generic allocator and mspec driver
Subsystems (2)
RFC 1/1 generic allocator and mspec driver
Hi,
I've been working on a generic allocator to be used for the uncached memory pools as needed by the mspec driver (former SGI fetchop).
I based the allocator on the one from the sym2 driver and then allowing one to setup whatever pools one would want. Everything works pretty well and we integrated it with the mspec (fetchop) driver which now also allows for uncached and cached mappings on non sn2 hardware. Right now the driver is located in arch/ia64/sn/kernel/ however I plan to move that to arch/ia64/kernel/ or drivers/char/.
The remaining issue I am facing is that for the uncached pool I want to make it node aware and we want to use the spill pages from the lower granules for this pool which is easily done on SN2. However I see no generic way to get from a physical address to node id for pages that do not have a struct page assigned to them. I am curious if anyone has any suggestions for how to solve this in a generic way?
Otherwise I'd also welcome comments if anyone has any objections or ideas for the current code before submitting it.
Patch attached includes the driver and the generic allocator in one big patch.
Cheers,
Jes
arch/ia64/Kconfig +9 -0
@@ -225,6 +225,15 @@ If you are compiling a kernel that will run under SGI's IA-64 simulator (Medusa) then say Y, otherwise say N. +config MSPEC + tristate "SGI SN2 Special Memory support" + depends on IA64 + help + This driver exports special memory capabilities of the SGI SN + architecture such as the fetchop facility to user processes. + Fetchops are atomic memory operations that are implemented in the + memory controller on SGI SN hardware. + config FORCE_MAX_ZONEORDER int default "18"
arch/ia64/configs/sn2_defconfig +1 -0
@@ -82,6 +82,7 @@ # CONFIG_IA64_CYCLONE is not set CONFIG_IOSAPIC=y CONFIG_IA64_SGI_SN_SIM=y +CONFIG_SGI_MSPEC=m CONFIG_FORCE_MAX_ZONEORDER CONFIG_SMP=y CONFIG_NR_CPUSQ2
arch/ia64/defconfig +1 -0
@@ -80,6 +80,7 @@ CONFIG_DISCONTIGMEM=y CONFIG_IA64_CYCLONE=y CONFIG_IOSAPIC=y +CONFIG_SGI_MSPEC=m CONFIG_FORCE_MAX_ZONEORDER CONFIG_SMP=y CONFIG_NR_CPUSQ2
arch/ia64/sn/kernel/Makefile +1 -0
@@ -10,3 +10,4 @@ obj-y += setup.o bte.o bte_error.o irq.o mca.o idle.o \ huberror.o io_init.o iomv.o klconflib.o sn2/ obj-$(CONFIG_IA64_GENERIC) += machvec.o +obj-$(CONFIG_SGI_MSPEC) += mspec.o
arch/ia64/sn/kernel/mspec.c +789 -0
789 lines not on this page · raw message
include/asm-ia64/sn/fetchop.h +0 -85
85 lines not on this page · raw message
include/asm-ia64/sn/mspec.h +72 -0
72 lines not on this page · raw message
include/linux/genalloc.h +46 -0
46 lines not on this page · raw message
init/main.c +2 -0
@@ -78,6 +78,7 @@static int init (void *); +extern void gen_pool_init (void );extern void init_IRQ (void );extern void sock_init (void );extern void fork_init (unsigned long );
@@ -482,6 +483,7 @@#endif vfs_caches_init_early ();mem_init (); +gen_pool_init ();kmem_cache_init ();numa_policy_init ();if (late_time_init)
lib/Makefile +1 -1
@@ -6,7 +6,7 @@ bust_spinlocks.o rbtree.o radix-tree.o dump_stack.o \ kobject.o kref.o idr.o div64.o parser.o int_sqrt.o \ bitmap.o extable.o kobject_uevent.o prio_tree.o \ - sha1.o halfmd4.o + sha1.o halfmd4.o genalloc.o ifeq ($(CONFIG_DEBUG_KOBJECT),y) CFLAGS_kobject.o += -DDEBUG
lib/genalloc.c +218 -0
218 lines not on this page · raw message
Quoted 6 lines · Jes Sorensen
The remaining issue I am facing is that for the uncached pool I want to make it node aware and we want to use the spill pages from the lower granules for this pool which is easily done on SN2. However I see no generic way to get from a physical address to node id for pages that do not have a struct page assigned to them. I am curious if anyone has any suggestions for how to solve this in a generic way?
paddr_to_nid?
Jesse
On Wed, Feb 02, 2005 at 02:10:32PM -0500, Jes Sorensen wrote:
General comment:
1) I may be paranoid, but I'm nervous about using memory visible to the VM system for fetchops. If ANYTHING in the kernel makes a reference to the memory and causes a TLB dropin to occur, then we are exposed to data corruption. If memory being used for fetchops is loaded into the cache, either directly or by speculation, then data corruption of the uncached fetchop memory can occur.
Am I being overly paranoid? How can we be certain that nothing will ever reference the fetchop memory alocated from the general VM pool. lcrash, for example.
2) Is there a limit on the number of mspec pages that can be allocated? Is there a shaker that will cause unused mspec granules to be freed? What prevents a malicious/stupid/buggy user from filling the system with mspec pages?
3) Is an "mspec" address a physical address or an uncached virtual address? Some places in the code appear inconsistent. For example:
mspec_free_page(TO_PHYS(maddr)) vs. maddr; /* phys addr of start of mspecs. */
A few code specific issues:
...
+ printk(KERN_WARNING "smp_call_function failed for "
+ "mspec_ipi_visibility! (%i)\n", status);
+ }
+
+ sn_flush_all_caches((unsigned long)tmp, IA64_GRANULE_SIZE);
Don't the TLBs need to be flushed before you flush caches. Otherwise, the cpu may reload data via speculation.
I dont see any TLB flushing of the kernel TLB entries that map the chunks. That needs to be done.
...
+ /*
+ * The kernel requires a page structure to be returned upon + * success, but there are no page structures for low granule pages. + * remap_page_range() creates the pte for us and we return a + * bogus page back to the kernel fault handler to keep it happy + * (the page is freed immediately there). + */
Ugly hack. Isn't there a better way? (I know this isn't your code & you probably don't like this either. I had hoped for a cleaner solution in 2.6....)
...
+ /*
+ * Use the bte to ensure cache lines + * are actually pulled from the + * processor back to the md. + */
+
This doesn't need to be done if the memory was being used for fetchops or uncached memory.
+ s <<= 1;
+ }
+ a = (unsigned long) h[j].next;
It appears that you are keeping a linked list of free memory WITHIN the mspec memory itself. If I'm reading this correctly, all the addresses are uncached virtual addresses so that should be ok. However, it might be good to add debugging code to make sure that you never cause a cachable reference to be made to any of the fetchop memory. The resulting data corruption problems are almost impossible to debug.
Signature · 4 lines
Thanks Jack Steiner ([email protected]) 651-683-5302 Principal Engineer SGI - Silicon Graphics, Inc.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jesse" = Jesse Barnes <[email protected]> writes:
Quoted 7 lines · Jes Sorensen
Quoted 7 lines
The remaining issue I am facing is that for the uncached pool I want to make it node aware and we want to use the spill pages from the lower granules for this pool which is easily done on SN2. However I see no generic way to get from a physical address to node id for pages that do not have a struct page assigned to them. I am curious if anyone has any suggestions for how to solve this in a generic way?
Jesse> paddr_to_nid?
It's not clear to me this covers the spill memory found in the lower granules and second it's really really slow, though that may not be a real problem.
Cheers,
Jes
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jack" = Jack Steiner <[email protected]> writes:
Jack> On Wed, Feb 02, 2005 at 02:10:32PM -0500, Jes Sorensen wrote:
Jack> General comment:
Jack, thanks for the comments, I'll look at it, however I have the following comments (which may or may not be correct from my side):
Jack> 1) I may be paranoid, but I'm nervous about using memory visible Jack> to the VM system for fetchops. If ANYTHING in the kernel makes a Jack> reference to the memory and causes a TLB dropin to occur, then Jack> we are exposed to data corruption. If memory being used for Jack> fetchops is loaded into the cache, either directly or by Jack> speculation, then data corruption of the uncached fetchop memory Jack> can occur.
Jack> Am I being overly paranoid? How can we be certain that nothing
Jack> will ever reference the fetchop memory alocated from the general Jack> VM pool. lcrash, for example.
Once a page is handed out using alloc_pages, the kernel won't touch it again unless you explicitly map it etc. or if some process touches memory at random, which could also happen with the spill pages. So I don't think the situation is any worse than it is for the spill pages in the lower granules.
Jack> 2) Is there a limit on the number of mspec pages that can be Jack> allocated? Is there a shaker that will cause unused mspec Jack> granules to be freed? What prevents a malicious/stupid/buggy Jack> user from filling the system with mspec pages?
Currently there is no limit on this, however it could easily be imposed either by having a max number of granules allocated per node or system wide.
I could add code to free granules when it's all released but I believe the amount of memory being pulled in for this in real life situations is so limited it's not really worth the complexity. Adding a hard limit for how much is allowed to be allocated seems simpler.
Jack> 3) Is an "mspec" address a physical address or an uncached Jack> virtual address? Some places in the code appear Jack> inconsistent. For example:
Jack> mspec_free_page(TO_PHYS(maddr)) vs. maddr; /* phys addr of
Jack> start of mspecs. */
Uncached virtual, the comments you point out are leftovers from the old version of the driver.
Jack> A few code specific issues:
Jack> ... + printk(KERN_WARNING "smp_call_function failed for " + Jack> "mspec_ipi_visibility! (%i)\n", status); + } + + Jack> sn_flush_all_caches((unsigned long)tmp, IA64_GRANULE_SIZE);
Jack> Don't the TLBs need to be flushed before you flush Jack> caches. Otherwise, the cpu may reload data via speculation.
Jack> I dont see any TLB flushing of the kernel TLB entries that map Jack> the chunks. That needs to be done. ...
I thought about this one a fair bit after reading your comments and I don't think it's an issue. The pages in the kernel's cached mapping are identity mapped which means we shouldn't see any tlbs for this, which leaves us with just tlbs for pages that have explicitly been mapped somewhere - user tlbs should be removed when a process is shot down or pages unmapped and vfree() calls flush_tlb_all(). Or, am I missing something?
Jack> + /* + * The kernel requires a page structure to be returned Jack> upon + * success, but there are no page structures for low Jack> granule pages. + * remap_page_range() creates the pte for us Jack> and we return a + * bogus page back to the kernel fault handler Jack> to keep it happy + * (the page is freed immediately there). + Jack> */
Jack> Ugly hack. Isn't there a better way? (I know this isn't your Jack> code & you probably don't like this either. I had hoped for a Jack> cleaner solution in 2.6....)
It's gross, ugly and I hate it ... not sure if there's a simpler way. Maybe we can use the same approach as the fbmem driver and do it all in the mmap() function, I will have to investigate that.
Jack> + /* + * Use the bte to ensure cache lines + * are actually Jack> pulled from the + * processor back to the md. + */ +
Jack> This doesn't need to be done if the memory was being used for Jack> fetchops or uncached memory.
I'll check.
Jack> + s <<= 1; + } + a = (unsigned long) h[j].next;
Jack> It appears that you are keeping a linked list of free memory Jack> WITHIN the mspec memory itself. If I'm reading this correctly, Jack> all the addresses are uncached virtual addresses so that should Jack> be ok. However, it might be good to add debugging code to make Jack> sure that you never cause a cachable reference to be made to any Jack> of the fetchop memory. The resulting data corruption problems Jack> are almost impossible to debug.
You are correct that I keep the lists in the memory. I may change the allocator at a later stage to use descriptors instead, but for now I think this should be ok. I'll add a check to make sure we never receive a cached address back into mspec_free_page.
Thanks,
Jes
Quoted 7 lines
Jack> Ugly hack. Isn't there a better way? (I know this isn't your Jack> code & you probably don't like this either. I had hoped for a Jack> cleaner solution in 2.6....)
It's gross, ugly and I hate it ... not sure if there's a simpler way. Maybe we can use the same approach as the fbmem driver and do it all in the
mmap()function, I will have to investigate that.
If you do it in the mmap, all the pages will be allocated and mapped on the node doing the map. This will result in large applications using multiple threads to incur _LARGE_ amounts of numa traffic. The first fault is critical for performance.
Quoted 8 lines
Jack> + /* + * Use the bte to ensure cache lines + * are actually Jack> pulled from the + * processor back to the md. + */ +
Jack> This doesn't need to be done if the memory was being used for Jack> fetchops or uncached memory.
I'll check.
The bte zero is needed for memory mapped cached (one of the mechanisms here) and also ensures the memory is zereod out when returned to the memory pool.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Robin" = Robin Holt <[email protected]> writes:
Jack> Ugly hack. Isn't there a better way? (I know this isn't your Jack> code & you probably don't like this either. I had hoped for a Jack> cleaner solution in 2.6....)
Quoted 3 lines
Quoted 3 lines
It's gross, ugly and I hate it ... not sure if there's a simpler
way. Maybe we can use the same approach as the fbmem driver and do it all in themmap()function, I will have to investigate that.
Robin> If you do it in the mmap, all the pages will be allocated and Robin> mapped on the node doing the map. This will result in large Robin> applications using multiple threads to incur _LARGE_ amounts of Robin> numa traffic. The first fault is critical for performance.
Robin,
Is this because the applications will normally allocate their fetchops in the main thread before spawning off the threads? If the mmap is done by the thread that will 'own' the individual fetchop this wouldn't be a problem? Sorry, just trying to understand the nature of how these applications work, my knowledge of MPI etc. is ehm ... limited ;)
Jack> + /* + * Use the bte to ensure cache lines + * are actually Jack> pulled from the + * processor back to the md. + */ +
Quoted 1 line
Quoted 1 line
Jack> This doesn't need to be done if the memory was being used for Jack> fetchops or uncached memory.
Quoted 1 line
Quoted 1 line
I'll check.
Robin> The bte zero is needed for memory mapped cached (one of the Robin> mechanisms here) and also ensures the memory is zereod out when Robin> returned to the memory pool.
I already updated the comment to state that we want the memory zeroed in the pool.
Thanks,
Jes
Quoted 10 lines · Jes Sorensen
Jack> ... +
printk(KERN_WARNING "smp_call_function failed for " + Jack> "mspec_ipi_visibility! (%i)\n", status); + } + + Jack>sn_flush_all_caches((unsigned long)tmp, IA64_GRANULE_SIZE);Jack> Don't the TLBs need to be flushed before you flush Jack> caches. Otherwise, the cpu may reload data via speculation.
Jack> I dont see any TLB flushing of the kernel TLB entries that map Jack> the chunks. That needs to be done. ...
Quoted 7 lines
I thought about this one a fair bit after reading your comments and I don't think it's an issue. The pages in the kernel's cached mapping are identity mapped which means we shouldn't see any tlbs for this, which leaves us with just tlbs for pages that have explicitly been mapped somewhere - user tlbs should be removed when a process is shot down or pages unmapped and
vfree()callsflush_tlb_all(). Or, am I missing something?
Even identity mapped regions have TLB entries associated with them. The translation registers only cover the code and static data section, afaik. When we take a miss on an identity mapped region, the kernel still does an 'itc', so you'll still need to purge the TLB.
Jesse
Quoted 15 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
"Robin" = Robin Holt <[email protected]> writes:
Jack> Ugly hack. Isn't there a better way? (I know this isn't your Jack> code & you probably don't like this either. I had hoped for a Jack> cleaner solution in 2.6....)
Quoted 3 lines
Quoted 3 lines
It's gross, ugly and I hate it ... not sure if there's a simpler
way. Maybe we can use the same approach as the fbmem driver and do it all in themmap()function, I will have to investigate that.Robin> If you do it in the mmap, all the pages will be allocated and Robin> mapped on the node doing the map. This will result in large Robin> applications using multiple threads to incur _LARGE_ amounts of Robin> numa traffic. The first fault is critical for performance.
Quoted 8 lines
Robin,
Is this because the applications will normally allocate their fetchops in the main thread before spawning off the threads? If the mmap is done by the thread that will 'own' the individual fetchop this wouldn't be a problem? Sorry, just trying to understand the nature of how these applications work, my knowledge of MPI etc. is ehm ... limited ;)
You are correct. The parent process mmaps one segment of fetchop space and forks off all the worker processes. The worker processes inherit the fetchop mapping across the fork and each have their portion of the segment they "own". Now every process can read/write every other process' fetchop area.
Dean
---
Dean Roe
Silicon Graphics, Inc.
[email protected]
Quoted 6 lines · Jes Sorensen
Is this because the applications will normally allocate their fetchops in the main thread before spawning off the threads? If the mmap is done by the thread that will 'own' the individual fetchop this wouldn't be a problem? Sorry, just trying to understand the nature of how these applications work, my knowledge of MPI etc. is ehm ... limited ;)
That is the normal way of doing things. Nearly all large applications have been operating under that assumption for a very long time and changing all those apps to do the mmap after spawning workers is unlikely.
Robin
Quoted 11 lines · Jes Sorensen
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jesse" = Jesse Barnes <[email protected]> writes:
Quoted 7 lines · Jes Sorensen
Quoted 7 lines
The remaining issue I am facing is that for the uncached pool I want to make it node aware and we want to use the spill pages from the lower granules for this pool which is easily done on SN2. However I see no generic way to get from a physical address to node id for pages that do not have a struct page assigned to them. I am curious if anyone has any suggestions for how to solve this in a generic way?
Quoted 5 lines
Jesse>
paddr_to_nid?It's not clear to me this covers the spill memory found in the lower granules and second it's really really slow, though that may not be a real problem.
You're right, it won't deal with low addresses, only regular, cachable ones afaict.
Jesse
Quoted 1 line · Jes Sorensen
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jack" = Jack Steiner <[email protected]> writes:
sorry - I mised your reply. Apparantly, it looks like SPAM:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Subject: ***** SUSPECTED SPAM ***** Re: [rfc] generic allocator and mspec driver
From: Jes Sorensen <[email protected]>
X-Virus-Scanned: by cuda.sgi.com at sgi.com
X-Barracuda-Spam-Score: 0.60
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
X-Barracuda-Spam-Status: Yes, SCORE=0.60 using per-user scores of TAG_LEVEL=0.2 QUARANTINE_LEVEL=2.3 KILL_LEVEL00.0 tests=FORGED_RCVD_HELO,
MARKETING_SUBJECT
X-Barracuda-Spam-Report: Code version 2.64, rules version 2.1.1028
Rule breakdown below pts rule name description
---- ---------------------- -------------------------------------------
0.60MARKETING_SUBJECTSubject contains popular marketing words
0.00FORGED_RCVD_HELOReceived: contains a forged HELO
X-Priority: 5 (Lowest)
Oh well....
Quoted 19 lines
Jack> On Wed, Feb 02, 2005 at 02:10:32PM -0500, Jes Sorensen wrote:
Jack> General comment:Jack, thanks for the comments, I'll look at it, however I have the following comments (which may or may not be correct from my side):
Jack> 1) I may be paranoid, but I'm nervous about using memory visible Jack> to the VM system for fetchops. If ANYTHING in the kernel makes a Jack> reference to the memory and causes a TLB dropin to occur, then Jack> we are exposed to data corruption. If memory being used for Jack> fetchops is loaded into the cache, either directly or by Jack> speculation, then data corruption of the uncached fetchop memory Jack> can occur.
Jack> Am I being overly paranoid? How can we be certain that nothing
Jack> will ever reference the fetchop memory alocated from the general Jack> VM pool. lcrash, for example.
Quoted 5 lines
Once a page is handed out using
alloc_pages, the kernel won't touch it again unless you explicitly map it etc. or if some process touches memory at random, which could also happen with the spill pages. So I don't think the situation is any worse than it is for the spill pages in the lower granules.
In theory, you are correct & maybe I'm being overly paranoid. However, a failure is almost impossible to debug.
Using the UC area in the low granules seems safe but it still took us a long time to get it right. The kernel is unaware of the memory & with the exception of the fetchop code, nothing in the kernel ever references the spill areas.
I can't find any specific place that will fail using kernel memory for mspec but my gut feeling is that we are more exposed to errors using memory the kernel knows about than in using the spill areas.
For example, although I don't see any problems here because of it's limited use, virt_addr_valid() & pfn_valid() is FALSE for the spill area but TRUE for kernel memory.
What prevents lcrash (or /dev/kmem or /proc/kcore) from referencing special memory being used for fetchops? Granted, this takes root privilege but the consequences of a bad reference can cause silent data corruption that is impossible to debug. Should we add code to prohibit these area from referencing granules being used for mspec memory?
Forgive the paranoia but several of use spent a long time debugging some of these issues. Maybe all I'm asking is that everyone spend a little extra time thinking of ways that the kernel could cause a TLB entry to be made to a granule being used for mspec memory.
Quoted 6 lines
Jack> 2) Is there a limit on the number of mspec pages that can be Jack> allocated? Is there a shaker that will cause unused mspec Jack> granules to be freed? What prevents a malicious/stupid/buggy Jack> user from filling the system with mspec pages?
Quoted 8 lines
Currently there is no limit on this, however it could easily be imposed either by having a max number of granules allocated per node or system wide.
I could add code to free granules when it's all released but I believe the amount of memory being pulled in for this in real life situations is so limited it's not really worth the complexity. Adding a hard limit for how much is allowed to be allocated seems simpler.
Seems like some sort of limit is needed. I agree - something simple is all that is needed.
Quoted 23 lines
Jack> 3) Is an "mspec" address a physical address or an uncached Jack> virtual address? Some places in the code appear Jack> inconsistent. For example:
Jack> mspec_free_page(TO_PHYS(maddr)) vs. maddr; /* phys addr of
Jack> start of mspecs. */Uncached virtual, the comments you point out are leftovers from the old version of the driver.
Jack> A few code specific issues:
Jack> ... +
printk(KERN_WARNING "smp_call_function failed for " + Jack> "mspec_ipi_visibility! (%i)\n", status); + } + + Jack>sn_flush_all_caches((unsigned long)tmp, IA64_GRANULE_SIZE);Jack> Don't the TLBs need to be flushed before you flush Jack> caches. Otherwise, the cpu may reload data via speculation.
Jack> I dont see any TLB flushing of the kernel TLB entries that map Jack> the chunks. That needs to be done. ...
Quoted 7 lines
I thought about this one a fair bit after reading your comments and I don't think it's an issue. The pages in the kernel's cached mapping are identity mapped which means we shouldn't see any tlbs for this, which leaves us with just tlbs for pages that have explicitly been mapped somewhere - user tlbs should be removed when a process is shot down or pages unmapped and
vfree()callsflush_tlb_all(). Or, am I missing something?
Identity mapped memory still requires a TLB entry. Somewhere, these entries need to be purged before using a newly allocated granule for fetchops or uncached memory. Also, the TLB entries need to be purged before the cache is flushed. And the cache flushing can't require a cacheable TLB entry to be made.
Quoted 41 lines
Jack> + /* + * The kernel requires a page structure to be returned Jack> upon + * success, but there are no page structures for low Jack> granule pages. + *
remap_page_range()creates the pte for us Jack> and we return a + * bogus page back to the kernel fault handler Jack> to keep it happy + * (the page is freed immediately there). + Jack> */Jack> Ugly hack. Isn't there a better way? (I know this isn't your Jack> code & you probably don't like this either. I had hoped for a Jack> cleaner solution in 2.6....)
It's gross, ugly and I hate it ... not sure if there's a simpler way. Maybe we can use the same approach as the fbmem driver and do it all in the
mmap()function, I will have to investigate that.Jack> + /* + * Use the bte to ensure cache lines + * are actually Jack> pulled from the + * processor back to the md. + */ +
Jack> This doesn't need to be done if the memory was being used for Jack> fetchops or uncached memory.
I'll check.
Jack> + s <<= 1; + } + a = (unsigned long) h[j].next;
Jack> It appears that you are keeping a linked list of free memory Jack> WITHIN the mspec memory itself. If I'm reading this correctly, Jack> all the addresses are uncached virtual addresses so that should Jack> be ok. However, it might be good to add debugging code to make Jack> sure that you never cause a cachable reference to be made to any Jack> of the fetchop memory. The resulting data corruption problems Jack> are almost impossible to debug.
You are correct that I keep the lists in the memory. I may change the allocator at a later stage to use descriptors instead, but for now I think this should be ok. I'll add a check to make sure we never receive a cached address back into
mspec_free_page.Thanks,
Jes
Signature · 4 lines
Thanks Jack Steiner ([email protected]) 651-683-5302 Principal Engineer SGI - Silicon Graphics, Inc.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Jack Steiner
Jack> Am I being overly paranoid?
I don't think so. Your concerns are hitting the nail on the head...
--david
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jack" = Jack Steiner <[email protected]> writes:
Jack> On Thu, Feb 03, 2005 at 03:38:45AM -0500, Jes Sorensen wrote:
Jack> sorry - I mised your reply. Apparantly, it looks like SPAM:
Urgh, Barracuda! No comment ;-(
Jack> On Wed, Feb 02, 2005 at 02:10:32PM -0500, Jes Sorensen wrote:
Jack> General comment:
Quoted 3 lines
Quoted 3 lines
Jack, thanks for the comments, I'll look at it, however I have the
following comments (which may or may not be correct from my side):
Jack> 1) I may be paranoid, but I'm nervous about using memory visible Jack> to the VM system for fetchops. If ANYTHING in the kernel makes a Jack> reference to the memory and causes a TLB dropin to occur, then Jack> we are exposed to data corruption. If memory being used for Jack> fetchops is loaded into the cache, either directly or by Jack> speculation, then data corruption of the uncached fetchop memory Jack> can occur.
Jack> Am I being overly paranoid? How can we be certain that nothing Jack> will ever reference the fetchop memory alocated from the general Jack> VM pool. lcrash, for example.
Jack,
I hear your concerns! However, at the same time, if something within the kernel starts mocking with memory it doesn't own, thats a bug. Admittedly that can a royal pain in the b*tt to debug since a simple load can trigger it.
I was more concerned that there would be a case where prefetching or speculation would spill into a page in another granule and thereby cause a cacheable operation on the memory. However I don't quite understand the hardware to this level to be 110% sure.
Quoted 5 lines
Quoted 5 lines
Once a page is handed out using
alloc_pages, the kernel won't
touch it again unless you explicitly map it etc. or if some process touches memory at random, which could also happen with the spill pages. So I don't think the situation is any worse than it is for the spill pages in the lower granules.
Jack> In theory, you are correct & maybe I'm being overly Jack> paranoid. However, a failure is almost impossible to debug.
I wonder if it's something one could run a test on by running all of the kernel data memory outside the identity mapped area and then read-protecting the pages handed to the uncached allocator. Would require a fair bit of instrumentation though so I am not sure it would be feasible to try out.
Jack> Using the UC area in the low granules seems safe but it still Jack> took us a long time to get it right. The kernel is unaware of Jack> the memory & with the exception of the fetchop code, nothing in Jack> the kernel ever references the spill areas.
Jack> I can't find any specific place that will fail using kernel Jack> memory for mspec but my gut feeling is that we are more exposed Jack> to errors using memory the kernel knows about than in using the Jack> spill areas. For example, although I don't see any problems Jack> here because of it's limited use, virt_addr_valid() & Jack> pfn_valid() is FALSE for the spill area but TRUE for kernel Jack> memory.
I added support for converting pages since it was my understanding this was something there was a wish for. We can limit the uncached allocator to just use the spill pages but then we're back in the exact same situation we had with the old allocator there was in fetchop.c.
Jack> What prevents lcrash (or /dev/kmem or /proc/kcore) from Jack> referencing special memory being used for fetchops? Granted, Jack> this takes root privilege but the consequences of a bad Jack> reference can cause silent data corruption that is impossible to Jack> debug. Should we add code to prohibit these area from Jack> referencing granules being used for mspec memory?
I believe you will have the same problem with anyone messing with /dev/mem at random on the spill pages. I can't see us doing anything here besides saying that fetchop isn't supported if someone plays that kinds of games on their system as root.
Jack> Forgive the paranoia but several of use spent a long time Jack> debugging some of these issues. Maybe all I'm asking is that Jack> everyone spend a little extra time thinking of ways that the Jack> kernel could cause a TLB entry to be made to a granule being Jack> used for mspec memory.
Perfectly reasonable, which is also why I posted it to the list. More brains thinking about a problem are more likely to find any potential caveats.
Quoted 7 lines
Quoted 7 lines
I thought about this one a fair bit after reading your comments
and I don't think it's an issue. The pages in the kernel's cached mapping are identity mapped which means we shouldn't see any tlbs for this, which leaves us with just tlbs for pages that have explicitly been mapped somewhere - user tlbs should be removed when a process is shot down or pages unmapped andvfree()callsflush_tlb_all(). Or, am I missing something?
Jack> Identity mapped memory still requires a TLB entry. Somewhere, Jack> these entries need to be purged before using a newly allocated Jack> granule for fetchops or uncached memory. Also, the TLB entries Jack> need to be purged before the cache is flushed. And the cache Jack> flushing can't require a cacheable TLB entry to be made.
Aha, I wasn't aware of this, I thought the region registers worked like some giant TLB. I'll add a flush for the granule when it's pulled into the allocator.
Cheers,
Jes
Quoted 4 lines
Aha, I wasn't aware of this, I thought the region registers worked like some giant TLB. I'll add a flush for the granule when it's pulled into the allocator.
No, the Alt-DTLB miss handler in ivt.S will blindly insert TLB entries to fix any faults in region 6 or 7 ... giving the illusion that it's all mapped 1:1, but actually only a few granules will have mappings at any given time.
Maybe this handler should consult a "granule bitmap" before the insert to make sure it doesn't compound an error from a wild access by creating a TLB entry pointing at something that should be left unmapped? But that would be a big piece of memory for a sparse machine like an Altix, and an extra memory load in the Alt-DTLB path.
-Tony
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Jes Sorensen
Jes> I hear your concerns! However, at the same time, if something Jes> within the kernel starts mocking with memory it doesn't own, Jes> thats a bug.
I'm not convinced of that. Jack already mentioned lcrash...
Jes> I was more concerned that there would be a case where Jes> prefetching or speculation would spill into a page in another Jes> granule and thereby cause a cacheable operation on the Jes> memory.
This is not an issue. The AltTLB miss-handlers ignore misses triggered by speculative accesses (including lfetch) or accesses triggered by user-level.
Jack> What prevents lcrash (or /dev/kmem or /proc/kcore) from Jack> referencing special memory being used for fetchops? Granted, Jack> this takes root privilege but the consequences of a bad Jack> reference can cause silent data corruption that is impossible Jack> to debug. Should we add code to prohibit these area from Jack> referencing granules being used for mspec memory?
Jes> I believe you will have the same problem with anyone messing Jes> with /dev/mem at random on the spill pages.
/dev/mem does check the memory attributes and accesses the memory uncached if EFI_MEMORY_WB is _not_ set. This probably needs to be extended to check what the current usage of the page is (as opposed to what the potential usages are). Now that the kernel has it's own copy of the memory-map, that's probably something we can do more easily.
OTOH, for pages which have a page-descriptor, a flag-bit would be the most logical choice and that would make it possible to efficiently check whether it is OK to access a page via a cacheable reference.
Jes> Aha, I wasn't aware of this, I thought the region registers Jes> worked like some giant TLB. I'll add a flush for the granule Jes> when it's pulled into the allocator.
Don't you have to purge the all TLBs completely anyhow?
--david
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
"David" = David Mosberger <[email protected]> writes:
On 03 Feb 2005 15:58:16 -0500, Jes Sorensen <[email protected]> said:
Jes> I hear your concerns! However, at the same time, if something Jes> within the kernel starts mocking with memory it doesn't own, Jes> thats a bug.
David> I'm not convinced of that. Jack already mentioned lcrash...
Mmmmm, I took a look at lcrash and I'll refrain from commenting on it's method of operation here to avoid too much public swearing. Yuck!
Anyway, it seems that lcrash at least only accesses the memory using read() and write()? which means it is relatively easy to using a struct page flag to do uncached vs cached access on the fly. mmap on the other hand would be a bit of a nightmare as it would require a custom fault handler that would know when to go cached and when not to.
How would you feel about using PG_arch_1 for this on ia64?
Cheers,
Jes
Quoted 8 lines
Anyway, it seems that lcrash at least only accesses the memory using
read()andwrite()? which means it is relatively easy to using a struct page flag to do uncached vs cached access on the fly. mmap on the other hand would be a bit of a nightmare as it would require a custom fault handler that would know when to go cached and when not to.How would you feel about using
PG_arch_1for this on ia64?
Is ia64 the only architecture that has uncached vs. cached access issues? If so, the PG_arch_1 might have to be the solution, but surely others have cache coherence problems too if there are mixed cacheable and uncacheable access to the same memory? In which case a new generic bit would be more appropriate.
A few weeks back there was also some discussion about allocating one or more bits to mark bad pages (those containing ECC errors) as "do not touch".
-Tony
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Tony
Tony> If so, the PG_arch_1 might have to be the solution
Note that PG_arch_1 is already being used as a "i-cache coherent" flag.
--david
Quoted 10 lines
Quoted 10 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Tony" = Luck, Tony <[email protected]> writes:
Anyway, it seems that lcrash at least only accesses the memory using
read()andwrite()? which means it is relatively easy to using a struct page flag to do uncached vs cached access on the fly. mmap on the other hand would be a bit of a nightmare as it would require a custom fault handler that would know when to go cached and when not to.How would you feel about using
PG_arch_1for this on ia64?
Tony> Is ia64 the only architecture that has uncached vs. cached Tony> access issues? If so, the PG_arch_1 might have to be the Tony> solution, but surely others have cache coherence problems too if Tony> there are mixed cacheable and uncacheable access to the same Tony> memory? In which case a new generic bit would be more Tony> appropriate.
None of the ones I know well have this problem, but I have little knowledge about this level of stuff on most architectures. The ones that could have issues would probably be like PPC, PARISC and maybe Alpha .....
Tony> A few weeks back there was also some discussion about allocating Tony> one or more bits to mark bad pages (those containing ECC errors) Tony> as "do not touch".
Probably something which needs to be done sooner or later.
Cheers,
Jes
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
"David" = David Mosberger <[email protected]> writes:
On Mon, 14 Feb 2005 11:12:54 -0800, "Luck, Tony" <[email protected]> said:
Tony> If so, the PG_arch_1 might have to be the solution
David> Note that PG_arch_1 is already being used as a "i-cache David> coherent" flag.
Ahh, so not totally off. It's also used for internal debugging in reiser4, but I'll happily ignore that while waving the 'friends don't let friends run reiser4' flag ;)
Cheers,
Jes
Quoted 4 lines · Jes Sorensen
None of the ones I know well have this problem, but I have little knowledge about this level of stuff on most architectures. The ones that could have issues would probably be like PPC, PARISC and maybe Alpha .....
With parisc 2.0 CPU, one could map the same physical address as both cacheable and uncacheable via two entries in the page table. AFAIK, the uncacheable bit is only use to map MMIO address space. In practice, I'm not sure what happens since I'm not aware of any need for both mappings at the same time because we can use "LDWA" (Load Word Absolute). LDWA can load from (almost) any physical address and is by nature uncached (caches are VIVT).
Older PA1.x CPU had hardwired 0xf0000000-0xffffffff to be uncacheable. I.e. it's not possible to have an overlap.
hth,
grant
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines · Jes Sorensen
"Tony" = Luck, Tony <[email protected]> writes:
Tony> Is ia64 the only architecture that has uncached vs. cached Tony> access issues? If so, the PG_arch_1 might have to be the Tony> solution, but surely others have cache coherence problems too Tony> if there are mixed cacheable and uncacheable access to the Tony> same memory? In which case a new generic bit would be more Tony> appropriate.
Jes> None of the ones I know well have this problem, but I have Jes> little knowledge about this level of stuff on most Jes> architectures. The ones that could have issues would probably Jes> be like PPC, PARISC and maybe Alpha .....
Well, any CPU that allows overlapping mappings and does _any_ sort of speculative accesses will have a problem. The only question is whether you'll get an explicit error notification (e.g., MCA) or silent data corruption.
--david
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Jes Sorensen
Jes> Ahh, so not totally off. It's also used for internal debugging Jes> in reiser4, but I'll happily ignore that while waving the Jes> 'friends don't let friends run reiser4' flag ;)
What are they smoking? Which part of PG_arch_1 do they not understand?
^^^^ --david
Quoted 16 lines · David Mosberger
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines · Jes Sorensen
"Tony" = Luck, Tony <[email protected]> writes:
Tony> Is ia64 the only architecture that has uncached vs. cached Tony> access issues? If so, the
PG_arch_1might have to be the Tony> solution, but surely others have cache coherence problems too Tony> if there are mixed cacheable and uncacheable access to the Tony> same memory? In which case a new generic bit would be more Tony> appropriate.Jes> None of the ones I know well have this problem, but I have Jes> little knowledge about this level of stuff on most Jes> architectures. The ones that could have issues would probably Jes> be like PPC, PARISC and maybe Alpha .....
Quoted 4 lines
Well, any CPU that allows overlapping mappings and does _any_ sort of speculative accesses will have a problem. The only question is whether you'll get an explicit error notification (e.g., MCA) or silent data corruption.
We saw many silent data corruptions. The SN2 hardware will give an MCA if both types of references are made by the cpu at about the same time, but that depends on both transactions being on the bus in close relationship to each other.
Robin
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Robin Holt
Robin> We saw many silent data corruptions. The SN2 hardware will Robin> give an MCA if both types of references are made by the cpu Robin> at about the same time, but that depends on both transactions Robin> being on the bus in close relationship to each other.
Yes, but the point is it affects _most_ of today's machines, including x86. A while ago, there were some nasty data corruption issues on AMD chips caused by the AGP GART and the large mappings used by the Linux kernel, for example.
--david
Quoted 11 lines · David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Robin Holt
Robin> We saw many silent data corruptions. The SN2 hardware will Robin> give an MCA if both types of references are made by the cpu Robin> at about the same time, but that depends on both transactions Robin> being on the bus in close relationship to each other.
Yes, but the point is it affects _most_ of today's machines, including x86. A while ago, there were some nasty data corruption issues on AMD chips caused by the AGP GART and the large mappings used by the Linux kernel, for example.
One thing I had not thought about before this is SN2 hardware will also allow us to change the memory protections for the cachelines in the pages of this granule to not allow cached references. I am not sure if this has been tested aside from design verification testing and probably some of the offline diagnostic tests.
Would anybody have a concern with marking those cache lines as processor uncached only? Would there ever be a time that we would expect _ANY_ type of cached reference when the granule is being used for uncached that would be an acceptable use?
Thanks,
Robin
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
"David" = David Mosberger <[email protected]> writes:
On 15 Feb 2005 03:44:26 -0500, Jes Sorensen <[email protected]> said:
Jes> Ahh, so not totally off. It's also used for internal debugging in Jes> reiser4, but I'll happily ignore that while waving the 'friends Jes> don't let friends run reiser4' flag ;)
David> What are they smoking? Which part of PG_arch_1 do they not David> understand? ^^^^
I suspect it's some sort of garden weed, the one labelled 'if we use this magic flag, then nobody will notice us sneaking something in under the rug'.
Gotta love 'em ;-)
Jes
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Robin" = Robin Holt <[email protected]> writes:
Robin> Would anybody have a concern with marking those cache lines as Robin> processor uncached only? Would there ever be a time that we Robin> would expect _ANY_ type of cached reference when the granule is Robin> being used for uncached that would be an acceptable use?
Robin,
I don't see any valid such cases. In principle one has to go through the reverse procedure to convert something from uncached to cached and since we don't do that in any place I don't see any valid reason to do so. Anyone doing it in kdb etc. would be in for what they are getting, so I don't think thats valid concern.
Cheers,
Jes
Quoted 7 lines · Jes Sorensen
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Robin" = Robin Holt <[email protected]> writes:
Robin> Would anybody have a concern with marking those cache lines as Robin> processor uncached only? Would there ever be a time that we Robin> would expect _ANY_ type of cached reference when the granule is Robin> being used for uncached that would be an acceptable use?
Quoted 7 lines
Robin,
I don't see any valid such cases. In principle one has to go through the reverse procedure to convert something from uncached to cached and since we don't do that in any place I don't see any valid reason to do so. Anyone doing it in kdb etc. would be in for what they are getting, so I don't think thats valid concern.
I guess I didn't follow you. Are you saying you would have no concern with changing the memory protections to prevent that type of reference?
Thanks,
Robin
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Robin" = Robin Holt <[email protected]> writes:
Quoted 5 lines · Jes Sorensen
Quoted 5 lines
I don't see any valid such cases. In principle one has to go through the reverse procedure to convert something from uncached to cached and since we don't do that in any place I don't see any valid reason to do so. Anyone doing it in kdb etc. would be in for what they are getting, so I don't think thats valid concern.
Robin> I guess I didn't follow you. Are you saying you would have no Robin> concern with changing the memory protections to prevent that Robin> type of reference?
Yes
Jes
Hi
Here's another version of the mspec driver with the generic allocator. I have tried to implement all the features we discussed on the list earlier, including marking pages as uncached using PG_arch_1 and making /dev/mem honor it. In order to do so I had to move our /dev/mem read/write functions into arch/ia64/kernel to avoid totally cluttering drivers/char/mem.c with #ifdefs. I also moved the mspec.c driver to arch/ia64/kernel to reflect that it now works on non-sn hardware for uncached and cached mappings.
Any more ideas/objections for this one?
The patch is still against 2.6.11-rc2-mm2 however it won't compile straight out of the box as it depends on another patch I dumped on Jesse which moved shubio.h to include/asm/sn
Cheers,
Jes
arch/ia64/Kconfig +9 -0
@@ -225,6 +225,15 @@ If you are compiling a kernel that will run under SGI's IA-64 simulator (Medusa) then say Y, otherwise say N. +config MSPEC + tristate "Special Memory support" + help + This driver allows for cached and uncached mappings of memory + to user processes. On SGI SN hardware it will also export the + special fetchop memory facility. + Fetchops are atomic memory operations that are implemented in the + memory controller on SGI SN hardware. + config FORCE_MAX_ZONEORDER int default "18"
arch/ia64/configs/sn2_defconfig +1 -0
@@ -82,6 +82,7 @@ # CONFIG_IA64_CYCLONE is not set CONFIG_IOSAPIC=y CONFIG_IA64_SGI_SN_SIM=y +CONFIG_MSPEC=m CONFIG_FORCE_MAX_ZONEORDER CONFIG_SMP=y CONFIG_NR_CPUSQ2
arch/ia64/defconfig +1 -0
@@ -80,6 +80,7 @@ CONFIG_DISCONTIGMEM=y CONFIG_IA64_CYCLONE=y CONFIG_IOSAPIC=y +CONFIG_MSPEC=m CONFIG_FORCE_MAX_ZONEORDER CONFIG_SMP=y CONFIG_NR_CPUSQ2
arch/ia64/kernel/Makefile +2 -1
@@ -7,7 +7,7 @@ obj-y := acpi.o entry.o efi.o efi_stub.o gate-data.o fsys.o ia64_ksyms.o irq.o irq_ia64.o \ irq_lsapic.o ivt.o machvec.o pal.o patch.o process.o perfmon.o ptrace.o sal.o \ salinfo.o semaphore.o setup.o signal.o sys_ia64.o time.o traps.o unaligned.o \ - unwind.o mca.o mca_asm.o topology.o + unwind.o mca.o mca_asm.o topology.o mem.o obj-$(CONFIG_IA64_BRL_EMU) += brl_emu.o obj-$(CONFIG_IA64_GENERIC) += acpi-ext.o
@@ -20,6 +20,7 @@ obj-$(CONFIG_PERFMON) += perfmon_default_smpl.o obj-$(CONFIG_IA64_CYCLONE) += cyclone.o obj-$(CONFIG_IA64_MCA_RECOVERY) += mca_recovery.o +obj-$(CONFIG_MSPEC) += mspec.o mca_recovery-y += mca_drv.o mca_drv_asm.o # The gate DSO image is built using a special linker script.
arch/ia64/kernel/mem.c +151 -0
151 lines not on this page · raw message
arch/ia64/kernel/mspec.c +804 -0
804 lines not on this page · raw message
drivers/char/mem.c +38 -35
104 lines not on this page · raw message
include/asm-ia64/io.h +2 -0
@@ -481,4 +481,6 @@#define BIO_VMERGE_BOUNDARY (ia64_max_iommu_merge_mask + 1)#endif +#define ARCH_HAS_DEV_MEM +#endif /* _ASM_IA64_IO_H */
include/asm-ia64/sn/fetchop.h +0 -85
85 lines not on this page · raw message
include/asm-ia64/sn/mspec.h +72 -0
72 lines not on this page · raw message
include/linux/genalloc.h +46 -0
46 lines not on this page · raw message
init/main.c +2 -0
@@ -78,6 +78,7 @@static int init (void *); +extern void gen_pool_init (void );extern void init_IRQ (void );extern void sock_init (void );extern void fork_init (unsigned long );
@@ -482,6 +483,7 @@#endif vfs_caches_init_early ();mem_init (); +gen_pool_init ();kmem_cache_init ();numa_policy_init ();if (late_time_init)
lib/Makefile +1 -1
@@ -6,7 +6,7 @@ bust_spinlocks.o rbtree.o radix-tree.o dump_stack.o \ kobject.o kref.o idr.o div64.o parser.o int_sqrt.o \ bitmap.o extable.o kobject_uevent.o prio_tree.o \ - sha1.o halfmd4.o + sha1.o halfmd4.o genalloc.o ifeq ($(CONFIG_DEBUG_KOBJECT),y) CFLAGS_kobject.o += -DDEBUG
lib/genalloc.c +218 -0
218 lines not on this page · raw message
Quoted 4 lines
Here's another version of the mspec driver with the generic allocator. I have tried to implement all the features we discussed on the list earlier, including marking pages as uncached using
PG_arch_1and making/dev/memhonor it.
Quoted 1 line · Did you see David's e-mail? The one where he
Note that
PG_arch_1is already being used as a "i-cache coherent" flag.
Won't bad things happen if someone tried to read a page from /dev/mem that was marked PG_arch_1 by swiotlb code?
I'm also not thrilled about mem.c being cloned into arch/ia64.
-Tony
Quoted 1 line
I'm also not thrilled about
mem.cbeing cloned intoarch/ia64.
Looking more closely at what you did, things aren't as bad as I first thought. Strike that comment from the record.
But the PG_arch_1 dual usage issue remains.
-Tony
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Tony" = Luck, Tony <[email protected]> writes:
I'm also not thrilled about
mem.cbeing cloned intoarch/ia64.
Tony> Looking more closely at what you did, things aren't as bad as I Tony> first thought. Strike that comment from the record.
Tony,
No problem. I really didn't want to do it, but it seems that the -mm series already introduced this for Xen. It was that or 3 zillion #ifdefs ;-(
Tony> But the PG_arch_1 dual usage issue remains.
Hmmm for some reason I had the impression PG_arch_1 was only used on PPC/PPC64 for icache current status, but it looks like you're right. Darn ;-(
Can't we just drop swiotlb and tell people to buy real hardware? ;-)))
Ok, I'll invent a new flag!
Thanks for the input!
Cheers,
Jes
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Jes Sorensen
Jes> Hmmm for some reason I had the impression PG_arch_1 was only Jes> used on PPC/PPC64 for icache current status, but it looks like Jes> you're right. Darn ;-(
Jes> Can't we just drop swiotlb and tell people to buy real Jes> hardware? ;-)))
The use of PG_arch_1 has nothing to do with swiotlb. It lets us optimize away expensive cache-flushing by taking advantage of the fact that DMA transactions automatically establish coherency.
--david