| Message ID | 1639193588-7027-1-git-send-email-yangtiezhu@loongson.cn (mailing list archive) |
|---|---|
| Headers | show |
| Series | kdump: simplify code | expand |
From: Tiezhu Yang > Sent: 11 December 2021 03:33 > > v2: > -- add copy_to_user_or_kernel() in lib/usercopy.c > -- define userbuf as bool type Instead of having a flag to indicate whether the buffer is user or kernel, would it be better to have two separate buffer pointers. One for a user space buffer, the other for a kernel space buffer. Exactly one of the buffers should always be NULL. That way the flag is never incorrectly set. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
Le 11/12/2021 à 18:53, David Laight a écrit : > From: Tiezhu Yang >> Sent: 11 December 2021 03:33 >> >> v2: >> -- add copy_to_user_or_kernel() in lib/usercopy.c >> -- define userbuf as bool type > > Instead of having a flag to indicate whether the buffer is user or kernel, > would it be better to have two separate buffer pointers. > One for a user space buffer, the other for a kernel space buffer. > Exactly one of the buffers should always be NULL. > > That way the flag is never incorrectly set. > It's a very good idea. I was worried about the casts forcing the __user property away and back. With that approach we will preserve the __user tags on user buffers and enable sparse checking. The only little drawback I see is that apparently GCC doesn't consider the NULL value as a constant and therefore doesn't perform constant folding on pointers. Not sure if this is a problem here. Christophe
On Sat, Dec 11, 2021 at 05:53:46PM +0000, David Laight wrote: > From: Tiezhu Yang > > Sent: 11 December 2021 03:33 > > > > v2: > > -- add copy_to_user_or_kernel() in lib/usercopy.c > > -- define userbuf as bool type > > Instead of having a flag to indicate whether the buffer is user or kernel, > would it be better to have two separate buffer pointers. > One for a user space buffer, the other for a kernel space buffer. > Exactly one of the buffers should always be NULL. No. You should be using an iov_iter instead. See https://lore.kernel.org/all/Ya4bdB0UBJCZhUSo@casper.infradead.org/ for a start on this.
From: Matthew Wilcox > Sent: 12 December 2021 11:48 > > On Sat, Dec 11, 2021 at 05:53:46PM +0000, David Laight wrote: > > From: Tiezhu Yang > > > Sent: 11 December 2021 03:33 > > > > > > v2: > > > -- add copy_to_user_or_kernel() in lib/usercopy.c > > > -- define userbuf as bool type > > > > Instead of having a flag to indicate whether the buffer is user or kernel, > > would it be better to have two separate buffer pointers. > > One for a user space buffer, the other for a kernel space buffer. > > Exactly one of the buffers should always be NULL. > > No. You should be using an iov_iter instead. See > https://lore.kernel.org/all/Ya4bdB0UBJCZhUSo@casper.infradead.org/ > for a start on this. iov_iter gets horribly expensive... David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Mon, Dec 13, 2021 at 08:30:33AM +0000, David Laight wrote: > From: Matthew Wilcox > > Sent: 12 December 2021 11:48 > > > > On Sat, Dec 11, 2021 at 05:53:46PM +0000, David Laight wrote: > > > From: Tiezhu Yang > > > > Sent: 11 December 2021 03:33 > > > > > > > > v2: > > > > -- add copy_to_user_or_kernel() in lib/usercopy.c > > > > -- define userbuf as bool type > > > > > > Instead of having a flag to indicate whether the buffer is user or kernel, > > > would it be better to have two separate buffer pointers. > > > One for a user space buffer, the other for a kernel space buffer. > > > Exactly one of the buffers should always be NULL. > > > > No. You should be using an iov_iter instead. See > > https://lore.kernel.org/all/Ya4bdB0UBJCZhUSo@casper.infradead.org/ > > for a start on this. > > iov_iter gets horribly expensive... Oh, right. Reading the kcore is a high-performance path, my mistake.
On 12/13/2021 10:43 PM, Matthew Wilcox wrote: > On Mon, Dec 13, 2021 at 08:30:33AM +0000, David Laight wrote: >> From: Matthew Wilcox >>> Sent: 12 December 2021 11:48 >>> >>> On Sat, Dec 11, 2021 at 05:53:46PM +0000, David Laight wrote: >>>> From: Tiezhu Yang >>>>> Sent: 11 December 2021 03:33 >>>>> >>>>> v2: >>>>> -- add copy_to_user_or_kernel() in lib/usercopy.c >>>>> -- define userbuf as bool type >>>> >>>> Instead of having a flag to indicate whether the buffer is user or kernel, >>>> would it be better to have two separate buffer pointers. >>>> One for a user space buffer, the other for a kernel space buffer. >>>> Exactly one of the buffers should always be NULL. >>> >>> No. You should be using an iov_iter instead. See >>> https://lore.kernel.org/all/Ya4bdB0UBJCZhUSo@casper.infradead.org/ >>> for a start on this. >> >> iov_iter gets horribly expensive... > > Oh, right. Reading the kcore is a high-performance path, my mistake. > Hi, Thank you for your discussions. The intention of this patchset is to simplify the related code with no functional changes and no side effects. At this moment, if you are OK, I will send v3 used with inline function copy_to_user_or_kernel() to keep it simple, maybe other more changes can be done in the future if no any side effect. The v3 will contain the following three patches to make the changes more clear: kdump: vmcore: remove copy_to() and add copy_to_user_or_kernel() kdump: crashdump: use copy_to_user_or_kernel() to simplify code kdump: vmcore: crashdump: make variable type of userbuf as bool
On Tue, Dec 14, 2021 at 06:03:11PM +0800, Tiezhu Yang wrote: > On 12/13/2021 10:43 PM, Matthew Wilcox wrote: > > On Mon, Dec 13, 2021 at 08:30:33AM +0000, David Laight wrote: > > > From: Matthew Wilcox > > > > Sent: 12 December 2021 11:48 > > > > > > > > On Sat, Dec 11, 2021 at 05:53:46PM +0000, David Laight wrote: > > > > > From: Tiezhu Yang > > > > > > Sent: 11 December 2021 03:33 > > > > > > > > > > > > v2: > > > > > > -- add copy_to_user_or_kernel() in lib/usercopy.c > > > > > > -- define userbuf as bool type > > > > > > > > > > Instead of having a flag to indicate whether the buffer is user or kernel, > > > > > would it be better to have two separate buffer pointers. > > > > > One for a user space buffer, the other for a kernel space buffer. > > > > > Exactly one of the buffers should always be NULL. > > > > > > > > No. You should be using an iov_iter instead. See > > > > https://lore.kernel.org/all/Ya4bdB0UBJCZhUSo@casper.infradead.org/ > > > > for a start on this. > > > > > > iov_iter gets horribly expensive... > > > > Oh, right. Reading the kcore is a high-performance path, my mistake. > > > > Hi, > > Thank you for your discussions. > > The intention of this patchset is to simplify the related code with no > functional changes and no side effects. > > At this moment, if you are OK, I will send v3 used with inline function > copy_to_user_or_kernel() to keep it simple, maybe other more changes can > be done in the future if no any side effect. That would be pointless. I already sent a series to remove this, which you were cc'd on.