On Thu, Oct 12, 2000 at 04:48:55PM -0700, lamont@icopyright.com wrote:
>
> if you aren't comfortable with dropping a lot of the 2.2.18preX stuff onto
> a production box, there is also the 2.2.18pre2aa2 kernel that andrea made
2.2.18pre2aa2 should be kind of rock solid (the only problem it has
as every other rawio capable kernel out there is rawio so if you don't use rawio
you're just fine for using it in production).
The latest aa patchkit (2.2.18pre15aa1) also fixes the couple of severe rawio
bugs that can generate memleaks in SMP memory corruption and a deadlock
condition that leads to processes unkillable in D state. Obviously nobody
before me ever used rawio to read from disk to a completly swapped out area.
The swapped out buffer case is the one that two of those bugs trivially (just
after one second of runtime the task goes in D state) but if the machine is
under a little memory pressure bad things could happen anyways (infact the
reason I looked into it the first time was a bugreport about 2.2.18pre2aa2
using a real DBMS on a 256Mbyte RAM machine under some memory pressure, the
kernel was returning -EFAULT and writing error messages).
The fix also reduces the number of times the kernel walks pagetable
for each page that we read/write from/to (it reduces from two walks to
1 walk in the common case where the page is just mapped in memory).
I was waiting confirm that there are no problems anymore using DBMS (as I can't
definitely reproduce here anymore with my rawio regression test tool) before
sending the fix to SCT. But since it's taking some time and I don't want to
delay the fixes further I'm extracting the fixes here. 2.4.x needs the most
important part of these fixes too.
diff -urN rawio-ref/fs/buffer.c rawio/fs/buffer.c
--- rawio-ref/fs/buffer.c Fri Oct 13 15:43:59 2000
+++ rawio/fs/buffer.c Fri Oct 13 15:46:06 2000
@@ -1276,6 +1276,7 @@
int i;
int err;
struct buffer_head *tmp;
+ unsigned long flags;
dprintk ("do_kio start\n");
@@ -1293,9 +1294,10 @@
iosize = 0;
}
- free_async_buffers(tmp);
+ put_unused_buffer_head(tmp);
iosize += size;
}
+ wake_up(&buffer_wait);
dprintk ("do_kio end %d %d\n", iosize, err);
@@ -1358,6 +1360,7 @@
unsigned long bounce;
struct page * map;
struct buffer_head *tmp, *bh[KIO_MAX_SECTORS];
+ unsigned long flags;
/*
* First, do some alignment and validity checks
@@ -1472,9 +1475,9 @@
error:
/* We got an error allocation the bh'es. Just free the current
buffer_heads and exit. */
- for (i = bhind; --i >= 0; ) {
- free_async_buffers(bh[bhind]);
- }
+ for (i = bhind; --i >= 0; )
+ put_unused_buffer_head(bh[i]);
+ wake_up(&buffer_wait);
clear_kiobuf_bounce_pages(iobuf);
diff -urN rawio-ref/mm/memory.c rawio/mm/memory.c
--- rawio-ref/mm/memory.c Fri Oct 13 15:43:59 2000
+++ rawio/mm/memory.c Fri Oct 13 15:46:06 2000
@@ -401,7 +401,7 @@
/*
* Do a quick page-table lookup for a single page.
*/
-static unsigned long get_page(unsigned long address)
+static unsigned long get_page(unsigned long address, int write)
{
pgd_t *pgd;
pmd_t *pmd;
@@ -411,11 +411,11 @@
if (pmd) {
pte_t * pte = pte_offset(pmd, address);
if (pte && pte_present(*pte)) {
- return pte_page(*pte);
+ if (!write || pte_write(*pte))
+ return pte_page(*pte);
}
}
- printk(KERN_ERR "Missing page in lock_down_page\n");
return 0;
}
@@ -448,12 +448,15 @@
unsigned long ptr, end;
int err;
struct mm_struct * mm;
- struct vm_area_struct * vma = 0;
+ struct vm_area_struct * vma = 0, * prev_vma;
unsigned long page;
struct page * map;
int doublepage = 0;
int repeat = 0;
int i;
+ int write = (rw == READ); /* if we read from disk
+ it means we write
+ to memory */
/* Make sure the iobuf is not already mapped somewhere. */
if (iobuf->nr_pages)
@@ -483,16 +486,36 @@
*/
while (ptr < end) {
if (!vma || ptr >= vma->vm_end) {
- vma = find_vma(current->mm, ptr);
- if (!vma)
+ refind:
+ vma = find_vma_prev(mm, ptr, &prev_vma);
+ if (!vma)
goto out_unlock;
+ if (vma->vm_start > ptr) {
+ if (!(vma->vm_flags & VM_GROWSDOWN))
+ goto out_unlock;
+ if (expand_stack(vma, ptr, prev_vma))
+ goto out_unlock;
+ }
+ err = -EACCES;
+ if (write) {
+ if (!(vma->vm_flags & VM_WRITE))
+ goto out_unlock;
+ } else {
+ if (!(vma->vm_flags & VM_READ))
+ goto out_unlock;
+ }
+ err = -EFAULT;
}
- if (!handle_mm_fault(current, vma, ptr, (rw==READ)))
- goto out_unlock;
- page = get_page(ptr);
- if (!page) {
- printk (KERN_ERR "Missing page in map_user_kiobuf\n");
- goto out_unlock;
+ while (!(page = get_page(ptr, write))) {
+ char * user_ptr = (char *) ptr;
+ char c;
+ up(&mm->mmap_sem);
+ if (get_user(c, user_ptr))
+ goto failed_pagein;
+ if (write && put_user(c, user_ptr))
+ goto failed_pagein;
+ down(&mm->mmap_sem);
+ goto refind;
}
map = get_page_map(page);
if (map) {
@@ -515,9 +538,15 @@
out_unlock:
up(&mm->mmap_sem);
+ out:
unmap_kiobuf(iobuf);
dprintk ("map_user_kiobuf: end %d\n", err);
return err;
+
+ failed_pagein:
+ printk(KERN_ERR "map_user_kiobuf: failed pagein\n");
+ goto out;
+
retry:
diff -urN rawio-ref/mm/page_alloc.c rawio/mm/page_alloc.c
--- rawio-ref/mm/page_alloc.c Tue Sep 5 02:28:50 2000
+++ rawio/mm/page_alloc.c Fri Oct 13 15:46:06 2000
@@ -357,6 +357,7 @@
struct page *new_page;
unsigned long offset = SWP_OFFSET(entry);
struct swap_info_struct *swapdev = SWP_TYPE(entry) + swap_info;
+ unsigned long readaround_entry;
offset = (offset >> page_cluster) << page_cluster;
@@ -376,8 +377,11 @@
if (test_bit(offset, swapdev->swap_lockmap))
break;
- /* Ok, do the async read-ahead now */
- new_page = read_swap_cache_async(SWP_ENTRY(SWP_TYPE(entry), offset), 0);
+ readaround_entry = SWP_ENTRY(SWP_TYPE(entry), offset);
+ new_page = find_page(&swapper_inode, ulong2pgoff(readaround_entry));
+ if (!new_page)
+ /* Ok, do the async read-ahead now */
+ new_page = read_swap_cache_async(readaround_entry, 0);
if (new_page != NULL)
__free_page(new_page);
offset++;
2.2.18pre15 is better than 2.2.18pre2 so there should be no worry in upgrading
to it. About the incremental 2.2.18pre15aa1 the only experimental thing
included in 2.2.18pre15aa1 is nfsv3 both client and server and both converted
to be lfs capable as the aa patchkit is lfs capable as 2.4.x. So you should
keep using nfsv2 for production right now when running 2.2.18pre15aa1.
nfsv3 is been reported to return an error on close if used with 4096 large udp
packets (OTOH Andi seems not able reproduce that). In the next version I'll
take 4096 as default so that bug if not yet fixed should trigger even more
easily, but the real reason of the change is that if somebody needs to deal
with broken TCP stacks he can just override with lower values on the mount
command line and otherwise people complains for the too low performance.
Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Oct 15 2000 - 21:00:25 EST