_unmaps++; +WARN_ON(map->kunmap_ops[offset + i].status !=
GNTST_okay&& map->kunmap_ops[offset + i].handle !=INVALID_GRANT_HANDLE); pr_debug("kunmap handle=%u st=%d\n",drivers/xen/gntdev.c +10
-1@@ -411,11 +417,15 @@ static void __unmap_grant_pages_done(int result, map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE; } } + /* * Decrease the live-grant counter. This must happen after the loop to * prevent premature reuse of the grants by gnttab_mmap(). */ - atomic_sub(data->count, &map->live_grants); + live_grants = atomic_sub_return(successful_unmaps, &map->live_grants); + if (WARN_ON(live_grants < 0)) + pr_err("%s: live_grants became negative (%d) after unmapping %d pages!\n", + __func__, live_grants, successful_unmaps); /* Release reference taken by __unmap_grant_pages */ gntdev_put_map(NULL, map);@@ -424,6 +434,8 @@ static void __unmap_grant_pages_done(int result, static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, int pages) { + int idx; + if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) { int pgno = (map->notify.addr >> PAGE_SHIFT);@@ -436,6 +448,16 @@ static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, } } + /* Set all unmap/kunmap status fields to an arbitrary positive value, +* so that it is possible to determine which grants were successfully +* unmapped by inspecting the status fields. drivers/xen/gntdev.c static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, · line 462
Quoted 10 lines
drivers/xen/gntdev.c +7
-0+*/ + for (idx = offset; idx < offset + pages; idx++) { + map->unmap_ops[idx].status = 1; + if (use_ptemod) + map->kunmap_ops[idx].status = 1; + } + map->unmap_data.unmap_ops = map->unmap_ops + offset; map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL; map->unmap_data.pages = map->pages + offset;Juergen
PATCH 2/2 xen/gntdev: Accommodate VMA splitting
In reply to M. Vefa BicakciPrior to this commit, the gntdev driver code did not handle the following scenario correctly:
* User process sets up a gntdev mapping composed of two grant mappings (i.e., two pages shared by another Xen domain).
* User process munmap()s one of the pages.
* User process munmap()s the remaining page.
* User process exits.
In the scenario above, the user process would cause the kernel to log the following messages in dmesg for the first munmap(), and the second munmap() call would result in similar log messages:
BUG: Bad page map in process doublemap.test pte:... pmd:...
page:0000000057c97bff refcount:1 mapcount:-1 \
mapping:0000000000000000 index:0x0 pfn:...
...
page dumped because: bad pte
...
file:gntdev fault:0x0 mmap:gntdev_mmap [xen_gntdev] readpage:0x0
...
Call Trace:
<TASK>
dump_stack_lvl+0x46/0x5e
print_bad_pte.cold+0x66/0xb6
unmap_page_range+0x7e5/0xdc0
unmap_vmas+0x78/0xf0
unmap_region+0xa8/0x110
__do_munmap+0x1ea/0x4e0
__vm_munmap+0x75/0x120
__x64_sys_munmap+0x28/0x40
do_syscall_64+0x38/0x90
entry_SYSCALL_64_after_hwframe+0x61/0xcb
...For each munmap() call, the Xen hypervisor (if built with CONFIG_DEBUG) would print out the following and trigger a general protection fault in dom0:
(XEN) d0v... Attempt to implicitly unmap d0's grant PTE ... (XEN) d0v... Attempt to implicitly unmap d0's grant PTE ...
As of this writing, gntdev_grant_map structure's vma field (referred to as map->vma below) is mainly used for checking the start and end addresses of mappings. However, with split VMAs, these may change, and there could be more than one VMA associated with a gntdev mapping. Hence, remove the use of map->vma and rely on map->pages_vm_start for the original start address and on (map->count << PAGE_SHIFT) for the original mapping size. Let the invalidate() and find_special_page() hooks use these.
Also, given that there can be multiple VMAs associated with a gntdev mapping, move the "mmu_interval_notifier_remove(&map->notifier)" call to the end of gntdev_put_map, so that the MMU notifier is only removed after the closing of the last remaining VMA.
Finally, use an atomic to prevent inadvertent gntdev mapping re-use, instead of using the map->live_grants atomic counter and/or the map->vma pointer (the latter of which is now removed). This prevents the userspace from mmap()'ing (with MAP_FIXED) a gntdev mapping over the same address range as a previously set up gntdev mapping. This scenario can be summarized with the following call-trace, which was valid prior to this commit:
mmap
gntdev_mmap
mmap (repeat mmap with MAP_FIXED over the same address range)
gntdev_invalidate
unmap_grant_pages (sets 'being_removed' entries to true)
gnttab_unmap_refs_async
unmap_single_vma
gntdev_mmap (maps the shared pages again)
munmap
gntdev_invalidate
unmap_grant_pages
(no-op because 'being_removed' entries are true)
unmap_single_vma (Xen reports that a granted page is being
unmapped and triggers a general protection fault in dom0 if Xen was built with CONFIG_DEBUG)
The fix for this last scenario could be worth its own commit, but we opted for a single commit, because removing the gntdev_grant_map structure's vma field requires guarding the entry to gntdev_mmap(), and the live_grants atomic counter is not sufficient on its own to prevent the mmap() over a pre-existing mapping.
---
Note for reviewers:
I am not 100% sure if the "Fixes" tag is correct. Based on a quick look at the history of the modified file, I am under the impression that VMA splits could be broken for the Xen gntdev driver since day 1 (i.e., v2.6.38), but I did not yet attempt to verify this by testing older kernels where the gntdev driver's code is sufficiently similar.
Also, resetting the being_removed flags to false after the completion of unmap operation could be another potential solution (that I have not yet tested in the context of this change) to the mmap and MAP_FIXED issue discussed at the end of the patch description.
--- drivers/xen/gntdev-common.h | 3 +- drivers/xen/gntdev.c | 58 ++++++++++++++++--------------------- 2 files changed, 27 insertions(+), 34 deletions(-)
drivers/xen/gntdev-common.h +2 -1
@@ -44,9 +44,10 @@ struct gntdev_unmap_notify { };struct gntdev_grant_map { +atomic_t in_use;struct mmu_interval_notifier notifier; +bool notifier_init;struct list_head next; -struct vm_area_struct *vma;int index;int count;int flags;
drivers/xen/gntdev.c +25 -33
131 lines not on this page · raw message
Signature · 1 line
2.37.3
Re: [PATCH 2/2] xen/gntdev: Accommodate VMA splitting
Quoted 86 lines · M. Vefa Bicakci
Prior to this commit, the gntdev driver code did not handle the following scenario correctly:
* User process sets up a gntdev mapping composed of two grant mappings (i.e., two pages shared by another Xen domain).
* User process munmap()s one of the pages.
* User process munmap()s the remaining page.
* User process exits.In the scenario above, the user process would cause the kernel to log the following messages in dmesg for the first
munmap(), and the secondmunmap()call would result in similar log messages:BUG: Bad page map in process doublemap.test pte:... pmd:... page:0000000057c97bff refcount:1 mapcount:-1 \ mapping:0000000000000000 index:0x0 pfn:... ... page dumped because: bad pte ... file:gntdev fault:0x0 mmap:gntdev_mmap [xen_gntdev] readpage:0x0 ... Call Trace: <TASK> dump_stack_lvl+0x46/0x5e print_bad_pte.cold+0x66/0xb6 unmap_page_range+0x7e5/0xdc0 unmap_vmas+0x78/0xf0 unmap_region+0xa8/0x110 __do_munmap+0x1ea/0x4e0 __vm_munmap+0x75/0x120 __x64_sys_munmap+0x28/0x40 do_syscall_64+0x38/0x90 entry_SYSCALL_64_after_hwframe+0x61/0xcb ...For each
munmap()call, the Xen hypervisor (if built withCONFIG_DEBUG) would print out the following and trigger a general protection fault in dom0:(XEN) d0v... Attempt to implicitly unmap d0's grant PTE ... (XEN) d0v... Attempt to implicitly unmap d0's grant PTE ...
As of this writing,
gntdev_grant_mapstructure's vma field (referred to asmap->vmabelow) is mainly used for checking the start and end addresses of mappings. However, with split VMAs, these may change, and there could be more than one VMA associated with a gntdev mapping. Hence, remove the use ofmap->vmaand rely onmap->pages_vm_startfor the original start address and on (map->count<<PAGE_SHIFT) for the original mapping size. Let theinvalidate()andfind_special_page()hooks use these.Also, given that there can be multiple VMAs associated with a gntdev mapping, move the "
mmu_interval_notifier_remove(&map->notifier)" call to the end ofgntdev_put_map, so that the MMU notifier is only removed after the closing of the last remaining VMA.Finally, use an atomic to prevent inadvertent gntdev mapping re-use, instead of using the
map->live_grantsatomic counter and/or themap->vmapointer (the latter of which is now removed). This prevents the userspace from mmap()'ing (withMAP_FIXED) a gntdev mapping over the same address range as a previously set up gntdev mapping. This scenario can be summarized with the following call-trace, which was valid prior to this commit:mmap
gntdev_mmap
mmap (repeat mmap withMAP_FIXEDover the same address range)
gntdev_invalidate
unmap_grant_pages(sets 'being_removed' entries to true)
gnttab_unmap_refs_async
unmap_single_vma
gntdev_mmap(maps the shared pages again)
munmap
gntdev_invalidate
unmap_grant_pages
(no-op because 'being_removed' entries are true)
unmap_single_vma(Xen reports that a granted page is being
unmapped and triggers a general protection fault in dom0 if Xen was built withCONFIG_DEBUG)The fix for this last scenario could be worth its own commit, but we opted for a single commit, because removing the
gntdev_grant_mapstructure's vma field requires guarding the entry togntdev_mmap(), and thelive_grantsatomic counter is not sufficient on its own to prevent themmap()over a pre-existing mapping.
Quoted 4 lines
Fixes ab31523c2fca ("xen/gntdev: allow usermode to map granted pages")Signed-off-by M. Vefa Bicakci <[email protected]>
Juergen